mz_compute_types/plan/

reduce.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.

//! Reduction execution planning and dataflow construction.

//! We build `ReducePlan`s to manage the complexity of planning the generated dataflow for a
//! given reduce expression. The intent here is that each creating a `ReducePlan` should capture
//! all of the decision making about what kind of dataflow do we need to render and what each
//! operator needs to do, and then actually rendering the plan can be a relatively simple application
//! of (as much as possible) straight line code.
//!
//! Materialize needs to be able to maintain reductions incrementally (roughly, using
//! time proportional to the number of changes in the input) and ideally, with a
//! memory footprint proportional to the number of reductions being computed. We have to employ
//! several tricks to achieve that, and these tricks constitute most of the complexity involved
//! with planning and rendering reduce expressions. There's some additional complexity involved
//! in handling aggregations with `DISTINCT` correctly so that we can efficiently suppress
//! duplicate updates.
//!
//! In order to optimize the performance of our rendered dataflow, we divide all aggregations
//! into three distinct types. Each type gets rendered separately, with its own specialized plan
//! and dataflow. The three types are as follows:
//!
//! 1. Accumulable:
//!    Accumulable reductions can be computed inline in a Differential update's `difference`
//!    field because they basically boil down to tracking counts of things. `sum()` is an
//!    example of an accumulable reduction, and when some element `x` is removed from the set
//!    of elements being summed, we can introduce `-x` to incrementally maintain the sum. More
//!    formally, accumulable reductions correspond to instances of commutative Abelian groups.
//! 2. Hierarchical:
//!    Hierarchical reductions don't have a meaningful negation like accumulable reductions do, but
//!    they are still commutative and associative, which lets us compute the reduction over subsets
//!    of the input, and then compute the reduction again on those results. For example:
//!    `min[2, 5, 1, 10]` is the same as `min[ min[2, 5], min[1, 10]]`. When we compute hierarchical
//!    reductions this way, we can maintain the computation in sublinear time with respect to
//!    the overall input. `min` and `max` are two examples of hierarchical reductions. More formally,
//!    hierarchical reductions correspond to instances of semigroups, in that they are associative,
//!    but in order to benefit from being computed hierarchically, they need to have some reduction
//!    in data size as well. A function like "concat-everything-to-a-string" wouldn't benefit from
//!    hierarchical evaluation.
//!
//!    When the input is append-only, or monotonic, reductions that would otherwise have to be computed
//!    hierarchically can instead be computed in-place, because we only need to keep the value that's
//!    better than the "best" (minimal or maximal for min and max) seen so far.
//! 3. Basic:
//!    Basic reductions are a bit like the Hufflepuffs of this trifecta. They are neither accumulable nor
//!    hierarchical (most likely they are associative but don't involve any data reduction) and so for these
//!    we can't do much more than just defer to Differential's reduce operator and eat a large maintenance cost.
//!
//! When we render these reductions we want to limit the number of arrangements we produce. When we build a
//! dataflow for a reduction containing multiple types of reductions, we have no choice but to divide up the
//! requested aggregations by type, render each type separately and then take those results and collate them
//! back in the requested output order. However, if we only need to perform aggregations of a single reduction
//! type, we can specialize and render the dataflow to compute those aggregations in the correct order, and
//! return the output arrangement directly and avoid the extra collation arrangement.

use std::fmt::Display;

use mz_expr::explain::{HumanizeDisplay, HumanizedExpr, HumanizerMode};
use mz_expr::{
    AggregateExpr, AggregateFunc, MapFilterProject, MirScalarExpr, UnmaterializableFunc,
    permutation_for_arrangement,
};
use mz_ore::soft_assert_or_log;
use serde::{Deserialize, Serialize};

use crate::plan::scalar::LirScalarExpr;
use crate::plan::{AvailableCollections, bucketing_of_expected_group_size};

/// This enum represents the three potential types of aggregations.
#[derive(
    Copy,
    Clone,
    Debug,
    Deserialize,
    Eq,
    Hash,
    Ord,
    PartialEq,
    PartialOrd,
    Serialize
)]
pub enum ReductionType {
    /// Accumulable functions can be subtracted from (are invertible), and associative.
    /// We can compute these results by moving some data to the diff field under arbitrary
    /// changes to inputs. Examples include sum or count.
    Accumulable,
    /// Hierarchical functions are associative, which means we can split up the work of
    /// computing them across subsets. Note that hierarchical reductions should also
    /// reduce the data in some way, as otherwise rendering them hierarchically is not
    /// worth it. Examples include min or max.
    Hierarchical,
    /// Basic, for lack of a better word, are functions that are neither accumulable
    /// nor hierarchical. Examples include jsonb_agg.
    Basic,
}

impl TryFrom<&ReducePlan> for ReductionType {
    type Error = ();

    fn try_from(plan: &ReducePlan) -> Result<Self, Self::Error> {
        match plan {
            ReducePlan::Hierarchical(_) => Ok(ReductionType::Hierarchical),
            ReducePlan::Accumulable(_) => Ok(ReductionType::Accumulable),
            ReducePlan::Basic(_) => Ok(ReductionType::Basic),
            ReducePlan::Distinct => Err(()),
        }
    }
}

/// A `ReducePlan` provides a concise description for how we will
/// execute a given reduce expression.
///
/// The provided reduce expression can have no
/// aggregations, in which case its just a `Distinct` and otherwise
/// it's composed of a combination of accumulable, hierarchical and
/// basic aggregations.
///
/// We want to try to centralize as much decision making about the
/// shape / general computation of the rendered dataflow graph
/// in this plan, and then make actually rendering the graph
/// be as simple (and compiler verifiable) as possible.
#[derive(Clone, Debug, Serialize, Deserialize, Eq, PartialEq, Ord, PartialOrd)]
pub enum ReducePlan {
    /// Plan for not computing any aggregations, just determining the set of
    /// distinct keys.
    Distinct,
    /// Plan for computing only accumulable aggregations.
    Accumulable(AccumulablePlan),
    /// Plan for computing only hierarchical aggregations.
    Hierarchical(HierarchicalPlan),
    /// Plan for computing only basic aggregations.
    Basic(BasicPlan),
}

/// All reduce plans depend on a notion of aggregation.
///
/// We could use `mz_expr::AggregateExpr`, but it explicitly names
/// `MirScalarExpr`, while LIR aggregates contain `LirScalarExpr`. This
/// duplication exists because the old MzReflect test framework could not
/// accommodate type parameters. TODO: with that framework gone, this will
/// be resolved shortly by parameterizing over the scalar expression type.
///
/// We don't build a separate `AggregateFunc`, since we'd only eliminate one variant
/// and need to duplicate the evaluation code.
#[derive(
    Clone,
    Debug,
    Eq,
    PartialEq,
    Ord,
    PartialOrd,
    Serialize,
    Deserialize,
    Hash
)]
pub struct LirAggregateExpr {
    /// Names the aggregation function.
    pub func: AggregateFunc,
    /// An expression which extracts from each row the input to `func`.
    pub expr: LirScalarExpr,
    /// Should the aggregation be applied only to distinct results in each group.
    #[serde(default)]
    pub distinct: bool,
}

impl LirAggregateExpr {
    /// Translates an aggregate from MIR to LIR.
    ///
    /// Panics on unmaterializable functions.
    pub fn from_mir(mir: AggregateExpr) -> Self {
        Self::try_from(mir).expect("no unmaterializable functions in aggregates")
    }

    /// Determines whether this aggregate is `COUNT(*)`.
    pub fn is_count_asterisk(&self) -> bool {
        self.func == AggregateFunc::Count && self.expr.is_literal_true() && !self.distinct
    }
}

impl TryFrom<AggregateExpr> for LirAggregateExpr {
    type Error = Vec<UnmaterializableFunc>;

    fn try_from(mir: AggregateExpr) -> Result<Self, Self::Error> {
        let func = mir.func;
        let expr = LirScalarExpr::try_from(&mir.expr)?;
        let distinct = mir.distinct;

        Ok(LirAggregateExpr {
            func,
            expr,
            distinct,
        })
    }
}

impl HumanizeDisplay for LirAggregateExpr {
    fn humanize<'a, M: HumanizerMode>(
        e: &HumanizedExpr<'a, Self, M>,
        f: &mut std::fmt::Formatter<'_>,
    ) -> std::fmt::Result {
        if e.expr.is_count_asterisk() {
            return write!(f, "count(*)");
        }

        write!(
            f,
            "{}({}",
            e.child(&e.expr.func),
            if e.expr.distinct { "distinct " } else { "" }
        )?;

        e.child(&e.expr.expr).fmt(f)?;
        write!(f, ")")
    }
}

/// Plan for computing a set of accumulable aggregations.
///
/// We fuse all of the accumulable aggregations together
/// and compute them with one dataflow fragment. We need to
/// be careful to separate out the aggregations that
/// apply only to the distinct set of values. We need
/// to apply a distinct operator to those before we
/// combine them with everything else.
#[derive(Clone, Debug, Serialize, Deserialize, Eq, PartialEq, Ord, PartialOrd)]
pub struct AccumulablePlan {
    /// All of the aggregations we were asked to compute, stored
    /// in order.
    pub full_aggrs: Vec<LirAggregateExpr>,
    /// All of the non-distinct accumulable aggregates.
    /// Each element represents:
    /// (index of the datum among inputs, aggregation expr)
    /// These will all be rendered together in one dataflow fragment.
    pub simple_aggrs: Vec<(usize, LirAggregateExpr)>,
    /// Same as above but for all of the `DISTINCT` accumulable aggregations.
    pub distinct_aggrs: Vec<(usize, LirAggregateExpr)>,
}

/// Plan for computing a set of hierarchical aggregations.
///
/// In the append-only setting we can render them in-place
/// with monotonic plans, but otherwise, we need to render
/// them with a reduction tree that splits the inputs into
/// small, and then progressively larger, buckets
#[derive(Clone, Debug, Serialize, Deserialize, Eq, PartialEq, Ord, PartialOrd)]
pub enum HierarchicalPlan {
    /// Plan hierarchical aggregations under monotonic inputs.
    Monotonic(MonotonicPlan),
    /// Plan for hierarchical aggregations under non-monotonic inputs.
    Bucketed(BucketedPlan),
}

impl HierarchicalPlan {
    /// Returns the set of aggregations computed by this plan.
    pub fn aggr_funcs(&self) -> &[AggregateFunc] {
        match self {
            HierarchicalPlan::Monotonic(plan) => &plan.aggr_funcs,
            HierarchicalPlan::Bucketed(plan) => &plan.aggr_funcs,
        }
    }

    /// Upgrades from a bucketed plan to a monotonic plan, if necessary,
    /// and sets consolidation requirements.
    pub fn as_monotonic(&mut self, must_consolidate: bool) {
        match self {
            HierarchicalPlan::Bucketed(bucketed) => {
                // TODO: ideally we would not have the `clone()` but ownership
                // seems fraught here as we are behind a `&mut self` reference.
                *self =
                    HierarchicalPlan::Monotonic(bucketed.clone().into_monotonic(must_consolidate));
            }
            HierarchicalPlan::Monotonic(monotonic) => {
                monotonic.must_consolidate = must_consolidate;
            }
        }
    }
}

/// Plan for computing a set of hierarchical aggregations with a
/// monotonic input.
///
/// Here, the aggregations will be rendered in place. We don't
/// need to worry about retractions because the inputs are
/// append only, so we can change our computation to
/// only retain the "best" value in the diff field, instead
/// of holding onto all values.
#[derive(Clone, Debug, Serialize, Deserialize, Eq, PartialEq, Ord, PartialOrd)]
pub struct MonotonicPlan {
    /// All of the aggregations we were asked to compute.
    pub aggr_funcs: Vec<AggregateFunc>,
    /// True if the input is not physically monotonic, and the operator must perform
    /// consolidation to remove potential negations. The operator implementation is
    /// free to consolidate as late as possible while ensuring correctness, so it is
    /// not a requirement that the input be directly subjected to consolidation.
    /// More details in the monotonic one-shot `SELECT`s design doc.[^1]
    ///
    /// [^1]: <https://github.com/MaterializeInc/materialize/blob/main/doc/developer/design/20230421_stabilize_monotonic_select.md>
    pub must_consolidate: bool,
}

/// Plan for computing a set of hierarchical aggregations
/// with non-monotonic inputs.
///
/// To perform hierarchical aggregations with stable runtimes
/// under updates we'll subdivide the group key into buckets, compute
/// the reduction in each of those subdivided buckets and then combine
/// the results into a coarser bucket (one that represents a larger
/// fraction of the original input) and redo the reduction in another
/// layer. Effectively, we'll construct a min / max heap out of a series
/// of reduce operators (each one is a separate layer).
#[derive(Clone, Debug, Serialize, Deserialize, Eq, PartialEq, Ord, PartialOrd)]
pub struct BucketedPlan {
    /// All of the aggregations we were asked to compute.
    pub aggr_funcs: Vec<AggregateFunc>,
    /// The number of buckets in each layer of the reduction tree. Should
    /// be decreasing, and ideally, a power of two so that we can easily
    /// distribute values to buckets with `value.hashed() % buckets[layer]`.
    pub buckets: Vec<u64>,
}

impl BucketedPlan {
    /// Convert to a monotonic plan, indicate whether the operator must apply
    /// consolidation to its input.
    fn into_monotonic(self, must_consolidate: bool) -> MonotonicPlan {
        MonotonicPlan {
            aggr_funcs: self.aggr_funcs,
            must_consolidate,
        }
    }
}

/// Plan for computing a set of basic aggregations.
///
/// There's much less complexity when rendering basic aggregations.
/// Each aggregation corresponds to one Differential reduce operator.
/// That's it. However, we still want to present one final arrangement
/// so basic aggregations present results with the same interface
/// (one arrangement containing a row with all results) that accumulable
/// and hierarchical aggregations do. To provide that, we render an
/// additional reduce operator whenever we have multiple reduce aggregates
/// to combine and present results in the appropriate order. If we
/// were only asked to compute a single aggregation, we can skip
/// that step and return the arrangement provided by computing the aggregation
/// directly.
#[derive(Clone, Debug, Serialize, Deserialize, Eq, PartialEq, Ord, PartialOrd)]
pub enum BasicPlan {
    /// Plan for rendering a single basic aggregation.
    Single(SingleBasicPlan),
    /// Plan for rendering multiple basic aggregations.
    /// These need to then be collated together in an additional
    /// reduction. Each element represents the:
    /// `(index of the set of the input we are aggregating over,
    ///   the aggregation function)`
    Multiple(Vec<LirAggregateExpr>),
}

/// Plan for rendering a single basic aggregation, with possibly fusing a `FlatMap UnnestList` with
/// this aggregation.
#[derive(Clone, Debug, Serialize, Deserialize, Eq, PartialEq, Ord, PartialOrd)]
pub struct SingleBasicPlan {
    /// The aggregation that we should perform.
    pub expr: LirAggregateExpr,
    /// Whether we fused a `FlatMap UnnestList` with this aggregation.
    pub fused_unnest_list: bool,
}

/// Plan for collating the results of computing multiple aggregation
/// types.
///
/// TODO: could we express this as a delta join
#[derive(
    Clone,
    Debug,
    Default,
    Serialize,
    Deserialize,
    Eq,
    PartialEq,
    Ord,
    PartialOrd
)]
pub struct CollationPlan {
    /// Accumulable aggregation results to collate, if any.
    pub accumulable: Option<AccumulablePlan>,
    /// Hierarchical aggregation results to collate, if any.
    pub hierarchical: Option<HierarchicalPlan>,
    /// Basic aggregation results to collate, if any.
    pub basic: Option<BasicPlan>,
    /// When we get results back from each of the different
    /// aggregation types, they will be subsequences of
    /// the sequence aggregations in the original reduce expression.
    /// We keep a map from output position -> reduction type
    /// to easily merge results back into the requested order.
    pub aggregate_types: Vec<ReductionType>,
}

impl CollationPlan {
    /// Upgrades the hierarchical component of the collation plan to monotonic, if necessary,
    /// and sets consolidation requirements.
    pub fn as_monotonic(&mut self, must_consolidate: bool) {
        self.hierarchical
            .as_mut()
            .map(|plan| plan.as_monotonic(must_consolidate));
    }
}

impl ReducePlan {
    /// Generate a plan for computing the supplied aggregations.
    ///
    /// The resulting plan summarizes what the dataflow to be created
    /// and how the aggregations will be executed.
    pub fn create_from(
        aggregates: Vec<AggregateExpr>,
        monotonic: bool,
        expected_group_size: Option<u64>,
        fused_unnest_list: bool,
    ) -> Self {
        // We need to make sure that all aggregates have the same type.
        let mut aggregates_list = Vec::with_capacity(aggregates.len());
        let mut aggregates = aggregates.into_iter();
        if let Some(aggregate) = aggregates.next() {
            let typ = reduction_type(&aggregate.func);
            aggregates_list.push(LirAggregateExpr::from_mir(aggregate));

            for aggregate in aggregates {
                assert_eq!(
                    typ,
                    reduction_type(&aggregate.func),
                    "Multiple reduction types detected"
                );
                aggregates_list.push(LirAggregateExpr::from_mir(aggregate));
            }
            ReducePlan::create_inner(
                typ,
                aggregates_list,
                monotonic,
                expected_group_size,
                fused_unnest_list,
            )
        } else {
            // If we don't have any aggregations we are just computing a distinct.
            ReducePlan::Distinct
        }
    }

    /// Generate a plan for computing the specified type of aggregations.
    ///
    /// This function assumes that all of the supplied aggregates are
    /// actually of the correct reduction type.
    fn create_inner(
        typ: ReductionType,
        aggregates_list: Vec<LirAggregateExpr>,
        monotonic: bool,
        expected_group_size: Option<u64>,
        fused_unnest_list: bool,
    ) -> Self {
        if fused_unnest_list {
            assert!(matches!(typ, ReductionType::Basic) && aggregates_list.len() == 1);
        }
        assert!(
            aggregates_list.len() > 0,
            "error: tried to render a reduce dataflow with no aggregates"
        );

        match typ {
            ReductionType::Accumulable => {
                let mut simple_aggrs = vec![];
                let mut distinct_aggrs = vec![];
                let full_aggrs = aggregates_list.clone();
                for (datum_index, aggr) in aggregates_list.into_iter().enumerate() {
                    // Accumulable aggregations need to do extra per-aggregate work
                    // for aggregations with the distinct bit set, so we'll separate
                    // those out now.
                    if aggr.distinct {
                        distinct_aggrs.push((datum_index, aggr));
                    } else {
                        simple_aggrs.push((datum_index, aggr));
                    };
                }
                ReducePlan::Accumulable(AccumulablePlan {
                    full_aggrs,
                    simple_aggrs,
                    distinct_aggrs,
                })
            }
            ReductionType::Hierarchical => {
                let aggr_funcs = aggregates_list
                    .iter()
                    .map(|aggr| aggr.func.clone())
                    .collect();

                if monotonic {
                    let monotonic = MonotonicPlan {
                        aggr_funcs,
                        must_consolidate: false,
                    };
                    ReducePlan::Hierarchical(HierarchicalPlan::Monotonic(monotonic))
                } else {
                    let buckets = bucketing_of_expected_group_size(expected_group_size);
                    let bucketed = BucketedPlan {
                        aggr_funcs,
                        buckets,
                    };

                    ReducePlan::Hierarchical(HierarchicalPlan::Bucketed(bucketed))
                }
            }
            ReductionType::Basic => match <_ as TryInto<[_; 1]>>::try_into(aggregates_list) {
                Ok([expr]) => ReducePlan::Basic(BasicPlan::Single(SingleBasicPlan {
                    expr,
                    fused_unnest_list,
                })),
                Err(aggregates_list) => ReducePlan::Basic(BasicPlan::Multiple(aggregates_list)),
            },
        }
    }

    /// Reports all keys of produced arrangements.
    ///
    /// This is likely either an empty vector, for no arrangement,
    /// or a singleton vector containing the list of expressions
    /// that key a single arrangement.
    pub fn keys(&self, key_arity: usize, arity: usize) -> AvailableCollections {
        let key = (0..key_arity)
            .map(LirScalarExpr::column)
            .collect::<Vec<_>>();
        let (permutation, thinning) = permutation_for_arrangement(&key, arity);
        AvailableCollections::new_arranged(vec![(key, permutation, thinning)])
    }

    /// Extracts a fusable MFP for the reduction from the given `mfp` along with a residual
    /// non-fusable MFP and potentially revised output arity. The provided `mfp` must be the
    /// one sitting on top of the reduction.
    ///
    /// Non-fusable parts include temporal predicates or any other parts that cannot be
    /// conservatively asserted to not increase the memory requirements of the output
    /// arrangement for the reduction. Either the fusable or non-fusable parts may end up
    /// being the identity MFP.
    pub fn extract_mfp_after(
        &self,
        mut mfp: MapFilterProject,
        key_arity: usize,
    ) -> (MapFilterProject, MapFilterProject, usize) {
        // Extract temporal predicates, as we cannot push them into `Reduce`.
        let temporal_mfp = mfp.extract_temporal();
        let non_temporal = mfp;
        mfp = temporal_mfp;

        // We ensure we do not attempt to project away the key, as we cannot accomplish
        // this. This is done by a simple analysis of the non-temporal part of `mfp` to
        // check if can be directly absorbed; if it can't, we then default to a general
        // strategy that unpacks the MFP to absorb only the filter and supporting map
        // parts, followed by a post-MFP step.
        let input_arity = non_temporal.input_arity;
        let key = Vec::from_iter(0..key_arity);
        let mut mfp_push;
        let output_arity;

        if non_temporal.projection.len() <= input_arity
            && non_temporal.projection.iter().all(|c| *c < input_arity)
            && non_temporal.projection.starts_with(&key)
        {
            // Special case: The key is preserved as a prefix and the projection is only
            // of output fields from the reduction. So we know that: (a) We can process the
            // fused MFP per-key; (b) The MFP application gets rid of all mapped columns;
            // and (c) The output projection is at most as wide as the output that would be
            // produced by the reduction, so we are sure to never regress the memory
            // requirements of the output arrangement.
            // Note that this strategy may change the arity of the output arrangement.
            output_arity = non_temporal.projection.len();
            mfp_push = non_temporal;
        } else {
            // General strategy: Unpack MFP as MF followed by P' that removes all M
            // columns, then MP afterwards.
            // Note that this strategy does not result in any changes to the arity of
            // the output arrangement.
            let (m, f, p) = non_temporal.into_map_filter_project();
            mfp_push = MapFilterProject::new(input_arity)
                .map(m.clone())
                .filter(f)
                .project(0..input_arity);
            output_arity = input_arity;

            // We still need to perform the map and projection for the actual output.
            let mfp_left = MapFilterProject::new(input_arity).map(m).project(p);

            // Compose the non-pushed MFP components.
            mfp = MapFilterProject::compose(mfp_left, mfp);
        }
        mfp_push.optimize();
        mfp.optimize();
        (mfp_push, mfp, output_arity)
    }
}

/// Plan for extracting keys and values in preparation for a reduction.
#[derive(Clone, Debug, Serialize, Deserialize, Eq, PartialEq, Ord, PartialOrd)]
pub struct KeyValPlan {
    /// Extracts the columns used as the key.
    pub key_plan: mz_expr::SafeMfpPlan<LirScalarExpr>,
    /// Extracts the columns used to feed the aggregations.
    pub val_plan: mz_expr::SafeMfpPlan<LirScalarExpr>,
}

impl KeyValPlan {
    /// Create a new [KeyValPlan] from aggregation arguments.
    pub fn new(
        input_arity: usize,
        group_key: &[MirScalarExpr],
        aggregates: &[AggregateExpr],
        input_permutation_and_new_arity: Option<(Vec<usize>, usize)>,
    ) -> Self {
        // Form an operator for evaluating key expressions.
        let mut key_mfp = MapFilterProject::new(input_arity)
            .map(group_key.iter().cloned())
            .project(input_arity..(input_arity + group_key.len()));
        if let Some((input_permutation, new_arity)) = input_permutation_and_new_arity.clone() {
            key_mfp.permute_fn(|c| input_permutation[c], new_arity);
        }

        // Form an operator for evaluating value expressions.
        let mut val_mfp = MapFilterProject::new(input_arity)
            .map(aggregates.iter().map(|a| a.expr.clone()))
            .project(input_arity..(input_arity + aggregates.len()));
        if let Some((input_permutation, new_arity)) = input_permutation_and_new_arity {
            val_mfp.permute_fn(|c| input_permutation[c], new_arity);
        }

        key_mfp.optimize();
        let key_plan = crate::plan::scalar::safe_mfp_mir_to_lir(
            key_mfp.into_plan().unwrap().into_nontemporal().unwrap(),
        );
        val_mfp.optimize();
        let val_plan = crate::plan::scalar::safe_mfp_mir_to_lir(
            val_mfp.into_plan().unwrap().into_nontemporal().unwrap(),
        );

        Self { key_plan, val_plan }
    }

    /// The arity of the key plan
    pub fn key_arity(&self) -> usize {
        self.key_plan.projection.len()
    }
}

/// Transforms a vector containing indexes of needed columns into one containing
/// the "skips" an iterator over a Row would need to perform to see those values.
///
/// This function requires that all of the elements in `indexes` are strictly
/// increasing.
///
/// # Examples
///
/// ```
/// use mz_compute_types::plan::reduce::convert_indexes_to_skips;
/// assert_eq!(convert_indexes_to_skips(vec![3, 6, 10, 15]), [3, 2, 3, 4])
/// ```
pub fn convert_indexes_to_skips(mut indexes: Vec<usize>) -> Vec<usize> {
    for i in 1..indexes.len() {
        soft_assert_or_log!(
            indexes[i - 1] < indexes[i],
            "convert_indexes_to_skip needs indexes to be strictly increasing. Received: {:?}",
            indexes,
        );
    }

    for i in (1..indexes.len()).rev() {
        indexes[i] -= indexes[i - 1];
        indexes[i] -= 1;
    }

    indexes
}

/// Determines whether a function can be accumulated in an update's "difference" field,
/// and whether it can be subjected to recursive (hierarchical) aggregation.
///
/// Accumulable aggregations will be packed into differential dataflow's "difference" field,
/// which can be accumulated in-place using the addition operation on the type. Aggregations
/// that indicate they are accumulable will still need to provide an action that takes their
/// data and introduces it as a difference, and the post-processing when the accumulated value
/// is presented as data.
///
/// Hierarchical aggregations will be subjected to repeated aggregation on initially small but
/// increasingly large subsets of each key. This has the intended property that no invocation
/// is on a significantly large set of values (and so, no incremental update needs to reform
/// significant input data). Hierarchical aggregates can be rendered more efficiently if the
/// input stream is append-only as then we only need to retain the "currently winning" value.
/// Every hierarchical aggregate needs to supply a corresponding ReductionMonoid implementation.
pub fn reduction_type(func: &AggregateFunc) -> ReductionType {
    match func {
        AggregateFunc::SumInt16
        | AggregateFunc::SumInt32
        | AggregateFunc::SumInt64
        | AggregateFunc::SumUInt16
        | AggregateFunc::SumUInt32
        | AggregateFunc::SumUInt64
        | AggregateFunc::SumFloat32
        | AggregateFunc::SumFloat64
        | AggregateFunc::SumNumeric
        | AggregateFunc::Count
        | AggregateFunc::Any
        | AggregateFunc::All
        | AggregateFunc::Dummy => ReductionType::Accumulable,
        AggregateFunc::MaxNumeric
        | AggregateFunc::MaxInt16
        | AggregateFunc::MaxInt32
        | AggregateFunc::MaxInt64
        | AggregateFunc::MaxUInt16
        | AggregateFunc::MaxUInt32
        | AggregateFunc::MaxUInt64
        | AggregateFunc::MaxMzTimestamp
        | AggregateFunc::MaxFloat32
        | AggregateFunc::MaxFloat64
        | AggregateFunc::MaxBool
        | AggregateFunc::MaxString
        | AggregateFunc::MaxDate
        | AggregateFunc::MaxTimestamp
        | AggregateFunc::MaxTimestampTz
        | AggregateFunc::MaxInterval
        | AggregateFunc::MaxTime
        | AggregateFunc::MinNumeric
        | AggregateFunc::MinInt16
        | AggregateFunc::MinInt32
        | AggregateFunc::MinInt64
        | AggregateFunc::MinUInt16
        | AggregateFunc::MinUInt32
        | AggregateFunc::MinUInt64
        | AggregateFunc::MinMzTimestamp
        | AggregateFunc::MinInterval
        | AggregateFunc::MinFloat32
        | AggregateFunc::MinFloat64
        | AggregateFunc::MinBool
        | AggregateFunc::MinString
        | AggregateFunc::MinDate
        | AggregateFunc::MinTimestamp
        | AggregateFunc::MinTimestampTz
        | AggregateFunc::MinTime => ReductionType::Hierarchical,
        AggregateFunc::JsonbAgg { .. }
        | AggregateFunc::JsonbObjectAgg { .. }
        | AggregateFunc::MapAgg { .. }
        | AggregateFunc::ArrayConcat { .. }
        | AggregateFunc::ListConcat { .. }
        | AggregateFunc::StringAgg { .. }
        | AggregateFunc::RowNumber { .. }
        | AggregateFunc::Rank { .. }
        | AggregateFunc::DenseRank { .. }
        | AggregateFunc::LagLead { .. }
        | AggregateFunc::FirstValue { .. }
        | AggregateFunc::LastValue { .. }
        | AggregateFunc::WindowAggregate { .. }
        | AggregateFunc::FusedValueWindowFunc { .. }
        | AggregateFunc::FusedWindowAggregate { .. } => ReductionType::Basic,
    }
}