mz_sql/plan/statement/

ddl.rs

// Copyright Materialize, Inc. and contributors. All rights reserved.
//
// Use of this software is governed by the Business Source License
// included in the LICENSE file.
//
// As of the Change Date specified in that file, in accordance with
// the Business Source License, use of this software will be governed
// by the Apache License, Version 2.0.

//! Data definition language (DDL).
//!
//! This module houses the handlers for statements that modify the catalog, like
//! `ALTER`, `CREATE`, and `DROP`.

use std::collections::{BTreeMap, BTreeSet};
use std::fmt::Write;
use std::iter;
use std::num::NonZeroU32;
use std::time::Duration;

use chrono::DateTime;
use itertools::Itertools;
use mz_adapter_types::compaction::{CompactionWindow, DEFAULT_LOGICAL_COMPACTION_WINDOW_DURATION};
use mz_arrow_util::builder::ArrowBuilder;
use mz_auth::password::Password;
use mz_controller_types::{ClusterId, DEFAULT_REPLICA_LOGGING_INTERVAL, ReplicaId};
use mz_expr::{CollectionPlan, UnmaterializableFunc};
use mz_interchange::avro::{AvroSchemaGenerator, DocTarget};
use mz_ore::cast::{CastFrom, TryCastFrom};
use mz_ore::collections::{CollectionExt, HashSet};
use mz_ore::num::NonNeg;
use mz_ore::str::StrExt;
use mz_ore::{soft_assert_or_log, soft_panic_or_log};
use mz_proto::RustType;
use mz_repr::adt::interval::Interval;
use mz_repr::adt::mz_acl_item::{MzAclItem, PrivilegeMap};
use mz_repr::adt::timestamp::CheckedTimestamp;
use mz_repr::network_policy_id::NetworkPolicyId;
use mz_repr::optimize::OptimizerFeatureOverrides;
use mz_repr::refresh_schedule::{RefreshEvery, RefreshSchedule};
use mz_repr::role_id::RoleId;
use mz_repr::{
    CatalogItemId, ColumnName, RelationDesc, RelationVersion, RelationVersionSelector,
    SqlColumnType, SqlRelationType, SqlScalarType, Timestamp, VersionedRelationDesc,
    preserves_order, strconv,
};
use mz_sql_parser::ast::{
    self, AlterClusterAction, AlterClusterStatement, AlterConnectionAction, AlterConnectionOption,
    AlterConnectionOptionName, AlterConnectionStatement, AlterIndexAction, AlterIndexStatement,
    AlterMaterializedViewApplyReplacementStatement, AlterNetworkPolicyStatement,
    AlterObjectRenameStatement, AlterObjectSwapStatement, AlterRetainHistoryStatement,
    AlterRoleOption, AlterRoleStatement, AlterSecretStatement, AlterSetClusterStatement,
    AlterSinkAction, AlterSinkStatement, AlterSourceAction, AlterSourceAddSubsourceOption,
    AlterSourceAddSubsourceOptionName, AlterSourceStatement, AlterSystemResetAllStatement,
    AlterSystemResetStatement, AlterSystemSetStatement, AlterTableAddColumnStatement, AvroSchema,
    AvroSchemaOption, AvroSchemaOptionName, ClusterAlterOption, ClusterAlterOptionName,
    ClusterAlterOptionValue, ClusterAlterUntilReadyOption, ClusterAlterUntilReadyOptionName,
    ClusterAutoScalingStrategyOptionValue, ClusterFeature, ClusterFeatureName, ClusterOption,
    ClusterOptionName, ClusterScheduleOptionValue, ColumnDef, ColumnOption, CommentObjectType,
    CommentStatement, ConnectionOption, ConnectionOptionName, CreateClusterReplicaStatement,
    CreateClusterStatement, CreateConnectionOption, CreateConnectionOptionName,
    CreateConnectionStatement, CreateConnectionType, CreateDatabaseStatement, CreateIndexStatement,
    CreateMaterializedViewStatement, CreateMetricSinkOption, CreateMetricSinkOptionName,
    CreateMetricSinkStatement, CreateNetworkPolicyStatement, CreateRoleStatement,
    CreateSchemaStatement, CreateSecretStatement, CreateSinkConnection, CreateSinkOption,
    CreateSinkOptionName, CreateSinkStatement, CreateSourceConnection, CreateSourceOption,
    CreateSourceOptionName, CreateSourceStatement, CreateSubsourceOption,
    CreateSubsourceOptionName, CreateSubsourceStatement, CreateTableFromSourceStatement,
    CreateTableStatement, CreateTypeAs, CreateTypeListOption, CreateTypeListOptionName,
    CreateTypeMapOption, CreateTypeMapOptionName, CreateTypeStatement, CreateViewStatement,
    CreateWebhookSourceStatement, CsrConfigOption, CsrConfigOptionName, CsrConnection,
    CsrConnectionAvro, CsrConnectionProtobuf, CsrSeedProtobuf, CsvColumns, DeferredItemName,
    DocOnIdentifier, DocOnSchema, DropObjectsStatement, DropOwnedStatement, Expr, Format,
    FormatSpecifier, GlueAvroOption, GlueAvroOptionName, IcebergSinkConfigOption, Ident,
    IfExistsBehavior, IndexOption, IndexOptionName, KafkaSinkConfigOption, KeyConstraint,
    LoadGeneratorOption, LoadGeneratorOptionName, MaterializedViewOption,
    MaterializedViewOptionName, MySqlConfigOption, MySqlConfigOptionName, NetworkPolicyOption,
    NetworkPolicyOptionName, NetworkPolicyRuleDefinition, NetworkPolicyRuleOption,
    NetworkPolicyRuleOptionName, OnHydrationOptionValue, PgConfigOption, PgConfigOptionName,
    ProtobufSchema, QualifiedReplica, RefreshAtOptionValue, RefreshEveryOptionValue,
    RefreshOptionValue, ReplicaDefinition, ReplicaOption, ReplicaOptionName, RoleAttribute,
    SetRoleVar, SourceErrorPolicy, SourceIncludeMetadata, SqlServerConfigOption,
    SqlServerConfigOptionName, Statement, TableConstraint, TableFromSourceColumns,
    TableFromSourceOption, TableFromSourceOptionName, TableOption, TableOptionName,
    UnresolvedDatabaseName, UnresolvedItemName, UnresolvedObjectName, UnresolvedSchemaName, Value,
    ViewDefinition, WithOptionValue,
};
use mz_sql_parser::ident;
use mz_sql_parser::parser::StatementParseResult;
use mz_storage_types::connections::inline::ReferencedConnection;
use mz_storage_types::connections::{Connection, KafkaTopicOptions};
use mz_storage_types::sinks::{
    IcebergSinkConnection, KafkaIdStyle, KafkaSinkConnection, KafkaSinkFormat, KafkaSinkFormatType,
    SinkEnvelope, StorageSinkConnection, iceberg_type_overrides,
};
use mz_storage_types::sources::encoding::{
    AvroEncoding, ColumnSpec, CsvEncoding, DataEncoding, ProtobufEncoding, RegexEncoding,
    SourceDataEncoding, included_column_desc,
};
use mz_storage_types::sources::envelope::{
    KeyEnvelope, NoneEnvelope, SourceEnvelope, UnplannedSourceEnvelope, UpsertStyle,
};
use mz_storage_types::sources::kafka::{
    KafkaMetadataKind, KafkaSourceConnection, KafkaSourceExportDetails, kafka_metadata_columns_desc,
};
use mz_storage_types::sources::load_generator::{
    KeyValueLoadGenerator, LOAD_GENERATOR_KEY_VALUE_OFFSET_DEFAULT, LoadGenerator,
    LoadGeneratorOutput, LoadGeneratorSourceConnection, LoadGeneratorSourceExportDetails,
};
use mz_storage_types::sources::mysql::{
    MySqlSourceConnection, MySqlSourceDetails, ProtoMySqlSourceDetails,
};
use mz_storage_types::sources::postgres::{
    PostgresSourceConnection, PostgresSourcePublicationDetails,
    ProtoPostgresSourcePublicationDetails,
};
use mz_storage_types::sources::sql_server::{
    ProtoSqlServerSourceExtras, SqlServerSourceExportDetails,
};
use mz_storage_types::sources::{
    GenericSourceConnection, MySqlSourceExportDetails, PostgresSourceExportDetails,
    ProtoSourceExportStatementDetails, SourceConnection, SourceDesc, SourceExportDataConfig,
    SourceExportDetails, SourceExportStatementDetails, SqlServerSourceConnection,
    SqlServerSourceExtras, Timeline,
};
use mz_storage_types::wire_format::WireFormat;
use prost::Message;

use crate::ast::display::AstDisplay;
use crate::catalog::{
    CatalogCluster, CatalogDatabase, CatalogError, CatalogItem, CatalogItemType,
    CatalogRecordField, CatalogType, CatalogTypeDetails, ObjectType, SystemObjectType,
};
use crate::iceberg::IcebergSinkConfigOptionExtracted;
use crate::kafka_util::{KafkaSinkConfigOptionExtracted, KafkaSourceConfigOptionExtracted};
use crate::names::{
    Aug, CommentObjectId, DatabaseId, DependencyIds, ObjectId, PartialItemName, QualifiedItemName,
    ResolvedClusterName, ResolvedColumnReference, ResolvedDataType, ResolvedDatabaseSpecifier,
    ResolvedItemName, ResolvedNetworkPolicyName, SchemaSpecifier, SystemObjectId,
};
use crate::normalize::{self, ident};
use crate::plan::error::PlanError;
use crate::plan::query::{
    ExprContext, QueryLifetime, TypeResolutionBudget, plan_expr, scalar_type_from_sql,
};
use crate::plan::scope::Scope;
use crate::plan::statement::ddl::connection::{INALTERABLE_OPTIONS, MUTUALLY_EXCLUSIVE_SETS};
use crate::plan::statement::{StatementContext, StatementDesc, scl};
use crate::plan::typeconv::CastContext;
use crate::plan::with_options::{OptionalDuration, OptionalString, TryFromValue};
use crate::plan::{
    AlterClusterPlan, AlterClusterPlanStrategy, AlterClusterRenamePlan,
    AlterClusterReplicaRenamePlan, AlterClusterSwapPlan, AlterConnectionPlan, AlterItemRenamePlan,
    AlterMaterializedViewApplyReplacementPlan, AlterNetworkPolicyPlan, AlterNoopPlan,
    AlterOptionParameter, AlterRetainHistoryPlan, AlterRolePlan, AlterSchemaRenamePlan,
    AlterSchemaSwapPlan, AlterSecretPlan, AlterSetClusterPlan, AlterSinkPlan,
    AlterSourceTimestampIntervalPlan, AlterSystemResetAllPlan, AlterSystemResetPlan,
    AlterSystemSetPlan, AlterTablePlan, AutoScalingStrategy, ClusterSchedule, CommentPlan,
    ComputeReplicaConfig, ComputeReplicaIntrospectionConfig, ConnectionDetails,
    CreateClusterManagedPlan, CreateClusterPlan, CreateClusterReplicaPlan,
    CreateClusterUnmanagedPlan, CreateClusterVariant, CreateConnectionPlan, CreateDatabasePlan,
    CreateIndexPlan, CreateMaterializedViewPlan, CreateMetricSinkPlan, CreateNetworkPolicyPlan,
    CreateRolePlan, CreateSchemaPlan, CreateSecretPlan, CreateSinkPlan, CreateSourcePlan,
    CreateTablePlan, CreateTypePlan, CreateViewPlan, DataSourceDesc, DropObjectsPlan,
    DropOwnedPlan, HirRelationExpr, Index, MaterializedView, MetricSink, NetworkPolicyRule,
    NetworkPolicyRuleAction, NetworkPolicyRuleDirection, OnHydration, Plan, PlanClusterOption,
    PlanNotice, PolicyAddress, QueryContext, ReplicaConfig, Secret, Sink, Source, Table,
    TableDataSource, Type, VariableValue, View, WebhookBodyFormat, WebhookHeaderFilters,
    WebhookHeaders, WebhookValidation, literal, plan_utils, query, transform_ast,
};
use crate::session::vars::{
    self, ENABLE_AUTO_SCALING_STRATEGY, ENABLE_CLUSTER_SCHEDULE_REFRESH,
    ENABLE_COLLECTION_PARTITION_BY, ENABLE_CREATE_TABLE_FROM_SOURCE, ENABLE_KAFKA_SINK_HEADERS,
    ENABLE_METRIC_SINK, ENABLE_REFRESH_EVERY_MVS, ENABLE_REPLICA_TARGETED_MATERIALIZED_VIEWS,
    VarInput,
};
use crate::{names, parse};

mod connection;

// TODO: Figure out what the maximum number of columns we can actually support is, and set that.
//
// The real max is probably higher than this, but it's easier to relax a constraint than make it
// more strict.
const MAX_NUM_COLUMNS: usize = 256;

const MAX_KAFKA_TOPIC_METADATA_REFRESH_INTERVAL: Duration = Duration::from_secs(60 * 60);
const MIN_KAFKA_TOPIC_METADATA_REFRESH_INTERVAL: Duration = Duration::from_secs(1);

static MANAGED_REPLICA_PATTERN: std::sync::LazyLock<regex::Regex> =
    std::sync::LazyLock::new(|| regex::Regex::new(r"^r(\d)+$").unwrap());

/// Given a relation desc and a column list, checks that:
/// - the column list is a prefix of the desc;
/// - all the listed columns are types that have meaningful Persist-level ordering.
fn check_partition_by(desc: &RelationDesc, mut partition_by: Vec<Ident>) -> Result<(), PlanError> {
    if partition_by.len() > desc.len() {
        tracing::error!(
            "PARTITION BY contains more columns than the relation. (expected at most {}, got {})",
            desc.len(),
            partition_by.len()
        );
        partition_by.truncate(desc.len());
    }

    let desc_prefix = desc.iter().take(partition_by.len());
    for (idx, ((desc_name, desc_type), partition_name)) in
        desc_prefix.zip_eq(partition_by).enumerate()
    {
        let partition_name = normalize::column_name(partition_name);
        if *desc_name != partition_name {
            sql_bail!(
                "PARTITION BY columns should be a prefix of the relation's columns (expected {desc_name} at index {idx}, got {partition_name})"
            );
        }
        if !preserves_order(&desc_type.scalar_type) {
            sql_bail!("PARTITION BY column {partition_name} has unsupported type");
        }
    }
    Ok(())
}

pub fn describe_create_database(
    _: &StatementContext,
    _: CreateDatabaseStatement,
) -> Result<StatementDesc, PlanError> {
    Ok(StatementDesc::new(None))
}

pub fn plan_create_database(
    _: &StatementContext,
    CreateDatabaseStatement {
        name,
        if_not_exists,
    }: CreateDatabaseStatement,
) -> Result<Plan, PlanError> {
    Ok(Plan::CreateDatabase(CreateDatabasePlan {
        name: normalize::ident(name.0),
        if_not_exists,
    }))
}

pub fn describe_create_schema(
    _: &StatementContext,
    _: CreateSchemaStatement,
) -> Result<StatementDesc, PlanError> {
    Ok(StatementDesc::new(None))
}

pub fn plan_create_schema(
    scx: &StatementContext,
    CreateSchemaStatement {
        mut name,
        if_not_exists,
    }: CreateSchemaStatement,
) -> Result<Plan, PlanError> {
    if name.0.len() > 2 {
        sql_bail!("schema name {} has more than two components", name);
    }
    let schema_name = normalize::ident(
        name.0
            .pop()
            .expect("names always have at least one component"),
    );
    let database_spec = match name.0.pop() {
        None => match scx.catalog.active_database() {
            Some(id) => ResolvedDatabaseSpecifier::Id(id.clone()),
            None => sql_bail!("no database specified and no active database"),
        },
        Some(n) => match scx.resolve_database(&UnresolvedDatabaseName(n.clone())) {
            Ok(database) => ResolvedDatabaseSpecifier::Id(database.id()),
            Err(_) => sql_bail!("invalid database {}", n.as_str()),
        },
    };
    Ok(Plan::CreateSchema(CreateSchemaPlan {
        database_spec,
        schema_name,
        if_not_exists,
    }))
}

pub fn describe_create_table(
    _: &StatementContext,
    _: CreateTableStatement<Aug>,
) -> Result<StatementDesc, PlanError> {
    Ok(StatementDesc::new(None))
}

pub fn plan_create_table(
    scx: &StatementContext,
    stmt: CreateTableStatement<Aug>,
) -> Result<Plan, PlanError> {
    let CreateTableStatement {
        name,
        columns,
        constraints,
        if_not_exists,
        temporary,
        with_options,
    } = &stmt;

    let names: Vec<_> = columns
        .iter()
        .filter(|c| {
            // This set of `names` is used to create the initial RelationDesc.
            // Columns that have been added at later versions of the table will
            // get added further below.
            let is_versioned = c
                .options
                .iter()
                .any(|o| matches!(o.option, ColumnOption::Versioned { .. }));
            !is_versioned
        })
        .map(|c| normalize::column_name(c.name.clone()))
        .collect();

    if let Some(dup) = names.iter().duplicates().next() {
        sql_bail!("column {} specified more than once", dup.quoted());
    }

    // Build initial relation type that handles declared data types
    // and NOT NULL constraints.
    let mut column_types = Vec::with_capacity(columns.len());
    let mut defaults = Vec::with_capacity(columns.len());
    let mut changes = BTreeMap::new();
    let mut keys = Vec::new();

    for (i, c) in columns.into_iter().enumerate() {
        if c.collation
            .as_ref()
            .is_some_and(|collation| !query::is_supported_collation(collation))
        {
            bail_unsupported!("COLLATE");
        }
        let aug_data_type = &c.data_type;
        let ty = query::scalar_type_from_sql(scx, aug_data_type)?;
        let mut nullable = true;
        let mut default = Expr::null();
        let mut versioned = false;
        for option in &c.options {
            match &option.option {
                ColumnOption::NotNull => nullable = false,
                ColumnOption::Default(expr) => {
                    // Ensure expression can be planned and yields the correct
                    // type.
                    let mut expr = expr.clone();
                    transform_ast::transform(scx, &mut expr)?;
                    let _ = query::plan_default_expr(scx, &expr, &ty)?;
                    default = expr.clone();
                }
                ColumnOption::Unique { is_primary } => {
                    keys.push(vec![i]);
                    if *is_primary {
                        nullable = false;
                    }
                }
                ColumnOption::Versioned { action, version } => {
                    let version = normalize::relation_version(*version);
                    versioned = true;

                    let name = normalize::column_name(c.name.clone());
                    let typ = ty.clone().nullable(nullable);

                    changes.insert(version, (action.clone(), name, typ));
                }
                other => {
                    bail_unsupported!(format!("CREATE TABLE with column constraint: {}", other))
                }
            }
        }
        // TODO(alter_table): This assumes all versioned columns are at the
        // end. This will no longer be true when we support dropping columns.
        if !versioned {
            column_types.push(ty.nullable(nullable));
        }
        defaults.push(default);
    }

    let mut seen_primary = false;
    'c: for constraint in constraints {
        match constraint {
            TableConstraint::Unique {
                name: _,
                columns,
                is_primary,
                nulls_not_distinct,
            } => {
                if seen_primary && *is_primary {
                    sql_bail!(
                        "multiple primary keys for table {} are not allowed",
                        name.to_ast_string_stable()
                    );
                }
                seen_primary = *is_primary || seen_primary;

                let mut key = vec![];
                for column in columns {
                    let column = normalize::column_name(column.clone());
                    match names.iter().position(|name| *name == column) {
                        None => sql_bail!("unknown column in constraint: {}", column),
                        Some(i) => {
                            let nullable = &mut column_types[i].nullable;
                            if *is_primary {
                                if *nulls_not_distinct {
                                    sql_bail!(
                                        "[internal error] PRIMARY KEY does not support NULLS NOT DISTINCT"
                                    );
                                }

                                *nullable = false;
                            } else if !(*nulls_not_distinct || !*nullable) {
                                // Non-primary key unique constraints are only keys if all of their
                                // columns are `NOT NULL` or the constraint is `NULLS NOT DISTINCT`.
                                break 'c;
                            }

                            key.push(i);
                        }
                    }
                }

                if *is_primary {
                    keys.insert(0, key);
                } else {
                    keys.push(key);
                }
            }
            TableConstraint::ForeignKey { .. } => {
                // Foreign key constraints are not presently enforced. We allow
                // them with feature flags for sqllogictest's sake.
                scx.require_feature_flag(&vars::UNSAFE_ENABLE_TABLE_FOREIGN_KEY)?
            }
            TableConstraint::Check { .. } => {
                // Check constraints are not presently enforced. We allow them
                // with feature flags for sqllogictest's sake.
                scx.require_feature_flag(&vars::UNSAFE_ENABLE_TABLE_CHECK_CONSTRAINT)?
            }
        }
    }

    if !keys.is_empty() {
        // Unique constraints are not presently enforced. We allow them with feature flags for
        // sqllogictest's sake.
        scx.require_feature_flag(&vars::UNSAFE_ENABLE_TABLE_KEYS)?
    }

    let typ = SqlRelationType::new(column_types).with_keys(keys);

    let temporary = *temporary;
    let name = if temporary {
        scx.allocate_temporary_qualified_name(normalize::unresolved_item_name(name.to_owned())?)?
    } else {
        scx.allocate_qualified_name(normalize::unresolved_item_name(name.to_owned())?)?
    };

    // Check for an object in the catalog with this same name
    let full_name = scx.catalog.resolve_full_name(&name);
    let partial_name = PartialItemName::from(full_name.clone());
    // For PostgreSQL compatibility, we need to prevent creating tables when
    // there is an existing object *or* type of the same name.
    if let (false, Ok(item)) = (
        if_not_exists,
        scx.catalog.resolve_item_or_type(&partial_name),
    ) {
        return Err(PlanError::ItemAlreadyExists {
            name: full_name.to_string(),
            item_type: item.item_type(),
        });
    }

    let desc = RelationDesc::new(typ, names);
    let mut desc = VersionedRelationDesc::new(desc);
    for (version, (_action, name, typ)) in changes.into_iter() {
        let new_version = desc.add_column(name, typ);
        if version != new_version {
            return Err(PlanError::InvalidTable {
                name: full_name.item,
            });
        }
    }

    let create_sql = normalize::create_statement(scx, Statement::CreateTable(stmt.clone()))?;

    // Table options should only consider the original columns, since those
    // were the only ones in scope when the table was created.
    //
    // TODO(alter_table): Will need to reconsider this when we support ALTERing
    // the PARTITION BY columns.
    let original_desc = desc.at_version(RelationVersionSelector::Specific(RelationVersion::root()));
    let options = plan_table_options(scx, &original_desc, with_options.clone())?;

    let compaction_window = options.iter().find_map(|o| {
        #[allow(irrefutable_let_patterns)]
        if let crate::plan::TableOption::RetainHistory(lcw) = o {
            Some(lcw.clone())
        } else {
            None
        }
    });

    let table = Table {
        create_sql,
        desc,
        temporary,
        compaction_window,
        data_source: TableDataSource::TableWrites { defaults },
    };
    Ok(Plan::CreateTable(CreateTablePlan {
        name,
        table,
        if_not_exists: *if_not_exists,
    }))
}

pub fn describe_create_table_from_source(
    _: &StatementContext,
    _: CreateTableFromSourceStatement<Aug>,
) -> Result<StatementDesc, PlanError> {
    Ok(StatementDesc::new(None))
}

pub fn describe_create_webhook_source(
    _: &StatementContext,
    _: CreateWebhookSourceStatement<Aug>,
) -> Result<StatementDesc, PlanError> {
    Ok(StatementDesc::new(None))
}

pub fn describe_create_source(
    _: &StatementContext,
    _: CreateSourceStatement<Aug>,
) -> Result<StatementDesc, PlanError> {
    Ok(StatementDesc::new(None))
}

pub fn describe_create_subsource(
    _: &StatementContext,
    _: CreateSubsourceStatement<Aug>,
) -> Result<StatementDesc, PlanError> {
    Ok(StatementDesc::new(None))
}

generate_extracted_config!(
    CreateSourceOption,
    (TimestampInterval, Duration),
    (RetainHistory, OptionalDuration)
);

generate_extracted_config!(
    PgConfigOption,
    (Details, String),
    (Publication, String),
    (Failover, bool, Default(false)),
    (TextColumns, Vec::<UnresolvedItemName>, Default(vec![])),
    (ExcludeColumns, Vec::<UnresolvedItemName>, Default(vec![]))
);

generate_extracted_config!(
    MySqlConfigOption,
    (Details, String),
    (TextColumns, Vec::<UnresolvedItemName>, Default(vec![])),
    (ExcludeColumns, Vec::<UnresolvedItemName>, Default(vec![]))
);

generate_extracted_config!(
    SqlServerConfigOption,
    (Details, String),
    (TextColumns, Vec::<UnresolvedItemName>, Default(vec![])),
    (ExcludeColumns, Vec::<UnresolvedItemName>, Default(vec![]))
);

pub fn plan_create_webhook_source(
    scx: &StatementContext,
    mut stmt: CreateWebhookSourceStatement<Aug>,
) -> Result<Plan, PlanError> {
    if stmt.is_table {
        scx.require_feature_flag(&ENABLE_CREATE_TABLE_FROM_SOURCE)?;
    }

    // We will rewrite the cluster if one is not provided, so we must use the `in_cluster` value
    // we plan to normalize when we canonicalize the create statement.
    let in_cluster = source_sink_cluster_config(scx, &mut stmt.in_cluster)?;
    let create_sql =
        normalize::create_statement(scx, Statement::CreateWebhookSource(stmt.clone()))?;

    let CreateWebhookSourceStatement {
        name,
        if_not_exists,
        body_format,
        include_headers,
        validate_using,
        is_table,
        // We resolved `in_cluster` above, so we want to ignore it here.
        in_cluster: _,
    } = stmt;

    let validate_using = validate_using
        .map(|stmt| query::plan_webhook_validate_using(scx, stmt))
        .transpose()?;
    if let Some(WebhookValidation { expression, .. }) = &validate_using {
        // If the validation expression doesn't reference any part of the request, then we should
        // return an error because it's almost definitely wrong.
        if !expression.contains_column() {
            return Err(PlanError::WebhookValidationDoesNotUseColumns);
        }
        // Validation expressions cannot contain unmaterializable functions, except `now()`. We
        // allow calls to `now()` because some webhook providers recommend rejecting requests that
        // are older than a certain threshold.
        if expression.contains_unmaterializable_except(&[UnmaterializableFunc::CurrentTimestamp]) {
            return Err(PlanError::WebhookValidationNonDeterministic);
        }
    }

    let body_format = match body_format {
        Format::Bytes => WebhookBodyFormat::Bytes,
        Format::Json { array } => WebhookBodyFormat::Json { array },
        Format::Text => WebhookBodyFormat::Text,
        // TODO(parkmycar): Make an issue to support more types, or change this to NeverSupported.
        ty => {
            return Err(PlanError::Unsupported {
                feature: format!("{ty} is not a valid BODY FORMAT for a WEBHOOK source"),
                discussion_no: None,
            });
        }
    };

    let mut column_ty = vec![
        // Always include the body of the request as the first column.
        SqlColumnType {
            scalar_type: SqlScalarType::from(body_format),
            nullable: false,
        },
    ];
    let mut column_names = vec!["body".to_string()];

    let mut headers = WebhookHeaders::default();

    // Include a `headers` column, possibly filtered.
    if let Some(filters) = include_headers.column {
        column_ty.push(SqlColumnType {
            scalar_type: SqlScalarType::Map {
                value_type: Box::new(SqlScalarType::String),
                custom_id: None,
            },
            nullable: false,
        });
        column_names.push("headers".to_string());

        let (allow, block): (BTreeSet<_>, BTreeSet<_>) =
            filters.into_iter().partition_map(|filter| {
                if filter.block {
                    itertools::Either::Right(filter.header_name)
                } else {
                    itertools::Either::Left(filter.header_name)
                }
            });
        headers.header_column = Some(WebhookHeaderFilters { allow, block });
    }

    // Map headers to specific columns.
    for header in include_headers.mappings {
        let scalar_type = header
            .use_bytes
            .then_some(SqlScalarType::Bytes)
            .unwrap_or(SqlScalarType::String);
        column_ty.push(SqlColumnType {
            scalar_type,
            nullable: true,
        });
        column_names.push(header.column_name.into_string());

        let column_idx = column_ty.len() - 1;
        // Double check we're consistent with column names.
        assert_eq!(
            column_idx,
            column_names.len() - 1,
            "header column names and types don't match"
        );
        headers
            .mapped_headers
            .insert(column_idx, (header.header_name, header.use_bytes));
    }

    // Validate our columns.
    let mut unique_check = HashSet::with_capacity(column_names.len());
    for name in &column_names {
        if !unique_check.insert(name) {
            return Err(PlanError::AmbiguousColumn(name.clone().into()));
        }
    }
    if column_names.len() > MAX_NUM_COLUMNS {
        return Err(PlanError::TooManyColumns {
            max_num_columns: MAX_NUM_COLUMNS,
            req_num_columns: column_names.len(),
        });
    }

    let typ = SqlRelationType::new(column_ty);
    let desc = RelationDesc::new(typ, column_names);

    // Check for an object in the catalog with this same name
    let name = scx.allocate_qualified_name(normalize::unresolved_item_name(name)?)?;
    let full_name = scx.catalog.resolve_full_name(&name);
    let partial_name = PartialItemName::from(full_name.clone());
    if let (false, Ok(item)) = (if_not_exists, scx.catalog.resolve_item(&partial_name)) {
        return Err(PlanError::ItemAlreadyExists {
            name: full_name.to_string(),
            item_type: item.item_type(),
        });
    }

    // Note(parkmycar): We don't currently support specifying a timeline for Webhook sources. As
    // such, we always use a default of EpochMilliseconds.
    let timeline = Timeline::EpochMilliseconds;

    let plan = if is_table {
        let data_source = DataSourceDesc::Webhook {
            validate_using,
            body_format,
            headers,
            cluster_id: Some(in_cluster.id()),
        };
        let data_source = TableDataSource::DataSource {
            desc: data_source,
            timeline,
        };
        Plan::CreateTable(CreateTablePlan {
            name,
            if_not_exists,
            table: Table {
                create_sql,
                desc: VersionedRelationDesc::new(desc),
                temporary: false,
                compaction_window: None,
                data_source,
            },
        })
    } else {
        let data_source = DataSourceDesc::Webhook {
            validate_using,
            body_format,
            headers,
            // Important: The cluster is set at the `Source` level.
            cluster_id: None,
        };
        Plan::CreateSource(CreateSourcePlan {
            name,
            source: Source {
                create_sql,
                data_source,
                desc,
                compaction_window: None,
            },
            if_not_exists,
            timeline,
            in_cluster: Some(in_cluster.id()),
        })
    };

    Ok(plan)
}

pub fn plan_create_source(
    scx: &StatementContext,
    mut stmt: CreateSourceStatement<Aug>,
) -> Result<Plan, PlanError> {
    let CreateSourceStatement {
        name,
        in_cluster: _,
        col_names,
        connection: source_connection,
        envelope,
        if_not_exists,
        format,
        key_constraint,
        include_metadata,
        with_options,
        external_references: referenced_subsources,
        progress_subsource,
    } = &stmt;

    mz_ore::soft_assert_or_log!(
        referenced_subsources.is_none(),
        "referenced subsources must be cleared in purification"
    );

    let force_source_table_syntax = scx.catalog.system_vars().enable_create_table_from_source()
        && scx.catalog.system_vars().force_source_table_syntax();

    // If the new source table syntax is forced all the options related to the primary
    // source output should be un-set.
    if force_source_table_syntax {
        if envelope.is_some() || format.is_some() || !include_metadata.is_empty() {
            Err(PlanError::UseTablesForSources(
                "CREATE SOURCE (ENVELOPE|FORMAT|INCLUDE)".to_string(),
            ))?;
        }
    }

    let envelope = envelope.clone().unwrap_or(ast::SourceEnvelope::None);

    if !matches!(source_connection, CreateSourceConnection::Kafka { .. })
        && include_metadata
            .iter()
            .any(|sic| matches!(sic, SourceIncludeMetadata::Headers { .. }))
    {
        // TODO(guswynn): should this be `bail_unsupported!`?
        sql_bail!("INCLUDE HEADERS with non-Kafka sources not supported");
    }
    if !matches!(
        source_connection,
        CreateSourceConnection::Kafka { .. } | CreateSourceConnection::LoadGenerator { .. }
    ) && !include_metadata.is_empty()
    {
        bail_unsupported!("INCLUDE metadata with non-Kafka sources");
    }

    if !include_metadata.is_empty()
        && !matches!(
            envelope,
            ast::SourceEnvelope::Upsert { .. }
                | ast::SourceEnvelope::None
                | ast::SourceEnvelope::Debezium
        )
    {
        sql_bail!("INCLUDE <metadata> requires ENVELOPE (NONE|UPSERT|DEBEZIUM)");
    }

    let external_connection =
        plan_generic_source_connection(scx, source_connection, include_metadata)?;

    let CreateSourceOptionExtracted {
        timestamp_interval,
        retain_history,
        seen: _,
    } = CreateSourceOptionExtracted::try_from(with_options.clone())?;

    let metadata_columns_desc = match external_connection {
        GenericSourceConnection::Kafka(KafkaSourceConnection {
            ref metadata_columns,
            ..
        }) => kafka_metadata_columns_desc(metadata_columns),
        _ => vec![],
    };

    // Generate the relation description for the primary export of the source.
    let (mut desc, envelope, encoding) = apply_source_envelope_encoding(
        scx,
        &envelope,
        format,
        Some(external_connection.default_key_desc()),
        external_connection.default_value_desc(),
        include_metadata,
        metadata_columns_desc,
        &external_connection,
    )?;
    plan_utils::maybe_rename_columns(format!("source {}", name), &mut desc, col_names)?;

    let names: Vec<_> = desc.iter_names().cloned().collect();
    if let Some(dup) = names.iter().duplicates().next() {
        sql_bail!("column {} specified more than once", dup.quoted());
    }

    // Apply user-specified key constraint
    if let Some(KeyConstraint::PrimaryKeyNotEnforced { columns }) = key_constraint.clone() {
        // Don't remove this without addressing
        // https://github.com/MaterializeInc/database-issues/issues/4371.
        scx.require_feature_flag(&vars::ENABLE_PRIMARY_KEY_NOT_ENFORCED)?;

        let key_columns = columns
            .into_iter()
            .map(normalize::column_name)
            .collect::<Vec<_>>();

        let mut uniq = BTreeSet::new();
        for col in key_columns.iter() {
            if !uniq.insert(col) {
                sql_bail!("Repeated column name in source key constraint: {}", col);
            }
        }

        let key_indices = key_columns
            .iter()
            .map(|col| {
                let name_idx = desc
                    .get_by_name(col)
                    .map(|(idx, _type)| idx)
                    .ok_or_else(|| sql_err!("No such column in source key constraint: {}", col))?;
                if desc.get_unambiguous_name(name_idx).is_none() {
                    sql_bail!("Ambiguous column in source key constraint: {}", col);
                }
                Ok(name_idx)
            })
            .collect::<Result<Vec<_>, _>>()?;

        if !desc.typ().keys.is_empty() {
            return Err(key_constraint_err(&desc, &key_columns));
        } else {
            desc = desc.with_key(key_indices);
        }
    }

    let timestamp_interval = match timestamp_interval {
        Some(duration) => {
            // Only validate bounds for new statements (pcx is Some), not during
            // catalog deserialization (pcx is None). Previously persisted sources
            // may have intervals that no longer fall within the current bounds.
            if scx.pcx.is_some() {
                let min = scx.catalog.system_vars().min_timestamp_interval();
                let max = scx.catalog.system_vars().max_timestamp_interval();
                if duration < min || duration > max {
                    return Err(PlanError::InvalidTimestampInterval {
                        min,
                        max,
                        requested: duration,
                    });
                }
            }
            duration
        }
        None => scx.catalog.system_vars().default_timestamp_interval(),
    };

    let (desc, data_source) = match progress_subsource {
        Some(name) => {
            let DeferredItemName::Named(name) = name else {
                sql_bail!("[internal error] progress subsource must be named during purification");
            };
            let ResolvedItemName::Item { id, .. } = name else {
                sql_bail!("[internal error] invalid target id");
            };

            let details = match external_connection {
                GenericSourceConnection::Kafka(ref c) => {
                    SourceExportDetails::Kafka(KafkaSourceExportDetails {
                        metadata_columns: c.metadata_columns.clone(),
                    })
                }
                GenericSourceConnection::LoadGenerator(ref c) => match c.load_generator {
                    LoadGenerator::Auction
                    | LoadGenerator::Marketing
                    | LoadGenerator::Tpch { .. } => SourceExportDetails::None,
                    LoadGenerator::Counter { .. }
                    | LoadGenerator::Clock
                    | LoadGenerator::Datums
                    | LoadGenerator::KeyValue(_) => {
                        SourceExportDetails::LoadGenerator(LoadGeneratorSourceExportDetails {
                            output: LoadGeneratorOutput::Default,
                        })
                    }
                },
                GenericSourceConnection::Postgres(_)
                | GenericSourceConnection::MySql(_)
                | GenericSourceConnection::SqlServer(_) => SourceExportDetails::None,
            };

            let data_source = DataSourceDesc::OldSyntaxIngestion {
                desc: SourceDesc {
                    connection: external_connection,
                    timestamp_interval,
                },
                progress_subsource: *id,
                data_config: SourceExportDataConfig {
                    encoding,
                    envelope: envelope.clone(),
                },
                details,
            };
            (desc, data_source)
        }
        None => {
            let desc = external_connection.timestamp_desc();
            let data_source = DataSourceDesc::Ingestion(SourceDesc {
                connection: external_connection,
                timestamp_interval,
            });
            (desc, data_source)
        }
    };

    let if_not_exists = *if_not_exists;
    let name = scx.allocate_qualified_name(normalize::unresolved_item_name(name.clone())?)?;

    // Check for an object in the catalog with this same name
    let full_name = scx.catalog.resolve_full_name(&name);
    let partial_name = PartialItemName::from(full_name.clone());
    // For PostgreSQL compatibility, we need to prevent creating sources when
    // there is an existing object *or* type of the same name.
    if let (false, Ok(item)) = (
        if_not_exists,
        scx.catalog.resolve_item_or_type(&partial_name),
    ) {
        return Err(PlanError::ItemAlreadyExists {
            name: full_name.to_string(),
            item_type: item.item_type(),
        });
    }

    // We will rewrite the cluster if one is not provided, so we must use the
    // `in_cluster` value we plan to normalize when we canonicalize the create
    // statement.
    let in_cluster = source_sink_cluster_config(scx, &mut stmt.in_cluster)?;

    let create_sql = normalize::create_statement(scx, Statement::CreateSource(stmt))?;

    // Determine a default timeline for the source.
    let timeline = match envelope {
        SourceEnvelope::CdcV2 => {
            Timeline::External(scx.catalog.resolve_full_name(&name).to_string())
        }
        _ => Timeline::EpochMilliseconds,
    };

    let compaction_window = plan_retain_history_option(scx, retain_history)?;

    let source = Source {
        create_sql,
        data_source,
        desc,
        compaction_window,
    };

    Ok(Plan::CreateSource(CreateSourcePlan {
        name,
        source,
        if_not_exists,
        timeline,
        in_cluster: Some(in_cluster.id()),
    }))
}

pub fn plan_generic_source_connection(
    scx: &StatementContext<'_>,
    source_connection: &CreateSourceConnection<Aug>,
    include_metadata: &Vec<SourceIncludeMetadata>,
) -> Result<GenericSourceConnection<ReferencedConnection>, PlanError> {
    Ok(match source_connection {
        CreateSourceConnection::Kafka {
            connection,
            options,
        } => GenericSourceConnection::Kafka(plan_kafka_source_connection(
            scx,
            connection,
            options,
            include_metadata,
        )?),
        CreateSourceConnection::Postgres {
            connection,
            options,
        } => GenericSourceConnection::Postgres(plan_postgres_source_connection(
            scx, connection, options,
        )?),
        CreateSourceConnection::SqlServer {
            connection,
            options,
        } => GenericSourceConnection::SqlServer(plan_sqlserver_source_connection(
            scx, connection, options,
        )?),
        CreateSourceConnection::MySql {
            connection,
            options,
        } => {
            GenericSourceConnection::MySql(plan_mysql_source_connection(scx, connection, options)?)
        }
        CreateSourceConnection::LoadGenerator { generator, options } => {
            GenericSourceConnection::LoadGenerator(plan_load_generator_source_connection(
                scx,
                generator,
                options,
                include_metadata,
            )?)
        }
    })
}

fn plan_load_generator_source_connection(
    scx: &StatementContext<'_>,
    generator: &ast::LoadGenerator,
    options: &Vec<LoadGeneratorOption<Aug>>,
    include_metadata: &Vec<SourceIncludeMetadata>,
) -> Result<LoadGeneratorSourceConnection, PlanError> {
    let load_generator =
        load_generator_ast_to_generator(scx, generator, options, include_metadata)?;
    let LoadGeneratorOptionExtracted {
        tick_interval,
        as_of,
        up_to,
        ..
    } = options.clone().try_into()?;
    let tick_micros = match tick_interval {
        Some(interval) => Some(interval.as_micros().try_into()?),
        None => None,
    };
    if up_to < as_of {
        sql_bail!("UP TO cannot be less than AS OF");
    }
    Ok(LoadGeneratorSourceConnection {
        load_generator,
        tick_micros,
        as_of,
        up_to,
    })
}

fn plan_mysql_source_connection(
    scx: &StatementContext<'_>,
    connection: &ResolvedItemName,
    options: &Vec<MySqlConfigOption<Aug>>,
) -> Result<MySqlSourceConnection<ReferencedConnection>, PlanError> {
    let connection_item = scx.get_item_by_resolved_name(connection)?;
    match connection_item.connection()? {
        Connection::MySql(connection) => connection,
        _ => sql_bail!(
            "{} is not a MySQL connection",
            scx.catalog.resolve_full_name(connection_item.name())
        ),
    };
    let MySqlConfigOptionExtracted {
        details,
        // text/exclude columns are already part of the source-exports and are only included
        // in these options for round-tripping of a `CREATE SOURCE` statement. This should
        // be removed once we drop support for implicitly created subsources.
        text_columns: _,
        exclude_columns: _,
        seen: _,
    } = options.clone().try_into()?;
    let details = details
        .as_ref()
        .ok_or_else(|| internal_err!("MySQL source missing details"))?;
    let details = hex::decode(details).map_err(|e| sql_err!("{}", e))?;
    let details = ProtoMySqlSourceDetails::decode(&*details).map_err(|e| sql_err!("{}", e))?;
    let details = MySqlSourceDetails::from_proto(details).map_err(|e| sql_err!("{}", e))?;
    Ok(MySqlSourceConnection {
        connection: connection_item.id(),
        connection_id: connection_item.id(),
        details,
    })
}

fn plan_sqlserver_source_connection(
    scx: &StatementContext<'_>,
    connection: &ResolvedItemName,
    options: &Vec<SqlServerConfigOption<Aug>>,
) -> Result<SqlServerSourceConnection<ReferencedConnection>, PlanError> {
    let connection_item = scx.get_item_by_resolved_name(connection)?;
    match connection_item.connection()? {
        Connection::SqlServer(connection) => connection,
        _ => sql_bail!(
            "{} is not a SQL Server connection",
            scx.catalog.resolve_full_name(connection_item.name())
        ),
    };
    let SqlServerConfigOptionExtracted { details, .. } = options.clone().try_into()?;
    let details = details
        .as_ref()
        .ok_or_else(|| internal_err!("SQL Server source missing details"))?;
    let extras = hex::decode(details)
        .map_err(|e| sql_err!("{e}"))
        .and_then(|raw| ProtoSqlServerSourceExtras::decode(&*raw).map_err(|e| sql_err!("{e}")))
        .and_then(|proto| SqlServerSourceExtras::from_proto(proto).map_err(|e| sql_err!("{e}")))?;
    Ok(SqlServerSourceConnection {
        connection_id: connection_item.id(),
        connection: connection_item.id(),
        extras,
    })
}

fn plan_postgres_source_connection(
    scx: &StatementContext<'_>,
    connection: &ResolvedItemName,
    options: &Vec<PgConfigOption<Aug>>,
) -> Result<PostgresSourceConnection<ReferencedConnection>, PlanError> {
    let connection_item = scx.get_item_by_resolved_name(connection)?;
    let PgConfigOptionExtracted {
        details,
        publication,
        failover,
        // text columns are already part of the source-exports and are only included
        // in these options for round-tripping of a `CREATE SOURCE` statement. This should
        // be removed once we drop support for implicitly created subsources.
        text_columns: _,
        // exclude columns are already part of the source-exports and are only included
        // in these options for round-tripping of a `CREATE SOURCE` statement. This should
        // be removed once we drop support for implicitly created subsources.
        exclude_columns: _,
        seen,
    } = options.clone().try_into()?;
    if seen.contains(&PgConfigOptionName::Failover) {
        scx.require_feature_flag(&vars::ENABLE_POSTGRES_SOURCE_FAILOVER)?;
    }
    let details = details
        .as_ref()
        .ok_or_else(|| internal_err!("Postgres source missing details"))?;
    let details = hex::decode(details).map_err(|e| sql_err!("{}", e))?;
    let details =
        ProtoPostgresSourcePublicationDetails::decode(&*details).map_err(|e| sql_err!("{}", e))?;
    let publication_details =
        PostgresSourcePublicationDetails::from_proto(details).map_err(|e| sql_err!("{}", e))?;
    Ok(PostgresSourceConnection {
        connection: connection_item.id(),
        connection_id: connection_item.id(),
        // Validated during purification.
        publication: publication.ok_or_else(|| internal_err!("PUBLICATION option is required"))?,
        publication_details,
        failover,
    })
}

fn plan_kafka_source_connection(
    scx: &StatementContext<'_>,
    connection_name: &ResolvedItemName,
    options: &Vec<ast::KafkaSourceConfigOption<Aug>>,
    include_metadata: &Vec<SourceIncludeMetadata>,
) -> Result<KafkaSourceConnection<ReferencedConnection>, PlanError> {
    let connection_item = scx.get_item_by_resolved_name(connection_name)?;
    if !matches!(connection_item.connection()?, Connection::Kafka(_)) {
        sql_bail!(
            "{} is not a kafka connection",
            scx.catalog.resolve_full_name(connection_item.name())
        )
    }
    let KafkaSourceConfigOptionExtracted {
        group_id_prefix,
        topic,
        topic_metadata_refresh_interval,
        start_timestamp: _, // purified into `start_offset`
        start_offset,
        seen: _,
    }: KafkaSourceConfigOptionExtracted = options.clone().try_into()?;
    // Validated during purification.
    let topic = topic.ok_or_else(|| internal_err!("TOPIC option is required"))?;
    let mut start_offsets = BTreeMap::new();
    if let Some(offsets) = start_offset {
        for (part, offset) in offsets.iter().enumerate() {
            if *offset < 0 {
                sql_bail!("START OFFSET must be a nonnegative integer");
            }
            start_offsets.insert(i32::try_from(part)?, *offset);
        }
    }
    if topic_metadata_refresh_interval > Duration::from_secs(60 * 60) {
        // This is a librdkafka-enforced restriction that, if violated,
        // would result in a runtime error for the source.
        sql_bail!("TOPIC METADATA REFRESH INTERVAL cannot be greater than 1 hour");
    }
    if topic_metadata_refresh_interval < Duration::from_secs(1) {
        // This is a librdkafka-enforced restriction that, if violated,
        // would result in a runtime error for the source.
        sql_bail!("TOPIC METADATA REFRESH INTERVAL must be at least 1 second");
    }
    let metadata_columns = include_metadata
        .into_iter()
        .flat_map(|item| match item {
            SourceIncludeMetadata::Timestamp { alias } => {
                let name = match alias {
                    Some(name) => name.to_string(),
                    None => "timestamp".to_owned(),
                };
                Some((name, KafkaMetadataKind::Timestamp))
            }
            SourceIncludeMetadata::Partition { alias } => {
                let name = match alias {
                    Some(name) => name.to_string(),
                    None => "partition".to_owned(),
                };
                Some((name, KafkaMetadataKind::Partition))
            }
            SourceIncludeMetadata::Offset { alias } => {
                let name = match alias {
                    Some(name) => name.to_string(),
                    None => "offset".to_owned(),
                };
                Some((name, KafkaMetadataKind::Offset))
            }
            SourceIncludeMetadata::Headers { alias } => {
                let name = match alias {
                    Some(name) => name.to_string(),
                    None => "headers".to_owned(),
                };
                Some((name, KafkaMetadataKind::Headers))
            }
            SourceIncludeMetadata::Header {
                alias,
                key,
                use_bytes,
            } => Some((
                alias.to_string(),
                KafkaMetadataKind::Header {
                    key: key.clone(),
                    use_bytes: *use_bytes,
                },
            )),
            SourceIncludeMetadata::Key { .. } => {
                // handled below
                None
            }
        })
        .collect();
    Ok(KafkaSourceConnection {
        connection: connection_item.id(),
        connection_id: connection_item.id(),
        topic,
        start_offsets,
        group_id_prefix,
        topic_metadata_refresh_interval,
        metadata_columns,
    })
}

fn apply_source_envelope_encoding(
    scx: &StatementContext,
    envelope: &ast::SourceEnvelope,
    format: &Option<FormatSpecifier<Aug>>,
    key_desc: Option<RelationDesc>,
    value_desc: RelationDesc,
    include_metadata: &[SourceIncludeMetadata],
    metadata_columns_desc: Vec<(&str, SqlColumnType)>,
    source_connection: &GenericSourceConnection<ReferencedConnection>,
) -> Result<
    (
        RelationDesc,
        SourceEnvelope,
        Option<SourceDataEncoding<ReferencedConnection>>,
    ),
    PlanError,
> {
    let encoding = match format {
        Some(format) => Some(get_encoding(scx, format, envelope)?),
        None => None,
    };

    let (key_desc, value_desc) = match &encoding {
        Some(encoding) => {
            // If we are applying an encoding we need to ensure that the incoming value_desc is a
            // single column of type bytes.
            match value_desc.typ().columns() {
                [typ] => match typ.scalar_type {
                    SqlScalarType::Bytes => {}
                    _ => sql_bail!(
                        "The schema produced by the source is incompatible with format decoding"
                    ),
                },
                _ => sql_bail!(
                    "The schema produced by the source is incompatible with format decoding"
                ),
            }

            let (key_desc, value_desc) = encoding.desc()?;

            // TODO(petrosagg): This piece of code seems to be making a statement about the
            // nullability of the NONE envelope when the source is Kafka. As written, the code
            // misses opportunities to mark columns as not nullable and is over conservative. For
            // example in the case of `FORMAT BYTES ENVELOPE NONE` the output is indeed
            // non-nullable but we will mark it as nullable anyway. This kind of crude reasoning
            // should be replaced with precise type-level reasoning.
            let key_desc = key_desc.map(|desc| {
                let is_kafka = matches!(source_connection, GenericSourceConnection::Kafka(_));
                let is_envelope_none = matches!(envelope, ast::SourceEnvelope::None);
                if is_kafka && is_envelope_none {
                    RelationDesc::from_names_and_types(
                        desc.into_iter()
                            .map(|(name, typ)| (name, typ.nullable(true))),
                    )
                } else {
                    desc
                }
            });
            (key_desc, value_desc)
        }
        None => (key_desc, value_desc),
    };

    // KEY VALUE load generators are the only UPSERT source that
    // has no encoding but defaults to `INCLUDE KEY`.
    //
    // As discussed
    // <https://github.com/MaterializeInc/materialize/pull/26246#issuecomment-2023558097>,
    // removing this special case amounts to deciding how to handle null keys
    // from sources, in a holistic way. We aren't yet prepared to do this, so we leave
    // this special case in.
    //
    // Note that this is safe because this generator is
    // 1. The only source with no encoding that can have its key included.
    // 2. Never produces null keys (or values, for that matter).
    let key_envelope_no_encoding = matches!(
        source_connection,
        GenericSourceConnection::LoadGenerator(LoadGeneratorSourceConnection {
            load_generator: LoadGenerator::KeyValue(_),
            ..
        })
    );
    let mut key_envelope = get_key_envelope(
        include_metadata,
        encoding.as_ref(),
        key_envelope_no_encoding,
    )?;

    match (&envelope, &key_envelope) {
        (ast::SourceEnvelope::Debezium, KeyEnvelope::None) => {}
        (ast::SourceEnvelope::Debezium, _) => sql_bail!(
            "Cannot use INCLUDE KEY with ENVELOPE DEBEZIUM: Debezium values include all keys."
        ),
        _ => {}
    };

    // Not all source envelopes are compatible with all source connections.
    // Whoever constructs the source ingestion pipeline is responsible for
    // choosing compatible envelopes and connections.
    //
    // TODO(guswynn): ambiguously assert which connections and envelopes are
    // compatible in typechecking
    //
    // TODO: remove bails as more support for upsert is added.
    let envelope = match &envelope {
        // TODO: fixup key envelope
        ast::SourceEnvelope::None => UnplannedSourceEnvelope::None(key_envelope),
        ast::SourceEnvelope::Debezium => {
            //TODO check that key envelope is not set
            let after_idx = match typecheck_debezium(&value_desc) {
                Ok((_before_idx, after_idx)) => Ok(after_idx),
                Err(type_err) => match encoding.as_ref().map(|e| &e.value) {
                    Some(DataEncoding::Avro(_)) => Err(type_err),
                    _ => Err(sql_err!(
                        "ENVELOPE DEBEZIUM requires that VALUE FORMAT is set to AVRO"
                    )),
                },
            }?;

            UnplannedSourceEnvelope::Upsert {
                style: UpsertStyle::Debezium { after_idx },
            }
        }
        ast::SourceEnvelope::Upsert {
            value_decode_err_policy,
        } => {
            let key_encoding = match encoding.as_ref().and_then(|e| e.key.as_ref()) {
                None => {
                    if !key_envelope_no_encoding {
                        bail_unsupported!(format!(
                            "UPSERT requires a key/value format: {:?}",
                            format
                        ))
                    }
                    None
                }
                Some(key_encoding) => Some(key_encoding),
            };
            // `ENVELOPE UPSERT` implies `INCLUDE KEY`, if it is not explicitly
            // specified.
            if key_envelope == KeyEnvelope::None {
                key_envelope = get_unnamed_key_envelope(key_encoding)?;
            }
            // If the value decode error policy is not set we use the default upsert style.
            let style = match value_decode_err_policy.as_slice() {
                [] => UpsertStyle::Default(key_envelope),
                [SourceErrorPolicy::Inline { alias }] => {
                    scx.require_feature_flag(&vars::ENABLE_ENVELOPE_UPSERT_INLINE_ERRORS)?;
                    UpsertStyle::ValueErrInline {
                        key_envelope,
                        error_column: alias
                            .as_ref()
                            .map_or_else(|| "error".to_string(), |a| a.to_string()),
                    }
                }
                _ => {
                    bail_unsupported!("ENVELOPE UPSERT with unsupported value decode error policy")
                }
            };

            UnplannedSourceEnvelope::Upsert { style }
        }
        ast::SourceEnvelope::CdcV2 => {
            scx.require_feature_flag(&vars::ENABLE_ENVELOPE_MATERIALIZE)?;
            //TODO check that key envelope is not set
            match format {
                Some(FormatSpecifier::Bare(Format::Avro(_))) => {}
                _ => bail_unsupported!("non-Avro-encoded ENVELOPE MATERIALIZE"),
            }
            UnplannedSourceEnvelope::CdcV2
        }
    };

    let metadata_desc = included_column_desc(metadata_columns_desc);
    let (envelope, desc) = envelope.desc(key_desc, value_desc, metadata_desc)?;

    Ok((desc, envelope, encoding))
}

/// Plans the RelationDesc for a source export (subsource or table) that has a defined list
/// of columns and constraints.
fn plan_source_export_desc(
    scx: &StatementContext,
    name: &UnresolvedItemName,
    columns: &Vec<ColumnDef<Aug>>,
    constraints: &Vec<TableConstraint<Aug>>,
) -> Result<RelationDesc, PlanError> {
    let names: Vec<_> = columns
        .iter()
        .map(|c| normalize::column_name(c.name.clone()))
        .collect();

    if let Some(dup) = names.iter().duplicates().next() {
        sql_bail!("column {} specified more than once", dup.quoted());
    }

    // Build initial relation type that handles declared data types
    // and NOT NULL constraints.
    let mut column_types = Vec::with_capacity(columns.len());
    let mut keys = Vec::new();

    for (i, c) in columns.into_iter().enumerate() {
        let aug_data_type = &c.data_type;
        let ty = query::scalar_type_from_sql(scx, aug_data_type)?;
        let mut nullable = true;
        for option in &c.options {
            match &option.option {
                ColumnOption::NotNull => nullable = false,
                ColumnOption::Default(_) => {
                    bail_unsupported!("Source export with default value")
                }
                ColumnOption::Unique { is_primary } => {
                    keys.push(vec![i]);
                    if *is_primary {
                        nullable = false;
                    }
                }
                other => {
                    bail_unsupported!(format!("Source export with column constraint: {}", other))
                }
            }
        }
        column_types.push(ty.nullable(nullable));
    }

    let mut seen_primary = false;
    'c: for constraint in constraints {
        match constraint {
            TableConstraint::Unique {
                name: _,
                columns,
                is_primary,
                nulls_not_distinct,
            } => {
                if seen_primary && *is_primary {
                    sql_bail!(
                        "multiple primary keys for source export {} are not allowed",
                        name.to_ast_string_stable()
                    );
                }
                seen_primary = *is_primary || seen_primary;

                let mut key = vec![];
                for column in columns {
                    let column = normalize::column_name(column.clone());
                    match names.iter().position(|name| *name == column) {
                        None => sql_bail!("unknown column in constraint: {}", column),
                        Some(i) => {
                            let nullable = &mut column_types[i].nullable;
                            if *is_primary {
                                if *nulls_not_distinct {
                                    sql_bail!(
                                        "[internal error] PRIMARY KEY does not support NULLS NOT DISTINCT"
                                    );
                                }
                                *nullable = false;
                            } else if !(*nulls_not_distinct || !*nullable) {
                                // Non-primary key unique constraints are only keys if all of their
                                // columns are `NOT NULL` or the constraint is `NULLS NOT DISTINCT`.
                                break 'c;
                            }

                            key.push(i);
                        }
                    }
                }

                if *is_primary {
                    keys.insert(0, key);
                } else {
                    keys.push(key);
                }
            }
            TableConstraint::ForeignKey { .. } => {
                bail_unsupported!("Source export with a foreign key")
            }
            TableConstraint::Check { .. } => {
                bail_unsupported!("Source export with a check constraint")
            }
        }
    }

    let typ = SqlRelationType::new(column_types).with_keys(keys);
    let desc = RelationDesc::new(typ, names);
    Ok(desc)
}

generate_extracted_config!(
    CreateSubsourceOption,
    (Progress, bool, Default(false)),
    (ExternalReference, UnresolvedItemName),
    (RetainHistory, OptionalDuration),
    (TextColumns, Vec::<Ident>, Default(vec![])),
    (ExcludeColumns, Vec::<Ident>, Default(vec![])),
    (Details, String)
);

pub fn plan_create_subsource(
    scx: &StatementContext,
    stmt: CreateSubsourceStatement<Aug>,
) -> Result<Plan, PlanError> {
    let CreateSubsourceStatement {
        name,
        columns,
        of_source,
        constraints,
        if_not_exists,
        with_options,
    } = &stmt;

    let CreateSubsourceOptionExtracted {
        progress,
        retain_history,
        external_reference,
        text_columns,
        exclude_columns,
        details,
        seen: _,
    } = with_options.clone().try_into()?;

    // This invariant is enforced during purification; we are responsible for
    // creating the AST for subsources as a response to CREATE SOURCE
    // statements, so this would fire in integration testing if we failed to
    // uphold it.
    if !(progress ^ (external_reference.is_some() && of_source.is_some())) {
        bail_internal!(
            "CREATE SUBSOURCE statement must specify either PROGRESS or REFERENCES option"
        );
    }

    let desc = plan_source_export_desc(scx, name, columns, constraints)?;

    let data_source = if let Some(source_reference) = of_source {
        // If the new source table syntax is forced we should not be creating any non-progress
        // subsources.
        if scx.catalog.system_vars().enable_create_table_from_source()
            && scx.catalog.system_vars().force_source_table_syntax()
        {
            Err(PlanError::UseTablesForSources(
                "CREATE SUBSOURCE".to_string(),
            ))?;
        }

        // This is a subsource with the "natural" dependency order, i.e. it is
        // not a legacy subsource with the inverted structure.
        let ingestion_id = *source_reference.item_id();
        let external_reference = external_reference.ok_or_else(|| {
            sql_err!("CREATE SUBSOURCE with REFERENCES requires EXTERNAL REFERENCE option")
        })?;

        // Decode the details option stored on the subsource statement, which contains information
        // created during the purification process.
        let details = details
            .as_ref()
            .ok_or_else(|| internal_err!("source-export subsource missing details"))?;
        let details = hex::decode(details).map_err(|e| sql_err!("{}", e))?;
        let details =
            ProtoSourceExportStatementDetails::decode(&*details).map_err(|e| sql_err!("{}", e))?;
        let details =
            SourceExportStatementDetails::from_proto(details).map_err(|e| sql_err!("{}", e))?;
        let details = match details {
            SourceExportStatementDetails::Postgres {
                table,
                cast_oid_full_range,
                initial_lsn,
            } => SourceExportDetails::Postgres(PostgresSourceExportDetails {
                column_casts: crate::pure::postgres::generate_column_casts(
                    scx,
                    &table,
                    &text_columns,
                    cast_oid_full_range,
                )?,
                table,
                initial_lsn,
            }),
            SourceExportStatementDetails::MySql {
                table,
                initial_gtid_set,
                binlog_full_metadata,
            } => SourceExportDetails::MySql(MySqlSourceExportDetails {
                table,
                initial_gtid_set,
                text_columns: text_columns.into_iter().map(|c| c.into_string()).collect(),
                exclude_columns: exclude_columns
                    .into_iter()
                    .map(|c| c.into_string())
                    .collect(),
                binlog_full_metadata,
            }),
            SourceExportStatementDetails::SqlServer {
                table,
                capture_instance,
                initial_lsn,
            } => SourceExportDetails::SqlServer(SqlServerSourceExportDetails {
                capture_instance,
                table,
                initial_lsn,
                text_columns: text_columns.into_iter().map(|c| c.into_string()).collect(),
                exclude_columns: exclude_columns
                    .into_iter()
                    .map(|c| c.into_string())
                    .collect(),
            }),
            SourceExportStatementDetails::LoadGenerator { output } => {
                SourceExportDetails::LoadGenerator(LoadGeneratorSourceExportDetails { output })
            }
            SourceExportStatementDetails::Kafka {} => {
                bail_unsupported!("subsources cannot reference Kafka sources")
            }
        };
        DataSourceDesc::IngestionExport {
            ingestion_id,
            external_reference,
            details,
            // Subsources don't currently support non-default envelopes / encoding
            data_config: SourceExportDataConfig {
                envelope: SourceEnvelope::None(NoneEnvelope {
                    key_envelope: KeyEnvelope::None,
                    key_arity: 0,
                }),
                encoding: None,
            },
        }
    } else if progress {
        DataSourceDesc::Progress
    } else {
        sql_bail!("CREATE SUBSOURCE must specify one of PROGRESS or REFERENCES option")
    };

    let if_not_exists = *if_not_exists;
    let name = scx.allocate_qualified_name(normalize::unresolved_item_name(name.clone())?)?;

    let create_sql = normalize::create_statement(scx, Statement::CreateSubsource(stmt))?;

    let compaction_window = plan_retain_history_option(scx, retain_history)?;
    let source = Source {
        create_sql,
        data_source,
        desc,
        compaction_window,
    };

    Ok(Plan::CreateSource(CreateSourcePlan {
        name,
        source,
        if_not_exists,
        timeline: Timeline::EpochMilliseconds,
        in_cluster: None,
    }))
}

generate_extracted_config!(
    TableFromSourceOption,
    (TextColumns, Vec::<Ident>, Default(vec![])),
    (ExcludeColumns, Vec::<Ident>, Default(vec![])),
    (ExcludeConstraints, Vec::<String>, Default(vec![])),
    (ExcludeAllConstraints, bool, Default(false)),
    (ExcludeAllNotNullConstraints, bool, Default(false)),
    (PartitionBy, Vec<Ident>),
    (RetainHistory, OptionalDuration),
    (Details, String)
);

pub fn plan_create_table_from_source(
    scx: &StatementContext,
    stmt: CreateTableFromSourceStatement<Aug>,
) -> Result<Plan, PlanError> {
    if !scx.catalog.system_vars().enable_create_table_from_source() {
        sql_bail!("CREATE TABLE ... FROM SOURCE is not supported");
    }

    let CreateTableFromSourceStatement {
        name,
        columns,
        constraints,
        if_not_exists,
        source,
        external_reference,
        envelope,
        format,
        include_metadata,
        with_options,
    } = &stmt;

    let envelope = envelope.clone().unwrap_or(ast::SourceEnvelope::None);

    let TableFromSourceOptionExtracted {
        text_columns,
        exclude_columns,
        exclude_constraints: _,
        exclude_all_constraints: _,
        exclude_all_not_null_constraints: _,
        retain_history,
        partition_by,
        details,
        seen: _,
    } = with_options.clone().try_into()?;

    let source_item = scx.get_item_by_resolved_name(source)?;
    let ingestion_id = source_item.id();

    // Decode the details option stored on the statement, which contains information
    // created during the purification process.
    let details = details
        .as_ref()
        .ok_or_else(|| internal_err!("source-export missing details"))?;
    let details = hex::decode(details).map_err(|e| sql_err!("{}", e))?;
    let details =
        ProtoSourceExportStatementDetails::decode(&*details).map_err(|e| sql_err!("{}", e))?;
    let details =
        SourceExportStatementDetails::from_proto(details).map_err(|e| sql_err!("{}", e))?;

    if !matches!(details, SourceExportStatementDetails::Kafka { .. })
        && include_metadata
            .iter()
            .any(|sic| matches!(sic, SourceIncludeMetadata::Headers { .. }))
    {
        // TODO(guswynn): should this be `bail_unsupported!`?
        sql_bail!("INCLUDE HEADERS with non-Kafka source table not supported");
    }
    if !matches!(
        details,
        SourceExportStatementDetails::Kafka { .. }
            | SourceExportStatementDetails::LoadGenerator { .. }
    ) && !include_metadata.is_empty()
    {
        bail_unsupported!("INCLUDE metadata with non-Kafka source table");
    }

    let details = match details {
        SourceExportStatementDetails::Postgres {
            table,
            cast_oid_full_range,
            initial_lsn,
        } => SourceExportDetails::Postgres(PostgresSourceExportDetails {
            column_casts: crate::pure::postgres::generate_column_casts(
                scx,
                &table,
                &text_columns,
                cast_oid_full_range,
            )?,
            table,
            initial_lsn,
        }),
        SourceExportStatementDetails::MySql {
            table,
            initial_gtid_set,
            binlog_full_metadata,
        } => SourceExportDetails::MySql(MySqlSourceExportDetails {
            table,
            initial_gtid_set,
            text_columns: text_columns.into_iter().map(|c| c.into_string()).collect(),
            exclude_columns: exclude_columns
                .into_iter()
                .map(|c| c.into_string())
                .collect(),
            binlog_full_metadata,
        }),
        SourceExportStatementDetails::SqlServer {
            table,
            capture_instance,
            initial_lsn,
        } => SourceExportDetails::SqlServer(SqlServerSourceExportDetails {
            table,
            capture_instance,
            initial_lsn,
            text_columns: text_columns.into_iter().map(|c| c.into_string()).collect(),
            exclude_columns: exclude_columns
                .into_iter()
                .map(|c| c.into_string())
                .collect(),
        }),
        SourceExportStatementDetails::LoadGenerator { output } => {
            SourceExportDetails::LoadGenerator(LoadGeneratorSourceExportDetails { output })
        }
        SourceExportStatementDetails::Kafka {} => {
            if !include_metadata.is_empty()
                && !matches!(
                    envelope,
                    ast::SourceEnvelope::Upsert { .. }
                        | ast::SourceEnvelope::None
                        | ast::SourceEnvelope::Debezium
                )
            {
                // TODO(guswynn): should this be `bail_unsupported!`?
                sql_bail!("INCLUDE <metadata> requires ENVELOPE (NONE|UPSERT|DEBEZIUM)");
            }

            let metadata_columns = include_metadata
                .into_iter()
                .flat_map(|item| match item {
                    SourceIncludeMetadata::Timestamp { alias } => {
                        let name = match alias {
                            Some(name) => name.to_string(),
                            None => "timestamp".to_owned(),
                        };
                        Some((name, KafkaMetadataKind::Timestamp))
                    }
                    SourceIncludeMetadata::Partition { alias } => {
                        let name = match alias {
                            Some(name) => name.to_string(),
                            None => "partition".to_owned(),
                        };
                        Some((name, KafkaMetadataKind::Partition))
                    }
                    SourceIncludeMetadata::Offset { alias } => {
                        let name = match alias {
                            Some(name) => name.to_string(),
                            None => "offset".to_owned(),
                        };
                        Some((name, KafkaMetadataKind::Offset))
                    }
                    SourceIncludeMetadata::Headers { alias } => {
                        let name = match alias {
                            Some(name) => name.to_string(),
                            None => "headers".to_owned(),
                        };
                        Some((name, KafkaMetadataKind::Headers))
                    }
                    SourceIncludeMetadata::Header {
                        alias,
                        key,
                        use_bytes,
                    } => Some((
                        alias.to_string(),
                        KafkaMetadataKind::Header {
                            key: key.clone(),
                            use_bytes: *use_bytes,
                        },
                    )),
                    SourceIncludeMetadata::Key { .. } => {
                        // handled below
                        None
                    }
                })
                .collect();

            SourceExportDetails::Kafka(KafkaSourceExportDetails { metadata_columns })
        }
    };

    let source_connection = &source_item
        .source_desc()?
        .ok_or_else(|| sql_err!("item is not a source"))?
        .connection;

    // Some source-types (e.g. postgres, mysql, multi-output load-gen sources) define a value_schema
    // during purification and define the `columns` and `constraints` fields for the statement,
    // whereas other source-types (e.g. kafka, single-output load-gen sources) do not, so instead
    // we use the source connection's default schema.
    let (key_desc, value_desc) =
        if matches!(columns, TableFromSourceColumns::Defined(_)) || !constraints.is_empty() {
            let columns = match columns {
                TableFromSourceColumns::Defined(columns) => columns,
                _ => bail_internal!("expected column definitions to be present"),
            };
            let desc = plan_source_export_desc(scx, name, columns, constraints)?;
            (None, desc)
        } else {
            let key_desc = source_connection.default_key_desc();
            let value_desc = source_connection.default_value_desc();
            (Some(key_desc), value_desc)
        };

    let metadata_columns_desc = match &details {
        SourceExportDetails::Kafka(KafkaSourceExportDetails {
            metadata_columns, ..
        }) => kafka_metadata_columns_desc(metadata_columns),
        _ => vec![],
    };

    let (mut desc, envelope, encoding) = apply_source_envelope_encoding(
        scx,
        &envelope,
        format,
        key_desc,
        value_desc,
        include_metadata,
        metadata_columns_desc,
        source_connection,
    )?;
    if let TableFromSourceColumns::Named(col_names) = columns {
        plan_utils::maybe_rename_columns(format!("source table {}", name), &mut desc, col_names)?;
    }

    let names: Vec<_> = desc.iter_names().cloned().collect();
    if let Some(dup) = names.iter().duplicates().next() {
        sql_bail!("column {} specified more than once", dup.quoted());
    }

    let name = scx.allocate_qualified_name(normalize::unresolved_item_name(name.clone())?)?;

    // Allow users to specify a timeline. If they do not, determine a default
    // timeline for the source.
    let timeline = match envelope {
        SourceEnvelope::CdcV2 => {
            Timeline::External(scx.catalog.resolve_full_name(&name).to_string())
        }
        _ => Timeline::EpochMilliseconds,
    };

    if let Some(partition_by) = partition_by {
        scx.require_feature_flag(&ENABLE_COLLECTION_PARTITION_BY)?;
        check_partition_by(&desc, partition_by)?;
    }

    let data_source = DataSourceDesc::IngestionExport {
        ingestion_id,
        // Populated during purification.
        external_reference: external_reference
            .as_ref()
            .ok_or_else(|| sql_err!("EXTERNAL REFERENCE is required"))?
            .clone(),
        details,
        data_config: SourceExportDataConfig { envelope, encoding },
    };

    let if_not_exists = *if_not_exists;

    let create_sql = normalize::create_statement(scx, Statement::CreateTableFromSource(stmt))?;

    let compaction_window = plan_retain_history_option(scx, retain_history)?;
    let table = Table {
        create_sql,
        desc: VersionedRelationDesc::new(desc),
        temporary: false,
        compaction_window,
        data_source: TableDataSource::DataSource {
            desc: data_source,
            timeline,
        },
    };

    Ok(Plan::CreateTable(CreateTablePlan {
        name,
        table,
        if_not_exists,
    }))
}

generate_extracted_config!(
    LoadGeneratorOption,
    (TickInterval, Duration),
    (AsOf, u64, Default(0_u64)),
    (UpTo, u64, Default(u64::MAX)),
    (ScaleFactor, f64),
    (MaxCardinality, u64),
    (Keys, u64),
    (SnapshotRounds, u64),
    (TransactionalSnapshot, bool),
    (ValueSize, u64),
    (Seed, u64),
    (Partitions, u64),
    (BatchSize, u64)
);

impl LoadGeneratorOptionExtracted {
    pub(super) fn ensure_only_valid_options(
        &self,
        loadgen: &ast::LoadGenerator,
    ) -> Result<(), PlanError> {
        use mz_sql_parser::ast::LoadGeneratorOptionName::*;

        let mut options = self.seen.clone();

        let permitted_options: &[_] = match loadgen {
            ast::LoadGenerator::Auction => &[TickInterval, AsOf, UpTo],
            ast::LoadGenerator::Clock => &[TickInterval, AsOf, UpTo],
            ast::LoadGenerator::Counter => &[TickInterval, AsOf, UpTo, MaxCardinality],
            ast::LoadGenerator::Marketing => &[TickInterval, AsOf, UpTo],
            ast::LoadGenerator::Datums => &[TickInterval, AsOf, UpTo],
            ast::LoadGenerator::Tpch => &[TickInterval, AsOf, UpTo, ScaleFactor],
            ast::LoadGenerator::KeyValue => &[
                TickInterval,
                Keys,
                SnapshotRounds,
                TransactionalSnapshot,
                ValueSize,
                Seed,
                Partitions,
                BatchSize,
            ],
        };

        for o in permitted_options {
            options.remove(o);
        }

        if !options.is_empty() {
            sql_bail!(
                "{} load generators do not support {} values",
                loadgen,
                options.iter().join(", ")
            )
        }

        Ok(())
    }
}

pub(crate) fn load_generator_ast_to_generator(
    scx: &StatementContext,
    loadgen: &ast::LoadGenerator,
    options: &[LoadGeneratorOption<Aug>],
    include_metadata: &[SourceIncludeMetadata],
) -> Result<LoadGenerator, PlanError> {
    let extracted: LoadGeneratorOptionExtracted = options.to_vec().try_into()?;
    extracted.ensure_only_valid_options(loadgen)?;

    if loadgen != &ast::LoadGenerator::KeyValue && !include_metadata.is_empty() {
        sql_bail!("INCLUDE metadata only supported with `KEY VALUE` load generators");
    }

    let load_generator = match loadgen {
        ast::LoadGenerator::Auction => LoadGenerator::Auction,
        ast::LoadGenerator::Clock => {
            scx.require_feature_flag(&vars::ENABLE_LOAD_GENERATOR_CLOCK)?;
            LoadGenerator::Clock
        }
        ast::LoadGenerator::Counter => {
            scx.require_feature_flag(&vars::ENABLE_LOAD_GENERATOR_COUNTER)?;
            let LoadGeneratorOptionExtracted {
                max_cardinality, ..
            } = extracted;
            LoadGenerator::Counter { max_cardinality }
        }
        ast::LoadGenerator::Marketing => LoadGenerator::Marketing,
        ast::LoadGenerator::Datums => {
            scx.require_feature_flag(&vars::ENABLE_LOAD_GENERATOR_DATUMS)?;
            LoadGenerator::Datums
        }
        ast::LoadGenerator::Tpch => {
            let LoadGeneratorOptionExtracted { scale_factor, .. } = extracted;

            // Default to 0.01 scale factor (=10MB).
            let sf: f64 = scale_factor.unwrap_or(0.01);
            if !sf.is_finite() || sf < 0.0 {
                sql_bail!("unsupported scale factor {sf}");
            }

            let f_to_i = |multiplier: f64| -> Result<i64, PlanError> {
                let total = (sf * multiplier).floor();
                let mut i = i64::try_cast_from(total)
                    .ok_or_else(|| sql_err!("unsupported scale factor {sf}"))?;
                if i < 1 {
                    i = 1;
                }
                Ok(i)
            };

            // The multiplications here are safely unchecked because they will
            // overflow to infinity, which will be caught by f64_to_i64.
            let count_supplier = f_to_i(10_000f64)?;
            let count_part = f_to_i(200_000f64)?;
            let count_customer = f_to_i(150_000f64)?;
            let count_orders = f_to_i(150_000f64 * 10f64)?;
            let count_clerk = f_to_i(1_000f64)?;

            LoadGenerator::Tpch {
                count_supplier,
                count_part,
                count_customer,
                count_orders,
                count_clerk,
            }
        }
        mz_sql_parser::ast::LoadGenerator::KeyValue => {
            scx.require_feature_flag(&vars::ENABLE_LOAD_GENERATOR_KEY_VALUE)?;
            let LoadGeneratorOptionExtracted {
                keys,
                snapshot_rounds,
                transactional_snapshot,
                value_size,
                tick_interval,
                seed,
                partitions,
                batch_size,
                ..
            } = extracted;

            let mut include_offset = None;
            for im in include_metadata {
                match im {
                    SourceIncludeMetadata::Offset { alias } => {
                        include_offset = match alias {
                            Some(alias) => Some(alias.to_string()),
                            None => Some(LOAD_GENERATOR_KEY_VALUE_OFFSET_DEFAULT.to_string()),
                        }
                    }
                    SourceIncludeMetadata::Key { .. } => continue,

                    _ => {
                        sql_bail!("only `INCLUDE OFFSET` and `INCLUDE KEY` is supported");
                    }
                };
            }

            let lgkv = KeyValueLoadGenerator {
                keys: keys.ok_or_else(|| sql_err!("LOAD GENERATOR KEY VALUE requires KEYS"))?,
                snapshot_rounds: snapshot_rounds
                    .ok_or_else(|| sql_err!("LOAD GENERATOR KEY VALUE requires SNAPSHOT ROUNDS"))?,
                // Defaults to true.
                transactional_snapshot: transactional_snapshot.unwrap_or(true),
                value_size: value_size
                    .ok_or_else(|| sql_err!("LOAD GENERATOR KEY VALUE requires VALUE SIZE"))?,
                partitions: partitions
                    .ok_or_else(|| sql_err!("LOAD GENERATOR KEY VALUE requires PARTITIONS"))?,
                tick_interval,
                batch_size: batch_size
                    .ok_or_else(|| sql_err!("LOAD GENERATOR KEY VALUE requires BATCH SIZE"))?,
                seed: seed.ok_or_else(|| sql_err!("LOAD GENERATOR KEY VALUE requires SEED"))?,
                include_offset,
            };

            if lgkv.keys == 0
                || lgkv.partitions == 0
                || lgkv.value_size == 0
                || lgkv.batch_size == 0
            {
                sql_bail!("LOAD GENERATOR KEY VALUE options must be non-zero")
            }

            if lgkv.keys % lgkv.partitions != 0 {
                sql_bail!("KEYS must be a multiple of PARTITIONS")
            }

            if lgkv.batch_size > lgkv.keys {
                sql_bail!("KEYS must be larger than BATCH SIZE")
            }

            // This constraints simplifies the source implementation.
            // We can lift it later.
            if (lgkv.keys / lgkv.partitions) % lgkv.batch_size != 0 {
                sql_bail!("PARTITIONS * BATCH SIZE must be a divisor of KEYS")
            }

            if lgkv.snapshot_rounds == 0 {
                sql_bail!("SNAPSHOT ROUNDS must be larger than 0")
            }

            LoadGenerator::KeyValue(lgkv)
        }
    };

    Ok(load_generator)
}

fn typecheck_debezium(value_desc: &RelationDesc) -> Result<(Option<usize>, usize), PlanError> {
    let before = value_desc.get_by_name(&"before".into());
    let (after_idx, after_ty) = value_desc
        .get_by_name(&"after".into())
        .ok_or_else(|| sql_err!("'after' column missing from debezium input"))?;
    let before_idx = if let Some((before_idx, before_ty)) = before {
        if !matches!(before_ty.scalar_type, SqlScalarType::Record { .. }) {
            sql_bail!("'before' column must be of type record");
        }
        if before_ty != after_ty {
            sql_bail!("'before' type differs from 'after' column");
        }
        Some(before_idx)
    } else {
        None
    };
    Ok((before_idx, after_idx))
}

fn get_encoding(
    scx: &StatementContext,
    format: &FormatSpecifier<Aug>,
    envelope: &ast::SourceEnvelope,
) -> Result<SourceDataEncoding<ReferencedConnection>, PlanError> {
    let encoding = match format {
        FormatSpecifier::Bare(format) => get_encoding_inner(scx, format)?,
        FormatSpecifier::KeyValue { key, value } => {
            let key = {
                let encoding = get_encoding_inner(scx, key)?;
                Some(encoding.key.unwrap_or(encoding.value))
            };
            let value = get_encoding_inner(scx, value)?.value;
            SourceDataEncoding { key, value }
        }
    };

    let requires_keyvalue = matches!(
        envelope,
        ast::SourceEnvelope::Debezium | ast::SourceEnvelope::Upsert { .. }
    );
    let is_keyvalue = encoding.key.is_some();
    if requires_keyvalue && !is_keyvalue {
        sql_bail!("ENVELOPE [DEBEZIUM] UPSERT requires that KEY FORMAT be specified");
    };

    Ok(encoding)
}

/// Determine the cluster ID to use for this item.
///
/// If `in_cluster` is `None` we will update it to refer to the default cluster.
/// Because of this, do not normalize/canonicalize the create SQL statement
/// until after calling this function.
fn source_sink_cluster_config<'a, 'ctx>(
    scx: &'a StatementContext<'ctx>,
    in_cluster: &mut Option<ResolvedClusterName>,
) -> Result<&'a dyn CatalogCluster<'ctx>, PlanError> {
    let cluster = match in_cluster {
        None => {
            let cluster = scx.catalog.resolve_cluster(None)?;
            *in_cluster = Some(ResolvedClusterName {
                id: cluster.id(),
                print_name: None,
            });
            cluster
        }
        Some(in_cluster) => scx.catalog.get_cluster(in_cluster.id),
    };

    Ok(cluster)
}

generate_extracted_config!(AvroSchemaOption, (ConfluentWireFormat, bool, Default(true)));

// `SchemaName` is source-only (a source reads one value schema by name). The
// `Key`/`Value`-prefixed options are sink-only. Which options are valid in
// which context is enforced by the source purifier and the sink planner.
generate_extracted_config!(
    GlueAvroOption,
    (SchemaName, String),
    (KeySchemaName, String),
    (ValueSchemaName, String),
    (KeyCompatibilityLevel, String),
    (ValueCompatibilityLevel, String)
);

#[derive(Debug)]
pub struct Schema {
    pub key_schema: Option<String>,
    pub value_schema: String,
    /// Reference schemas for the key schema, in dependency order.
    pub key_reference_schemas: Vec<String>,
    /// Reference schemas for the value schema, in dependency order.
    pub value_reference_schemas: Vec<String>,
    /// Wire-format dispatch and the registry connection to fetch writer
    /// schemas from. Built directly by each `AvroSchema` variant rather
    /// than reconstructed from a (csr_connection, bool) pair downstream.
    pub wire_format: WireFormat<ReferencedConnection>,
}

fn get_encoding_inner(
    scx: &StatementContext,
    format: &Format<Aug>,
) -> Result<SourceDataEncoding<ReferencedConnection>, PlanError> {
    let value = match format {
        Format::Bytes => DataEncoding::Bytes,
        Format::Avro(schema) => {
            let Schema {
                key_schema,
                value_schema,
                key_reference_schemas,
                value_reference_schemas,
                wire_format,
            } = match schema {
                // TODO(jldlaughlin): we need a way to pass in primary key information
                // when building a source from a string or file.
                AvroSchema::InlineSchema {
                    schema: ast::Schema { schema },
                    with_options,
                } => {
                    let AvroSchemaOptionExtracted {
                        confluent_wire_format,
                        ..
                    } = with_options.clone().try_into()?;
                    let wire_format = if confluent_wire_format {
                        WireFormat::Confluent { registry: None }
                    } else {
                        WireFormat::None
                    };
                    Schema {
                        key_schema: None,
                        value_schema: schema.clone(),
                        key_reference_schemas: vec![],
                        value_reference_schemas: vec![],
                        wire_format,
                    }
                }
                AvroSchema::Csr {
                    csr_connection:
                        CsrConnectionAvro {
                            connection,
                            seed,
                            key_strategy: _,
                            value_strategy: _,
                        },
                } => {
                    let item = scx.get_item_by_resolved_name(&connection.connection)?;
                    let csr_connection = match item.connection()? {
                        Connection::Csr(_) => item.id(),
                        _ => {
                            sql_bail!(
                                "{} is not a Confluent Schema Registry connection",
                                scx.catalog
                                    .resolve_full_name(item.name())
                                    .to_string()
                                    .quoted()
                            )
                        }
                    };

                    if let Some(seed) = seed {
                        Schema {
                            key_schema: seed.key_schema.clone(),
                            value_schema: seed.value_schema.clone(),
                            key_reference_schemas: seed.key_reference_schemas.clone(),
                            value_reference_schemas: seed.value_reference_schemas.clone(),
                            wire_format: WireFormat::Confluent {
                                registry: Some(csr_connection),
                            },
                        }
                    } else {
                        sql_bail!("Avro CSR seed resolution has not been performed")
                    }
                }
                AvroSchema::Glue {
                    connection,
                    with_options: _,
                    seed,
                } => {
                    let item = scx.get_item_by_resolved_name(connection)?;
                    let glue_connection = match item.connection()? {
                        Connection::GlueSchemaRegistry(_) => item.id(),
                        _ => {
                            sql_bail!(
                                "{} is not an AWS Glue Schema Registry connection",
                                scx.catalog
                                    .resolve_full_name(item.name())
                                    .to_string()
                                    .quoted()
                            )
                        }
                    };

                    // `SCHEMA NAME` requiredness is enforced during
                    // purification, which also populates `seed`. By the time
                    // planning runs the option is guaranteed present.
                    let Some(seed) = seed else {
                        sql_bail!("Avro Glue seed resolution has not been performed");
                    };

                    Schema {
                        key_schema: None,
                        value_schema: seed.value_schema.clone(),
                        key_reference_schemas: vec![],
                        value_reference_schemas: vec![],
                        wire_format: WireFormat::Glue {
                            registry: Some(glue_connection),
                        },
                    }
                }
            };

            if let Some(key_schema) = key_schema {
                return Ok(SourceDataEncoding {
                    key: Some(DataEncoding::Avro(AvroEncoding {
                        schema: key_schema,
                        reference_schemas: key_reference_schemas,
                        wire_format: wire_format.clone(),
                    })),
                    value: DataEncoding::Avro(AvroEncoding {
                        schema: value_schema,
                        reference_schemas: value_reference_schemas,
                        wire_format,
                    }),
                });
            } else {
                DataEncoding::Avro(AvroEncoding {
                    schema: value_schema,
                    reference_schemas: value_reference_schemas,
                    wire_format,
                })
            }
        }
        Format::Protobuf(schema) => match schema {
            ProtobufSchema::Csr {
                csr_connection:
                    CsrConnectionProtobuf {
                        connection:
                            CsrConnection {
                                connection,
                                options,
                            },
                        seed,
                    },
            } => {
                if let Some(CsrSeedProtobuf { key, value }) = seed {
                    let item = scx.get_item_by_resolved_name(connection)?;
                    let _ = match item.connection()? {
                        Connection::Csr(connection) => connection,
                        _ => {
                            sql_bail!(
                                "{} is not a schema registry connection",
                                scx.catalog
                                    .resolve_full_name(item.name())
                                    .to_string()
                                    .quoted()
                            )
                        }
                    };

                    if !options.is_empty() {
                        sql_bail!("Protobuf CSR connections do not support any options");
                    }

                    let value = DataEncoding::Protobuf(ProtobufEncoding {
                        descriptors: strconv::parse_bytes(&value.schema)?,
                        message_name: value.message_name.clone(),
                        confluent_wire_format: true,
                    });
                    if let Some(key) = key {
                        return Ok(SourceDataEncoding {
                            key: Some(DataEncoding::Protobuf(ProtobufEncoding {
                                descriptors: strconv::parse_bytes(&key.schema)?,
                                message_name: key.message_name.clone(),
                                confluent_wire_format: true,
                            })),
                            value,
                        });
                    }
                    value
                } else {
                    sql_bail!("Protobuf CSR seed resolution has not been performed")
                }
            }
            ProtobufSchema::InlineSchema {
                message_name,
                schema: ast::Schema { schema },
            } => {
                let descriptors = strconv::parse_bytes(schema)?;

                DataEncoding::Protobuf(ProtobufEncoding {
                    descriptors,
                    message_name: message_name.to_owned(),
                    confluent_wire_format: false,
                })
            }
        },
        Format::Regex(regex) => DataEncoding::Regex(RegexEncoding {
            regex: mz_repr::adt::regex::Regex::new(regex, false)
                .map_err(|e| sql_err!("parsing regex: {e}"))?,
        }),
        Format::Csv { columns, delimiter } => {
            let columns = match columns {
                CsvColumns::Header { names } => {
                    if names.is_empty() {
                        sql_bail!("[internal error] column spec should get names in purify")
                    }
                    ColumnSpec::Header {
                        names: names.iter().cloned().map(|n| n.into_string()).collect(),
                    }
                }
                CsvColumns::Count(n) => ColumnSpec::Count(usize::cast_from(*n)),
            };
            DataEncoding::Csv(CsvEncoding {
                columns,
                delimiter: u8::try_from(*delimiter)
                    .map_err(|_| sql_err!("CSV delimiter must be an ASCII character"))?,
            })
        }
        Format::Json { array: false } => DataEncoding::Json,
        Format::Json { array: true } => bail_unsupported!("JSON ARRAY format in sources"),
        Format::Text => DataEncoding::Text,
    };
    Ok(SourceDataEncoding { key: None, value })
}

/// Extract the key envelope, if it is requested
fn get_key_envelope(
    included_items: &[SourceIncludeMetadata],
    encoding: Option<&SourceDataEncoding<ReferencedConnection>>,
    key_envelope_no_encoding: bool,
) -> Result<KeyEnvelope, PlanError> {
    let key_definition = included_items
        .iter()
        .find(|i| matches!(i, SourceIncludeMetadata::Key { .. }));
    if let Some(SourceIncludeMetadata::Key { alias }) = key_definition {
        match (alias, encoding.and_then(|e| e.key.as_ref())) {
            (Some(name), Some(_)) => Ok(KeyEnvelope::Named(name.as_str().to_string())),
            (None, Some(key)) => get_unnamed_key_envelope(Some(key)),
            (Some(name), _) if key_envelope_no_encoding => {
                Ok(KeyEnvelope::Named(name.as_str().to_string()))
            }
            (None, _) if key_envelope_no_encoding => get_unnamed_key_envelope(None),
            (_, None) => {
                // `kd.alias` == `None` means `INCLUDE KEY`
                // `kd.alias` == `Some(_) means INCLUDE KEY AS ___`
                // These both make sense with the same error message
                sql_bail!(
                    "INCLUDE KEY requires specifying KEY FORMAT .. VALUE FORMAT, \
                        got bare FORMAT"
                );
            }
        }
    } else {
        Ok(KeyEnvelope::None)
    }
}

/// Gets the key envelope for a given key encoding when no name for the key has
/// been requested by the user.
fn get_unnamed_key_envelope(
    key: Option<&DataEncoding<ReferencedConnection>>,
) -> Result<KeyEnvelope, PlanError> {
    // If the key is requested but comes from an unnamed type then it gets the name "key"
    //
    // Otherwise it gets the names of the columns in the type
    let is_composite = match key {
        Some(DataEncoding::Bytes | DataEncoding::Json | DataEncoding::Text) => false,
        Some(
            DataEncoding::Avro(_)
            | DataEncoding::Csv(_)
            | DataEncoding::Protobuf(_)
            | DataEncoding::Regex { .. },
        ) => true,
        None => false,
    };

    if is_composite {
        Ok(KeyEnvelope::Flattened)
    } else {
        Ok(KeyEnvelope::Named("key".to_string()))
    }
}

pub fn describe_create_view(
    _: &StatementContext,
    _: CreateViewStatement<Aug>,
) -> Result<StatementDesc, PlanError> {
    Ok(StatementDesc::new(None))
}

pub fn plan_view(
    scx: &StatementContext,
    def: &mut ViewDefinition<Aug>,
    temporary: bool,
) -> Result<(QualifiedItemName, View), PlanError> {
    let create_sql = normalize::create_statement(
        scx,
        Statement::CreateView(CreateViewStatement {
            if_exists: IfExistsBehavior::Error,
            temporary,
            definition: def.clone(),
        }),
    )?;

    let ViewDefinition {
        name,
        columns,
        query,
    } = def;

    let query::PlannedRootQuery {
        expr,
        mut desc,
        finishing,
        scope: _,
    } = query::plan_root_query(scx, query.clone(), QueryLifetime::View)?;
    // We get back a trivial finishing, because `plan_root_query` applies the given finishing.
    // Note: Earlier, we were thinking to maybe persist the finishing information with the view
    // here to help with database-issues#236. However, in the meantime, there might be a better
    // approach to solve database-issues#236:
    // https://github.com/MaterializeInc/database-issues/issues/236#issuecomment-1688293709
    assert!(HirRelationExpr::is_trivial_row_set_finishing_hir(
        &finishing,
        expr.arity()
    ));
    if expr.contains_parameters()? {
        return Err(PlanError::ParameterNotAllowed("views".to_string()));
    }

    let dependencies = expr
        .depends_on()
        .into_iter()
        .map(|gid| scx.catalog.resolve_item_id(&gid))
        .collect();

    let name = if temporary {
        scx.allocate_temporary_qualified_name(normalize::unresolved_item_name(name.to_owned())?)?
    } else {
        scx.allocate_qualified_name(normalize::unresolved_item_name(name.to_owned())?)?
    };

    plan_utils::maybe_rename_columns_exact(
        scx.catalog,
        format!("view {}", scx.catalog.resolve_full_name(&name)),
        &mut desc,
        columns,
    )?;
    let names: Vec<ColumnName> = desc.iter_names().cloned().collect();

    if let Some(dup) = names.iter().duplicates().next() {
        sql_bail!("column {} specified more than once", dup.quoted());
    }

    let view = View {
        create_sql,
        expr,
        dependencies,
        column_names: names,
        temporary,
    };

    Ok((name, view))
}

pub fn plan_create_view(
    scx: &StatementContext,
    mut stmt: CreateViewStatement<Aug>,
) -> Result<Plan, PlanError> {
    let CreateViewStatement {
        temporary,
        if_exists,
        definition,
    } = &mut stmt;
    let (name, view) = plan_view(scx, definition, *temporary)?;

    // Override the statement-level IfExistsBehavior with Skip if this is
    // explicitly requested in the PlanContext (the default is `false`).
    let ignore_if_exists_errors = scx.pcx().map_or(false, |pcx| pcx.ignore_if_exists_errors);

    let replace = if *if_exists == IfExistsBehavior::Replace && !ignore_if_exists_errors {
        let if_exists = true;
        let cascade = false;
        let maybe_item_to_drop = plan_drop_item(
            scx,
            ObjectType::View,
            if_exists,
            definition.name.clone(),
            cascade,
        )?;

        // Check if the new View depends on the item that we would be replacing.
        if let Some(id) = maybe_item_to_drop {
            let dependencies = view.expr.depends_on();
            let invalid_drop = scx
                .get_item(&id)
                .global_ids()
                .any(|gid| dependencies.contains(&gid));
            if invalid_drop {
                let item = scx.catalog.get_item(&id);
                sql_bail!(
                    "cannot replace view {0}: depended upon by new {0} definition",
                    scx.catalog.resolve_full_name(item.name())
                );
            }

            Some(id)
        } else {
            None
        }
    } else {
        None
    };
    let drop_ids = replace
        .map(|id| {
            scx.catalog
                .item_dependents(id)
                .into_iter()
                .map(|id| id.unwrap_item_id())
                .collect()
        })
        .unwrap_or_default();

    validate_view_dependencies(scx, &view.dependencies.0)?;

    // Check for an object in the catalog with this same name
    let full_name = scx.catalog.resolve_full_name(&name);
    let partial_name = PartialItemName::from(full_name.clone());
    // For PostgreSQL compatibility, we need to prevent creating views when
    // there is an existing object *or* type of the same name.
    if let (Ok(item), IfExistsBehavior::Error, false) = (
        scx.catalog.resolve_item_or_type(&partial_name),
        *if_exists,
        ignore_if_exists_errors,
    ) {
        return Err(PlanError::ItemAlreadyExists {
            name: full_name.to_string(),
            item_type: item.item_type(),
        });
    }

    Ok(Plan::CreateView(CreateViewPlan {
        name,
        view,
        replace,
        drop_ids,
        if_not_exists: *if_exists == IfExistsBehavior::Skip,
        ambiguous_columns: *scx.ambiguous_columns.borrow(),
    }))
}

/// Validate the dependencies of a (materialized) view.
fn validate_view_dependencies(
    scx: &StatementContext,
    dependencies: &BTreeSet<CatalogItemId>,
) -> Result<(), PlanError> {
    for id in dependencies {
        let item = scx.catalog.get_item(id);
        if item.replacement_target().is_some() {
            let name = scx.catalog.minimal_qualification(item.name());
            return Err(PlanError::InvalidDependency {
                name: name.to_string(),
                item_type: format!("replacement {}", item.item_type()),
            });
        }
    }

    Ok(())
}

pub fn describe_create_materialized_view(
    _: &StatementContext,
    _: CreateMaterializedViewStatement<Aug>,
) -> Result<StatementDesc, PlanError> {
    Ok(StatementDesc::new(None))
}

pub fn describe_create_network_policy(
    _: &StatementContext,
    _: CreateNetworkPolicyStatement<Aug>,
) -> Result<StatementDesc, PlanError> {
    Ok(StatementDesc::new(None))
}

pub fn describe_alter_network_policy(
    _: &StatementContext,
    _: AlterNetworkPolicyStatement<Aug>,
) -> Result<StatementDesc, PlanError> {
    Ok(StatementDesc::new(None))
}

/// Rejects times that are too large.
fn check_refresh_time(option: &str, ts: Timestamp) -> Result<(), PlanError> {
    let renderable = i64::try_from(ts)
        .ok()
        .and_then(DateTime::from_timestamp_millis)
        .is_some_and(|dt| CheckedTimestamp::try_from(dt).is_ok());
    if !renderable {
        sql_bail!("{option} time too large: {ts}");
    }
    Ok(())
}

pub fn plan_create_materialized_view(
    scx: &StatementContext,
    mut stmt: CreateMaterializedViewStatement<Aug>,
) -> Result<Plan, PlanError> {
    let cluster_id =
        crate::plan::statement::resolve_cluster_for_materialized_view(scx.catalog, &stmt)?;
    stmt.in_cluster = Some(ResolvedClusterName {
        id: cluster_id,
        print_name: None,
    });

    let target_replica = match &stmt.in_cluster_replica {
        Some(replica_name) => {
            scx.require_feature_flag(&ENABLE_REPLICA_TARGETED_MATERIALIZED_VIEWS)?;

            let cluster = scx.catalog.get_cluster(cluster_id);
            let replica_id = cluster
                .replica_ids()
                .get(replica_name.as_str())
                .copied()
                .ok_or_else(|| {
                    CatalogError::UnknownClusterReplica(replica_name.as_str().to_string())
                })?;
            Some(replica_id)
        }
        None => None,
    };

    let create_sql =
        normalize::create_statement(scx, Statement::CreateMaterializedView(stmt.clone()))?;

    let partial_name = normalize::unresolved_item_name(stmt.name)?;
    let name = scx.allocate_qualified_name(partial_name.clone())?;

    let query::PlannedRootQuery {
        expr,
        mut desc,
        finishing,
        scope: _,
    } = query::plan_root_query(scx, stmt.query, QueryLifetime::MaterializedView)?;
    // We get back a trivial finishing, see comment in `plan_view`.
    assert!(HirRelationExpr::is_trivial_row_set_finishing_hir(
        &finishing,
        expr.arity()
    ));
    if expr.contains_parameters()? {
        return Err(PlanError::ParameterNotAllowed(
            "materialized views".to_string(),
        ));
    }

    plan_utils::maybe_rename_columns_exact(
        scx.catalog,
        format!("materialized view {}", scx.catalog.resolve_full_name(&name)),
        &mut desc,
        &stmt.columns,
    )?;
    let column_names: Vec<ColumnName> = desc.iter_names().cloned().collect();

    let MaterializedViewOptionExtracted {
        assert_not_null,
        partition_by,
        retain_history,
        refresh,
        seen: _,
    }: MaterializedViewOptionExtracted = stmt.with_options.try_into()?;

    if let Some(partition_by) = partition_by {
        scx.require_feature_flag(&ENABLE_COLLECTION_PARTITION_BY)?;
        check_partition_by(&desc, partition_by)?;
    }

    let refresh_schedule = {
        let mut refresh_schedule = RefreshSchedule::default();
        let mut on_commits_seen = 0;
        for refresh_option_value in refresh {
            if !matches!(refresh_option_value, RefreshOptionValue::OnCommit) {
                scx.require_feature_flag(&ENABLE_REFRESH_EVERY_MVS)?;
            }
            match refresh_option_value {
                RefreshOptionValue::OnCommit => {
                    on_commits_seen += 1;
                }
                RefreshOptionValue::AtCreation => {
                    soft_panic_or_log!("REFRESH AT CREATION should have been purified away");
                    bail_internal!("REFRESH AT CREATION should have been purified away")
                }
                RefreshOptionValue::At(RefreshAtOptionValue { mut time }) => {
                    transform_ast::transform(scx, &mut time)?; // Desugar the expression
                    let ecx = &ExprContext {
                        qcx: &QueryContext::root(scx, QueryLifetime::OneShot),
                        name: "REFRESH AT",
                        scope: &Scope::empty(),
                        relation_type: &SqlRelationType::empty(),
                        allow_aggregates: false,
                        allow_subqueries: false,
                        allow_parameters: false,
                        allow_windows: false,
                    };
                    let hir = plan_expr(ecx, &time)?.cast_to(
                        ecx,
                        CastContext::Assignment,
                        &SqlScalarType::MzTimestamp,
                    )?;
                    // (mz_now was purified away to a literal earlier)
                    let timestamp = hir
                        .into_literal_mz_timestamp()
                        .ok_or_else(|| PlanError::InvalidRefreshAt)?;
                    check_refresh_time("REFRESH AT", timestamp)?;
                    refresh_schedule.ats.push(timestamp);
                }
                RefreshOptionValue::Every(RefreshEveryOptionValue {
                    interval,
                    aligned_to,
                }) => {
                    let interval = Interval::try_from_value(Value::Interval(interval))?;
                    if interval.as_microseconds() <= 0 {
                        sql_bail!("REFRESH interval must be positive; got: {}", interval);
                    }
                    if interval.months != 0 {
                        // This limitation is because we want Intervals to be cleanly convertable
                        // to a unix epoch timestamp difference. When the interval involves months, then
                        // this is not true anymore, because months have variable lengths.
                        // See `Timestamp::round_up`.
                        sql_bail!("REFRESH interval must not involve units larger than days");
                    }
                    let interval = interval.duration()?;
                    // `Interval::from_duration` (needed to unparse the interval, e.g. for
                    // `mz_materialized_view_refresh_strategies`) requires the micros to fit
                    // in an i64, which is a tighter bound than `Duration::duration` enforces.
                    // Reject too-large intervals here to avoid panicking later.
                    if u64::try_from(interval.as_millis()).is_err()
                        || Interval::from_duration(&interval).is_err()
                    {
                        sql_bail!("REFRESH interval too large");
                    }
                    if interval.as_micros() < 1000 {
                        sql_bail!("REFRESH interval must be at least 1 ms")
                    }

                    let mut aligned_to = match aligned_to {
                        Some(aligned_to) => aligned_to,
                        None => {
                            soft_panic_or_log!(
                                "ALIGNED TO should have been filled in by purification"
                            );
                            sql_bail!(
                                "INTERNAL ERROR: ALIGNED TO should have been filled in by purification"
                            )
                        }
                    };

                    // Desugar the `aligned_to` expression
                    transform_ast::transform(scx, &mut aligned_to)?;

                    let ecx = &ExprContext {
                        qcx: &QueryContext::root(scx, QueryLifetime::OneShot),
                        name: "REFRESH EVERY ... ALIGNED TO",
                        scope: &Scope::empty(),
                        relation_type: &SqlRelationType::empty(),
                        allow_aggregates: false,
                        allow_subqueries: false,
                        allow_parameters: false,
                        allow_windows: false,
                    };
                    let aligned_to_hir = plan_expr(ecx, &aligned_to)?.cast_to(
                        ecx,
                        CastContext::Assignment,
                        &SqlScalarType::MzTimestamp,
                    )?;
                    // (mz_now was purified away to a literal earlier)
                    let aligned_to_const = aligned_to_hir
                        .into_literal_mz_timestamp()
                        .ok_or_else(|| PlanError::InvalidRefreshEveryAlignedTo)?;
                    check_refresh_time("REFRESH EVERY ... ALIGNED TO", aligned_to_const)?;

                    refresh_schedule.everies.push(RefreshEvery {
                        interval,
                        aligned_to: aligned_to_const,
                    });
                }
            }
        }

        if on_commits_seen > 1 {
            sql_bail!("REFRESH ON COMMIT cannot be specified multiple times");
        }
        if on_commits_seen > 0 && refresh_schedule != RefreshSchedule::default() {
            sql_bail!("REFRESH ON COMMIT is not compatible with any of the other REFRESH options");
        }

        if refresh_schedule == RefreshSchedule::default() {
            None
        } else {
            Some(refresh_schedule)
        }
    };

    let as_of = stmt.as_of.map(Timestamp::from);
    let compaction_window = plan_retain_history_option(scx, retain_history)?;
    let mut non_null_assertions = assert_not_null
        .into_iter()
        .map(normalize::column_name)
        .map(|assertion_name| {
            column_names
                .iter()
                .position(|col| col == &assertion_name)
                .ok_or_else(|| {
                    sql_err!(
                        "column {} in ASSERT NOT NULL option not found",
                        assertion_name.quoted()
                    )
                })
        })
        .collect::<Result<Vec<_>, _>>()?;
    non_null_assertions.sort();
    if let Some(dup) = non_null_assertions.iter().duplicates().next() {
        let dup = &column_names[*dup];
        sql_bail!("duplicate column {} in non-null assertions", dup.quoted());
    }

    if let Some(dup) = column_names.iter().duplicates().next() {
        sql_bail!("column {} specified more than once", dup.quoted());
    }

    // Override the statement-level IfExistsBehavior with Skip if this is
    // explicitly requested in the PlanContext (the default is `false`).
    let if_exists = match scx.pcx().map(|pcx| pcx.ignore_if_exists_errors) {
        Ok(true) => IfExistsBehavior::Skip,
        _ => stmt.if_exists,
    };

    let mut replace = None;
    let mut if_not_exists = false;
    match if_exists {
        IfExistsBehavior::Replace => {
            let if_exists = true;
            let cascade = false;
            let replace_id = plan_drop_item(
                scx,
                ObjectType::MaterializedView,
                if_exists,
                partial_name.clone().into(),
                cascade,
            )?;

            // Check if the new Materialized View depends on the item that we would be replacing.
            if let Some(id) = replace_id {
                let dependencies = expr.depends_on();
                let invalid_drop = scx
                    .get_item(&id)
                    .global_ids()
                    .any(|gid| dependencies.contains(&gid));
                if invalid_drop {
                    let item = scx.catalog.get_item(&id);
                    sql_bail!(
                        "cannot replace materialized view {0}: depended upon by new {0} definition",
                        scx.catalog.resolve_full_name(item.name())
                    );
                }
                replace = Some(id);
            }
        }
        IfExistsBehavior::Skip => if_not_exists = true,
        IfExistsBehavior::Error => (),
    }
    let drop_ids = replace
        .map(|id| {
            scx.catalog
                .item_dependents(id)
                .into_iter()
                .map(|id| id.unwrap_item_id())
                .collect()
        })
        .unwrap_or_default();
    let mut dependencies: BTreeSet<_> = expr
        .depends_on()
        .into_iter()
        .map(|gid| scx.catalog.resolve_item_id(&gid))
        .collect();

    // Validate the replacement target, if one is given.
    let mut replacement_target = None;
    if let Some(target_name) = &stmt.replacement_for {
        scx.require_feature_flag(&vars::ENABLE_REPLACEMENT_MATERIALIZED_VIEWS)?;

        let target = scx.get_item_by_resolved_name(target_name)?;
        if target.item_type() != CatalogItemType::MaterializedView {
            return Err(PlanError::InvalidReplacement {
                item_type: target.item_type(),
                item_name: scx.catalog.minimal_qualification(target.name()),
                replacement_type: CatalogItemType::MaterializedView,
                replacement_name: partial_name,
            });
        }
        if target.id().is_system() {
            sql_bail!(
                "cannot replace {} because it is required by the database system",
                scx.catalog.minimal_qualification(target.name()),
            );
        }

        // Check for dependency cycles.
        for dependent in scx.catalog.item_dependents(target.id()) {
            if let ObjectId::Item(id) = dependent
                && dependencies.contains(&id)
            {
                sql_bail!(
                    "replacement would cause {} to depend on itself",
                    scx.catalog.minimal_qualification(target.name()),
                );
            }
        }

        dependencies.insert(target.id());

        for use_id in target.used_by() {
            let use_item = scx.get_item(use_id);
            if use_item.replacement_target() == Some(target.id()) {
                sql_bail!(
                    "cannot replace {} because it already has a replacement: {}",
                    scx.catalog.minimal_qualification(target.name()),
                    scx.catalog.minimal_qualification(use_item.name()),
                );
            }
        }

        replacement_target = Some(target.id());
    }

    validate_view_dependencies(scx, &dependencies)?;

    // Check for an object in the catalog with this same name
    let full_name = scx.catalog.resolve_full_name(&name);
    let partial_name = PartialItemName::from(full_name.clone());
    // For PostgreSQL compatibility, we need to prevent creating materialized
    // views when there is an existing object *or* type of the same name.
    if let (IfExistsBehavior::Error, Ok(item)) =
        (if_exists, scx.catalog.resolve_item_or_type(&partial_name))
    {
        return Err(PlanError::ItemAlreadyExists {
            name: full_name.to_string(),
            item_type: item.item_type(),
        });
    }

    Ok(Plan::CreateMaterializedView(CreateMaterializedViewPlan {
        name,
        materialized_view: MaterializedView {
            create_sql,
            expr,
            dependencies: DependencyIds(dependencies),
            column_names,
            replacement_target,
            cluster_id,
            target_replica,
            non_null_assertions,
            compaction_window,
            refresh_schedule,
            as_of,
        },
        replace,
        drop_ids,
        if_not_exists,
        ambiguous_columns: *scx.ambiguous_columns.borrow(),
    }))
}

generate_extracted_config!(
    MaterializedViewOption,
    (AssertNotNull, Ident, AllowMultiple),
    (PartitionBy, Vec<Ident>),
    (RetainHistory, OptionalDuration),
    (Refresh, RefreshOptionValue<Aug>, AllowMultiple)
);

pub fn describe_create_sink(
    _: &StatementContext,
    _: CreateSinkStatement<Aug>,
) -> Result<StatementDesc, PlanError> {
    Ok(StatementDesc::new(None))
}

generate_extracted_config!(
    CreateSinkOption,
    (Snapshot, bool),
    (PartitionStrategy, String),
    (Version, u64),
    (CommitInterval, Duration)
);

pub fn plan_create_sink(
    scx: &StatementContext,
    stmt: CreateSinkStatement<Aug>,
) -> Result<Plan, PlanError> {
    // Check for an object in the catalog with this same name
    let Some(name) = stmt.name.clone() else {
        return Err(PlanError::MissingName(CatalogItemType::Sink));
    };
    let name = scx.allocate_qualified_name(normalize::unresolved_item_name(name)?)?;
    let full_name = scx.catalog.resolve_full_name(&name);
    let partial_name = PartialItemName::from(full_name.clone());
    if let (false, Ok(item)) = (stmt.if_not_exists, scx.catalog.resolve_item(&partial_name)) {
        return Err(PlanError::ItemAlreadyExists {
            name: full_name.to_string(),
            item_type: item.item_type(),
        });
    }

    plan_sink(scx, stmt)
}

/// This function will plan a sink as if it does not exist in the catalog. This is so the planning
/// logic is reused by both CREATE SINK and ALTER SINK planning. It is the responsibility of the
/// callers (plan_create_sink and plan_alter_sink) to check for name collisions if this is
/// important.
fn plan_sink(
    scx: &StatementContext,
    mut stmt: CreateSinkStatement<Aug>,
) -> Result<Plan, PlanError> {
    let CreateSinkStatement {
        name,
        in_cluster: _,
        from,
        connection,
        format,
        envelope,
        mode,
        if_not_exists,
        with_options,
    } = stmt.clone();

    let Some(name) = name else {
        return Err(PlanError::MissingName(CatalogItemType::Sink));
    };
    let name = scx.allocate_qualified_name(normalize::unresolved_item_name(name)?)?;

    let envelope = match (&connection, envelope, mode) {
        // Kafka sinks use ENVELOPE
        (CreateSinkConnection::Kafka { .. }, Some(ast::SinkEnvelope::Upsert), None) => {
            SinkEnvelope::Upsert
        }
        (CreateSinkConnection::Kafka { .. }, Some(ast::SinkEnvelope::Debezium), None) => {
            SinkEnvelope::Debezium
        }
        (CreateSinkConnection::Kafka { .. }, None, None) => {
            sql_bail!("ENVELOPE clause is required")
        }
        (CreateSinkConnection::Kafka { .. }, _, Some(_)) => {
            sql_bail!("MODE is not supported for Kafka sinks, use ENVELOPE instead")
        }
        // Iceberg sinks use MODE
        (CreateSinkConnection::Iceberg { .. }, None, Some(ast::IcebergSinkMode::Upsert)) => {
            SinkEnvelope::Upsert
        }
        (CreateSinkConnection::Iceberg { .. }, None, Some(ast::IcebergSinkMode::Append)) => {
            SinkEnvelope::Append
        }
        (CreateSinkConnection::Iceberg { .. }, None, None) => {
            sql_bail!("MODE clause is required")
        }
        (CreateSinkConnection::Iceberg { .. }, Some(_), _) => {
            sql_bail!("ENVELOPE is not supported for Iceberg sinks, use MODE instead")
        }
    };

    let from_name = &from;
    let from = scx.get_item_by_resolved_name(&from)?;

    {
        use CatalogItemType::*;
        match from.item_type() {
            Table | Source | MaterializedView => {
                if from.replacement_target().is_some() {
                    let name = scx.catalog.minimal_qualification(from.name());
                    return Err(PlanError::InvalidSinkFrom {
                        name: name.to_string(),
                        item_type: format!("replacement {}", from.item_type()),
                    });
                }
            }
            Sink | MetricSink | View | Index | Type | Func | Secret | Connection => {
                let name = scx.catalog.minimal_qualification(from.name());
                return Err(PlanError::InvalidSinkFrom {
                    name: name.to_string(),
                    item_type: from.item_type().to_string(),
                });
            }
        }
    }

    if from.id().is_system() {
        bail_unsupported!("creating a sink directly on a catalog object");
    }

    let desc = from
        .relation_desc()
        .ok_or_else(|| sql_err!("item does not have a relation description"))?;
    let key_indices = match &connection {
        CreateSinkConnection::Kafka { key: Some(key), .. }
        | CreateSinkConnection::Iceberg { key: Some(key), .. } => {
            let key_columns = key
                .key_columns
                .clone()
                .into_iter()
                .map(normalize::column_name)
                .collect::<Vec<_>>();
            let mut uniq = BTreeSet::new();
            for col in key_columns.iter() {
                if !uniq.insert(col) {
                    sql_bail!("duplicate column referenced in KEY: {}", col);
                }
            }
            let indices = key_columns
                .iter()
                .map(|col| {
                    let name_idx =
                        desc.get_by_name(col)
                            .map(|(idx, _type)| idx)
                            .ok_or_else(|| {
                                sql_err!("column referenced in KEY does not exist: {}", col)
                            })?;
                    if desc.get_unambiguous_name(name_idx).is_none() {
                        sql_bail!("column referenced in KEY is ambiguous: {}", col);
                    }
                    Ok(name_idx)
                })
                .collect::<Result<Vec<_>, _>>()?;

            // Iceberg equality deletes require primitive, non-float key columns.
            // Use an allow-list so that new types are rejected by default.
            if matches!(&connection, CreateSinkConnection::Iceberg { .. }) {
                let cols: Vec<_> = desc.iter().collect();
                for &idx in &indices {
                    let (col_name, col_type) = cols[idx];
                    let scalar = &col_type.scalar_type;
                    let is_valid = matches!(
                        scalar,
                        // integers
                        SqlScalarType::Bool
                            | SqlScalarType::Int16
                            | SqlScalarType::Int32
                            | SqlScalarType::Int64
                            | SqlScalarType::UInt16
                            | SqlScalarType::UInt32
                            | SqlScalarType::UInt64
                            // decimal / numeric
                            | SqlScalarType::Numeric { .. }
                            // date / time
                            | SqlScalarType::Date
                            | SqlScalarType::Time
                            | SqlScalarType::Timestamp { .. }
                            | SqlScalarType::TimestampTz { .. }
                            | SqlScalarType::Interval
                            | SqlScalarType::MzTimestamp
                            // string-like
                            | SqlScalarType::String
                            | SqlScalarType::Char { .. }
                            | SqlScalarType::VarChar { .. }
                            | SqlScalarType::PgLegacyChar
                            | SqlScalarType::PgLegacyName
                            | SqlScalarType::Bytes
                            | SqlScalarType::Jsonb
                            // identifiers
                            | SqlScalarType::Uuid
                            | SqlScalarType::Oid
                            | SqlScalarType::RegProc
                            | SqlScalarType::RegType
                            | SqlScalarType::RegClass
                            | SqlScalarType::MzAclItem
                            | SqlScalarType::AclItem
                            | SqlScalarType::Int2Vector
                    );
                    if !is_valid {
                        return Err(PlanError::IcebergSinkUnsupportedKeyType {
                            column: col_name.to_string(),
                            column_type: format!("{:?}", scalar),
                        });
                    }
                }
            }

            let is_valid_key = desc
                .typ()
                .keys
                .iter()
                .any(|key_columns| key_columns.iter().all(|column| indices.contains(column)));

            if !is_valid_key && envelope == SinkEnvelope::Upsert {
                // Persisted sinks (pcx is None) are re-planned against the keys
                // the running optimizer derives, which can be fewer than when
                // the sink was created. Rejecting the key there fails catalog
                // boot, so only new statements enforce it.
                if key.not_enforced || scx.pcx.is_none() {
                    scx.catalog
                        .add_notice(PlanNotice::UpsertSinkKeyNotEnforced {
                            key: key_columns.clone(),
                            name: name.item.clone(),
                        })
                } else {
                    return Err(PlanError::UpsertSinkWithInvalidKey {
                        name: from_name.full_name_str(),
                        desired_key: key_columns.iter().map(|c| c.to_string()).collect(),
                        valid_keys: desc
                            .typ()
                            .keys
                            .iter()
                            .map(|key| {
                                key.iter()
                                    .map(|col| desc.get_name(*col).as_str().into())
                                    .collect()
                            })
                            .collect(),
                    });
                }
            }
            Some(indices)
        }
        CreateSinkConnection::Kafka { key: None, .. }
        | CreateSinkConnection::Iceberg { key: None, .. } => None,
    };

    if key_indices.is_some() && envelope == SinkEnvelope::Append {
        sql_bail!("KEY is not supported for MODE APPEND Iceberg sinks");
    }

    // Reject input columns that clash with the columns MODE APPEND adds to the Iceberg table.
    if envelope == SinkEnvelope::Append {
        if let CreateSinkConnection::Iceberg { .. } = &connection {
            use mz_storage_types::sinks::{
                ICEBERG_APPEND_DIFF_COLUMN, ICEBERG_APPEND_TIMESTAMP_COLUMN,
            };
            for (col_name, _) in desc.iter() {
                if col_name.as_str() == ICEBERG_APPEND_DIFF_COLUMN
                    || col_name.as_str() == ICEBERG_APPEND_TIMESTAMP_COLUMN
                {
                    sql_bail!(
                        "column {} conflicts with the system column that MODE APPEND \
                         adds to the Iceberg table",
                        col_name.quoted()
                    );
                }
            }
        }
    }

    let headers_index = match &connection {
        CreateSinkConnection::Kafka {
            headers: Some(headers),
            ..
        } => {
            scx.require_feature_flag(&ENABLE_KAFKA_SINK_HEADERS)?;

            match envelope {
                SinkEnvelope::Upsert | SinkEnvelope::Append => (),
                SinkEnvelope::Debezium => {
                    sql_bail!("HEADERS option is not supported with ENVELOPE DEBEZIUM")
                }
            };

            let headers = normalize::column_name(headers.clone());
            let (idx, ty) = desc
                .get_by_name(&headers)
                .ok_or_else(|| sql_err!("HEADERS column ({}) is unknown", headers))?;

            if desc.get_unambiguous_name(idx).is_none() {
                sql_bail!("HEADERS column ({}) is ambiguous", headers);
            }

            match &ty.scalar_type {
                SqlScalarType::Map { value_type, .. }
                    if matches!(&**value_type, SqlScalarType::String | SqlScalarType::Bytes) => {}
                _ => sql_bail!(
                    "HEADERS column must have type map[text => text] or map[text => bytea]"
                ),
            }

            Some(idx)
        }
        _ => None,
    };

    // pick the first valid natural relation key, if any
    let relation_key_indices = desc.typ().keys.get(0).cloned();

    let key_desc_and_indices = key_indices.map(|key_indices| {
        let cols = desc
            .iter()
            .map(|(name, ty)| (name.clone(), ty.clone()))
            .collect::<Vec<_>>();
        let (names, types): (Vec<_>, Vec<_>) =
            key_indices.iter().map(|&idx| cols[idx].clone()).unzip();
        let typ = SqlRelationType::new(types);
        (RelationDesc::new(typ, names), key_indices)
    });

    if key_desc_and_indices.is_none() && envelope == SinkEnvelope::Upsert {
        return Err(PlanError::UpsertSinkWithoutKey);
    }

    let CreateSinkOptionExtracted {
        snapshot,
        version,
        partition_strategy: _,
        seen: _,
        commit_interval,
    } = with_options.try_into()?;

    let connection_builder = match connection {
        CreateSinkConnection::Kafka {
            connection,
            options,
            ..
        } => kafka_sink_builder(
            scx,
            connection,
            options,
            format,
            relation_key_indices,
            key_desc_and_indices,
            headers_index,
            desc.into_owned(),
            envelope,
            from.id(),
            commit_interval,
        )?,
        CreateSinkConnection::Iceberg {
            catalog_connection,
            aws_connection,
            options,
            ..
        } => iceberg_sink_builder(
            scx,
            catalog_connection,
            aws_connection,
            options,
            relation_key_indices,
            key_desc_and_indices,
            commit_interval,
            &desc,
        )?,
    };

    // WITH SNAPSHOT defaults to true
    let with_snapshot = snapshot.unwrap_or(true);
    // VERSION defaults to 0
    let version = version.unwrap_or(0);

    // We will rewrite the cluster if one is not provided, so we must use the
    // `in_cluster` value we plan to normalize when we canonicalize the create
    // statement.
    let in_cluster = source_sink_cluster_config(scx, &mut stmt.in_cluster)?;
    let create_sql = normalize::create_statement(scx, Statement::CreateSink(stmt))?;

    Ok(Plan::CreateSink(CreateSinkPlan {
        name,
        sink: Sink {
            create_sql,
            from: from.global_id(),
            connection: connection_builder,
            envelope,
            version,
            commit_interval,
        },
        with_snapshot,
        if_not_exists,
        in_cluster: in_cluster.id(),
    }))
}

fn key_constraint_err(desc: &RelationDesc, user_keys: &[ColumnName]) -> PlanError {
    let user_keys = user_keys.iter().map(|column| column.as_str()).join(", ");

    let existing_keys = desc
        .typ()
        .keys
        .iter()
        .map(|key_columns| {
            key_columns
                .iter()
                .map(|col| desc.get_name(*col).as_str())
                .join(", ")
        })
        .join(", ");

    sql_err!(
        "Key constraint ({}) conflicts with existing key ({})",
        user_keys,
        existing_keys
    )
}

/// Creating this by hand instead of using generate_extracted_config! macro
/// because the macro doesn't support parameterized enums. See <https://github.com/MaterializeInc/database-issues/issues/6698>
#[derive(Debug, Default, PartialEq, Clone)]
pub struct CsrConfigOptionExtracted {
    seen: ::std::collections::BTreeSet<CsrConfigOptionName<Aug>>,
    pub(crate) avro_key_fullname: Option<String>,
    pub(crate) avro_value_fullname: Option<String>,
    pub(crate) null_defaults: bool,
    pub(crate) value_doc_options: BTreeMap<DocTarget, String>,
    pub(crate) key_doc_options: BTreeMap<DocTarget, String>,
    pub(crate) key_compatibility_level: Option<mz_ccsr::CompatibilityLevel>,
    pub(crate) value_compatibility_level: Option<mz_ccsr::CompatibilityLevel>,
}

impl std::convert::TryFrom<Vec<CsrConfigOption<Aug>>> for CsrConfigOptionExtracted {
    type Error = crate::plan::PlanError;
    fn try_from(v: Vec<CsrConfigOption<Aug>>) -> Result<CsrConfigOptionExtracted, Self::Error> {
        let mut extracted = CsrConfigOptionExtracted::default();
        let mut common_doc_comments = BTreeMap::new();
        for option in v {
            if !extracted.seen.insert(option.name.clone()) {
                return Err(PlanError::Unstructured({
                    format!("{} specified more than once", option.name)
                }));
            }
            let option_name = option.name.clone();
            let option_name_str = option_name.to_ast_string_simple();
            let better_error = |e: PlanError| PlanError::InvalidOptionValue {
                option_name: option_name.to_ast_string_simple(),
                err: e.into(),
            };
            let to_compatibility_level = |val: Option<WithOptionValue<Aug>>| {
                val.map(|s| match s {
                    WithOptionValue::Value(Value::String(s)) => {
                        mz_ccsr::CompatibilityLevel::try_from(s.to_uppercase().as_str())
                    }
                    _ => Err("must be a string".to_string()),
                })
                .transpose()
                .map_err(PlanError::Unstructured)
                .map_err(better_error)
            };
            match option.name {
                CsrConfigOptionName::AvroKeyFullname => {
                    extracted.avro_key_fullname =
                        <Option<String>>::try_from_value(option.value).map_err(better_error)?;
                }
                CsrConfigOptionName::AvroValueFullname => {
                    extracted.avro_value_fullname =
                        <Option<String>>::try_from_value(option.value).map_err(better_error)?;
                }
                CsrConfigOptionName::NullDefaults => {
                    extracted.null_defaults =
                        <bool>::try_from_value(option.value).map_err(better_error)?;
                }
                CsrConfigOptionName::AvroDocOn(doc_on) => {
                    let value = String::try_from_value(option.value.ok_or_else(|| {
                        PlanError::InvalidOptionValue {
                            option_name: option_name_str,
                            err: Box::new(PlanError::Unstructured("cannot be empty".to_string())),
                        }
                    })?)
                    .map_err(better_error)?;
                    let key = match doc_on.identifier {
                        DocOnIdentifier::Column(ast::ColumnName {
                            relation: ResolvedItemName::Item { id, .. },
                            column: ResolvedColumnReference::Column { name, index: _ },
                        }) => DocTarget::Field {
                            object_id: id,
                            column_name: name,
                        },
                        DocOnIdentifier::Type(ResolvedItemName::Item { id, .. }) => {
                            DocTarget::Type(id)
                        }
                        _ => sql_bail!("invalid DOC ON identifier"),
                    };

                    match doc_on.for_schema {
                        DocOnSchema::KeyOnly => {
                            extracted.key_doc_options.insert(key, value);
                        }
                        DocOnSchema::ValueOnly => {
                            extracted.value_doc_options.insert(key, value);
                        }
                        DocOnSchema::All => {
                            common_doc_comments.insert(key, value);
                        }
                    }
                }
                CsrConfigOptionName::KeyCompatibilityLevel => {
                    extracted.key_compatibility_level = to_compatibility_level(option.value)?;
                }
                CsrConfigOptionName::ValueCompatibilityLevel => {
                    extracted.value_compatibility_level = to_compatibility_level(option.value)?;
                }
            }
        }

        for (key, value) in common_doc_comments {
            if !extracted.key_doc_options.contains_key(&key) {
                extracted.key_doc_options.insert(key.clone(), value.clone());
            }
            if !extracted.value_doc_options.contains_key(&key) {
                extracted.value_doc_options.insert(key, value);
            }
        }
        Ok(extracted)
    }
}

fn iceberg_sink_builder(
    scx: &StatementContext,
    catalog_connection: ResolvedItemName,
    storage_connection: Option<ResolvedItemName>,
    options: Vec<IcebergSinkConfigOption<Aug>>,
    relation_key_indices: Option<Vec<usize>>,
    key_desc_and_indices: Option<(RelationDesc, Vec<usize>)>,
    commit_interval: Option<Duration>,
    desc: &RelationDesc,
) -> Result<StorageSinkConnection<ReferencedConnection>, PlanError> {
    // Reject types that arrow-rs's parquet writer cannot handle, before
    // sink creation. Pass the iceberg overrides so types iceberg remaps
    // (e.g. interval -> string) don't trip the check.
    ArrowBuilder::validate_desc_for_parquet(desc, iceberg_type_overrides)
        .map_err(|e| sql_err!("{}", e))?;

    let catalog_connection_item = scx.get_item_by_resolved_name(&catalog_connection)?;
    let catalog_connection_id = catalog_connection_item.id();
    if !matches!(
        catalog_connection_item.connection()?,
        Connection::IcebergCatalog(_)
    ) {
        sql_bail!(
            "{} is not an iceberg catalog connection",
            scx.catalog
                .resolve_full_name(catalog_connection_item.name())
                .to_string()
                .quoted()
        );
    };

    let storage_connection_item = storage_connection
        .map(|c| scx.get_item_by_resolved_name(&c))
        .transpose()?;
    let storage_connection_id = storage_connection_item.as_ref().map(|c| c.id());
    if let Some(c) = &storage_connection_item
        && !matches!(c.connection()?, Connection::Aws(_))
    {
        sql_bail!(
            "{} is not an AWS connection",
            scx.catalog.resolve_full_name(c.name()).to_string().quoted()
        );
    }

    let IcebergSinkConfigOptionExtracted {
        table,
        namespace,
        seen: _,
    }: IcebergSinkConfigOptionExtracted = options.try_into()?;

    let Some(table) = table else {
        sql_bail!("Iceberg sink must specify TABLE");
    };
    let Some(namespace) = namespace else {
        sql_bail!("Iceberg sink must specify NAMESPACE");
    };
    match commit_interval {
        None => sql_bail!("Iceberg sink must specify COMMIT INTERVAL"),
        // The sink truncates the interval to whole milliseconds, and a
        // truncated interval of zero would make it mint zero-width batches in
        // a busy loop, never committing any data. Require a full second so
        // truncation is irrelevant, matching the TOPIC METADATA REFRESH
        // INTERVAL minimum.
        Some(interval) if interval < Duration::from_secs(1) => {
            sql_bail!("COMMIT INTERVAL must be at least 1 second")
        }
        Some(_) => {}
    }

    Ok(StorageSinkConnection::Iceberg(IcebergSinkConnection {
        catalog_connection_id,
        catalog_connection: catalog_connection_id,
        storage_connection_id,
        storage_connection: storage_connection_id,
        table,
        namespace,
        relation_key_indices,
        key_desc_and_indices,
    }))
}

fn kafka_sink_builder(
    scx: &StatementContext,
    connection: ResolvedItemName,
    options: Vec<KafkaSinkConfigOption<Aug>>,
    format: Option<FormatSpecifier<Aug>>,
    relation_key_indices: Option<Vec<usize>>,
    key_desc_and_indices: Option<(RelationDesc, Vec<usize>)>,
    headers_index: Option<usize>,
    value_desc: RelationDesc,
    envelope: SinkEnvelope,
    sink_from: CatalogItemId,
    commit_interval: Option<Duration>,
) -> Result<StorageSinkConnection<ReferencedConnection>, PlanError> {
    // Get Kafka connection.
    let connection_item = scx.get_item_by_resolved_name(&connection)?;
    let connection_id = connection_item.id();
    match connection_item.connection()? {
        Connection::Kafka(_) => (),
        _ => sql_bail!(
            "{} is not a kafka connection",
            scx.catalog.resolve_full_name(connection_item.name())
        ),
    };

    if commit_interval.is_some() {
        sql_bail!("COMMIT INTERVAL option is not supported with KAFKA sinks");
    }

    let KafkaSinkConfigOptionExtracted {
        topic,
        compression_type,
        partition_by,
        progress_group_id_prefix,
        transactional_id_prefix,
        legacy_ids,
        topic_config,
        topic_metadata_refresh_interval,
        topic_partition_count,
        topic_replication_factor,
        seen: _,
    }: KafkaSinkConfigOptionExtracted = options.try_into()?;

    let transactional_id = match (transactional_id_prefix, legacy_ids) {
        (Some(_), Some(true)) => {
            sql_bail!("LEGACY IDS cannot be used at the same time as TRANSACTIONAL ID PREFIX")
        }
        (None, Some(true)) => KafkaIdStyle::Legacy,
        (prefix, _) => KafkaIdStyle::Prefix(prefix),
    };

    let progress_group_id = match (progress_group_id_prefix, legacy_ids) {
        (Some(_), Some(true)) => {
            sql_bail!("LEGACY IDS cannot be used at the same time as PROGRESS GROUP ID PREFIX")
        }
        (None, Some(true)) => KafkaIdStyle::Legacy,
        (prefix, _) => KafkaIdStyle::Prefix(prefix),
    };

    let topic_name = topic.ok_or_else(|| sql_err!("KAFKA CONNECTION must specify TOPIC"))?;

    if topic_metadata_refresh_interval > MAX_KAFKA_TOPIC_METADATA_REFRESH_INTERVAL {
        // This is a librdkafka-enforced restriction that, if violated,
        // would result in a runtime error for the source.
        sql_bail!("TOPIC METADATA REFRESH INTERVAL cannot be greater than 1 hour");
    } else if topic_metadata_refresh_interval < MIN_KAFKA_TOPIC_METADATA_REFRESH_INTERVAL {
        // We enforce a minimum of 1 second here to prevent excessive refreshes, and ensure that
        // tokio::time::interval receives a valid (positive) duration.
        sql_bail!("TOPIC METADATA REFRESH INTERVAL must be at least 1 second");
    }

    let assert_positive = |val: Option<i32>, name: &str| {
        if let Some(val) = val {
            if val <= 0 {
                sql_bail!("{} must be a positive integer", name);
            }
        }
        val.map(NonNeg::try_from)
            .transpose()
            .map_err(|_| PlanError::Unstructured(format!("{} must be a positive integer", name)))
    };
    let topic_partition_count = assert_positive(topic_partition_count, "TOPIC PARTITION COUNT")?;
    let topic_replication_factor =
        assert_positive(topic_replication_factor, "TOPIC REPLICATION FACTOR")?;

    // Helper method to parse avro connection options for format specifiers that use avro
    // for either key or value encoding.
    let gen_avro_schema_options = |conn| {
        let CsrConnectionAvro {
            connection:
                CsrConnection {
                    connection,
                    options,
                },
            seed,
            key_strategy,
            value_strategy,
        } = conn;
        if seed.is_some() {
            sql_bail!("SEED option does not make sense with sinks");
        }
        if key_strategy.is_some() {
            sql_bail!("KEY STRATEGY option does not make sense with sinks");
        }
        if value_strategy.is_some() {
            sql_bail!("VALUE STRATEGY option does not make sense with sinks");
        }

        let item = scx.get_item_by_resolved_name(&connection)?;
        let csr_connection = match item.connection()? {
            Connection::Csr(_) => item.id(),
            _ => {
                sql_bail!(
                    "{} is not a schema registry connection",
                    scx.catalog
                        .resolve_full_name(item.name())
                        .to_string()
                        .quoted()
                )
            }
        };
        let extracted_options: CsrConfigOptionExtracted = options.try_into()?;

        if key_desc_and_indices.is_none() && extracted_options.avro_key_fullname.is_some() {
            sql_bail!("Cannot specify AVRO KEY FULLNAME without a corresponding KEY field");
        }

        if key_desc_and_indices.is_some()
            && (extracted_options.avro_key_fullname.is_some()
                ^ extracted_options.avro_value_fullname.is_some())
        {
            sql_bail!(
                "Must specify both AVRO KEY FULLNAME and AVRO VALUE FULLNAME when specifying generated schema names"
            );
        }

        Ok((csr_connection, extracted_options))
    };

    let map_format = |format: Format<Aug>, desc: &RelationDesc, is_key: bool| match format {
        Format::Json { array: false } => Ok::<_, PlanError>(KafkaSinkFormatType::Json),
        Format::Bytes if desc.arity() == 1 => {
            let col_type = &desc.typ().column_types[0].scalar_type;
            if !mz_pgrepr::Value::can_encode_binary(col_type) {
                bail_unsupported!(format!(
                    "BYTES format with non-encodable type: {:?}",
                    col_type
                ));
            }

            Ok(KafkaSinkFormatType::Bytes)
        }
        Format::Text if desc.arity() == 1 => Ok(KafkaSinkFormatType::Text),
        Format::Bytes | Format::Text => {
            bail_unsupported!("BYTES or TEXT format with multiple columns")
        }
        Format::Json { array: true } => bail_unsupported!("JSON ARRAY format in sinks"),
        Format::Avro(AvroSchema::Csr { csr_connection }) => {
            let (csr_connection, options) = gen_avro_schema_options(csr_connection)?;
            let schema = if is_key {
                AvroSchemaGenerator::new(
                    desc.clone(),
                    false,
                    options.key_doc_options,
                    options.avro_key_fullname.as_deref().unwrap_or("row"),
                    options.null_defaults,
                    Some(sink_from),
                    false,
                )?
                .schema()
                .to_string()
            } else {
                AvroSchemaGenerator::new(
                    desc.clone(),
                    matches!(envelope, SinkEnvelope::Debezium),
                    options.value_doc_options,
                    options.avro_value_fullname.as_deref().unwrap_or("envelope"),
                    options.null_defaults,
                    Some(sink_from),
                    true,
                )?
                .schema()
                .to_string()
            };
            Ok(KafkaSinkFormatType::Avro {
                schema,
                compatibility_level: if is_key {
                    options.key_compatibility_level
                } else {
                    options.value_compatibility_level
                },
                // Confluent always derives the subject from the topic.
                schema_name: None,
                wire_format: WireFormat::Confluent {
                    registry: Some(csr_connection),
                },
            })
        }
        Format::Avro(AvroSchema::Glue {
            connection,
            with_options,
            seed,
        }) => {
            if seed.is_some() {
                sql_bail!("SEED option does not make sense with sinks");
            }

            let extracted: GlueAvroOptionExtracted = with_options.try_into()?;
            let GlueAvroOptionExtracted {
                schema_name,
                key_schema_name,
                value_schema_name,
                key_compatibility_level,
                value_compatibility_level,
                seen: _,
            } = extracted;

            // The singular `SCHEMA NAME` is source-only: a source reads one
            // value schema by name, but a sink registers separate key and value
            // schemas that must not share a name. Steer users to the per-side
            // options instead.
            if schema_name.is_some() {
                sql_bail!(
                    "SCHEMA NAME is not supported for AWS Glue Schema Registry sinks, \
                     use KEY SCHEMA NAME and VALUE SCHEMA NAME instead"
                );
            }

            // The `KEY`-prefixed options only apply to the key schema, so they
            // are meaningless without a key. Mirror the CSR clause, which
            // rejects key-side options without a corresponding KEY field rather
            // than silently ignoring them. This is checked here rather than
            // per-side because the key branch runs only when a key exists.
            if key_desc_and_indices.is_none()
                && (key_schema_name.is_some() || key_compatibility_level.is_some())
            {
                sql_bail!(
                    "KEY SCHEMA NAME and KEY COMPATIBILITY LEVEL require a corresponding KEY field"
                );
            }

            let item = scx.get_item_by_resolved_name(&connection)?;
            let glue_connection = match item.connection()? {
                Connection::GlueSchemaRegistry(_) => item.id(),
                _ => {
                    sql_bail!(
                        "{} is not an AWS Glue Schema Registry connection",
                        scx.catalog
                            .resolve_full_name(item.name())
                            .to_string()
                            .quoted()
                    )
                }
            };

            // Parse the compatibility level through the CSR enum, which rejects
            // Glue-only levels like DISABLED. `None` leaves the level unset, so
            // a newly created schema defaults to Glue's `BACKWARD`.
            let compatibility_level = {
                let raw = if is_key {
                    key_compatibility_level
                } else {
                    value_compatibility_level
                };
                raw.map(|s| {
                    mz_ccsr::CompatibilityLevel::try_from(s.to_uppercase().as_str())
                        .map_err(PlanError::Unstructured)
                })
                .transpose()?
            };

            // The schema name defaults to the topic-derived subject and can be
            // overridden per side. The topic-derived fallback is applied in the
            // storage layer, which knows the resolved topic name.
            let schema_name = if is_key {
                key_schema_name
            } else {
                value_schema_name
            };

            // Unlike the CSR clause, the Glue clause carries no fullname or doc
            // options, so the schema is generated with defaults.
            let schema = if is_key {
                AvroSchemaGenerator::new(
                    desc.clone(),
                    false,
                    Default::default(),
                    "row",
                    false,
                    Some(sink_from),
                    false,
                )?
                .schema()
                .to_string()
            } else {
                AvroSchemaGenerator::new(
                    desc.clone(),
                    matches!(envelope, SinkEnvelope::Debezium),
                    Default::default(),
                    "envelope",
                    false,
                    Some(sink_from),
                    true,
                )?
                .schema()
                .to_string()
            };
            Ok(KafkaSinkFormatType::Avro {
                schema,
                compatibility_level,
                schema_name,
                wire_format: WireFormat::Glue {
                    registry: Some(glue_connection),
                },
            })
        }
        format => bail_unsupported!(format!("sink format {:?}", format)),
    };

    let partition_by = match &partition_by {
        Some(partition_by) => {
            let mut scope = Scope::from_source(None, value_desc.iter_names());

            match envelope {
                SinkEnvelope::Upsert | SinkEnvelope::Append => (),
                SinkEnvelope::Debezium => {
                    let key_indices: HashSet<_> = key_desc_and_indices
                        .as_ref()
                        .map(|(_desc, indices)| indices.as_slice())
                        .unwrap_or_default()
                        .into_iter()
                        .collect();
                    for (i, item) in scope.items.iter_mut().enumerate() {
                        if !key_indices.contains(&i) {
                            item.error_if_referenced = Some(|_table, column| {
                                PlanError::InvalidPartitionByEnvelopeDebezium {
                                    column_name: column.to_string(),
                                }
                            });
                        }
                    }
                }
            };

            let ecx = &ExprContext {
                qcx: &QueryContext::root(scx, QueryLifetime::OneShot),
                name: "PARTITION BY",
                scope: &scope,
                relation_type: value_desc.typ(),
                allow_aggregates: false,
                allow_subqueries: false,
                allow_parameters: false,
                allow_windows: false,
            };
            let expr = plan_expr(ecx, partition_by)?.cast_to(
                ecx,
                CastContext::Assignment,
                &SqlScalarType::UInt64,
            )?;
            let expr = expr.lower_uncorrelated(scx.catalog.system_vars())?;

            Some(expr)
        }
        _ => None,
    };

    // Map from the format specifier of the statement to the individual key/value formats for the sink.
    let format = match format {
        Some(FormatSpecifier::KeyValue { key, value }) => {
            let key_format = match key_desc_and_indices.as_ref() {
                Some((desc, _indices)) => Some(map_format(key, desc, true)?),
                None => None,
            };
            KafkaSinkFormat {
                value_format: map_format(value, &value_desc, false)?,
                key_format,
            }
        }
        Some(FormatSpecifier::Bare(format)) => {
            let key_format = match key_desc_and_indices.as_ref() {
                Some((desc, _indices)) => Some(map_format(format.clone(), desc, true)?),
                None => None,
            };
            KafkaSinkFormat {
                value_format: map_format(format, &value_desc, false)?,
                key_format,
            }
        }
        None => bail_unsupported!("sink without format"),
    };

    Ok(StorageSinkConnection::Kafka(KafkaSinkConnection {
        connection_id,
        connection: connection_id,
        format,
        topic: topic_name,
        relation_key_indices,
        key_desc_and_indices,
        headers_index,
        value_desc,
        partition_by,
        compression_type,
        progress_group_id,
        transactional_id,
        topic_options: KafkaTopicOptions {
            partition_count: topic_partition_count,
            replication_factor: topic_replication_factor,
            topic_config: topic_config.unwrap_or_default(),
        },
        topic_metadata_refresh_interval,
    }))
}

pub fn describe_create_metric_sink(
    _: &StatementContext,
    _: CreateMetricSinkStatement<Aug>,
) -> Result<StatementDesc, PlanError> {
    Ok(StatementDesc::new(None))
}

/// The columns a metric sink reads from its source, and the type each one has to be. Order
/// doesn't matter, and nullability isn't checked here: nulls are the operator's problem at
/// runtime.
const METRIC_SINK_SOURCE_COLUMNS: &[(&str, fn(&SqlScalarType) -> bool)] = &[
    ("metric_name", |t| matches!(t, SqlScalarType::String)),
    ("metric_type", |t| matches!(t, SqlScalarType::String)),
    (
        "labels",
        |t| matches!(t, SqlScalarType::Map { value_type, .. } if matches!(**value_type, SqlScalarType::String)),
    ),
    ("value", |t| matches!(t, SqlScalarType::Float64)),
    ("help", |t| matches!(t, SqlScalarType::String)),
];

generate_extracted_config!(CreateMetricSinkOption, (Prefix, String));

/// The reserved namespace every metric sink prefix must start with. It confines the family names a
/// sink publishes (`prefix + name`) to the `mz_metric_sink_` lane, which nothing else in a
/// replica's Prometheus registry writes: platform metrics use other `mz_*` names, third-party
/// collectors namespace away from `mz_*`, and the sink's own health gauges are
/// `mz_compute_metric_sink_*`. Requiring the marker rather than blocklisting platform names cannot
/// rot as new platform collectors are added.
const METRIC_SINK_PREFIX_MARKER: &str = "mz_metric_sink_";

/// The prefix reserved for the curated sinks (`mz_adapter::coord::metric_sink::CURATED`);
/// [`validate_user_metric_sink_prefix`] keeps user prefixes clear of it.
pub const METRIC_SINK_CURATED_PREFIX_MARKER: &str = "mz_metric_sink_curated_";

/// Rejects a prefix that is not a legal start of a Prometheus metric name, or that escapes the
/// reserved `mz_metric_sink_` lane (`METRIC_SINK_PREFIX_MARKER`). The reserved leading character
/// is what lets a row's bare `metric_name` start with a digit or be empty. User sinks use
/// [`validate_user_metric_sink_prefix`], which adds the curated reservation.
pub fn validate_metric_sink_prefix(prefix: &str) -> Result<(), PlanError> {
    if prefix.is_empty() {
        return Err(sql_err!("metric sink prefix must not be empty"));
    }
    let mut chars = prefix.chars();
    let valid = match chars.next() {
        Some(c) if c.is_ascii_alphabetic() || c == '_' || c == ':' => {
            chars.all(|c| c.is_ascii_alphanumeric() || c == '_' || c == ':')
        }
        _ => false,
    };
    if !valid {
        return Err(sql_err!(
            "metric sink prefix {:?} is not a valid start of a Prometheus metric name",
            prefix
        ));
    }
    if !prefix.starts_with(METRIC_SINK_PREFIX_MARKER) {
        return Err(sql_err!(
            "metric sink prefix {:?} must start with {:?}",
            prefix,
            METRIC_SINK_PREFIX_MARKER
        ));
    }
    Ok(())
}

/// [`validate_metric_sink_prefix`] plus the reservation that keeps user sinks clear of the curated
/// families ([`METRIC_SINK_CURATED_PREFIX_MARKER`]).
///
/// The check runs both ways so it is total: a prefix that starts with the curated prefix is
/// rejected, and so is one the curated prefix starts with (a bare `mz_metric_sink_` plus a
/// `curated_...` row would otherwise publish a curated family, which Prometheus merges silently).
/// It is static rather than a catalog scan like `Coordinator::ensure_metric_sink_prefix_is_free`,
/// because curated sinks are not catalog items.
pub fn validate_user_metric_sink_prefix(prefix: &str) -> Result<(), PlanError> {
    validate_metric_sink_prefix(prefix)?;
    if prefix.starts_with(METRIC_SINK_CURATED_PREFIX_MARKER)
        || METRIC_SINK_CURATED_PREFIX_MARKER.starts_with(prefix)
    {
        return Err(sql_err!(
            "metric sink prefix {:?} overlaps {:?}, which is reserved for the built-in metric sinks",
            prefix,
            METRIC_SINK_CURATED_PREFIX_MARKER
        ));
    }
    Ok(())
}

/// Checks that `desc` exposes the canonical metric-sink columns, the contract
/// `mz_adapter::optimize::metric_sink`'s row shaping and the compute-side operator both rely on.
///
/// Every metric-sink source has to pass this, whether it is the `FROM` relation of a
/// `CREATE METRIC SINK` or the query behind a coordinator-installed curated sink.
pub fn validate_metric_sink_desc(desc: &RelationDesc) -> Result<(), PlanError> {
    for (name, type_ok) in METRIC_SINK_SOURCE_COLUMNS {
        let col = ColumnName::from(*name);
        let (_, column_type) = desc
            .get_by_name(&col)
            .ok_or_else(|| sql_err!("metric sink source must expose column {:?}", name))?;
        if !type_ok(&column_type.scalar_type) {
            return Err(sql_err!(
                "metric sink source column {:?} is not of the required type",
                name
            ));
        }
    }
    Ok(())
}

/// Plans a metric sink over `stmt.from`, which must expose the columns in
/// `METRIC_SINK_SOURCE_COLUMNS`.
///
/// Every distinct `(metric_name, labels)` becomes its own Prometheus series, so high-cardinality
/// labels mean a lot of series. A null `value` is a gap in the exposition rather than a zero, and
/// the series stays missing until a non-null value shows up for that same key. Use
/// `coalesce(value, 0)` if you want zeroes instead.
pub fn plan_create_metric_sink(
    scx: &StatementContext,
    mut stmt: CreateMetricSinkStatement<Aug>,
) -> Result<Plan, PlanError> {
    scx.require_feature_flag(&ENABLE_METRIC_SINK)?;

    let CreateMetricSinkStatement {
        name,
        in_cluster,
        if_not_exists,
        from,
        with_options,
    } = &mut stmt;

    let if_not_exists = *if_not_exists;
    let CreateMetricSinkOptionExtracted { prefix, seen: _ } = with_options.clone().try_into()?;
    // Required, not defaulted. A name-derived default reintroduces the collision problem the
    // moment the sink is renamed, and an empty one never had it solved.
    let Some(prefix) = prefix else {
        sql_bail!("CREATE METRIC SINK requires a PREFIX option");
    };
    validate_user_metric_sink_prefix(&prefix)?;
    let name = scx.allocate_qualified_name(normalize::unresolved_item_name(name.clone())?)?;
    let full_name = scx.catalog.resolve_full_name(&name);
    let partial_name = PartialItemName::from(full_name.clone());
    if let (false, Ok(item)) = (if_not_exists, scx.catalog.resolve_item(&partial_name)) {
        return Err(PlanError::ItemAlreadyExists {
            name: full_name.to_string(),
            item_type: item.item_type(),
        });
    }

    let from_item = scx.get_item_by_resolved_name(from)?;
    // `relation_desc()` returns `None` for exactly the item types a metric sink cannot read
    // from (including `Index`, which lives on a cluster but exposes no relation description), so
    // its `None` branch doubles as the reject filter. Fold the per-type error message in here.
    let desc = from_item.relation_desc().ok_or_else(|| {
        sql_err!(
            "cannot create metric sink from {} because it is a {}",
            scx.catalog.minimal_qualification(from_item.name()),
            from_item.item_type(),
        )
    })?;
    validate_metric_sink_desc(&desc)?;

    let cluster_id = match in_cluster {
        None => scx.resolve_cluster(None)?.id(),
        Some(in_cluster) => in_cluster.id,
    };
    *in_cluster = Some(ResolvedClusterName {
        id: cluster_id,
        print_name: None,
    });
    let from_global_id = from_item.global_id();

    let create_sql = normalize::create_statement(scx, Statement::CreateMetricSink(stmt))?;

    Ok(Plan::CreateMetricSink(CreateMetricSinkPlan {
        name,
        metric_sink: MetricSink {
            create_sql,
            from: from_global_id,
            cluster_id,
            prefix,
        },
        if_not_exists,
    }))
}

pub fn describe_create_index(
    _: &StatementContext,
    _: CreateIndexStatement<Aug>,
) -> Result<StatementDesc, PlanError> {
    Ok(StatementDesc::new(None))
}

pub fn plan_create_index(
    scx: &StatementContext,
    mut stmt: CreateIndexStatement<Aug>,
) -> Result<Plan, PlanError> {
    let CreateIndexStatement {
        name,
        on_name,
        in_cluster,
        key_parts,
        with_options,
        if_not_exists,
    } = &mut stmt;
    let on = scx.get_item_by_resolved_name(on_name)?;

    {
        use CatalogItemType::*;
        match on.item_type() {
            Table | Source | View | MaterializedView => {
                if on.replacement_target().is_some() {
                    sql_bail!(
                        "index cannot be created on {} because it is a replacement {}",
                        on_name.full_name_str(),
                        on.item_type(),
                    );
                }
            }
            Sink | MetricSink | Index | Type | Func | Secret | Connection => {
                sql_bail!(
                    "index cannot be created on {} because it is a {}",
                    on_name.full_name_str(),
                    on.item_type(),
                );
            }
        }
    }

    let on_desc = on
        .relation_desc()
        .ok_or_else(|| sql_err!("item does not have a relation description"))?;

    let filled_key_parts = match key_parts {
        Some(kp) => kp.to_vec(),
        None => {
            // `key_parts` is None if we're creating a "default" index.
            // Precompute which column names are unambiguous in a single pass,
            // avoiding the O(n * k) cost of calling get_unambiguous_name per
            // key column.
            let mut name_counts = BTreeMap::new();
            for name in on_desc.iter_names() {
                *name_counts.entry(name).or_insert(0usize) += 1;
            }
            let key = on_desc.typ().default_key();
            key.iter()
                .map(|i| {
                    let name = on_desc.get_name(*i);
                    if name_counts.get(name).copied() == Some(1) {
                        Expr::Identifier(vec![normalize::column_name_ident(name)])
                    } else {
                        Expr::Value(Value::Number((i + 1).to_string()))
                    }
                })
                .collect()
        }
    };
    let keys = query::plan_index_exprs(scx, &on_desc, filled_key_parts.clone())?;

    let index_name = if let Some(name) = name {
        QualifiedItemName {
            qualifiers: on.name().qualifiers.clone(),
            item: normalize::ident(name.clone()),
        }
    } else {
        let mut idx_name = QualifiedItemName {
            qualifiers: on.name().qualifiers.clone(),
            item: on.name().item.clone(),
        };
        if key_parts.is_none() {
            // We're trying to create the "default" index.
            idx_name.item += "_primary_idx";
        } else {
            // Use PG schema for automatically naming indexes:
            // `<table>_<_-separated indexed expressions>_idx`
            let index_name_col_suffix = keys
                .iter()
                .map(|k| match k {
                    mz_expr::MirScalarExpr::Column(i, name) => {
                        match (on_desc.get_unambiguous_name(*i), &name.0) {
                            (Some(col_name), _) => col_name.to_string(),
                            (None, Some(name)) => name.to_string(),
                            (None, None) => format!("{}", i + 1),
                        }
                    }
                    _ => "expr".to_string(),
                })
                .join("_");
            write!(idx_name.item, "_{index_name_col_suffix}_idx")
                .expect("write on strings cannot fail");
            idx_name.item = normalize::ident(Ident::new(&idx_name.item)?)
        }

        if !*if_not_exists {
            scx.catalog.find_available_name(idx_name)
        } else {
            idx_name
        }
    };

    // Check for an object in the catalog with this same name
    let full_name = scx.catalog.resolve_full_name(&index_name);
    let partial_name = PartialItemName::from(full_name.clone());
    // For PostgreSQL compatibility, we need to prevent creating indexes when
    // there is an existing object *or* type of the same name.
    //
    // Technically, we only need to prevent coexistence of indexes and types
    // that have an associated relation (record types but not list/map types).
    // Enforcing that would be more complicated, though. It's backwards
    // compatible to weaken this restriction in the future.
    if let (Ok(item), false, false) = (
        scx.catalog.resolve_item_or_type(&partial_name),
        *if_not_exists,
        scx.pcx().map_or(false, |pcx| pcx.ignore_if_exists_errors),
    ) {
        return Err(PlanError::ItemAlreadyExists {
            name: full_name.to_string(),
            item_type: item.item_type(),
        });
    }

    let options = plan_index_options(scx, with_options.clone())?;
    let cluster_id = match in_cluster {
        None => scx.resolve_cluster(None)?.id(),
        Some(in_cluster) => in_cluster.id,
    };

    *in_cluster = Some(ResolvedClusterName {
        id: cluster_id,
        print_name: None,
    });

    // Normalize `stmt`.
    *name = Some(Ident::new(index_name.item.clone())?);
    *key_parts = Some(filled_key_parts);
    let if_not_exists = *if_not_exists;

    let create_sql = normalize::create_statement(scx, Statement::CreateIndex(stmt))?;
    let compaction_window = options.iter().find_map(|o| {
        #[allow(irrefutable_let_patterns)]
        if let crate::plan::IndexOption::RetainHistory(lcw) = o {
            Some(lcw.clone())
        } else {
            None
        }
    });

    Ok(Plan::CreateIndex(CreateIndexPlan {
        name: index_name,
        index: Index {
            create_sql,
            on: on.global_id(),
            keys,
            cluster_id,
            compaction_window,
        },
        if_not_exists,
    }))
}

pub fn describe_create_type(
    _: &StatementContext,
    _: CreateTypeStatement<Aug>,
) -> Result<StatementDesc, PlanError> {
    Ok(StatementDesc::new(None))
}

pub fn plan_create_type(
    scx: &StatementContext,
    stmt: CreateTypeStatement<Aug>,
) -> Result<Plan, PlanError> {
    let create_sql = normalize::create_statement(scx, Statement::CreateType(stmt.clone()))?;
    let CreateTypeStatement { name, as_type, .. } = stmt;

    // The type being created does not yet exist in the catalog, so its children
    // (list element, map value, record fields) are validated directly. They all
    // draw from one shared budget that also accounts for the root, so creation
    // rejects exactly the types a later direct `scalar_type_from_catalog` call
    // would reject. In particular a wide record whose fields are individually
    // valid but collectively enormous is rejected here rather than materializing
    // an unbounded type tree during sequencing.
    fn validate_data_type(
        scx: &StatementContext,
        data_type: ResolvedDataType,
        as_type: &str,
        key: &str,
        budget: &mut TypeResolutionBudget,
    ) -> Result<(CatalogItemId, Vec<i64>), PlanError> {
        let (id, modifiers) = match data_type {
            ResolvedDataType::Named { id, modifiers, .. } => (id, modifiers),
            _ => sql_bail!(
                "CREATE TYPE ... AS {}option {} can only use named data types, but \
                        found unnamed data type {}. Use CREATE TYPE to create a named type first",
                as_type,
                key,
                data_type.human_readable_name(),
            ),
        };

        let item = scx.catalog.get_item(&id);
        match item.type_details() {
            None => sql_bail!(
                "{} must be of class type, but received {} which is of class {}",
                key,
                scx.catalog.resolve_full_name(item.name()),
                item.item_type()
            ),
            Some(CatalogTypeDetails {
                typ: CatalogType::Char,
                ..
            }) => {
                bail_unsupported!("embedding char type in a list or map")
            }
            _ => {
                // Validate that the modifiers are actually valid, and that the
                // referenced type resolves within the shared budget.
                budget.resolve_child(scx.catalog, id, &modifiers)?;

                Ok((id, modifiers))
            }
        }
    }

    let mut budget = TypeResolutionBudget::for_root(scx.catalog);
    let inner = match as_type {
        CreateTypeAs::List { options } => {
            let CreateTypeListOptionExtracted {
                element_type,
                seen: _,
            } = CreateTypeListOptionExtracted::try_from(options)?;
            let element_type =
                element_type.ok_or_else(|| sql_err!("ELEMENT TYPE option is required"))?;
            let (id, modifiers) =
                validate_data_type(scx, element_type, "LIST ", "ELEMENT TYPE", &mut budget)?;
            CatalogType::List {
                element_reference: id,
                element_modifiers: modifiers,
            }
        }
        CreateTypeAs::Map { options } => {
            let CreateTypeMapOptionExtracted {
                key_type,
                value_type,
                seen: _,
            } = CreateTypeMapOptionExtracted::try_from(options)?;
            let key_type = key_type.ok_or_else(|| sql_err!("KEY TYPE option is required"))?;
            let value_type = value_type.ok_or_else(|| sql_err!("VALUE TYPE option is required"))?;
            // A map's resolved type ignores the key (map keys are always text at
            // runtime), so it is not part of the root's materialized tree and is
            // validated under its own budget.
            let (key_id, key_modifiers) = validate_data_type(
                scx,
                key_type,
                "MAP ",
                "KEY TYPE",
                &mut TypeResolutionBudget::for_root(scx.catalog),
            )?;
            let (value_id, value_modifiers) =
                validate_data_type(scx, value_type, "MAP ", "VALUE TYPE", &mut budget)?;
            CatalogType::Map {
                key_reference: key_id,
                key_modifiers,
                value_reference: value_id,
                value_modifiers,
            }
        }
        CreateTypeAs::Record { column_defs } => {
            let mut fields = vec![];
            for column_def in column_defs {
                let data_type = column_def.data_type;
                let key = ident(column_def.name.clone());
                let (id, modifiers) = validate_data_type(scx, data_type, "", &key, &mut budget)?;
                fields.push(CatalogRecordField {
                    name: ColumnName::from(key.clone()),
                    type_reference: id,
                    type_modifiers: modifiers,
                });
            }
            CatalogType::Record { fields }
        }
    };

    let name = scx.allocate_qualified_name(normalize::unresolved_item_name(name)?)?;

    // Check for an object in the catalog with this same name
    let full_name = scx.catalog.resolve_full_name(&name);
    let partial_name = PartialItemName::from(full_name.clone());
    // For PostgreSQL compatibility, we need to prevent creating types when
    // there is an existing object *or* type of the same name.
    if let Ok(item) = scx.catalog.resolve_item_or_type(&partial_name) {
        if item.item_type().conflicts_with_type() {
            return Err(PlanError::ItemAlreadyExists {
                name: full_name.to_string(),
                item_type: item.item_type(),
            });
        }
    }

    Ok(Plan::CreateType(CreateTypePlan {
        name,
        typ: Type { create_sql, inner },
    }))
}

generate_extracted_config!(CreateTypeListOption, (ElementType, ResolvedDataType));

generate_extracted_config!(
    CreateTypeMapOption,
    (KeyType, ResolvedDataType),
    (ValueType, ResolvedDataType)
);

#[derive(Debug)]
pub enum PlannedAlterRoleOption {
    Attributes(PlannedRoleAttributes),
    Variable(PlannedRoleVariable),
}

#[derive(Debug, Clone)]
pub struct PlannedRoleAttributes {
    pub inherit: Option<bool>,
    pub password: Option<Password>,
    pub scram_iterations: Option<NonZeroU32>,
    /// `nopassword` is set to true if the password is from the parser is None.
    /// This is semantically different than not supplying a password at all,
    /// to allow for unsetting a password.
    pub nopassword: Option<bool>,
    pub superuser: Option<bool>,
    pub login: Option<bool>,
}

fn plan_role_attributes(
    options: Vec<RoleAttribute>,
    scx: &StatementContext,
) -> Result<PlannedRoleAttributes, PlanError> {
    let mut planned_attributes = PlannedRoleAttributes {
        inherit: None,
        password: None,
        scram_iterations: None,
        superuser: None,
        login: None,
        nopassword: None,
    };

    for option in options {
        match option {
            RoleAttribute::Inherit | RoleAttribute::NoInherit
                if planned_attributes.inherit.is_some() =>
            {
                sql_bail!("conflicting or redundant options");
            }
            RoleAttribute::CreateCluster | RoleAttribute::NoCreateCluster => {
                bail_never_supported!(
                    "CREATECLUSTER attribute",
                    "sql/create-role/#details",
                    "Use system privileges instead."
                );
            }
            RoleAttribute::CreateDB | RoleAttribute::NoCreateDB => {
                bail_never_supported!(
                    "CREATEDB attribute",
                    "sql/create-role/#details",
                    "Use system privileges instead."
                );
            }
            RoleAttribute::CreateRole | RoleAttribute::NoCreateRole => {
                bail_never_supported!(
                    "CREATEROLE attribute",
                    "sql/create-role/#details",
                    "Use system privileges instead."
                );
            }
            RoleAttribute::Password(_) if planned_attributes.password.is_some() => {
                sql_bail!("conflicting or redundant options");
            }

            RoleAttribute::Inherit => planned_attributes.inherit = Some(true),
            RoleAttribute::NoInherit => planned_attributes.inherit = Some(false),
            RoleAttribute::Password(password) => {
                if let Some(password) = password {
                    planned_attributes.password = Some(password.into());
                    planned_attributes.scram_iterations =
                        Some(scx.catalog.system_vars().scram_iterations())
                } else {
                    planned_attributes.nopassword = Some(true);
                }
            }
            RoleAttribute::SuperUser => {
                if planned_attributes.superuser == Some(false) {
                    sql_bail!("conflicting or redundant options");
                }
                planned_attributes.superuser = Some(true);
            }
            RoleAttribute::NoSuperUser => {
                if planned_attributes.superuser == Some(true) {
                    sql_bail!("conflicting or redundant options");
                }
                planned_attributes.superuser = Some(false);
            }
            RoleAttribute::Login => {
                if planned_attributes.login == Some(false) {
                    sql_bail!("conflicting or redundant options");
                }
                planned_attributes.login = Some(true);
            }
            RoleAttribute::NoLogin => {
                if planned_attributes.login == Some(true) {
                    sql_bail!("conflicting or redundant options");
                }
                planned_attributes.login = Some(false);
            }
        }
    }
    if planned_attributes.inherit == Some(false) {
        bail_unsupported!("non inherit roles");
    }

    Ok(planned_attributes)
}

#[derive(Debug)]
pub enum PlannedRoleVariable {
    Set { name: String, value: VariableValue },
    Reset { name: String },
}

impl PlannedRoleVariable {
    pub fn name(&self) -> &str {
        match self {
            PlannedRoleVariable::Set { name, .. } => name,
            PlannedRoleVariable::Reset { name } => name,
        }
    }
}

fn plan_role_variable(
    scx: &StatementContext,
    variable: SetRoleVar,
) -> Result<PlannedRoleVariable, PlanError> {
    let plan = match variable {
        SetRoleVar::Set { name, value } => {
            let name = name.to_string();
            let value = scl::plan_set_variable_to(value)?;
            // Gate feature-flagged isolation levels, matching the `SET` and
            // connection-option paths in `SessionVars::set`.
            if let VariableValue::Values(values) = &value {
                vars::check_transaction_isolation_feature_flag(
                    &name,
                    VarInput::SqlSet(values),
                    scx.catalog.system_vars(),
                )?;
            }
            PlannedRoleVariable::Set { name, value }
        }
        SetRoleVar::Reset { name } => PlannedRoleVariable::Reset {
            name: name.to_string(),
        },
    };
    Ok(plan)
}

pub fn describe_create_role(
    _: &StatementContext,
    _: CreateRoleStatement,
) -> Result<StatementDesc, PlanError> {
    Ok(StatementDesc::new(None))
}

pub fn plan_create_role(
    scx: &StatementContext,
    CreateRoleStatement { name, options }: CreateRoleStatement,
) -> Result<Plan, PlanError> {
    let attributes = plan_role_attributes(options, scx)?;
    Ok(Plan::CreateRole(CreateRolePlan {
        name: normalize::ident(name),
        attributes: attributes.into(),
    }))
}

pub fn plan_create_network_policy(
    ctx: &StatementContext,
    CreateNetworkPolicyStatement { name, options }: CreateNetworkPolicyStatement<Aug>,
) -> Result<Plan, PlanError> {
    ctx.require_feature_flag(&vars::ENABLE_NETWORK_POLICIES)?;
    let policy_options: NetworkPolicyOptionExtracted = options.try_into()?;

    let Some(rule_defs) = policy_options.rules else {
        sql_bail!("RULES must be specified when creating network policies.");
    };

    let mut rules = vec![];
    for NetworkPolicyRuleDefinition { name, options } in rule_defs {
        let NetworkPolicyRuleOptionExtracted {
            seen: _,
            direction,
            action,
            address,
        } = options.try_into()?;
        let (direction, action, address) = match (direction, action, address) {
            (Some(direction), Some(action), Some(address)) => (
                NetworkPolicyRuleDirection::try_from(direction.as_str())?,
                NetworkPolicyRuleAction::try_from(action.as_str())?,
                PolicyAddress::try_from(address.as_str())?,
            ),
            (_, _, _) => {
                sql_bail!("Direction, Address, and Action must specified when creating a rule")
            }
        };
        rules.push(NetworkPolicyRule {
            name: normalize::ident(name),
            direction,
            action,
            address,
        });
    }

    if rules.len()
        > ctx
            .catalog
            .system_vars()
            .max_rules_per_network_policy()
            .try_into()?
    {
        sql_bail!("RULES count exceeds max_rules_per_network_policy.")
    }

    Ok(Plan::CreateNetworkPolicy(CreateNetworkPolicyPlan {
        name: normalize::ident(name),
        rules,
    }))
}

pub fn plan_alter_network_policy(
    ctx: &StatementContext,
    AlterNetworkPolicyStatement { name, options }: AlterNetworkPolicyStatement<Aug>,
) -> Result<Plan, PlanError> {
    ctx.require_feature_flag(&vars::ENABLE_NETWORK_POLICIES)?;

    let policy_options: NetworkPolicyOptionExtracted = options.try_into()?;
    let policy = ctx
        .catalog
        .resolve_network_policy(normalize::ident_ref(&name))?;

    let Some(rule_defs) = policy_options.rules else {
        sql_bail!("RULES must be specified when creating network policies.");
    };

    let mut rules = vec![];
    for NetworkPolicyRuleDefinition { name, options } in rule_defs {
        let NetworkPolicyRuleOptionExtracted {
            seen: _,
            direction,
            action,
            address,
        } = options.try_into()?;

        let (direction, action, address) = match (direction, action, address) {
            (Some(direction), Some(action), Some(address)) => (
                NetworkPolicyRuleDirection::try_from(direction.as_str())?,
                NetworkPolicyRuleAction::try_from(action.as_str())?,
                PolicyAddress::try_from(address.as_str())?,
            ),
            (_, _, _) => {
                sql_bail!("Direction, Address, and Action must specified when creating a rule")
            }
        };
        rules.push(NetworkPolicyRule {
            name: normalize::ident(name),
            direction,
            action,
            address,
        });
    }
    if rules.len()
        > ctx
            .catalog
            .system_vars()
            .max_rules_per_network_policy()
            .try_into()?
    {
        sql_bail!("RULES count exceeds max_rules_per_network_policy.")
    }

    Ok(Plan::AlterNetworkPolicy(AlterNetworkPolicyPlan {
        id: policy.id(),
        name: normalize::ident(name),
        rules,
    }))
}

pub fn describe_create_cluster(
    _: &StatementContext,
    _: CreateClusterStatement<Aug>,
) -> Result<StatementDesc, PlanError> {
    Ok(StatementDesc::new(None))
}

// WARNING:
// DO NOT set any `Default` value here using the built-in mechanism of `generate_extracted_config`!
// These options are also used in ALTER CLUSTER, where not giving an option means that the value of
// that option stays the same. If you were to give a default value here, then not giving that option
// to ALTER CLUSTER would always reset the value of that option to the default.
generate_extracted_config!(
    ClusterOption,
    (AutoScalingStrategy, ClusterAutoScalingStrategyOptionValue),
    (AvailabilityZones, Vec<String>),
    (Disk, bool),
    (ExperimentalArrangementCompression, bool),
    (IntrospectionDebugging, bool),
    (IntrospectionInterval, OptionalDuration),
    (Managed, bool),
    (Replicas, Vec<ReplicaDefinition<Aug>>),
    (ReplicationFactor, u32),
    (Size, String),
    (Schedule, ClusterScheduleOptionValue),
    (WorkloadClass, OptionalString)
);

generate_extracted_config!(
    NetworkPolicyOption,
    (Rules, Vec<NetworkPolicyRuleDefinition<Aug>>)
);

generate_extracted_config!(
    NetworkPolicyRuleOption,
    (Direction, String),
    (Action, String),
    (Address, String)
);

generate_extracted_config!(ClusterAlterOption, (Wait, ClusterAlterOptionValue<Aug>));

generate_extracted_config!(
    ClusterAlterUntilReadyOption,
    (Timeout, Duration),
    (OnTimeout, String)
);

generate_extracted_config!(
    ClusterFeature,
    (ReoptimizeImportedViews, Option<bool>, Default(None)),
    (EnableEagerDeltaJoins, Option<bool>, Default(None)),
    (EnableNewOuterJoinLowering, Option<bool>, Default(None)),
    (EnableVariadicLeftJoinLowering, Option<bool>, Default(None)),
    (EnableLetrecFixpointAnalysis, Option<bool>, Default(None)),
    (EnableJoinPrioritizeArranged, Option<bool>, Default(None)),
    (
        EnableProjectionPushdownAfterRelationCse,
        Option<bool>,
        Default(None)
    ),
    (
        EnableUnionCancellationAfterRelationCse,
        Option<bool>,
        Default(None)
    )
);

/// Convert a [`CreateClusterStatement`] into a [`Plan`].
///
/// The reverse of [`unplan_create_cluster`].
pub fn plan_create_cluster(
    scx: &StatementContext,
    stmt: CreateClusterStatement<Aug>,
) -> Result<Plan, PlanError> {
    let plan = plan_create_cluster_inner(scx, stmt)?;

    // Roundtrip through unplan and make sure that we end up with the same plan.
    if let CreateClusterVariant::Managed(_) = &plan.variant {
        let stmt = unplan_create_cluster(scx, plan.clone())
            .map_err(|e| PlanError::Replan(e.to_string()))?;
        let create_sql = stmt.to_ast_string_stable();
        let stmt = parse::parse(&create_sql)
            .map_err(|e| PlanError::Replan(e.to_string()))?
            .into_element()
            .ast;
        let (stmt, _resolved_ids) =
            names::resolve(scx.catalog, stmt).map_err(|e| PlanError::Replan(e.to_string()))?;
        let stmt = match stmt {
            Statement::CreateCluster(stmt) => stmt,
            stmt => {
                return Err(PlanError::Replan(format!(
                    "replan does not match: plan={plan:?}, create_sql={create_sql:?}, stmt={stmt:?}"
                )));
            }
        };
        let replan =
            plan_create_cluster_inner(scx, stmt).map_err(|e| PlanError::Replan(e.to_string()))?;
        if plan != replan {
            return Err(PlanError::Replan(format!(
                "replan does not match: plan={plan:?}, replan={replan:?}"
            )));
        }
    }

    Ok(Plan::CreateCluster(plan))
}

pub fn plan_create_cluster_inner(
    scx: &StatementContext,
    CreateClusterStatement {
        name,
        options,
        features,
        if_not_exists,
    }: CreateClusterStatement<Aug>,
) -> Result<CreateClusterPlan, PlanError> {
    let ClusterOptionExtracted {
        auto_scaling_strategy,
        availability_zones,
        experimental_arrangement_compression,
        introspection_debugging,
        introspection_interval,
        managed,
        replicas,
        replication_factor,
        seen: _,
        size,
        disk,
        schedule,
        workload_class,
    }: ClusterOptionExtracted = options.try_into()?;

    let managed = managed.unwrap_or_else(|| replicas.is_none());

    if !scx.catalog.active_role_id().is_system() {
        if !features.is_empty() {
            sql_bail!("FEATURES not supported for non-system users");
        }
        if workload_class.is_some() {
            sql_bail!("WORKLOAD CLASS not supported for non-system users");
        }
    }

    let schedule = schedule.unwrap_or(ClusterScheduleOptionValue::Manual);
    let workload_class = workload_class.and_then(|v| v.0);

    if managed {
        if replicas.is_some() {
            sql_bail!("REPLICAS not supported for managed clusters");
        }
        let Some(size) = size else {
            sql_bail!("SIZE must be specified for managed clusters");
        };

        if disk.is_some() {
            // The `DISK` option is a no-op for legacy cluster sizes and was never allowed for
            // `cc` sizes. The long term plan is to phase out the legacy sizes, at which point
            // we'll be able to remove the `DISK` option entirely.
            if scx.catalog.is_cluster_size_cc(&size) {
                sql_bail!(
                    "DISK option not supported for modern cluster sizes because disk is always enabled"
                );
            }

            scx.catalog
                .add_notice(PlanNotice::ReplicaDiskOptionDeprecated);
        }

        let compute = plan_compute_replica_config(
            introspection_interval,
            introspection_debugging.unwrap_or(false),
            experimental_arrangement_compression.unwrap_or(false),
        )?;

        let replication_factor = if matches!(schedule, ClusterScheduleOptionValue::Manual) {
            replication_factor.unwrap_or_else(|| {
                scx.catalog
                    .system_vars()
                    .default_cluster_replication_factor()
            })
        } else {
            scx.require_feature_flag(&ENABLE_CLUSTER_SCHEDULE_REFRESH)?;
            if replication_factor.is_some() {
                sql_bail!(
                    "REPLICATION FACTOR cannot be given together with any SCHEDULE other than MANUAL"
                );
            }
            // If we have a non-trivial schedule, then let's not have any replicas initially,
            // to avoid quickly going back and forth if the schedule doesn't want a replica
            // initially.
            0
        };
        let availability_zones = availability_zones.unwrap_or_default();

        if !availability_zones.is_empty() {
            scx.require_feature_flag(&vars::ENABLE_MANAGED_CLUSTER_AVAILABILITY_ZONES)?;
        }

        // Plan OptimizerFeatureOverrides.
        let ClusterFeatureExtracted {
            reoptimize_imported_views,
            enable_eager_delta_joins,
            enable_new_outer_join_lowering,
            enable_variadic_left_join_lowering,
            enable_letrec_fixpoint_analysis,
            enable_join_prioritize_arranged,
            enable_projection_pushdown_after_relation_cse,
            enable_union_cancellation_after_relation_cse,
            seen: _,
        } = ClusterFeatureExtracted::try_from(features)?;
        let optimizer_feature_overrides = OptimizerFeatureOverrides {
            reoptimize_imported_views,
            enable_eager_delta_joins,
            enable_new_outer_join_lowering,
            enable_variadic_left_join_lowering,
            enable_letrec_fixpoint_analysis,
            enable_join_prioritize_arranged,
            enable_projection_pushdown_after_relation_cse,
            enable_union_cancellation_after_relation_cse,
            ..Default::default()
        };

        // The gate applies to new DDL only: clusters are stored structurally,
        // not as SQL, so existing configs survive a flag rollback.
        let auto_scaling_strategy = match auto_scaling_strategy {
            Some(value) => {
                scx.require_feature_flag(&ENABLE_AUTO_SCALING_STRATEGY)?;
                let strategy = plan_auto_scaling_strategy(value)?;
                if let Some(strategy) = &strategy {
                    let schedule_non_manual =
                        !matches!(schedule, ClusterScheduleOptionValue::Manual);
                    validate_auto_scaling_strategy(strategy, Some(&size), schedule_non_manual)?;
                }
                strategy
            }
            None => None,
        };

        let schedule = plan_cluster_schedule(schedule)?;

        Ok(CreateClusterPlan {
            name: normalize::ident(name),
            variant: CreateClusterVariant::Managed(CreateClusterManagedPlan {
                replication_factor,
                size,
                availability_zones,
                compute,
                optimizer_feature_overrides,
                schedule,
                auto_scaling_strategy,
            }),
            workload_class,
            if_not_exists,
        })
    } else {
        let Some(replica_defs) = replicas else {
            sql_bail!("REPLICAS must be specified for unmanaged clusters");
        };
        if auto_scaling_strategy.is_some() {
            sql_bail!("AUTO SCALING STRATEGY not supported for unmanaged clusters");
        }
        if availability_zones.is_some() {
            sql_bail!("AVAILABILITY ZONES not supported for unmanaged clusters");
        }
        if replication_factor.is_some() {
            sql_bail!("REPLICATION FACTOR not supported for unmanaged clusters");
        }
        if introspection_debugging.is_some() {
            sql_bail!("INTROSPECTION DEBUGGING not supported for unmanaged clusters");
        }
        if introspection_interval.is_some() {
            sql_bail!("INTROSPECTION INTERVAL not supported for unmanaged clusters");
        }
        if experimental_arrangement_compression.is_some() {
            sql_bail!("EXPERIMENTAL ARRANGEMENT COMPRESSION not supported for unmanaged clusters");
        }
        if size.is_some() {
            sql_bail!("SIZE not supported for unmanaged clusters");
        }
        if disk.is_some() {
            sql_bail!("DISK not supported for unmanaged clusters");
        }
        if !features.is_empty() {
            sql_bail!("FEATURES not supported for unmanaged clusters");
        }
        if !matches!(schedule, ClusterScheduleOptionValue::Manual) {
            sql_bail!(
                "cluster schedules other than MANUAL are not supported for unmanaged clusters"
            );
        }

        let mut replicas = vec![];
        for ReplicaDefinition { name, options } in replica_defs {
            replicas.push((normalize::ident(name), plan_replica_config(scx, options)?));
        }

        Ok(CreateClusterPlan {
            name: normalize::ident(name),
            variant: CreateClusterVariant::Unmanaged(CreateClusterUnmanagedPlan { replicas }),
            workload_class,
            if_not_exists,
        })
    }
}

/// Convert a [`CreateClusterPlan`] into a [`CreateClusterStatement`].
///
/// The reverse of [`plan_create_cluster`], so this renders `IF NOT EXISTS`
/// when the plan carries it. A caller that wants the cluster's canonical
/// definition rather than a faithful reverse must pass `if_not_exists: false`,
/// which is what `Cluster::try_to_plan` produces for `SHOW CREATE CLUSTER`.
pub fn unplan_create_cluster(
    scx: &StatementContext,
    CreateClusterPlan {
        name,
        variant,
        workload_class,
        if_not_exists,
    }: CreateClusterPlan,
) -> Result<CreateClusterStatement<Aug>, PlanError> {
    match variant {
        CreateClusterVariant::Managed(CreateClusterManagedPlan {
            replication_factor,
            size,
            availability_zones,
            compute,
            optimizer_feature_overrides,
            schedule,
            auto_scaling_strategy,
        }) => {
            let schedule = unplan_cluster_schedule(schedule);
            let auto_scaling_strategy = auto_scaling_strategy
                .as_ref()
                .map(unplan_auto_scaling_strategy);
            let OptimizerFeatureOverrides {
                enable_reduce_mfp_fusion: _,
                enable_cardinality_estimates: _,
                persist_fast_path_limit: _,
                reoptimize_imported_views,
                enable_eager_delta_joins,
                enable_new_outer_join_lowering,
                enable_variadic_left_join_lowering,
                enable_letrec_fixpoint_analysis,
                enable_join_prioritize_arranged,
                enable_projection_pushdown_after_relation_cse,
                enable_union_cancellation_after_relation_cse,
                enable_less_reduce_in_eqprop: _,
                enable_dequadratic_eqprop_map: _,
                enable_eq_classes_withholding_errors: _,
                enable_fast_path_plan_insights: _,
                enable_cast_elimination: _,
                enable_case_literal_transform: _,
                enable_simplify_quantified_comparisons: _,
                enable_simplify_from_less_existence: _,
                enable_coalesce_case_transform: _,
                enable_will_distinct_propagation: _,
                enable_fixed_correlated_cte_lowering: _,
            } = optimizer_feature_overrides;
            // The ones from above that don't occur below are not wired up to cluster features.
            let features_extracted = ClusterFeatureExtracted {
                // Seen is ignored when unplanning.
                seen: Default::default(),
                reoptimize_imported_views,
                enable_eager_delta_joins,
                enable_new_outer_join_lowering,
                enable_variadic_left_join_lowering,
                enable_letrec_fixpoint_analysis,
                enable_join_prioritize_arranged,
                enable_projection_pushdown_after_relation_cse,
                enable_union_cancellation_after_relation_cse,
            };
            let features = features_extracted.into_values(scx.catalog);
            let availability_zones = if availability_zones.is_empty() {
                None
            } else {
                Some(availability_zones)
            };
            let (introspection_interval, introspection_debugging, arrangement_compression) =
                unplan_compute_replica_config(compute);
            // Replication factor cannot be explicitly specified with a refresh schedule, it's
            // always 1 or less.
            let replication_factor = match &schedule {
                ClusterScheduleOptionValue::Manual => Some(replication_factor),
                ClusterScheduleOptionValue::Refresh { .. } => {
                    // A cluster with a refresh schedule is turned On/Off by the cluster scheduling
                    // policy, so its replication factor should always be 0 or 1, and CREATE/ALTER
                    // reject setting both a non-MANUAL schedule and a higher replication factor. If
                    // we nevertheless find one (e.g., a cluster left in an invalid state by an
                    // older version), log loudly rather than crashing the coordinator: the
                    // replication factor is omitted from the rendered statement regardless.
                    soft_assert_or_log!(
                        replication_factor <= 1,
                        "replication factor, {replication_factor:?}, must be <= 1 with a refresh schedule"
                    );
                    None
                }
            };
            let workload_class = workload_class.map(|s| OptionalString(Some(s)));
            let options_extracted = ClusterOptionExtracted {
                // Seen is ignored when unplanning.
                seen: Default::default(),
                auto_scaling_strategy,
                availability_zones,
                disk: None,
                experimental_arrangement_compression: Some(arrangement_compression),
                introspection_debugging: Some(introspection_debugging),
                introspection_interval,
                managed: Some(true),
                replicas: None,
                replication_factor,
                size: Some(size),
                schedule: Some(schedule),
                workload_class,
            };
            let options = options_extracted.into_values(scx.catalog);
            let name = Ident::new_unchecked(name);
            Ok(CreateClusterStatement {
                name,
                options,
                features,
                if_not_exists,
            })
        }
        CreateClusterVariant::Unmanaged(_) => {
            bail_unsupported!("SHOW CREATE for unmanaged clusters")
        }
    }
}

generate_extracted_config!(
    ReplicaOption,
    (AvailabilityZone, String),
    (BilledAs, String),
    (ComputeAddresses, Vec<String>),
    (ComputectlAddresses, Vec<String>),
    (Disk, bool),
    (ExperimentalArrangementCompression, bool, Default(false)),
    (Internal, bool, Default(false)),
    (IntrospectionDebugging, bool, Default(false)),
    (IntrospectionInterval, OptionalDuration),
    (Size, String),
    (StorageAddresses, Vec<String>),
    (StoragectlAddresses, Vec<String>),
    (Workers, u16)
);

fn plan_replica_config(
    scx: &StatementContext,
    options: Vec<ReplicaOption<Aug>>,
) -> Result<ReplicaConfig, PlanError> {
    let ReplicaOptionExtracted {
        availability_zone,
        billed_as,
        computectl_addresses,
        disk,
        experimental_arrangement_compression,
        internal,
        introspection_debugging,
        introspection_interval,
        size,
        storagectl_addresses,
        ..
    }: ReplicaOptionExtracted = options.try_into()?;

    let compute = plan_compute_replica_config(
        introspection_interval,
        introspection_debugging,
        experimental_arrangement_compression,
    )?;

    match (
        size,
        availability_zone,
        billed_as,
        storagectl_addresses,
        computectl_addresses,
    ) {
        // Common cases we expect end users to hit.
        (None, _, None, None, None) => {
            // We don't mention the unmanaged options in the error message
            // because they are only available in unsafe mode.
            sql_bail!("SIZE option must be specified");
        }
        (Some(size), availability_zone, billed_as, None, None) => {
            if disk.is_some() {
                // The `DISK` option is a no-op for legacy cluster sizes and was never allowed for
                // `cc` sizes. The long term plan is to phase out the legacy sizes, at which point
                // we'll be able to remove the `DISK` option entirely.
                if scx.catalog.is_cluster_size_cc(&size) {
                    sql_bail!(
                        "DISK option not supported for modern cluster sizes because disk is always enabled"
                    );
                }

                scx.catalog
                    .add_notice(PlanNotice::ReplicaDiskOptionDeprecated);
            }

            Ok(ReplicaConfig::Orchestrated {
                size,
                availability_zone,
                compute,
                billed_as,
                internal,
            })
        }

        (None, None, None, storagectl_addresses, computectl_addresses) => {
            scx.require_feature_flag(&vars::UNSAFE_ENABLE_UNORCHESTRATED_CLUSTER_REPLICAS)?;

            // When manually testing Materialize in unsafe mode, it's easy to
            // accidentally omit one of these options, so we try to produce
            // helpful error messages.
            let Some(storagectl_addrs) = storagectl_addresses else {
                sql_bail!("missing STORAGECTL ADDRESSES option");
            };
            let Some(computectl_addrs) = computectl_addresses else {
                sql_bail!("missing COMPUTECTL ADDRESSES option");
            };

            if storagectl_addrs.len() != computectl_addrs.len() {
                sql_bail!(
                    "COMPUTECTL ADDRESSES and STORAGECTL ADDRESSES must have the same length"
                );
            }

            if disk.is_some() {
                sql_bail!("DISK can't be specified for unorchestrated clusters");
            }

            Ok(ReplicaConfig::Unorchestrated {
                storagectl_addrs,
                computectl_addrs,
                compute,
            })
        }
        _ => {
            // We don't bother trying to produce a more helpful error message
            // here because no user is likely to hit this path.
            sql_bail!("invalid mixture of orchestrated and unorchestrated replica options");
        }
    }
}

/// Convert an [`Option<OptionalDuration>`] and [`bool`] into a [`ComputeReplicaConfig`].
///
/// The reverse of [`unplan_compute_replica_config`].
fn plan_compute_replica_config(
    introspection_interval: Option<OptionalDuration>,
    introspection_debugging: bool,
    arrangement_compression: bool,
) -> Result<ComputeReplicaConfig, PlanError> {
    let introspection_interval = introspection_interval
        .map(|OptionalDuration(i)| i)
        .unwrap_or(Some(DEFAULT_REPLICA_LOGGING_INTERVAL));
    let introspection = match introspection_interval {
        Some(interval) => Some(ComputeReplicaIntrospectionConfig {
            interval,
            debugging: introspection_debugging,
        }),
        None if introspection_debugging => {
            sql_bail!("INTROSPECTION DEBUGGING cannot be specified without INTROSPECTION INTERVAL")
        }
        None => None,
    };
    let compute = ComputeReplicaConfig {
        introspection,
        arrangement_compression,
    };
    Ok(compute)
}

/// Convert a [`ComputeReplicaConfig`] into its introspection interval, introspection debugging,
/// and arrangement compression option values.
///
/// The reverse of [`plan_compute_replica_config`].
fn unplan_compute_replica_config(
    compute_replica_config: ComputeReplicaConfig,
) -> (Option<OptionalDuration>, bool, bool) {
    let ComputeReplicaConfig {
        introspection,
        arrangement_compression,
    } = compute_replica_config;
    match introspection {
        Some(ComputeReplicaIntrospectionConfig {
            debugging,
            interval,
        }) => (
            Some(OptionalDuration(Some(interval))),
            debugging,
            arrangement_compression,
        ),
        None => (Some(OptionalDuration(None)), false, arrangement_compression),
    }
}

/// Convert a [`ClusterScheduleOptionValue`] into a [`ClusterSchedule`].
///
/// The reverse of [`unplan_cluster_schedule`].
fn plan_cluster_schedule(
    schedule: ClusterScheduleOptionValue,
) -> Result<ClusterSchedule, PlanError> {
    Ok(match schedule {
        ClusterScheduleOptionValue::Manual => ClusterSchedule::Manual,
        // If `HYDRATION TIME ESTIMATE` is not explicitly given, we default to 0.
        ClusterScheduleOptionValue::Refresh {
            hydration_time_estimate: None,
        } => ClusterSchedule::Refresh {
            hydration_time_estimate: Duration::from_millis(0),
        },
        // Otherwise we convert the `IntervalValue` to a `Duration`.
        ClusterScheduleOptionValue::Refresh {
            hydration_time_estimate: Some(interval_value),
        } => {
            let interval = Interval::try_from_value(Value::Interval(interval_value))?;
            if interval.as_microseconds() < 0 {
                sql_bail!(
                    "HYDRATION TIME ESTIMATE must be non-negative; got: {}",
                    interval
                );
            }
            if interval.months != 0 {
                // This limitation is because we want this interval to be cleanly convertable
                // to a unix epoch timestamp difference. When the interval involves months, then
                // this is not true anymore, because months have variable lengths.
                sql_bail!("HYDRATION TIME ESTIMATE must not involve units larger than days");
            }
            let duration = interval.duration()?;
            if u64::try_from(duration.as_millis()).is_err()
                || Interval::from_duration(&duration).is_err()
            {
                sql_bail!("HYDRATION TIME ESTIMATE too large");
            }
            ClusterSchedule::Refresh {
                hydration_time_estimate: duration,
            }
        }
    })
}

/// Convert a [`ClusterSchedule`] into a [`ClusterScheduleOptionValue`].
///
/// The reverse of [`plan_cluster_schedule`].
fn unplan_cluster_schedule(schedule: ClusterSchedule) -> ClusterScheduleOptionValue {
    match schedule {
        ClusterSchedule::Manual => ClusterScheduleOptionValue::Manual,
        ClusterSchedule::Refresh {
            hydration_time_estimate,
        } => {
            let interval = Interval::from_duration(&hydration_time_estimate)
                .expect("planning ensured that this is convertible back to Interval");
            let interval_value = literal::unplan_interval(&interval);
            ClusterScheduleOptionValue::Refresh {
                hydration_time_estimate: Some(interval_value),
            }
        }
    }
}

/// Convert a [`ClusterAutoScalingStrategyOptionValue`] into an
/// [`AutoScalingStrategy`]. An empty block (no sub-policies) maps to `None`
/// (autoscaling disabled), so an empty `AUTO SCALING STRATEGY = ()` behaves like
/// a reset. Cross-config invariants are checked separately by
/// [`validate_auto_scaling_strategy`].
///
/// The reverse of [`unplan_auto_scaling_strategy`].
fn plan_auto_scaling_strategy(
    value: ClusterAutoScalingStrategyOptionValue,
) -> Result<Option<AutoScalingStrategy>, PlanError> {
    let ClusterAutoScalingStrategyOptionValue { on_hydration } = value;
    let Some(on_hydration) = on_hydration else {
        // An empty block disables autoscaling.
        return Ok(None);
    };

    let hydration_size = String::try_from_value(on_hydration.hydration_size)?;

    let linger_duration = on_hydration
        .linger_duration
        .map(Duration::try_from_value)
        .transpose()?;

    Ok(Some(AutoScalingStrategy {
        on_hydration: Some(OnHydration {
            hydration_size,
            linger_duration,
        }),
    }))
}

/// Validate an [`AutoScalingStrategy`] against the cluster's effective `SIZE`
/// and `SCHEDULE`. Rejects a burst `HYDRATION SIZE` equal to the cluster `SIZE`
/// (a no-op burst) and the `AUTO SCALING STRATEGY` + non-MANUAL `SCHEDULE`
/// combination.
///
/// `cluster_size` is `None` only when the effective size is unknown at plan time
/// (an unmanaged→managed `ALTER` that does not set `SIZE`, which fails later for
/// the missing size); the size equality check is then skipped.
fn validate_auto_scaling_strategy(
    strategy: &AutoScalingStrategy,
    cluster_size: Option<&str>,
    schedule_non_manual: bool,
) -> Result<(), PlanError> {
    if let (Some(on_hydration), Some(cluster_size)) = (&strategy.on_hydration, cluster_size) {
        if on_hydration.hydration_size == cluster_size {
            return Err(PlanError::HydrationSizeEqualsClusterSize {
                size: cluster_size.to_string(),
            });
        }
    }
    if schedule_non_manual {
        sql_bail!("AUTO SCALING STRATEGY cannot be combined with a SCHEDULE other than MANUAL");
    }
    Ok(())
}

/// Convert an [`AutoScalingStrategy`] back into a
/// [`ClusterAutoScalingStrategyOptionValue`] for `SHOW CREATE CLUSTER` rendering.
///
/// The reverse of [`plan_auto_scaling_strategy`].
fn unplan_auto_scaling_strategy(
    strategy: &AutoScalingStrategy,
) -> ClusterAutoScalingStrategyOptionValue {
    ClusterAutoScalingStrategyOptionValue {
        on_hydration: strategy
            .on_hydration
            .as_ref()
            .map(|on_hydration| OnHydrationOptionValue {
                hydration_size: Value::String(on_hydration.hydration_size.clone()),
                linger_duration: on_hydration.linger_duration.map(|d| {
                    let interval = Interval::from_duration(&d)
                        .expect("planning ensured this is convertible back to Interval");
                    Value::Interval(literal::unplan_interval(&interval))
                }),
            }),
    }
}

pub fn describe_create_cluster_replica(
    _: &StatementContext,
    _: CreateClusterReplicaStatement<Aug>,
) -> Result<StatementDesc, PlanError> {
    Ok(StatementDesc::new(None))
}

pub fn plan_create_cluster_replica(
    scx: &StatementContext,
    CreateClusterReplicaStatement {
        definition: ReplicaDefinition { name, options },
        of_cluster,
        if_not_exists,
    }: CreateClusterReplicaStatement<Aug>,
) -> Result<Plan, PlanError> {
    let cluster = scx
        .catalog
        .resolve_cluster(Some(&normalize::ident(of_cluster)))?;

    let config = plan_replica_config(scx, options)?;

    if let ReplicaConfig::Orchestrated { internal: true, .. } = &config {
        if MANAGED_REPLICA_PATTERN.is_match(name.as_str()) {
            return Err(PlanError::MangedReplicaName(name.into_string()));
        }
    } else {
        ensure_cluster_is_not_managed(scx, cluster.id())?;
    }

    Ok(Plan::CreateClusterReplica(CreateClusterReplicaPlan {
        name: normalize::ident(name),
        cluster_id: cluster.id(),
        config,
        if_not_exists,
    }))
}

pub fn describe_create_secret(
    _: &StatementContext,
    _: CreateSecretStatement<Aug>,
) -> Result<StatementDesc, PlanError> {
    Ok(StatementDesc::new(None))
}

pub fn plan_create_secret(
    scx: &StatementContext,
    stmt: CreateSecretStatement<Aug>,
) -> Result<Plan, PlanError> {
    let CreateSecretStatement {
        name,
        if_not_exists,
        value,
    } = &stmt;

    let name = scx.allocate_qualified_name(normalize::unresolved_item_name(name.to_owned())?)?;
    let mut create_sql_statement = stmt.clone();
    create_sql_statement.value = Expr::Value(Value::String("********".to_string()));
    let create_sql =
        normalize::create_statement(scx, Statement::CreateSecret(create_sql_statement))?;
    let secret_as = query::plan_secret_as(scx, value.clone())?;

    let secret = Secret {
        create_sql,
        secret_as,
    };

    Ok(Plan::CreateSecret(CreateSecretPlan {
        name,
        secret,
        if_not_exists: *if_not_exists,
    }))
}

pub fn describe_create_connection(
    _: &StatementContext,
    _: CreateConnectionStatement<Aug>,
) -> Result<StatementDesc, PlanError> {
    Ok(StatementDesc::new(None))
}

generate_extracted_config!(CreateConnectionOption, (Validate, bool));

pub fn plan_create_connection(
    scx: &StatementContext,
    mut stmt: CreateConnectionStatement<Aug>,
) -> Result<Plan, PlanError> {
    let CreateConnectionStatement {
        name,
        connection_type,
        values,
        if_not_exists,
        with_options,
    } = stmt.clone();
    let connection_options_extracted = connection::ConnectionOptionExtracted::try_from(values)?;
    let details = connection_options_extracted.try_into_connection_details(scx, connection_type)?;
    let name = scx.allocate_qualified_name(normalize::unresolved_item_name(name)?)?;

    let options = CreateConnectionOptionExtracted::try_from(with_options)?;
    if options.validate.is_some() {
        scx.require_feature_flag(&vars::ENABLE_CONNECTION_VALIDATION_SYNTAX)?;
    }
    let validate = match options.validate {
        Some(val) => val,
        None => {
            scx.catalog
                .system_vars()
                .enable_default_connection_validation()
                && details.to_connection().validate_by_default()
        }
    };

    // Check for an object in the catalog with this same name
    let full_name = scx.catalog.resolve_full_name(&name);
    let partial_name = PartialItemName::from(full_name.clone());
    if let (false, Ok(item)) = (if_not_exists, scx.catalog.resolve_item(&partial_name)) {
        return Err(PlanError::ItemAlreadyExists {
            name: full_name.to_string(),
            item_type: item.item_type(),
        });
    }

    // For SSH connections, overwrite the public key options based on the
    // connection details, in case we generated new keys during planning.
    if let ConnectionDetails::Ssh { key_1, key_2, .. } = &details {
        stmt.values.retain(|v| {
            v.name != ConnectionOptionName::PublicKey1 && v.name != ConnectionOptionName::PublicKey2
        });
        stmt.values.push(ConnectionOption {
            name: ConnectionOptionName::PublicKey1,
            value: Some(WithOptionValue::Value(Value::String(key_1.public_key()))),
        });
        stmt.values.push(ConnectionOption {
            name: ConnectionOptionName::PublicKey2,
            value: Some(WithOptionValue::Value(Value::String(key_2.public_key()))),
        });
    }
    let create_sql = normalize::create_statement(scx, Statement::CreateConnection(stmt))?;

    let plan = CreateConnectionPlan {
        name,
        if_not_exists,
        connection: crate::plan::Connection {
            create_sql,
            details,
        },
        validate,
    };
    Ok(Plan::CreateConnection(plan))
}

fn plan_drop_database(
    scx: &StatementContext,
    if_exists: bool,
    name: &UnresolvedDatabaseName,
    cascade: bool,
) -> Result<Option<DatabaseId>, PlanError> {
    Ok(match resolve_database(scx, name, if_exists)? {
        Some(database) => {
            if !cascade && database.has_schemas() {
                sql_bail!(
                    "database '{}' cannot be dropped with RESTRICT while it contains schemas",
                    name,
                );
            }
            Some(database.id())
        }
        None => None,
    })
}

pub fn describe_drop_objects(
    _: &StatementContext,
    _: DropObjectsStatement,
) -> Result<StatementDesc, PlanError> {
    Ok(StatementDesc::new(None))
}

pub fn plan_drop_objects(
    scx: &mut StatementContext,
    DropObjectsStatement {
        object_type,
        if_exists,
        names,
        cascade,
    }: DropObjectsStatement,
) -> Result<Plan, PlanError> {
    if object_type == mz_sql_parser::ast::ObjectType::Func {
        bail_unsupported!("DROP FUNCTION");
    }
    let object_type = object_type.into();

    let mut referenced_ids = Vec::new();
    for name in names {
        let id = match &name {
            UnresolvedObjectName::Cluster(name) => {
                plan_drop_cluster(scx, if_exists, name, cascade)?.map(ObjectId::Cluster)
            }
            UnresolvedObjectName::ClusterReplica(name) => {
                plan_drop_cluster_replica(scx, if_exists, name)?.map(ObjectId::ClusterReplica)
            }
            UnresolvedObjectName::Database(name) => {
                plan_drop_database(scx, if_exists, name, cascade)?.map(ObjectId::Database)
            }
            UnresolvedObjectName::Schema(name) => {
                plan_drop_schema(scx, if_exists, name, cascade)?.map(ObjectId::Schema)
            }
            UnresolvedObjectName::Role(name) => {
                plan_drop_role(scx, if_exists, name)?.map(ObjectId::Role)
            }
            UnresolvedObjectName::Item(name) => {
                // Defer the dependency check until all names are resolved, so a
                // dependent that is itself being dropped in this same statement
                // does not block a non-cascade drop.
                plan_drop_item_name(scx, object_type, if_exists, name.clone())?.map(ObjectId::Item)
            }
            UnresolvedObjectName::NetworkPolicy(name) => {
                plan_drop_network_policy(scx, if_exists, name)?.map(ObjectId::NetworkPolicy)
            }
        };
        match id {
            Some(id) => referenced_ids.push(id),
            None => scx.catalog.add_notice(PlanNotice::ObjectDoesNotExist {
                name: name.to_ast_string_simple(),
                object_type,
            }),
        }
    }

    // Now that the full set of explicitly-named items is known, run the
    // non-cascade dependency check. A dependent that is itself being dropped in
    // this statement does not block the drop, matching PostgreSQL.
    if !cascade {
        let dropped_items: BTreeSet<CatalogItemId> = referenced_ids
            .iter()
            .filter_map(|id| match id {
                ObjectId::Item(id) => Some(*id),
                _ => None,
            })
            .collect();
        for id in &dropped_items {
            let catalog_item = scx.catalog.get_item(id);
            ensure_no_blocking_dependents(scx, object_type, catalog_item, &dropped_items)?;
        }
    }

    let drop_ids = scx.catalog.object_dependents(&referenced_ids);

    Ok(Plan::DropObjects(DropObjectsPlan {
        referenced_ids,
        drop_ids,
        object_type,
    }))
}

fn plan_drop_schema(
    scx: &StatementContext,
    if_exists: bool,
    name: &UnresolvedSchemaName,
    cascade: bool,
) -> Result<Option<(ResolvedDatabaseSpecifier, SchemaSpecifier)>, PlanError> {
    // Special case for mz_temp: with lazy temporary schema creation, the temp
    // schema may not exist yet, but we still need to return the correct error.
    // Check the schema name directly against MZ_TEMP_SCHEMA.
    let normalized = normalize::unresolved_schema_name(name.clone())?;
    if normalized.database.is_none() && normalized.schema == mz_repr::namespaces::MZ_TEMP_SCHEMA {
        sql_bail!("cannot drop schema {name} because it is a temporary schema",)
    }

    Ok(match resolve_schema(scx, name.clone(), if_exists)? {
        Some((database_spec, schema_spec)) => {
            if let ResolvedDatabaseSpecifier::Ambient = database_spec {
                sql_bail!(
                    "cannot drop schema {name} because it is required by the database system",
                );
            }
            if let SchemaSpecifier::Temporary = schema_spec {
                sql_bail!("cannot drop schema {name} because it is a temporary schema",)
            }
            let schema = scx.get_schema(&database_spec, &schema_spec);
            if !cascade && schema.has_items() {
                let full_schema_name = scx.catalog.resolve_full_schema_name(schema.name());
                sql_bail!(
                    "schema '{}' cannot be dropped without CASCADE while it contains objects",
                    full_schema_name
                );
            }
            Some((database_spec, schema_spec))
        }
        None => None,
    })
}

fn plan_drop_role(
    scx: &StatementContext,
    if_exists: bool,
    name: &Ident,
) -> Result<Option<RoleId>, PlanError> {
    match scx.catalog.resolve_role(name.as_str()) {
        Ok(role) => {
            let id = role.id();
            if &id == scx.catalog.active_role_id() {
                sql_bail!("current role cannot be dropped");
            }
            for role in scx.catalog.get_roles() {
                for (member_id, grantor_id) in role.membership() {
                    if &id == grantor_id {
                        let member_role = scx.catalog.get_role(member_id);
                        sql_bail!(
                            "cannot drop role {}: still depended up by membership of role {} in role {}",
                            name.as_str(),
                            role.name(),
                            member_role.name()
                        );
                    }
                }
            }
            Ok(Some(role.id()))
        }
        Err(_) if if_exists => Ok(None),
        Err(e) => Err(e.into()),
    }
}

fn plan_drop_cluster(
    scx: &StatementContext,
    if_exists: bool,
    name: &Ident,
    cascade: bool,
) -> Result<Option<ClusterId>, PlanError> {
    Ok(match resolve_cluster(scx, name, if_exists)? {
        Some(cluster) => {
            if !cascade && !cluster.bound_objects().is_empty() {
                return Err(PlanError::DependentObjectsStillExist {
                    object_type: "cluster".to_string(),
                    object_name: cluster.name().to_string(),
                    dependents: Vec::new(),
                });
            }
            Some(cluster.id())
        }
        None => None,
    })
}

fn plan_drop_network_policy(
    scx: &StatementContext,
    if_exists: bool,
    name: &Ident,
) -> Result<Option<NetworkPolicyId>, PlanError> {
    match scx.catalog.resolve_network_policy(name.as_str()) {
        Ok(policy) => {
            // TODO(network_policy): When we support role based network policies, check if any role
            // currently has the specified policy set.
            if scx.catalog.system_vars().default_network_policy_name() == policy.name() {
                Err(PlanError::NetworkPolicyInUse)
            } else {
                Ok(Some(policy.id()))
            }
        }
        Err(_) if if_exists => Ok(None),
        Err(e) => Err(e.into()),
    }
}

fn plan_drop_cluster_replica(
    scx: &StatementContext,
    if_exists: bool,
    name: &QualifiedReplica,
) -> Result<Option<(ClusterId, ReplicaId)>, PlanError> {
    let cluster = resolve_cluster_replica(scx, name, if_exists)?;
    Ok(cluster.map(|(cluster, replica_id)| (cluster.id(), replica_id)))
}

/// Returns the [`CatalogItemId`] of the item we should drop, if it exists.
fn plan_drop_item(
    scx: &StatementContext,
    object_type: ObjectType,
    if_exists: bool,
    name: UnresolvedItemName,
    cascade: bool,
) -> Result<Option<CatalogItemId>, PlanError> {
    let Some(id) = plan_drop_item_name(scx, object_type, if_exists, name)? else {
        return Ok(None);
    };
    if !cascade {
        let catalog_item = scx.catalog.get_item(&id);
        ensure_no_blocking_dependents(scx, object_type, catalog_item, &BTreeSet::new())?;
    }
    Ok(Some(id))
}

/// Resolves `name` to the [`CatalogItemId`] of the item to drop, performing the
/// system-object check but *not* the dependency check. Returns `None` if the
/// item does not exist and `if_exists` is set.
fn plan_drop_item_name(
    scx: &StatementContext,
    object_type: ObjectType,
    if_exists: bool,
    name: UnresolvedItemName,
) -> Result<Option<CatalogItemId>, PlanError> {
    let resolved = match resolve_item_or_type(scx, object_type, name, if_exists) {
        Ok(r) => r,
        // Return a more helpful error on `DROP VIEW <materialized-view>`.
        Err(PlanError::MismatchedObjectType {
            name,
            is_type: ObjectType::MaterializedView,
            expected_type: ObjectType::View,
        }) => {
            return Err(PlanError::DropViewOnMaterializedView(name.to_string()));
        }
        e => e?,
    };

    Ok(match resolved {
        Some(catalog_item) => {
            if catalog_item.id().is_system() {
                sql_bail!(
                    "cannot drop {} {} because it is required by the database system",
                    catalog_item.item_type(),
                    scx.catalog.minimal_qualification(catalog_item.name()),
                );
            }
            Some(catalog_item.id())
        }
        None => None,
    })
}

/// Errors if dropping `catalog_item` would leave a dangling dependent, i.e. an
/// object that depends on it and is not itself being dropped. Dependents whose
/// ids are in `also_dropped` are ignored, since they are being dropped as part
/// of the same statement.
fn ensure_no_blocking_dependents(
    scx: &StatementContext,
    object_type: ObjectType,
    catalog_item: &dyn CatalogItem,
    also_dropped: &BTreeSet<CatalogItemId>,
) -> Result<(), PlanError> {
    for id in catalog_item.used_by() {
        if also_dropped.contains(id) {
            continue;
        }
        let dep = scx.catalog.get_item(id);
        if dependency_prevents_drop(object_type, dep) {
            return Err(PlanError::DependentObjectsStillExist {
                object_type: catalog_item.item_type().to_string(),
                object_name: scx
                    .catalog
                    .minimal_qualification(catalog_item.name())
                    .to_string(),
                dependents: vec![(
                    dep.item_type().to_string(),
                    scx.catalog.minimal_qualification(dep.name()).to_string(),
                )],
            });
        }
    }
    // TODO(jkosh44) It would be nice to also check if any active subscribe or pending peek
    //  relies on entry. Unfortunately, we don't have that information readily available.
    Ok(())
}

/// Does the dependency `dep` prevent a drop of a non-cascade query?
fn dependency_prevents_drop(object_type: ObjectType, dep: &dyn CatalogItem) -> bool {
    match object_type {
        ObjectType::Type => true,
        ObjectType::Table
        | ObjectType::View
        | ObjectType::MaterializedView
        | ObjectType::Source
        | ObjectType::Sink
        | ObjectType::MetricSink
        | ObjectType::Index
        | ObjectType::Role
        | ObjectType::Cluster
        | ObjectType::ClusterReplica
        | ObjectType::Secret
        | ObjectType::Connection
        | ObjectType::Database
        | ObjectType::Schema
        | ObjectType::Func
        | ObjectType::NetworkPolicy => match dep.item_type() {
            CatalogItemType::Func
            | CatalogItemType::Table
            | CatalogItemType::Source
            | CatalogItemType::View
            | CatalogItemType::MaterializedView
            | CatalogItemType::Sink
            | CatalogItemType::MetricSink
            | CatalogItemType::Type
            | CatalogItemType::Secret
            | CatalogItemType::Connection => true,
            CatalogItemType::Index => false,
        },
    }
}

pub fn describe_alter_index_options(
    _: &StatementContext,
    _: AlterIndexStatement<Aug>,
) -> Result<StatementDesc, PlanError> {
    Ok(StatementDesc::new(None))
}

pub fn describe_drop_owned(
    _: &StatementContext,
    _: DropOwnedStatement<Aug>,
) -> Result<StatementDesc, PlanError> {
    Ok(StatementDesc::new(None))
}

pub fn plan_drop_owned(
    scx: &StatementContext,
    drop: DropOwnedStatement<Aug>,
) -> Result<Plan, PlanError> {
    let cascade = drop.cascade();
    let role_ids: BTreeSet<_> = drop.role_names.into_iter().map(|role| role.id).collect();
    let mut drop_ids = Vec::new();
    let mut privilege_revokes = Vec::new();
    let mut default_privilege_revokes = Vec::new();

    fn update_privilege_revokes(
        object_id: SystemObjectId,
        privileges: &PrivilegeMap,
        role_ids: &BTreeSet<RoleId>,
        privilege_revokes: &mut Vec<(SystemObjectId, MzAclItem)>,
    ) {
        privilege_revokes.extend(iter::zip(
            iter::repeat(object_id),
            privileges
                .all_values()
                .filter(|privilege| role_ids.contains(&privilege.grantee))
                .cloned(),
        ));
    }

    // Replicas
    for replica in scx.catalog.get_cluster_replicas() {
        if role_ids.contains(&replica.owner_id()) {
            drop_ids.push((replica.cluster_id(), replica.replica_id()).into());
        }
    }

    // Clusters
    for cluster in scx.catalog.get_clusters() {
        if role_ids.contains(&cluster.owner_id()) {
            // Note: CASCADE is not required for replicas.
            if !cascade {
                let non_owned_bound_objects: Vec<_> = cluster
                    .bound_objects()
                    .into_iter()
                    .map(|item_id| scx.catalog.get_item(item_id))
                    .filter(|item| !role_ids.contains(&item.owner_id()))
                    .collect();
                if !non_owned_bound_objects.is_empty() {
                    let names: Vec<_> = non_owned_bound_objects
                        .into_iter()
                        .map(|item| {
                            (
                                item.item_type().to_string(),
                                scx.catalog.resolve_full_name(item.name()).to_string(),
                            )
                        })
                        .collect();
                    return Err(PlanError::DependentObjectsStillExist {
                        object_type: "cluster".to_string(),
                        object_name: cluster.name().to_string(),
                        dependents: names,
                    });
                }
            }
            drop_ids.push(cluster.id().into());
        }
        update_privilege_revokes(
            SystemObjectId::Object(cluster.id().into()),
            cluster.privileges(),
            &role_ids,
            &mut privilege_revokes,
        );
    }

    // Items
    for item in scx.catalog.get_items() {
        if role_ids.contains(&item.owner_id()) {
            if !cascade {
                // Checks if any items still depend on this one, returning an error if so.
                let check_if_dependents_exist = |used_by: &[CatalogItemId]| {
                    let non_owned_dependencies: Vec<_> = used_by
                        .into_iter()
                        .map(|item_id| scx.catalog.get_item(item_id))
                        .filter(|item| dependency_prevents_drop(item.item_type().into(), *item))
                        .filter(|item| !role_ids.contains(&item.owner_id()))
                        .collect();
                    if !non_owned_dependencies.is_empty() {
                        let names: Vec<_> = non_owned_dependencies
                            .into_iter()
                            .map(|item| {
                                let item_typ = item.item_type().to_string();
                                let item_name =
                                    scx.catalog.resolve_full_name(item.name()).to_string();
                                (item_typ, item_name)
                            })
                            .collect();
                        Err(PlanError::DependentObjectsStillExist {
                            object_type: item.item_type().to_string(),
                            object_name: scx
                                .catalog
                                .resolve_full_name(item.name())
                                .to_string()
                                .to_string(),
                            dependents: names,
                        })
                    } else {
                        Ok(())
                    }
                };

                // When this item gets dropped it will also drop its progress source, so we need to
                // check the users of those.
                if let Some(id) = item.progress_id() {
                    let progress_item = scx.catalog.get_item(&id);
                    check_if_dependents_exist(progress_item.used_by())?;
                }
                check_if_dependents_exist(item.used_by())?;
            }
            drop_ids.push(item.id().into());
        }
        update_privilege_revokes(
            SystemObjectId::Object(item.id().into()),
            item.privileges(),
            &role_ids,
            &mut privilege_revokes,
        );
    }

    // Schemas
    for schema in scx.catalog.get_schemas() {
        if !schema.id().is_temporary() {
            if role_ids.contains(&schema.owner_id()) {
                if !cascade {
                    let non_owned_dependencies: Vec<_> = schema
                        .item_ids()
                        .map(|item_id| scx.catalog.get_item(&item_id))
                        .filter(|item| dependency_prevents_drop(item.item_type().into(), *item))
                        .filter(|item| !role_ids.contains(&item.owner_id()))
                        .collect();
                    if !non_owned_dependencies.is_empty() {
                        let full_schema_name = scx.catalog.resolve_full_schema_name(schema.name());
                        sql_bail!(
                            "schema {} cannot be dropped without CASCADE while it contains non-owned objects",
                            full_schema_name.to_string().quoted()
                        );
                    }
                }
                drop_ids.push((*schema.database(), *schema.id()).into())
            }
            update_privilege_revokes(
                SystemObjectId::Object((*schema.database(), *schema.id()).into()),
                schema.privileges(),
                &role_ids,
                &mut privilege_revokes,
            );
        }
    }

    // Databases
    for database in scx.catalog.get_databases() {
        if role_ids.contains(&database.owner_id()) {
            if !cascade {
                let non_owned_schemas: Vec<_> = database
                    .schemas()
                    .into_iter()
                    .filter(|schema| !role_ids.contains(&schema.owner_id()))
                    .collect();
                if !non_owned_schemas.is_empty() {
                    sql_bail!(
                        "database {} cannot be dropped without CASCADE while it contains non-owned schemas",
                        database.name().quoted(),
                    );
                }
            }
            drop_ids.push(database.id().into());
        }
        update_privilege_revokes(
            SystemObjectId::Object(database.id().into()),
            database.privileges(),
            &role_ids,
            &mut privilege_revokes,
        );
    }

    // Network policies
    for network_policy in scx.catalog.get_network_policies() {
        if role_ids.contains(&network_policy.owner_id()) {
            drop_ids.push(ObjectId::NetworkPolicy(network_policy.id()));
        }
        update_privilege_revokes(
            SystemObjectId::Object(ObjectId::NetworkPolicy(network_policy.id())),
            network_policy.privileges(),
            &role_ids,
            &mut privilege_revokes,
        );
    }

    // System
    update_privilege_revokes(
        SystemObjectId::System,
        scx.catalog.get_system_privileges(),
        &role_ids,
        &mut privilege_revokes,
    );

    for (default_privilege_object, default_privilege_acl_items) in
        scx.catalog.get_default_privileges()
    {
        for default_privilege_acl_item in default_privilege_acl_items {
            if role_ids.contains(&default_privilege_object.role_id)
                || role_ids.contains(&default_privilege_acl_item.grantee)
            {
                default_privilege_revokes.push((
                    default_privilege_object.clone(),
                    default_privilege_acl_item.clone(),
                ));
            }
        }
    }

    let drop_ids = scx.catalog.object_dependents(&drop_ids);

    let system_ids: Vec<_> = drop_ids.iter().filter(|id| id.is_system()).collect();
    if !system_ids.is_empty() {
        let mut owners = system_ids
            .into_iter()
            .filter_map(|object_id| scx.catalog.get_owner_id(object_id))
            .collect::<BTreeSet<_>>()
            .into_iter()
            .map(|role_id| scx.catalog.get_role(&role_id).name().quoted());
        sql_bail!(
            "cannot drop objects owned by role {} because they are required by the database system",
            owners.join(", "),
        );
    }

    Ok(Plan::DropOwned(DropOwnedPlan {
        role_ids: role_ids.into_iter().collect(),
        drop_ids,
        privilege_revokes,
        default_privilege_revokes,
    }))
}

fn plan_retain_history_option(
    scx: &StatementContext,
    retain_history: Option<OptionalDuration>,
) -> Result<Option<CompactionWindow>, PlanError> {
    if let Some(OptionalDuration(lcw)) = retain_history {
        Ok(Some(plan_retain_history(scx, lcw)?))
    } else {
        Ok(None)
    }
}

// Convert a specified RETAIN HISTORY option into a compaction window. `None` corresponds to
// `DisableCompaction`. A zero duration will error. This is because the `OptionalDuration` type
// already converts the zero duration into `None`. This function must not be called in the `RESET
// (RETAIN HISTORY)` path, which should be handled by the outer `Option<OptionalDuration>` being
// `None`.
fn plan_retain_history(
    scx: &StatementContext,
    lcw: Option<Duration>,
) -> Result<CompactionWindow, PlanError> {
    scx.require_feature_flag(&vars::ENABLE_LOGICAL_COMPACTION_WINDOW)?;
    match lcw {
        // A zero duration has already been converted to `None` by `OptionalDuration` (and means
        // disable compaction), and should never occur here. Furthermore, some things actually do
        // break when this is set to real zero:
        // https://github.com/MaterializeInc/database-issues/issues/3798.
        Some(Duration::ZERO) => Err(PlanError::InvalidOptionValue {
            option_name: "RETAIN HISTORY".to_string(),
            err: Box::new(PlanError::Unstructured(
                "internal error: unexpectedly zero".to_string(),
            )),
        }),
        Some(duration) => {
            // Error if the duration is low and enable_unlimited_retain_history is not set (which
            // should only be possible during testing).
            if duration < DEFAULT_LOGICAL_COMPACTION_WINDOW_DURATION
                && scx
                    .require_feature_flag(&vars::ENABLE_UNLIMITED_RETAIN_HISTORY)
                    .is_err()
            {
                return Err(PlanError::RetainHistoryLow {
                    limit: DEFAULT_LOGICAL_COMPACTION_WINDOW_DURATION,
                });
            }
            Ok(duration.try_into()?)
        }
        // In the past `RETAIN HISTORY FOR '0'` meant disable compaction. Disabling compaction seems
        // to be a bad choice, so prevent it.
        None => {
            if scx
                .require_feature_flag(&vars::ENABLE_UNLIMITED_RETAIN_HISTORY)
                .is_err()
            {
                Err(PlanError::RetainHistoryRequired)
            } else {
                Ok(CompactionWindow::DisableCompaction)
            }
        }
    }
}

generate_extracted_config!(IndexOption, (RetainHistory, OptionalDuration));

fn plan_index_options(
    scx: &StatementContext,
    with_opts: Vec<IndexOption<Aug>>,
) -> Result<Vec<crate::plan::IndexOption>, PlanError> {
    if !with_opts.is_empty() {
        // Index options are not durable.
        scx.require_feature_flag(&vars::ENABLE_INDEX_OPTIONS)?;
    }

    let IndexOptionExtracted { retain_history, .. }: IndexOptionExtracted = with_opts.try_into()?;

    let mut out = Vec::with_capacity(1);
    if let Some(cw) = plan_retain_history_option(scx, retain_history)? {
        out.push(crate::plan::IndexOption::RetainHistory(cw));
    }
    Ok(out)
}

generate_extracted_config!(
    TableOption,
    (PartitionBy, Vec<Ident>),
    (RetainHistory, OptionalDuration),
    (RedactedTest, String)
);

fn plan_table_options(
    scx: &StatementContext,
    desc: &RelationDesc,
    with_opts: Vec<TableOption<Aug>>,
) -> Result<Vec<crate::plan::TableOption>, PlanError> {
    let TableOptionExtracted {
        partition_by,
        retain_history,
        redacted_test,
        ..
    }: TableOptionExtracted = with_opts.try_into()?;

    if let Some(partition_by) = partition_by {
        scx.require_feature_flag(&ENABLE_COLLECTION_PARTITION_BY)?;
        check_partition_by(desc, partition_by)?;
    }

    if redacted_test.is_some() {
        scx.require_feature_flag(&vars::ENABLE_REDACTED_TEST_OPTION)?;
    }

    let mut out = Vec::with_capacity(1);
    if let Some(cw) = plan_retain_history_option(scx, retain_history)? {
        out.push(crate::plan::TableOption::RetainHistory(cw));
    }
    Ok(out)
}

pub fn plan_alter_index_options(
    scx: &mut StatementContext,
    AlterIndexStatement {
        index_name,
        if_exists,
        action,
    }: AlterIndexStatement<Aug>,
) -> Result<Plan, PlanError> {
    let object_type = ObjectType::Index;
    match action {
        AlterIndexAction::ResetOptions(options) => {
            let mut options = options.into_iter();
            if let Some(opt) = options.next() {
                match opt {
                    IndexOptionName::RetainHistory => {
                        if options.next().is_some() {
                            sql_bail!("RETAIN HISTORY must be only option");
                        }
                        return alter_retain_history(
                            scx,
                            object_type,
                            if_exists,
                            UnresolvedObjectName::Item(index_name),
                            None,
                        );
                    }
                }
            }
            sql_bail!("expected option");
        }
        AlterIndexAction::SetOptions(options) => {
            let mut options = options.into_iter();
            if let Some(opt) = options.next() {
                match opt.name {
                    IndexOptionName::RetainHistory => {
                        if options.next().is_some() {
                            sql_bail!("RETAIN HISTORY must be only option");
                        }
                        return alter_retain_history(
                            scx,
                            object_type,
                            if_exists,
                            UnresolvedObjectName::Item(index_name),
                            opt.value,
                        );
                    }
                }
            }
            sql_bail!("expected option");
        }
    }
}

pub fn describe_alter_cluster_set_options(
    _: &StatementContext,
    _: AlterClusterStatement<Aug>,
) -> Result<StatementDesc, PlanError> {
    Ok(StatementDesc::new(None))
}

pub fn plan_alter_cluster(
    scx: &mut StatementContext,
    AlterClusterStatement {
        name,
        action,
        if_exists,
    }: AlterClusterStatement<Aug>,
) -> Result<Plan, PlanError> {
    let cluster = match resolve_cluster(scx, &name, if_exists)? {
        Some(entry) => entry,
        None => {
            scx.catalog.add_notice(PlanNotice::ObjectDoesNotExist {
                name: name.to_ast_string_simple(),
                object_type: ObjectType::Cluster,
            });

            return Ok(Plan::AlterNoop(AlterNoopPlan {
                object_type: ObjectType::Cluster,
            }));
        }
    };

    let mut options: PlanClusterOption = Default::default();
    let mut alter_strategy: AlterClusterPlanStrategy = AlterClusterPlanStrategy::None;

    match action {
        AlterClusterAction::SetOptions {
            options: set_options,
            with_options,
        } => {
            let ClusterOptionExtracted {
                auto_scaling_strategy,
                availability_zones,
                experimental_arrangement_compression,
                introspection_debugging,
                introspection_interval,
                managed,
                replicas: replica_defs,
                replication_factor,
                seen: _,
                size,
                disk,
                schedule,
                workload_class,
            }: ClusterOptionExtracted = set_options.try_into()?;

            if !scx.catalog.active_role_id().is_system() {
                if workload_class.is_some() {
                    sql_bail!("WORKLOAD CLASS not supported for non-system users");
                }
            }

            match managed.unwrap_or_else(|| cluster.is_managed()) {
                true => {
                    let alter_strategy_extracted =
                        ClusterAlterOptionExtracted::try_from(with_options)?;
                    alter_strategy = AlterClusterPlanStrategy::try_from(alter_strategy_extracted)?;

                    // WAIT applies to replica shape changes and RF retargets of
                    // an in-flight reconfiguration. For an RF-only change with
                    // nothing in flight, WAIT has no effect.
                    if !matches!(alter_strategy, AlterClusterPlanStrategy::None)
                        && size.is_none()
                        && replication_factor.is_none()
                        && availability_zones.is_none()
                        && introspection_debugging.is_none()
                        && introspection_interval.is_none()
                        && experimental_arrangement_compression.is_none()
                    {
                        sql_bail!(
                            "WAIT can only be used together with a SIZE, REPLICATION FACTOR, \
                            AVAILABILITY ZONES, INTROSPECTION, or EXPERIMENTAL ARRANGEMENT COMPRESSION change"
                        );
                    }

                    if replica_defs.is_some() {
                        sql_bail!("REPLICAS not supported for managed clusters");
                    }
                    if schedule.is_some()
                        && !matches!(schedule, Some(ClusterScheduleOptionValue::Manual))
                    {
                        scx.require_feature_flag(&ENABLE_CLUSTER_SCHEDULE_REFRESH)?;

                        // A cluster with a non-MANUAL schedule is automatically turned On/Off by
                        // the cluster scheduling policy, which means its replication factor is
                        // always 0 or 1. If the cluster currently has a higher replication factor
                        // and the user is not lowering it in the same statement (which would be
                        // rejected just below), then reject the schedule change: otherwise we'd
                        // leave the cluster in an invalid state with both a non-MANUAL schedule and
                        // a replication factor > 1 (which would, e.g., make SHOW CREATE CLUSTER
                        // panic).
                        if replication_factor.is_none()
                            && cluster.replication_factor().is_some_and(|rf| rf > 1)
                        {
                            sql_bail!(
                                "SCHEDULE cannot be set to anything other than MANUAL while the \
                                cluster's REPLICATION FACTOR is greater than 1; \
                                set the REPLICATION FACTOR to 1 first"
                            );
                        }
                    }

                    if replication_factor.is_some() {
                        if schedule.is_some()
                            && !matches!(schedule, Some(ClusterScheduleOptionValue::Manual))
                        {
                            sql_bail!(
                                "REPLICATION FACTOR cannot be given together with any SCHEDULE other than MANUAL"
                            );
                        }
                        if let Some(current_schedule) = cluster.schedule() {
                            if !matches!(current_schedule, ClusterSchedule::Manual) {
                                sql_bail!(
                                    "REPLICATION FACTOR cannot be set if the cluster SCHEDULE is anything other than MANUAL"
                                );
                            }
                        }
                    }

                    if let Some(value) = auto_scaling_strategy {
                        scx.require_feature_flag(&ENABLE_AUTO_SCALING_STRATEGY)?;
                        let strategy = plan_auto_scaling_strategy(value)?;
                        options.auto_scaling_strategy = AlterOptionParameter::Set(strategy);
                    }

                    // Validate against the *effective* config (the values this `ALTER`
                    // sets, else the cluster's current ones), so an invariant cannot be
                    // broken by changing the other side of its constraint.
                    let effective_strategy = match &options.auto_scaling_strategy {
                        AlterOptionParameter::Set(s) => s.clone(),
                        AlterOptionParameter::Reset => None,
                        AlterOptionParameter::Unchanged => cluster.auto_scaling_strategy().cloned(),
                    };
                    if let Some(effective_strategy) = &effective_strategy {
                        let effective_size = size.as_deref().or_else(|| cluster.managed_size());
                        let schedule_non_manual = match &schedule {
                            Some(s) => !matches!(s, ClusterScheduleOptionValue::Manual),
                            None => cluster
                                .schedule()
                                .is_some_and(|s| !matches!(s, ClusterSchedule::Manual)),
                        };
                        validate_auto_scaling_strategy(
                            effective_strategy,
                            effective_size,
                            schedule_non_manual,
                        )?;
                    }
                }
                false => {
                    if !with_options.is_empty() {
                        sql_bail!("ALTER... WITH not supported for unmanaged clusters");
                    }
                    if auto_scaling_strategy.is_some() {
                        sql_bail!("AUTO SCALING STRATEGY not supported for unmanaged clusters");
                    }
                    if availability_zones.is_some() {
                        sql_bail!("AVAILABILITY ZONES not supported for unmanaged clusters");
                    }
                    if replication_factor.is_some() {
                        sql_bail!("REPLICATION FACTOR not supported for unmanaged clusters");
                    }
                    if introspection_debugging.is_some() {
                        sql_bail!("INTROSPECTION DEBUGGING not supported for unmanaged clusters");
                    }
                    if introspection_interval.is_some() {
                        sql_bail!("INTROSPECTION INTERVAL not supported for unmanaged clusters");
                    }
                    if experimental_arrangement_compression.is_some() {
                        sql_bail!(
                            "EXPERIMENTAL ARRANGEMENT COMPRESSION not supported for unmanaged clusters"
                        );
                    }
                    if size.is_some() {
                        sql_bail!("SIZE not supported for unmanaged clusters");
                    }
                    if disk.is_some() {
                        sql_bail!("DISK not supported for unmanaged clusters");
                    }
                    if schedule.is_some()
                        && !matches!(schedule, Some(ClusterScheduleOptionValue::Manual))
                    {
                        sql_bail!(
                            "cluster schedules other than MANUAL are not supported for unmanaged clusters"
                        );
                    }
                    if let Some(current_schedule) = cluster.schedule() {
                        if !matches!(current_schedule, ClusterSchedule::Manual)
                            && schedule.is_none()
                        {
                            sql_bail!(
                                "when switching a cluster to unmanaged, if the managed \
                                cluster's SCHEDULE is anything other than MANUAL, you have to \
                                explicitly set the SCHEDULE to MANUAL"
                            );
                        }
                    }
                }
            }

            let mut replicas = vec![];
            for ReplicaDefinition { name, options } in
                replica_defs.into_iter().flat_map(Vec::into_iter)
            {
                replicas.push((normalize::ident(name), plan_replica_config(scx, options)?));
            }

            if let Some(managed) = managed {
                options.managed = AlterOptionParameter::Set(managed);
            }
            if let Some(replication_factor) = replication_factor {
                options.replication_factor = AlterOptionParameter::Set(replication_factor);
            } else if schedule
                .as_ref()
                .is_some_and(|s| !matches!(s, ClusterScheduleOptionValue::Manual))
                && managed != Some(true)
            {
                // Setting a non-MANUAL schedule hands the replica set to the
                // scheduler, so normalize the replication factor to 0 exactly
                // as CREATE CLUSTER does for a scheduled cluster. Giving
                // REPLICATION FACTOR together with a non-MANUAL SCHEDULE was
                // rejected above, so `replication_factor` is `None` here.
                //
                // Not when the same statement converts an unmanaged cluster to
                // managed: that conversion adopts the existing replicas, so the
                // sequencer requires a replication factor matching their count
                // and derives it when none is given. Forcing 0 would reject the
                // conversion whenever a replica exists. The controller
                // normalizes the adopted factor to 0 on its next tick.
                options.replication_factor = AlterOptionParameter::Set(0);
            }
            if let Some(size) = &size {
                options.size = AlterOptionParameter::Set(size.clone());
            }
            if let Some(availability_zones) = availability_zones {
                options.availability_zones = AlterOptionParameter::Set(availability_zones);
            }
            if let Some(introspection_debugging) = introspection_debugging {
                options.introspection_debugging =
                    AlterOptionParameter::Set(introspection_debugging);
            }
            if let Some(introspection_interval) = introspection_interval {
                options.introspection_interval = AlterOptionParameter::Set(introspection_interval);
            }
            if let Some(experimental_arrangement_compression) = experimental_arrangement_compression
            {
                options.arrangement_compression =
                    AlterOptionParameter::Set(experimental_arrangement_compression);
            }
            if disk.is_some() {
                // The `DISK` option is a no-op for legacy cluster sizes and was never allowed for
                // `cc` sizes. The long term plan is to phase out the legacy sizes, at which point
                // we'll be able to remove the `DISK` option entirely.
                let size = match size.as_deref() {
                    Some(s) => s,
                    None => cluster
                        .managed_size()
                        .ok_or_else(|| sql_err!("cluster is not managed"))?,
                };
                if scx.catalog.is_cluster_size_cc(size) {
                    sql_bail!(
                        "DISK option not supported for modern cluster sizes because disk is always enabled"
                    );
                }

                scx.catalog
                    .add_notice(PlanNotice::ReplicaDiskOptionDeprecated);
            }
            if !replicas.is_empty() {
                options.replicas = AlterOptionParameter::Set(replicas);
            }
            if let Some(schedule) = schedule {
                options.schedule = AlterOptionParameter::Set(plan_cluster_schedule(schedule)?);
            }
            if let Some(workload_class) = workload_class {
                options.workload_class = AlterOptionParameter::Set(workload_class.0);
            }
        }
        AlterClusterAction::ResetOptions(reset_options) => {
            use AlterOptionParameter::Reset;
            use ClusterOptionName::*;

            if !scx.catalog.active_role_id().is_system() {
                if reset_options.contains(&WorkloadClass) {
                    sql_bail!("WORKLOAD CLASS not supported for non-system users");
                }
            }

            // RESET (AUTO SCALING STRATEGY) is deliberately not feature-gated:
            // resetting to the default must stay available after a flag
            // rollback, or a cluster that picked up a policy while the flag was
            // on would carry it forever (and the strategy/schedule
            // compatibility check would then also block SCHEDULE changes).
            for option in reset_options {
                match option {
                    AutoScalingStrategy => options.auto_scaling_strategy = Reset,
                    AvailabilityZones => options.availability_zones = Reset,
                    Disk => scx
                        .catalog
                        .add_notice(PlanNotice::ReplicaDiskOptionDeprecated),
                    IntrospectionInterval => options.introspection_interval = Reset,
                    IntrospectionDebugging => options.introspection_debugging = Reset,
                    ExperimentalArrangementCompression => options.arrangement_compression = Reset,
                    Managed => options.managed = Reset,
                    Replicas => options.replicas = Reset,
                    ReplicationFactor => options.replication_factor = Reset,
                    Size => options.size = Reset,
                    Schedule => options.schedule = Reset,
                    WorkloadClass => options.workload_class = Reset,
                }
            }
        }
    }
    Ok(Plan::AlterCluster(AlterClusterPlan {
        id: cluster.id(),
        name: cluster.name().to_string(),
        options,
        strategy: alter_strategy,
    }))
}

pub fn describe_alter_set_cluster(
    _: &StatementContext,
    _: AlterSetClusterStatement<Aug>,
) -> Result<StatementDesc, PlanError> {
    Ok(StatementDesc::new(None))
}

pub fn plan_alter_item_set_cluster(
    scx: &StatementContext,
    AlterSetClusterStatement {
        if_exists,
        set_cluster: in_cluster_name,
        name,
        object_type,
    }: AlterSetClusterStatement<Aug>,
) -> Result<Plan, PlanError> {
    scx.require_feature_flag(&vars::ENABLE_ALTER_SET_CLUSTER)?;

    let object_type = object_type.into();

    // Prevent access to `SET CLUSTER` for unsupported objects.
    match object_type {
        ObjectType::MaterializedView => {}
        ObjectType::Index | ObjectType::Sink | ObjectType::MetricSink | ObjectType::Source => {
            bail_unsupported!(29606, format!("ALTER {object_type} SET CLUSTER"))
        }
        ObjectType::Table
        | ObjectType::View
        | ObjectType::Type
        | ObjectType::Role
        | ObjectType::Cluster
        | ObjectType::ClusterReplica
        | ObjectType::Secret
        | ObjectType::Connection
        | ObjectType::Database
        | ObjectType::Schema
        | ObjectType::Func
        | ObjectType::NetworkPolicy => {
            bail_never_supported!(
                format!("ALTER {object_type} SET CLUSTER"),
                "sql/alter-set-cluster/",
                format!("{object_type} has no associated cluster")
            )
        }
    }

    let in_cluster = scx.catalog.get_cluster(in_cluster_name.id);

    match resolve_item_or_type(scx, object_type, name.clone(), if_exists)? {
        Some(entry) => {
            let current_cluster = entry.cluster_id();
            let Some(current_cluster) = current_cluster else {
                sql_bail!("No cluster associated with {name}");
            };

            if current_cluster == in_cluster.id() {
                Ok(Plan::AlterNoop(AlterNoopPlan { object_type }))
            } else {
                Ok(Plan::AlterSetCluster(AlterSetClusterPlan {
                    id: entry.id(),
                    set_cluster: in_cluster.id(),
                }))
            }
        }
        None => {
            scx.catalog.add_notice(PlanNotice::ObjectDoesNotExist {
                name: name.to_ast_string_simple(),
                object_type,
            });

            Ok(Plan::AlterNoop(AlterNoopPlan { object_type }))
        }
    }
}

pub fn describe_alter_object_rename(
    _: &StatementContext,
    _: AlterObjectRenameStatement,
) -> Result<StatementDesc, PlanError> {
    Ok(StatementDesc::new(None))
}

pub fn plan_alter_object_rename(
    scx: &mut StatementContext,
    AlterObjectRenameStatement {
        name,
        object_type,
        to_item_name,
        if_exists,
    }: AlterObjectRenameStatement,
) -> Result<Plan, PlanError> {
    let object_type = object_type.into();
    match (object_type, name) {
        (
            ObjectType::View
            | ObjectType::MaterializedView
            | ObjectType::Table
            | ObjectType::Source
            | ObjectType::Index
            | ObjectType::Sink
            | ObjectType::Secret
            | ObjectType::Connection,
            UnresolvedObjectName::Item(name),
        ) => plan_alter_item_rename(scx, object_type, name, to_item_name, if_exists),
        (ObjectType::Cluster, UnresolvedObjectName::Cluster(name)) => {
            plan_alter_cluster_rename(scx, object_type, name, to_item_name, if_exists)
        }
        (ObjectType::ClusterReplica, UnresolvedObjectName::ClusterReplica(name)) => {
            plan_alter_cluster_replica_rename(scx, object_type, name, to_item_name, if_exists)
        }
        (ObjectType::Schema, UnresolvedObjectName::Schema(name)) => {
            plan_alter_schema_rename(scx, name, to_item_name, if_exists)
        }
        (object_type, name) => {
            // The earlier dispatch + name resolution should make this
            // combination impossible.
            bail_internal!("invalid object type '{object_type}' for ALTER RENAME with name {name}")
        }
    }
}

pub fn plan_alter_schema_rename(
    scx: &mut StatementContext,
    name: UnresolvedSchemaName,
    to_schema_name: Ident,
    if_exists: bool,
) -> Result<Plan, PlanError> {
    // Special case for mz_temp: with lazy temporary schema creation, the temp
    // schema may not exist yet, but we still need to return the correct error.
    // Check the schema name directly against MZ_TEMP_SCHEMA.
    let normalized = normalize::unresolved_schema_name(name.clone())?;
    if normalized.database.is_none() && normalized.schema == mz_repr::namespaces::MZ_TEMP_SCHEMA {
        sql_bail!(
            "cannot rename schemas in the ambient database: {:?}",
            mz_repr::namespaces::MZ_TEMP_SCHEMA
        );
    }

    let Some((db_spec, schema_spec)) = resolve_schema(scx, name.clone(), if_exists)? else {
        let object_type = ObjectType::Schema;
        scx.catalog.add_notice(PlanNotice::ObjectDoesNotExist {
            name: name.to_ast_string_simple(),
            object_type,
        });
        return Ok(Plan::AlterNoop(AlterNoopPlan { object_type }));
    };

    // Make sure the name is unique.
    if scx
        .resolve_schema_in_database(&db_spec, &to_schema_name)
        .is_ok()
    {
        return Err(PlanError::Catalog(CatalogError::SchemaAlreadyExists(
            to_schema_name.clone().into_string(),
        )));
    }

    // Prevent users from renaming system related schemas.
    let schema = scx.catalog.get_schema(&db_spec, &schema_spec);
    if schema.id().is_system() {
        bail_never_supported!(format!("renaming the {} schema", schema.name().schema))
    }

    Ok(Plan::AlterSchemaRename(AlterSchemaRenamePlan {
        cur_schema_spec: (db_spec, schema_spec),
        new_schema_name: to_schema_name.into_string(),
    }))
}

pub fn plan_alter_schema_swap<F>(
    scx: &mut StatementContext,
    name_a: UnresolvedSchemaName,
    name_b: Ident,
    if_exists: bool,
    gen_temp_suffix: F,
) -> Result<Plan, PlanError>
where
    F: Fn(&dyn Fn(&str) -> bool) -> Result<String, PlanError>,
{
    // Special case for mz_temp: with lazy temporary schema creation, the temp
    // schema may not exist yet, but we still need to return the correct error.
    // Check the schema name directly against MZ_TEMP_SCHEMA.
    let normalized_a = normalize::unresolved_schema_name(name_a.clone())?;
    if normalized_a.database.is_none() && normalized_a.schema == mz_repr::namespaces::MZ_TEMP_SCHEMA
    {
        sql_bail!("cannot swap schemas that are in the ambient database");
    }
    // Also check name_b (the target schema name)
    let name_b_str = normalize::ident_ref(&name_b);
    if name_b_str == mz_repr::namespaces::MZ_TEMP_SCHEMA {
        sql_bail!("cannot swap schemas that are in the ambient database");
    }

    let schema_a = match scx.resolve_schema(name_a.clone()) {
        Ok(schema) => schema,
        Err(_) if if_exists => {
            scx.catalog.add_notice(PlanNotice::ObjectDoesNotExist {
                name: name_a.to_ast_string_simple(),
                object_type: ObjectType::Schema,
            });
            return Ok(Plan::AlterNoop(AlterNoopPlan {
                object_type: ObjectType::Schema,
            }));
        }
        Err(e) => return Err(e),
    };

    let db_spec = schema_a.database().clone();
    if matches!(db_spec, ResolvedDatabaseSpecifier::Ambient) {
        sql_bail!("cannot swap schemas that are in the ambient database");
    };
    let schema_b = scx.resolve_schema_in_database(&db_spec, &name_b)?;

    // We cannot swap system schemas.
    if schema_a.id().is_system() || schema_b.id().is_system() {
        bail_never_supported!("swapping a system schema".to_string())
    }

    // Generate a temporary name we can swap schema_a to.
    //
    // 'check' returns if the temp schema name would be valid.
    const SCHEMA_SWAP_PREFIX: &str = "mz_schema_swap_";
    let check = |temp_suffix: &str| {
        let mut temp_name = ident!(SCHEMA_SWAP_PREFIX);
        temp_name.append_lossy(temp_suffix);
        scx.resolve_schema_in_database(&db_spec, &temp_name)
            .is_err()
    };
    let temp_suffix = gen_temp_suffix(&check)?;
    let name_temp = format!("{SCHEMA_SWAP_PREFIX}{temp_suffix}");

    Ok(Plan::AlterSchemaSwap(AlterSchemaSwapPlan {
        schema_a_spec: (*schema_a.database(), *schema_a.id()),
        schema_a_name: schema_a.name().schema.to_string(),
        schema_b_spec: (*schema_b.database(), *schema_b.id()),
        schema_b_name: schema_b.name().schema.to_string(),
        name_temp,
    }))
}

pub fn plan_alter_item_rename(
    scx: &mut StatementContext,
    object_type: ObjectType,
    name: UnresolvedItemName,
    to_item_name: Ident,
    if_exists: bool,
) -> Result<Plan, PlanError> {
    let resolved = match resolve_item_or_type(scx, object_type, name.clone(), if_exists) {
        Ok(r) => r,
        // Return a more helpful error on `DROP VIEW <materialized-view>`.
        Err(PlanError::MismatchedObjectType {
            name,
            is_type: ObjectType::MaterializedView,
            expected_type: ObjectType::View,
        }) => {
            return Err(PlanError::AlterViewOnMaterializedView(name.to_string()));
        }
        e => e?,
    };

    match resolved {
        Some(entry) => {
            let full_name = scx.catalog.resolve_full_name(entry.name());
            let item_type = entry.item_type();

            let proposed_name = QualifiedItemName {
                qualifiers: entry.name().qualifiers.clone(),
                item: to_item_name.clone().into_string(),
            };

            // For PostgreSQL compatibility, items and types cannot have
            // overlapping names in a variety of situations. See the comment on
            // `CatalogItemType::conflicts_with_type` for details.
            let conflicting_type_exists;
            let conflicting_item_exists;
            if item_type == CatalogItemType::Type {
                conflicting_type_exists = scx.catalog.get_type_by_name(&proposed_name).is_some();
                conflicting_item_exists = scx
                    .catalog
                    .get_item_by_name(&proposed_name)
                    .map(|item| item.item_type().conflicts_with_type())
                    .unwrap_or(false);
            } else {
                conflicting_type_exists = item_type.conflicts_with_type()
                    && scx.catalog.get_type_by_name(&proposed_name).is_some();
                conflicting_item_exists = scx.catalog.get_item_by_name(&proposed_name).is_some();
            };
            if conflicting_type_exists || conflicting_item_exists {
                sql_bail!("catalog item '{}' already exists", to_item_name);
            }

            Ok(Plan::AlterItemRename(AlterItemRenamePlan {
                id: entry.id(),
                current_full_name: full_name,
                to_name: normalize::ident(to_item_name),
                object_type,
            }))
        }
        None => {
            scx.catalog.add_notice(PlanNotice::ObjectDoesNotExist {
                name: name.to_ast_string_simple(),
                object_type,
            });

            Ok(Plan::AlterNoop(AlterNoopPlan { object_type }))
        }
    }
}

pub fn plan_alter_cluster_rename(
    scx: &mut StatementContext,
    object_type: ObjectType,
    name: Ident,
    to_name: Ident,
    if_exists: bool,
) -> Result<Plan, PlanError> {
    match resolve_cluster(scx, &name, if_exists)? {
        Some(entry) => Ok(Plan::AlterClusterRename(AlterClusterRenamePlan {
            id: entry.id(),
            name: entry.name().to_string(),
            to_name: ident(to_name),
        })),
        None => {
            scx.catalog.add_notice(PlanNotice::ObjectDoesNotExist {
                name: name.to_ast_string_simple(),
                object_type,
            });

            Ok(Plan::AlterNoop(AlterNoopPlan { object_type }))
        }
    }
}

pub fn plan_alter_cluster_swap<F>(
    scx: &mut StatementContext,
    name_a: Ident,
    name_b: Ident,
    if_exists: bool,
    gen_temp_suffix: F,
) -> Result<Plan, PlanError>
where
    F: Fn(&dyn Fn(&str) -> bool) -> Result<String, PlanError>,
{
    let cluster_a = match scx.resolve_cluster(Some(&name_a)) {
        Ok(cluster) => cluster,
        Err(_) if if_exists => {
            scx.catalog.add_notice(PlanNotice::ObjectDoesNotExist {
                name: name_a.to_ast_string_simple(),
                object_type: ObjectType::Cluster,
            });
            return Ok(Plan::AlterNoop(AlterNoopPlan {
                object_type: ObjectType::Cluster,
            }));
        }
        Err(e) => return Err(e),
    };
    let cluster_b = scx.resolve_cluster(Some(&name_b))?;

    const CLUSTER_SWAP_PREFIX: &str = "mz_cluster_swap_";
    let check = |temp_suffix: &str| {
        let mut temp_name = ident!(CLUSTER_SWAP_PREFIX);
        temp_name.append_lossy(temp_suffix);
        match scx.catalog.resolve_cluster(Some(temp_name.as_str())) {
            // Temp name does not exist, so we can use it.
            Err(CatalogError::UnknownCluster(_)) => true,
            // Temp name already exists!
            Ok(_) | Err(_) => false,
        }
    };
    let temp_suffix = gen_temp_suffix(&check)?;
    let name_temp = format!("{CLUSTER_SWAP_PREFIX}{temp_suffix}");

    Ok(Plan::AlterClusterSwap(AlterClusterSwapPlan {
        id_a: cluster_a.id(),
        id_b: cluster_b.id(),
        name_a: name_a.into_string(),
        name_b: name_b.into_string(),
        name_temp,
    }))
}

pub fn plan_alter_cluster_replica_rename(
    scx: &mut StatementContext,
    object_type: ObjectType,
    name: QualifiedReplica,
    to_item_name: Ident,
    if_exists: bool,
) -> Result<Plan, PlanError> {
    match resolve_cluster_replica(scx, &name, if_exists)? {
        Some((cluster, replica)) => {
            ensure_cluster_is_not_managed(scx, cluster.id())?;
            Ok(Plan::AlterClusterReplicaRename(
                AlterClusterReplicaRenamePlan {
                    cluster_id: cluster.id(),
                    replica_id: replica,
                    name: QualifiedReplica {
                        cluster: Ident::new(cluster.name())?,
                        replica: name.replica,
                    },
                    to_name: normalize::ident(to_item_name),
                },
            ))
        }
        None => {
            scx.catalog.add_notice(PlanNotice::ObjectDoesNotExist {
                name: name.to_ast_string_simple(),
                object_type,
            });

            Ok(Plan::AlterNoop(AlterNoopPlan { object_type }))
        }
    }
}

pub fn describe_alter_object_swap(
    _: &StatementContext,
    _: AlterObjectSwapStatement,
) -> Result<StatementDesc, PlanError> {
    Ok(StatementDesc::new(None))
}

pub fn plan_alter_object_swap(
    scx: &mut StatementContext,
    stmt: AlterObjectSwapStatement,
) -> Result<Plan, PlanError> {
    scx.require_feature_flag(&vars::ENABLE_ALTER_SWAP)?;

    let AlterObjectSwapStatement {
        object_type,
        if_exists,
        name_a,
        name_b,
    } = stmt;
    let object_type = object_type.into();

    // We'll try 10 times to generate a temporary suffix.
    let gen_temp_suffix = |check_fn: &dyn Fn(&str) -> bool| {
        let mut attempts = 0;
        let name_temp = loop {
            attempts += 1;
            if attempts > 10 {
                tracing::warn!("Unable to generate temp id for swapping");
                sql_bail!("unable to swap!");
            }

            // Call the provided closure to make sure this name is unique!
            let short_id = mz_ore::id_gen::temp_id();
            if check_fn(&short_id) {
                break short_id;
            }
        };

        Ok(name_temp)
    };

    match (object_type, name_a, name_b) {
        (ObjectType::Schema, UnresolvedObjectName::Schema(name_a), name_b) => {
            plan_alter_schema_swap(scx, name_a, name_b, if_exists, gen_temp_suffix)
        }
        (ObjectType::Cluster, UnresolvedObjectName::Cluster(name_a), name_b) => {
            plan_alter_cluster_swap(scx, name_a, name_b, if_exists, gen_temp_suffix)
        }
        (ObjectType::Schema | ObjectType::Cluster, _, _) => {
            bail_internal!("name type does not match object type for ALTER SWAP")
        }
        (
            ObjectType::Table
            | ObjectType::View
            | ObjectType::MaterializedView
            | ObjectType::Source
            | ObjectType::Sink
            | ObjectType::MetricSink
            | ObjectType::Index
            | ObjectType::Type
            | ObjectType::Role
            | ObjectType::ClusterReplica
            | ObjectType::Secret
            | ObjectType::Connection
            | ObjectType::Database
            | ObjectType::Func
            | ObjectType::NetworkPolicy,
            _,
            _,
        ) => Err(PlanError::Unsupported {
            feature: format!("ALTER {object_type} .. SWAP WITH ..."),
            discussion_no: None,
        }),
    }
}

pub fn describe_alter_retain_history(
    _: &StatementContext,
    _: AlterRetainHistoryStatement<Aug>,
) -> Result<StatementDesc, PlanError> {
    Ok(StatementDesc::new(None))
}

pub fn plan_alter_retain_history(
    scx: &StatementContext,
    AlterRetainHistoryStatement {
        object_type,
        if_exists,
        name,
        history,
    }: AlterRetainHistoryStatement<Aug>,
) -> Result<Plan, PlanError> {
    alter_retain_history(scx, object_type.into(), if_exists, name, history)
}

fn alter_retain_history(
    scx: &StatementContext,
    object_type: ObjectType,
    if_exists: bool,
    name: UnresolvedObjectName,
    history: Option<WithOptionValue<Aug>>,
) -> Result<Plan, PlanError> {
    let name = match (object_type, name) {
        (
            // View gets a special error below.
            ObjectType::View
            | ObjectType::MaterializedView
            | ObjectType::Table
            | ObjectType::Source
            | ObjectType::Index,
            UnresolvedObjectName::Item(name),
        ) => name,
        (object_type, _) => {
            bail_unsupported!(format!("RETAIN HISTORY on {object_type}"))
        }
    };
    match resolve_item_or_type(scx, object_type, name.clone(), if_exists)? {
        Some(entry) => {
            let full_name = scx.catalog.resolve_full_name(entry.name());
            let item_type = entry.item_type();

            // Return a more helpful error on `ALTER VIEW <materialized-view>`.
            if object_type == ObjectType::View && item_type == CatalogItemType::MaterializedView {
                return Err(PlanError::AlterViewOnMaterializedView(
                    full_name.to_string(),
                ));
            } else if object_type == ObjectType::View {
                sql_bail!("{object_type} does not support RETAIN HISTORY")
            } else if object_type != item_type {
                sql_bail!(
                    "\"{}\" is a {} not a {}",
                    full_name,
                    entry.item_type(),
                    format!("{object_type}").to_lowercase()
                )
            }

            // Save the original value so we can write it back down in the create_sql catalog item.
            let (value, lcw) = match &history {
                Some(WithOptionValue::RetainHistoryFor(value)) => {
                    let window = OptionalDuration::try_from_value(value.clone())?;
                    (Some(value.clone()), window.0)
                }
                // None is RESET, so use the default CW.
                None => (None, Some(DEFAULT_LOGICAL_COMPACTION_WINDOW_DURATION)),
                _ => sql_bail!("unexpected value type for RETAIN HISTORY"),
            };
            let window = plan_retain_history(scx, lcw)?;

            Ok(Plan::AlterRetainHistory(AlterRetainHistoryPlan {
                id: entry.id(),
                value,
                window,
                object_type,
            }))
        }
        None => {
            scx.catalog.add_notice(PlanNotice::ObjectDoesNotExist {
                name: name.to_ast_string_simple(),
                object_type,
            });

            Ok(Plan::AlterNoop(AlterNoopPlan { object_type }))
        }
    }
}

fn alter_source_timestamp_interval(
    scx: &StatementContext,
    if_exists: bool,
    source_name: UnresolvedItemName,
    value: Option<WithOptionValue<Aug>>,
) -> Result<Plan, PlanError> {
    let object_type = ObjectType::Source;
    match resolve_item_or_type(scx, object_type, source_name.clone(), if_exists)? {
        Some(entry) => {
            let full_name = scx.catalog.resolve_full_name(entry.name());
            if entry.item_type() != CatalogItemType::Source {
                sql_bail!(
                    "\"{}\" is a {} not a {}",
                    full_name,
                    entry.item_type(),
                    format!("{object_type}").to_lowercase()
                )
            }

            match value {
                Some(val) => {
                    let val = match val {
                        WithOptionValue::Value(v) => v,
                        _ => sql_bail!("TIMESTAMP INTERVAL requires an interval value"),
                    };
                    let duration = Duration::try_from_value(val.clone())?;

                    let min = scx.catalog.system_vars().min_timestamp_interval();
                    let max = scx.catalog.system_vars().max_timestamp_interval();
                    if duration < min || duration > max {
                        return Err(PlanError::InvalidTimestampInterval {
                            min,
                            max,
                            requested: duration,
                        });
                    }

                    Ok(Plan::AlterSourceTimestampInterval(
                        AlterSourceTimestampIntervalPlan {
                            id: entry.id(),
                            value: Some(val),
                            interval: duration,
                        },
                    ))
                }
                None => {
                    let interval = scx.catalog.system_vars().default_timestamp_interval();
                    Ok(Plan::AlterSourceTimestampInterval(
                        AlterSourceTimestampIntervalPlan {
                            id: entry.id(),
                            value: None,
                            interval,
                        },
                    ))
                }
            }
        }
        None => {
            scx.catalog.add_notice(PlanNotice::ObjectDoesNotExist {
                name: source_name.to_ast_string_simple(),
                object_type,
            });

            Ok(Plan::AlterNoop(AlterNoopPlan { object_type }))
        }
    }
}

pub fn describe_alter_secret_options(
    _: &StatementContext,
    _: AlterSecretStatement<Aug>,
) -> Result<StatementDesc, PlanError> {
    Ok(StatementDesc::new(None))
}

pub fn plan_alter_secret(
    scx: &mut StatementContext,
    stmt: AlterSecretStatement<Aug>,
) -> Result<Plan, PlanError> {
    let AlterSecretStatement {
        name,
        if_exists,
        value,
    } = stmt;
    let object_type = ObjectType::Secret;
    let id = match resolve_item_or_type(scx, object_type, name.clone(), if_exists)? {
        Some(entry) => entry.id(),
        None => {
            scx.catalog.add_notice(PlanNotice::ObjectDoesNotExist {
                name: name.to_string(),
                object_type,
            });

            return Ok(Plan::AlterNoop(AlterNoopPlan { object_type }));
        }
    };

    let secret_as = query::plan_secret_as(scx, value)?;

    Ok(Plan::AlterSecret(AlterSecretPlan { id, secret_as }))
}

pub fn describe_alter_connection(
    _: &StatementContext,
    _: AlterConnectionStatement<Aug>,
) -> Result<StatementDesc, PlanError> {
    Ok(StatementDesc::new(None))
}

generate_extracted_config!(AlterConnectionOption, (Validate, bool));

pub fn plan_alter_connection(
    scx: &StatementContext,
    stmt: AlterConnectionStatement<Aug>,
) -> Result<Plan, PlanError> {
    let AlterConnectionStatement {
        name,
        if_exists,
        actions,
        with_options,
    } = stmt;
    let conn_name = normalize::unresolved_item_name(name)?;
    let entry = match scx.catalog.resolve_item(&conn_name) {
        Ok(entry) => entry,
        Err(_) if if_exists => {
            scx.catalog.add_notice(PlanNotice::ObjectDoesNotExist {
                name: conn_name.to_string(),
                object_type: ObjectType::Connection,
            });

            return Ok(Plan::AlterNoop(AlterNoopPlan {
                object_type: ObjectType::Connection,
            }));
        }
        Err(e) => return Err(e.into()),
    };

    let connection = entry.connection()?;

    if actions
        .iter()
        .any(|action| matches!(action, AlterConnectionAction::RotateKeys))
    {
        if actions.len() > 1 {
            sql_bail!("cannot specify any other actions alongside ALTER CONNECTION...ROTATE KEYS");
        }

        if !with_options.is_empty() {
            sql_bail!(
                "ALTER CONNECTION...ROTATE KEYS does not support WITH ({})",
                with_options
                    .iter()
                    .map(|o| o.to_ast_string_simple())
                    .join(", ")
            );
        }

        if !matches!(connection, Connection::Ssh(_)) {
            sql_bail!(
                "{} is not an SSH connection",
                scx.catalog.resolve_full_name(entry.name())
            )
        }

        return Ok(Plan::AlterConnection(AlterConnectionPlan {
            id: entry.id(),
            action: crate::plan::AlterConnectionAction::RotateKeys,
        }));
    }

    let options = AlterConnectionOptionExtracted::try_from(with_options)?;
    if options.validate.is_some() {
        scx.require_feature_flag(&vars::ENABLE_CONNECTION_VALIDATION_SYNTAX)?;
    }

    let validate = match options.validate {
        Some(val) => val,
        None => {
            scx.catalog
                .system_vars()
                .enable_default_connection_validation()
                && connection.validate_by_default()
        }
    };

    let connection_type = match connection {
        Connection::Aws(_) => CreateConnectionType::Aws,
        Connection::AwsPrivatelink(_) => CreateConnectionType::AwsPrivatelink,
        Connection::Gcp(_) => CreateConnectionType::Gcp,
        Connection::Kafka(_) => CreateConnectionType::Kafka,
        Connection::Csr(_) => CreateConnectionType::Csr,
        Connection::GlueSchemaRegistry(_) => CreateConnectionType::GlueSchemaRegistry,
        Connection::Postgres(_) => CreateConnectionType::Postgres,
        Connection::Ssh(_) => CreateConnectionType::Ssh,
        Connection::MySql(_) => CreateConnectionType::MySql,
        Connection::SqlServer(_) => CreateConnectionType::SqlServer,
        Connection::IcebergCatalog(_) => CreateConnectionType::IcebergCatalog,
    };

    // Collect all options irrespective of action taken on them.
    let specified_options: BTreeSet<_> = actions
        .iter()
        .map(|action: &AlterConnectionAction<Aug>| match action {
            AlterConnectionAction::SetOption(option) => Ok(option.name.clone()),
            AlterConnectionAction::DropOption(name) => Ok(name.clone()),
            AlterConnectionAction::RotateKeys => {
                Err(internal_err!("RotateKeys is handled separately above"))
            }
        })
        .collect::<Result<_, PlanError>>()?;

    for invalid in INALTERABLE_OPTIONS {
        if specified_options.contains(invalid) {
            sql_bail!("cannot ALTER {} option {}", connection_type, invalid);
        }
    }

    connection::validate_options_per_connection_type(connection_type, specified_options)?;

    // Partition operations into set and drop.
    let mut set_options_vec: Vec<_> = Vec::new();
    let mut drop_options: BTreeSet<_> = BTreeSet::new();
    for action in actions {
        match action {
            AlterConnectionAction::SetOption(option) => set_options_vec.push(option),
            AlterConnectionAction::DropOption(name) => {
                drop_options.insert(name);
            }
            AlterConnectionAction::RotateKeys => {
                bail_internal!("RotateKeys is handled separately above")
            }
        }
    }

    let set_options: BTreeMap<_, _> = set_options_vec
        .clone()
        .into_iter()
        .map(|option| (option.name, option.value))
        .collect();

    // Type check values + avoid duplicates; we don't want to e.g. let users
    // drop and set the same option in the same statement, so treating drops as
    // sets here is fine.
    let connection_options_extracted =
        connection::ConnectionOptionExtracted::try_from(set_options_vec)?;

    let duplicates: Vec<_> = connection_options_extracted
        .seen
        .intersection(&drop_options)
        .collect();

    if !duplicates.is_empty() {
        sql_bail!(
            "cannot both SET and DROP/RESET options {}",
            duplicates
                .iter()
                .map(|option| option.to_string())
                .join(", ")
        )
    }

    for mutually_exclusive_options in MUTUALLY_EXCLUSIVE_SETS {
        let set_options_count = mutually_exclusive_options
            .iter()
            .filter(|o| set_options.contains_key(o))
            .count();
        let drop_options_count = mutually_exclusive_options
            .iter()
            .filter(|o| drop_options.contains(o))
            .count();

        // Disallow setting _and_ resetting mutually exclusive options
        if set_options_count > 0 && drop_options_count > 0 {
            sql_bail!(
                "cannot both SET and DROP/RESET mutually exclusive {} options {}",
                connection_type,
                mutually_exclusive_options
                    .iter()
                    .map(|option| option.to_string())
                    .join(", ")
            )
        }

        // If any option is either set or dropped, ensure all mutually exclusive
        // options are dropped. We do this "behind the scenes", even though we
        // disallow users from performing the same action because this is the
        // mechanism by which we overwrite values elsewhere in the code.
        if set_options_count > 0 || drop_options_count > 0 {
            drop_options.extend(mutually_exclusive_options.iter().cloned());
        }

        // n.b. if mutually exclusive options are set, those will error when we
        // try to replan the connection.
    }

    Ok(Plan::AlterConnection(AlterConnectionPlan {
        id: entry.id(),
        action: crate::plan::AlterConnectionAction::AlterOptions {
            set_options,
            drop_options,
            validate,
        },
    }))
}

pub fn describe_alter_sink(
    _: &StatementContext,
    _: AlterSinkStatement<Aug>,
) -> Result<StatementDesc, PlanError> {
    Ok(StatementDesc::new(None))
}

pub fn plan_alter_sink(
    scx: &mut StatementContext,
    stmt: AlterSinkStatement<Aug>,
) -> Result<Plan, PlanError> {
    let AlterSinkStatement {
        sink_name,
        if_exists,
        action,
    } = stmt;

    let object_type = ObjectType::Sink;
    let item = resolve_item_or_type(scx, object_type, sink_name.clone(), if_exists)?;

    let Some(item) = item else {
        scx.catalog.add_notice(PlanNotice::ObjectDoesNotExist {
            name: sink_name.to_string(),
            object_type,
        });

        return Ok(Plan::AlterNoop(AlterNoopPlan { object_type }));
    };
    // Always ALTER objects from their latest version.
    let item = item.at_version(RelationVersionSelector::Latest);

    // First we reconstruct the original CREATE SINK statement
    let create_sql = item.create_sql();
    let stmts = mz_sql_parser::parser::parse_statements(create_sql)?;
    let [stmt]: [StatementParseResult; 1] = stmts
        .try_into()
        .map_err(|_| internal_err!("create SQL of sink was not exactly one statement"))?;
    let Statement::CreateSink(stmt) = stmt.ast else {
        bail_internal!("create SQL of sink is not a CREATE SINK statement");
    };
    let (mut stmt, _) = crate::names::resolve(scx.catalog, stmt)?;

    // Then apply the requested change to the statement
    let mut set_options = vec![];
    let mut reset_options = vec![];
    match action {
        AlterSinkAction::ChangeRelation(new_from) => {
            stmt.from = new_from;
        }
        AlterSinkAction::SetOptions(options) => {
            for option in &options {
                match &option.name {
                    CreateSinkOptionName::CommitInterval => {}
                    name => bail_unsupported!(format!(
                        "ALTER SINK ... SET ({})",
                        name.to_ast_string_simple()
                    )),
                }
            }
            // Setting every option to its current value would restart the
            // sink dataflow without changing its behavior, so make it a
            // no-op instead. The values are compared as ASTs, so spelling
            // the same value differently (`'60s'` vs `'1m'`) still counts
            // as a change.
            //
            // NOTE: This check races with other `ALTER SINK` statements,
            // because `ALTER SINK` does not take the DDL lock. Example: the
            // interval is '1s', we plan `SET (COMMIT INTERVAL = '1s')` as a
            // no-op, and before we respond a concurrent `ALTER SINK` changes
            // the interval to '2s'. We then report success even though the
            // interval is now '2s', not the '1s' we were asked for. An alter
            // that is not a no-op detects this in sequencing by checking the
            // sink's version, but a no-op is never re-checked. We accept this
            // because alters are last-writer-wins anyway.
            if options.iter().all(|o| stmt.with_options.contains(o)) {
                return Ok(Plan::AlterNoop(AlterNoopPlan { object_type }));
            }
            set_options = options;
        }
        AlterSinkAction::ResetOptions(names) => {
            for name in &names {
                match name {
                    CreateSinkOptionName::CommitInterval => {}
                    name => bail_unsupported!(format!(
                        "ALTER SINK ... RESET ({})",
                        name.to_ast_string_simple()
                    )),
                }
                // Resetting an option that is not set would still restart the
                // sink dataflow, so reject it instead of silently no-oping.
                if !stmt.with_options.iter().any(|o| o.name == *name) {
                    sql_bail!(
                        "cannot RESET {}: option is not set",
                        name.to_ast_string_simple()
                    );
                }
            }
            reset_options = names;
        }
    }
    crate::plan::apply_sink_option_edits(&mut stmt.with_options, &set_options, &reset_options);

    // Finally re-plan the modified create sink statement to verify the new configuration is valid
    let Plan::CreateSink(mut plan) = plan_sink(scx, stmt)? else {
        bail_internal!("plan_sink did not produce a CreateSink plan");
    };

    plan.sink.version += 1;

    Ok(Plan::AlterSink(AlterSinkPlan {
        item_id: item.id(),
        global_id: item.global_id(),
        sink: plan.sink,
        with_snapshot: plan.with_snapshot,
        in_cluster: plan.in_cluster,
        set_options,
        reset_options,
    }))
}

pub fn describe_alter_source(
    _: &StatementContext,
    _: AlterSourceStatement<Aug>,
) -> Result<StatementDesc, PlanError> {
    // TODO: put the options here, right?
    Ok(StatementDesc::new(None))
}

generate_extracted_config!(
    AlterSourceAddSubsourceOption,
    (TextColumns, Vec::<UnresolvedItemName>, Default(vec![])),
    (ExcludeColumns, Vec::<UnresolvedItemName>, Default(vec![])),
    (Details, String)
);

pub fn plan_alter_source(
    scx: &mut StatementContext,
    stmt: AlterSourceStatement<Aug>,
) -> Result<Plan, PlanError> {
    let AlterSourceStatement {
        source_name,
        if_exists,
        action,
    } = stmt;
    let object_type = ObjectType::Source;

    if resolve_item_or_type(scx, object_type, source_name.clone(), if_exists)?.is_none() {
        scx.catalog.add_notice(PlanNotice::ObjectDoesNotExist {
            name: source_name.to_string(),
            object_type,
        });

        return Ok(Plan::AlterNoop(AlterNoopPlan { object_type }));
    }

    match action {
        AlterSourceAction::SetOptions(options) => {
            let mut options = options.into_iter();
            let option = options
                .next()
                .ok_or_else(|| sql_err!("ALTER SOURCE SET requires at least one option"))?;
            if option.name == CreateSourceOptionName::RetainHistory {
                if options.next().is_some() {
                    sql_bail!("RETAIN HISTORY must be only option");
                }
                return alter_retain_history(
                    scx,
                    object_type,
                    if_exists,
                    UnresolvedObjectName::Item(source_name),
                    option.value,
                );
            }
            if option.name == CreateSourceOptionName::TimestampInterval {
                if options.next().is_some() {
                    sql_bail!("TIMESTAMP INTERVAL must be only option");
                }
                return alter_source_timestamp_interval(scx, if_exists, source_name, option.value);
            }
            // n.b we use this statement in purification in a way that cannot be
            // planned directly.
            sql_bail!(
                "Cannot modify the {} of a SOURCE.",
                option.name.to_ast_string_simple()
            );
        }
        AlterSourceAction::ResetOptions(reset) => {
            let mut options = reset.into_iter();
            let option = options
                .next()
                .ok_or_else(|| sql_err!("ALTER SOURCE RESET requires at least one option"))?;
            if option == CreateSourceOptionName::RetainHistory {
                if options.next().is_some() {
                    sql_bail!("RETAIN HISTORY must be only option");
                }
                return alter_retain_history(
                    scx,
                    object_type,
                    if_exists,
                    UnresolvedObjectName::Item(source_name),
                    None,
                );
            }
            if option == CreateSourceOptionName::TimestampInterval {
                if options.next().is_some() {
                    sql_bail!("TIMESTAMP INTERVAL must be only option");
                }
                return alter_source_timestamp_interval(scx, if_exists, source_name, None);
            }
            sql_bail!(
                "Cannot modify the {} of a SOURCE.",
                option.to_ast_string_simple()
            );
        }
        AlterSourceAction::DropSubsources { .. } => {
            sql_bail!("ALTER SOURCE...DROP SUBSOURCE no longer supported; use DROP SOURCE")
        }
        AlterSourceAction::AddSubsources { .. } => {
            sql_bail!("ALTER SOURCE...ADD SUBSOURCE must be purified before planning")
        }
        AlterSourceAction::RefreshReferences => {
            sql_bail!("ALTER SOURCE...REFRESH REFERENCES must be purified before planning")
        }
    };
}

pub fn describe_alter_system_set(
    _: &StatementContext,
    _: AlterSystemSetStatement,
) -> Result<StatementDesc, PlanError> {
    Ok(StatementDesc::new(None))
}

pub fn plan_alter_system_set(
    _: &StatementContext,
    AlterSystemSetStatement { name, to }: AlterSystemSetStatement,
) -> Result<Plan, PlanError> {
    let name = name.to_string();
    Ok(Plan::AlterSystemSet(AlterSystemSetPlan {
        name,
        value: scl::plan_set_variable_to(to)?,
    }))
}

pub fn describe_alter_system_reset(
    _: &StatementContext,
    _: AlterSystemResetStatement,
) -> Result<StatementDesc, PlanError> {
    Ok(StatementDesc::new(None))
}

pub fn plan_alter_system_reset(
    _: &StatementContext,
    AlterSystemResetStatement { name }: AlterSystemResetStatement,
) -> Result<Plan, PlanError> {
    let name = name.to_string();
    Ok(Plan::AlterSystemReset(AlterSystemResetPlan { name }))
}

pub fn describe_alter_system_reset_all(
    _: &StatementContext,
    _: AlterSystemResetAllStatement,
) -> Result<StatementDesc, PlanError> {
    Ok(StatementDesc::new(None))
}

pub fn plan_alter_system_reset_all(
    _: &StatementContext,
    _: AlterSystemResetAllStatement,
) -> Result<Plan, PlanError> {
    Ok(Plan::AlterSystemResetAll(AlterSystemResetAllPlan {}))
}

pub fn describe_alter_role(
    _: &StatementContext,
    _: AlterRoleStatement<Aug>,
) -> Result<StatementDesc, PlanError> {
    Ok(StatementDesc::new(None))
}

pub fn plan_alter_role(
    scx: &StatementContext,
    AlterRoleStatement { name, option }: AlterRoleStatement<Aug>,
) -> Result<Plan, PlanError> {
    let option = match option {
        AlterRoleOption::Attributes(attrs) => {
            let attrs = plan_role_attributes(attrs, scx)?;
            PlannedAlterRoleOption::Attributes(attrs)
        }
        AlterRoleOption::Variable(variable) => {
            let var = plan_role_variable(scx, variable)?;
            PlannedAlterRoleOption::Variable(var)
        }
    };

    Ok(Plan::AlterRole(AlterRolePlan {
        id: name.id,
        name: name.name,
        option,
    }))
}

pub fn describe_alter_table_add_column(
    _: &StatementContext,
    _: AlterTableAddColumnStatement<Aug>,
) -> Result<StatementDesc, PlanError> {
    Ok(StatementDesc::new(None))
}

pub fn plan_alter_table_add_column(
    scx: &StatementContext,
    stmt: AlterTableAddColumnStatement<Aug>,
) -> Result<Plan, PlanError> {
    let AlterTableAddColumnStatement {
        if_exists,
        name,
        if_col_not_exist,
        column_name,
        data_type,
    } = stmt;
    let object_type = ObjectType::Table;

    scx.require_feature_flag(&vars::ENABLE_ALTER_TABLE_ADD_COLUMN)?;

    let (relation_id, item_name, desc) =
        match resolve_item_or_type(scx, object_type, name.clone(), if_exists)? {
            Some(item) => {
                // Always add columns to the latest version of the item.
                let item_name = scx.catalog.resolve_full_name(item.name());
                let item = item.at_version(RelationVersionSelector::Latest);
                let desc = item
                    .relation_desc()
                    .ok_or_else(|| sql_err!("item does not have a relation description"))?
                    .into_owned();
                (item.id(), item_name, desc)
            }
            None => {
                scx.catalog.add_notice(PlanNotice::ObjectDoesNotExist {
                    name: name.to_ast_string_simple(),
                    object_type,
                });
                return Ok(Plan::AlterNoop(AlterNoopPlan { object_type }));
            }
        };

    let column_name = ColumnName::from(column_name.as_str());
    if desc.get_by_name(&column_name).is_some() {
        if if_col_not_exist {
            scx.catalog.add_notice(PlanNotice::ColumnAlreadyExists {
                column_name: column_name.to_string(),
                object_name: item_name.item,
            });
            return Ok(Plan::AlterNoop(AlterNoopPlan { object_type }));
        } else {
            return Err(PlanError::ColumnAlreadyExists {
                column_name,
                object_name: item_name.item,
            });
        }
    }

    let scalar_type = scalar_type_from_sql(scx, &data_type)?;
    // TODO(alter_table): Support non-nullable columns with default values.
    let column_type = scalar_type.nullable(true);
    // "unresolve" our data type so we can later update the persisted create_sql.
    let raw_sql_type = mz_sql_parser::parser::parse_data_type(&data_type.to_ast_string_stable())?;

    Ok(Plan::AlterTableAddColumn(AlterTablePlan {
        relation_id,
        column_name,
        column_type,
        raw_sql_type,
    }))
}

pub fn describe_alter_materialized_view_apply_replacement(
    _: &StatementContext,
    _: AlterMaterializedViewApplyReplacementStatement,
) -> Result<StatementDesc, PlanError> {
    Ok(StatementDesc::new(None))
}

pub fn plan_alter_materialized_view_apply_replacement(
    scx: &StatementContext,
    stmt: AlterMaterializedViewApplyReplacementStatement,
) -> Result<Plan, PlanError> {
    let AlterMaterializedViewApplyReplacementStatement {
        if_exists,
        name,
        replacement_name,
    } = stmt;

    scx.require_feature_flag(&vars::ENABLE_REPLACEMENT_MATERIALIZED_VIEWS)?;

    let object_type = ObjectType::MaterializedView;
    let Some(mv) = resolve_item_or_type(scx, object_type, name.clone(), if_exists)? else {
        scx.catalog.add_notice(PlanNotice::ObjectDoesNotExist {
            name: name.to_ast_string_simple(),
            object_type,
        });
        return Ok(Plan::AlterNoop(AlterNoopPlan { object_type }));
    };

    let replacement = resolve_item_or_type(scx, object_type, replacement_name, false)?
        .ok_or_else(|| sql_err!("replacement materialized view does not exist"))?;

    if replacement.replacement_target() != Some(mv.id()) {
        return Err(PlanError::InvalidReplacement {
            item_type: mv.item_type(),
            item_name: scx.catalog.minimal_qualification(mv.name()),
            replacement_type: replacement.item_type(),
            replacement_name: scx.catalog.minimal_qualification(replacement.name()),
        });
    }

    Ok(Plan::AlterMaterializedViewApplyReplacement(
        AlterMaterializedViewApplyReplacementPlan {
            id: mv.id(),
            replacement_id: replacement.id(),
        },
    ))
}

pub fn describe_comment(
    _: &StatementContext,
    _: CommentStatement<Aug>,
) -> Result<StatementDesc, PlanError> {
    Ok(StatementDesc::new(None))
}

pub fn plan_comment(
    scx: &mut StatementContext,
    stmt: CommentStatement<Aug>,
) -> Result<Plan, PlanError> {
    const MAX_COMMENT_LENGTH: usize = 1024;

    let CommentStatement { object, comment } = stmt;

    // TODO(parkmycar): Make max comment length configurable.
    if let Some(c) = &comment {
        if c.len() > 1024 {
            return Err(PlanError::CommentTooLong {
                length: c.len(),
                max_size: MAX_COMMENT_LENGTH,
            });
        }
    }

    let (object_id, column_pos) = match &object {
        com_ty @ CommentObjectType::Table { name }
        | com_ty @ CommentObjectType::View { name }
        | com_ty @ CommentObjectType::MaterializedView { name }
        | com_ty @ CommentObjectType::Index { name }
        | com_ty @ CommentObjectType::Func { name }
        | com_ty @ CommentObjectType::Connection { name }
        | com_ty @ CommentObjectType::Source { name }
        | com_ty @ CommentObjectType::Sink { name }
        | com_ty @ CommentObjectType::Secret { name } => {
            let item = scx.get_item_by_resolved_name(name)?;
            match (com_ty, item.item_type()) {
                (CommentObjectType::Table { .. }, CatalogItemType::Table) => {
                    (CommentObjectId::Table(item.id()), None)
                }
                (CommentObjectType::View { .. }, CatalogItemType::View) => {
                    (CommentObjectId::View(item.id()), None)
                }
                (CommentObjectType::MaterializedView { .. }, CatalogItemType::MaterializedView) => {
                    (CommentObjectId::MaterializedView(item.id()), None)
                }
                (CommentObjectType::Index { .. }, CatalogItemType::Index) => {
                    (CommentObjectId::Index(item.id()), None)
                }
                (CommentObjectType::Func { .. }, CatalogItemType::Func) => {
                    (CommentObjectId::Func(item.id()), None)
                }
                (CommentObjectType::Connection { .. }, CatalogItemType::Connection) => {
                    (CommentObjectId::Connection(item.id()), None)
                }
                (CommentObjectType::Source { .. }, CatalogItemType::Source) => {
                    (CommentObjectId::Source(item.id()), None)
                }
                (CommentObjectType::Sink { .. }, CatalogItemType::Sink) => {
                    (CommentObjectId::Sink(item.id()), None)
                }
                (CommentObjectType::Secret { .. }, CatalogItemType::Secret) => {
                    (CommentObjectId::Secret(item.id()), None)
                }
                (com_ty, cat_ty) => {
                    let expected_type = match com_ty {
                        CommentObjectType::Table { .. } => ObjectType::Table,
                        CommentObjectType::View { .. } => ObjectType::View,
                        CommentObjectType::MaterializedView { .. } => ObjectType::MaterializedView,
                        CommentObjectType::Index { .. } => ObjectType::Index,
                        CommentObjectType::Func { .. } => ObjectType::Func,
                        CommentObjectType::Connection { .. } => ObjectType::Connection,
                        CommentObjectType::Source { .. } => ObjectType::Source,
                        CommentObjectType::Sink { .. } => ObjectType::Sink,
                        CommentObjectType::Secret { .. } => ObjectType::Secret,
                        _ => sql_bail!("cannot comment on this object type"),
                    };

                    return Err(PlanError::InvalidObjectType {
                        expected_type: SystemObjectType::Object(expected_type),
                        actual_type: SystemObjectType::Object(cat_ty.into()),
                        object_name: item.name().item.clone(),
                    });
                }
            }
        }
        CommentObjectType::Type { ty } => match ty {
            ResolvedDataType::AnonymousList(_) | ResolvedDataType::AnonymousMap { .. } => {
                sql_bail!("cannot comment on anonymous list or map type");
            }
            ResolvedDataType::Named { id, modifiers, .. } => {
                if !modifiers.is_empty() {
                    sql_bail!("cannot comment on type with modifiers");
                }
                (CommentObjectId::Type(*id), None)
            }
            ResolvedDataType::Error => bail_internal!("unresolved data type"),
        },
        CommentObjectType::Column { name } => {
            let (item, pos) = scx.get_column_by_resolved_name(name)?;
            match item.item_type() {
                CatalogItemType::Table => (CommentObjectId::Table(item.id()), Some(pos + 1)),
                CatalogItemType::Source => (CommentObjectId::Source(item.id()), Some(pos + 1)),
                CatalogItemType::View => (CommentObjectId::View(item.id()), Some(pos + 1)),
                CatalogItemType::MaterializedView => {
                    (CommentObjectId::MaterializedView(item.id()), Some(pos + 1))
                }
                CatalogItemType::Type => (CommentObjectId::Type(item.id()), Some(pos + 1)),
                r => {
                    return Err(PlanError::Unsupported {
                        feature: format!("Specifying comments on a column of {r}"),
                        discussion_no: None,
                    });
                }
            }
        }
        CommentObjectType::Role { name } => (CommentObjectId::Role(name.id), None),
        CommentObjectType::Database { name } => {
            (CommentObjectId::Database(*name.database_id()), None)
        }
        CommentObjectType::Schema { name } => {
            // Temporary schemas cannot have comments - they are connection-specific
            // and transient. With lazy temporary schema creation, the temp schema
            // may not exist yet, but we still need to return the correct error.
            if matches!(name.schema_spec(), SchemaSpecifier::Temporary) {
                sql_bail!(
                    "cannot comment on schema {} because it is a temporary schema",
                    mz_repr::namespaces::MZ_TEMP_SCHEMA
                );
            }
            (
                CommentObjectId::Schema((*name.database_spec(), *name.schema_spec())),
                None,
            )
        }
        CommentObjectType::Cluster { name } => (CommentObjectId::Cluster(name.id), None),
        CommentObjectType::ClusterReplica { name } => {
            let replica = scx.catalog.resolve_cluster_replica(name)?;
            (
                CommentObjectId::ClusterReplica((replica.cluster_id(), replica.replica_id())),
                None,
            )
        }
        CommentObjectType::NetworkPolicy { name } => {
            (CommentObjectId::NetworkPolicy(name.id), None)
        }
    };

    // Note: the `mz_comments` table uses an `Int4` for the column position, but in the catalog storage we
    // store a `usize` which would be a `Uint8`. We guard against a safe conversion here because
    // it's the easiest place to raise an error.
    //
    // TODO(parkmycar): https://github.com/MaterializeInc/database-issues/issues/6711.
    if let Some(p) = column_pos {
        i32::try_from(p).map_err(|_| PlanError::TooManyColumns {
            max_num_columns: MAX_NUM_COLUMNS,
            req_num_columns: p,
        })?;
    }

    Ok(Plan::Comment(CommentPlan {
        object_id,
        sub_component: column_pos,
        comment,
    }))
}

pub(crate) fn resolve_cluster<'a>(
    scx: &'a StatementContext,
    name: &'a Ident,
    if_exists: bool,
) -> Result<Option<&'a dyn CatalogCluster<'a>>, PlanError> {
    match scx.resolve_cluster(Some(name)) {
        Ok(cluster) => Ok(Some(cluster)),
        Err(_) if if_exists => Ok(None),
        Err(e) => Err(e),
    }
}

pub(crate) fn resolve_cluster_replica<'a>(
    scx: &'a StatementContext,
    name: &QualifiedReplica,
    if_exists: bool,
) -> Result<Option<(&'a dyn CatalogCluster<'a>, ReplicaId)>, PlanError> {
    match scx.resolve_cluster(Some(&name.cluster)) {
        Ok(cluster) => match cluster.replica_ids().get(name.replica.as_str()) {
            Some(replica_id) => Ok(Some((cluster, *replica_id))),
            None if if_exists => Ok(None),
            None => Err(sql_err!(
                "CLUSTER {} has no CLUSTER REPLICA named {}",
                cluster.name(),
                name.replica.as_str().quoted(),
            )),
        },
        Err(_) if if_exists => Ok(None),
        Err(e) => Err(e),
    }
}

pub(crate) fn resolve_database<'a>(
    scx: &'a StatementContext,
    name: &'a UnresolvedDatabaseName,
    if_exists: bool,
) -> Result<Option<&'a dyn CatalogDatabase>, PlanError> {
    match scx.resolve_database(name) {
        Ok(database) => Ok(Some(database)),
        Err(_) if if_exists => Ok(None),
        Err(e) => Err(e),
    }
}

pub(crate) fn resolve_schema<'a>(
    scx: &'a StatementContext,
    name: UnresolvedSchemaName,
    if_exists: bool,
) -> Result<Option<(ResolvedDatabaseSpecifier, SchemaSpecifier)>, PlanError> {
    match scx.resolve_schema(name) {
        Ok(schema) => Ok(Some((schema.database().clone(), schema.id().clone()))),
        Err(_) if if_exists => Ok(None),
        Err(e) => Err(e),
    }
}

pub(crate) fn resolve_network_policy<'a>(
    scx: &'a StatementContext,
    name: Ident,
    if_exists: bool,
) -> Result<Option<ResolvedNetworkPolicyName>, PlanError> {
    match scx
        .catalog
        .resolve_network_policy(normalize::ident_ref(&name))
    {
        Ok(policy) => Ok(Some(ResolvedNetworkPolicyName {
            id: policy.id(),
            name: policy.name().to_string(),
        })),
        Err(_) if if_exists => Ok(None),
        Err(e) => Err(e.into()),
    }
}

pub(crate) fn resolve_item_or_type<'a>(
    scx: &'a StatementContext,
    object_type: ObjectType,
    name: UnresolvedItemName,
    if_exists: bool,
) -> Result<Option<&'a dyn CatalogItem>, PlanError> {
    let name = normalize::unresolved_item_name(name)?;
    let catalog_item = match object_type {
        ObjectType::Type => scx.catalog.resolve_type(&name),
        ObjectType::Table
        | ObjectType::View
        | ObjectType::MaterializedView
        | ObjectType::Source
        | ObjectType::Sink
        | ObjectType::MetricSink
        | ObjectType::Index
        | ObjectType::Role
        | ObjectType::Cluster
        | ObjectType::ClusterReplica
        | ObjectType::Secret
        | ObjectType::Connection
        | ObjectType::Database
        | ObjectType::Schema
        | ObjectType::Func
        | ObjectType::NetworkPolicy => scx.catalog.resolve_item(&name),
    };

    match catalog_item {
        Ok(item) => {
            let is_type = ObjectType::from(item.item_type());
            if object_type == is_type {
                Ok(Some(item))
            } else {
                Err(PlanError::MismatchedObjectType {
                    name: scx.catalog.minimal_qualification(item.name()),
                    is_type,
                    expected_type: object_type,
                })
            }
        }
        Err(_) if if_exists => Ok(None),
        Err(e) => Err(e.into()),
    }
}

/// Returns an error if the given cluster is a managed cluster
fn ensure_cluster_is_not_managed(
    scx: &StatementContext,
    cluster_id: ClusterId,
) -> Result<(), PlanError> {
    let cluster = scx.catalog.get_cluster(cluster_id);
    if cluster.is_managed() {
        Err(PlanError::ManagedCluster {
            cluster_name: cluster.name().to_string(),
        })
    } else {
        Ok(())
    }
}

#[cfg(test)]
mod tests {
    use super::{
        METRIC_SINK_CURATED_PREFIX_MARKER, METRIC_SINK_PREFIX_MARKER,
        validate_user_metric_sink_prefix,
    };

    #[mz_ore::test]
    fn user_metric_sink_prefix_cannot_reach_the_curated_prefix() {
        // Starts with the curated prefix.
        assert!(validate_user_metric_sink_prefix(METRIC_SINK_CURATED_PREFIX_MARKER).is_err());
        assert!(validate_user_metric_sink_prefix("mz_metric_sink_curated_x").is_err());
        // A proper prefix of the curated prefix, which a `curated_...` row would complete.
        assert!(validate_user_metric_sink_prefix(METRIC_SINK_PREFIX_MARKER).is_err());
        assert!(validate_user_metric_sink_prefix("mz_metric_sink_cur").is_err());
        // Still outside the reserved lane entirely.
        assert!(validate_user_metric_sink_prefix("myapp_").is_err());
        // A prefix that diverges from the curated prefix is fine, including one that merely contains
        // the marker further along.
        assert!(validate_user_metric_sink_prefix("mz_metric_sink_myapp_").is_ok());
        assert!(validate_user_metric_sink_prefix("mz_metric_sink_curatex_").is_ok());
        assert!(validate_user_metric_sink_prefix("mz_metric_sink_my_curated_").is_ok());
    }
}