mz_compute/

metrics.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.

use std::sync::{Arc, Mutex};

use mz_compute_client::metrics::{CommandMetrics, HistoryMetrics};
use mz_ore::cast::CastFrom;
use mz_ore::metric;
use mz_ore::metrics::{MetricsRegistry, UIntGauge, raw};
use mz_repr::{GlobalId, SharedRow};
use prometheus::core::{AtomicF64, GenericCounter};
use prometheus::proto::LabelPair;
use prometheus::{Histogram, HistogramVec};

/// Metrics exposed by compute replicas.
//
// Most of the metrics here use the `raw` implementations, rather than the `DeleteOnDrop` wrappers
// because their labels are fixed throughout the lifetime of the replica process. For example, any
// metric labeled only by `worker_id` can be `raw` since the number of workers cannot change.
//
// Metrics that are labelled by a dimension that can change throughout the lifetime of the process
// (such as `collection_id`) MUST NOT use the `raw` metric types and must use the `DeleteOnDrop`
// types instead, to avoid memory leaks.
#[derive(Clone, Debug)]
pub struct ComputeMetrics {
    // Optional workload class label to apply to all metrics in registry.
    workload_class: Arc<Mutex<Option<String>>>,

    // command history
    history_command_count: raw::UIntGaugeVec,
    history_dataflow_count: raw::UIntGaugeVec,

    // reconciliation
    reconciliation_reused_dataflows_count_total: raw::IntCounterVec,
    reconciliation_replaced_dataflows_count_total: raw::IntCounterVec,

    // arrangements
    arrangement_maintenance_seconds_total: raw::CounterVec,
    arrangement_maintenance_active_info: raw::UIntGaugeVec,

    // timings
    //
    // Note that this particular metric unfortunately takes some care to
    // interpret. It measures the duration of step_or_park calls, which
    // undesirably includes the parking. This is probably fine because we
    // regularly send progress information through persist sources, which likely
    // means the parking is capped at a second or two in practice. It also
    // doesn't do anything to let you pinpoint _which_ operator or worker isn't
    // yielding, but it should hopefully alert us when there is something to
    // look at.
    timely_step_duration_seconds: HistogramVec,
    persist_peek_seconds: HistogramVec,
    stashed_peek_seconds: HistogramVec,
    handle_command_duration_seconds: HistogramVec,

    // memory usage
    shared_row_heap_capacity_bytes: raw::UIntGaugeVec,

    // replica expiration
    replica_expiration_timestamp_seconds: raw::UIntGaugeVec,
    replica_expiration_remaining_seconds: raw::GaugeVec,

    // collections
    collection_count: raw::UIntGaugeVec,
}

impl ComputeMetrics {
    pub fn register_with(registry: &MetricsRegistry) -> Self {
        let workload_class = Arc::new(Mutex::new(None));

        // Apply a `workload_class` label to all metrics in the registry when we
        // have a known workload class.
        registry.register_postprocessor({
            let workload_class = Arc::clone(&workload_class);
            move |metrics| {
                let workload_class: Option<String> =
                    workload_class.lock().expect("lock poisoned").clone();
                let Some(workload_class) = workload_class else {
                    return;
                };
                for metric in metrics {
                    for metric in metric.mut_metric() {
                        let mut label = LabelPair::default();
                        label.set_name("workload_class".into());
                        label.set_value(workload_class.clone());

                        let mut labels = metric.take_label();
                        labels.push(label);
                        metric.set_label(labels);
                    }
                }
            }
        });

        Self {
            workload_class,
            history_command_count: registry.register(metric!(
                name: "mz_compute_replica_history_command_count",
                help: "The number of commands in the replica's command history.",
                var_labels: ["worker_id", "command_type"],
            )),
            history_dataflow_count: registry.register(metric!(
                name: "mz_compute_replica_history_dataflow_count",
                help: "The number of dataflows in the replica's command history.",
                var_labels: ["worker_id"],
            )),
            reconciliation_reused_dataflows_count_total: registry.register(metric!(
                name: "mz_compute_reconciliation_reused_dataflows_count_total",
                help: "The total number of dataflows that were reused during compute reconciliation.",
                var_labels: ["worker_id"],
            )),
            reconciliation_replaced_dataflows_count_total: registry.register(metric!(
                name: "mz_compute_reconciliation_replaced_dataflows_count_total",
                help: "The total number of dataflows that were replaced during compute reconciliation.",
                var_labels: ["worker_id", "reason"],
            )),
            arrangement_maintenance_seconds_total: registry.register(metric!(
                name: "mz_arrangement_maintenance_seconds_total",
                help: "The total time spent maintaining arrangements.",
                var_labels: ["worker_id"],
            )),
            arrangement_maintenance_active_info: registry.register(metric!(
                name: "mz_arrangement_maintenance_active_info",
                help: "Whether maintenance is currently occuring.",
                var_labels: ["worker_id"],
            )),
            timely_step_duration_seconds: registry.register(metric!(
                name: "mz_timely_step_duration_seconds",
                help: "The time spent in each compute step_or_park call",
                const_labels: {"cluster" => "compute"},
                var_labels: ["worker_id"],
                buckets: mz_ore::stats::histogram_seconds_buckets(0.000_128, 32.0),
            )),
            shared_row_heap_capacity_bytes: registry.register(metric!(
                name: "mz_dataflow_shared_row_heap_capacity_bytes",
                help: "The heap capacity of the shared row.",
                var_labels: ["worker_id"],
            )),
            persist_peek_seconds: registry.register(metric!(
                name: "mz_persist_peek_seconds",
                help: "Time spent in (experimental) Persist fast-path peeks.",
                var_labels: ["worker_id"],
                buckets: mz_ore::stats::histogram_seconds_buckets(0.000_128, 8.0),
            )),
            stashed_peek_seconds: registry.register(metric!(
                name: "mz_stashed_peek_seconds",
                help: "Time spent reading a peek result and stashing it in the peek result stash (aka. persist blob).",
                var_labels: ["worker_id"],
                buckets: mz_ore::stats::histogram_seconds_buckets(0.000_128, 8.0),
            )),
            handle_command_duration_seconds: registry.register(metric!(
                name: "mz_cluster_handle_command_duration_seconds",
                help: "Time spent in handling commands.",
                const_labels: {"cluster" => "compute"},
                var_labels: ["worker_id", "command_type"],
                buckets: mz_ore::stats::histogram_seconds_buckets(0.000_128, 8.0),
            )),
            replica_expiration_timestamp_seconds: registry.register(metric!(
                name: "mz_dataflow_replica_expiration_timestamp_seconds",
                help: "The replica expiration timestamp in seconds since epoch.",
                var_labels: ["worker_id"],
            )),
            replica_expiration_remaining_seconds: registry.register(metric!(
                name: "mz_dataflow_replica_expiration_remaining_seconds",
                help: "The remaining seconds until replica expiration. Can go negative, can lag behind.",
                var_labels: ["worker_id"],
            )),
            collection_count: registry.register(metric!(
                name: "mz_compute_collection_count",
                help: "The number and hydration status of maintained compute collections.",
                var_labels: ["worker_id", "type", "hydrated"],
            )),
        }
    }

    /// Sets the workload class for the compute metrics.
    pub fn set_workload_class(&self, workload_class: Option<String>) {
        let mut guard = self.workload_class.lock().expect("lock poisoned");
        *guard = workload_class
    }

    pub fn for_worker(&self, worker_id: usize) -> WorkerMetrics {
        let worker = worker_id.to_string();
        let arrangement_maintenance_seconds_total = self
            .arrangement_maintenance_seconds_total
            .with_label_values(&[&worker]);
        let arrangement_maintenance_active_info = self
            .arrangement_maintenance_active_info
            .with_label_values(&[&worker]);
        let timely_step_duration_seconds = self
            .timely_step_duration_seconds
            .with_label_values(&[&worker]);
        let persist_peek_seconds = self.persist_peek_seconds.with_label_values(&[&worker]);
        let stashed_peek_seconds = self.stashed_peek_seconds.with_label_values(&[&worker]);
        let handle_command_duration_seconds = CommandMetrics::build(|typ| {
            self.handle_command_duration_seconds
                .with_label_values(&[&worker, typ])
        });
        let replica_expiration_timestamp_seconds = self
            .replica_expiration_timestamp_seconds
            .with_label_values(&[&worker]);
        let replica_expiration_remaining_seconds = self
            .replica_expiration_remaining_seconds
            .with_label_values(&[&worker]);
        let shared_row_heap_capacity_bytes = self
            .shared_row_heap_capacity_bytes
            .with_label_values(&[&worker]);

        WorkerMetrics {
            worker_label: worker,
            metrics: self.clone(),
            arrangement_maintenance_seconds_total,
            arrangement_maintenance_active_info,
            timely_step_duration_seconds,
            persist_peek_seconds,
            stashed_peek_seconds,
            handle_command_duration_seconds,
            replica_expiration_timestamp_seconds,
            replica_expiration_remaining_seconds,
            shared_row_heap_capacity_bytes,
        }
    }
}

/// Per-worker metrics.
#[derive(Clone, Debug)]
pub struct WorkerMetrics {
    worker_label: String,
    metrics: ComputeMetrics,

    /// The amount of time spent in arrangement maintenance.
    pub(crate) arrangement_maintenance_seconds_total: GenericCounter<AtomicF64>,
    /// 1 if this worker is currently doing maintenance.
    ///
    /// If maintenance turns out to take a very long time, this will allow us
    /// to gain a sense that Materialize is stuck on maintenance before the
    /// maintenance completes
    pub(crate) arrangement_maintenance_active_info: UIntGauge,
    /// Histogram of Timely step timings.
    pub(crate) timely_step_duration_seconds: Histogram,
    /// Histogram of persist peek durations.
    pub(crate) persist_peek_seconds: Histogram,
    /// Histogram of stashed peek durations.
    pub(crate) stashed_peek_seconds: Histogram,
    /// Histogram of command handling durations.
    pub(crate) handle_command_duration_seconds: CommandMetrics<Histogram>,
    /// The timestamp of replica expiration.
    pub(crate) replica_expiration_timestamp_seconds: UIntGauge,
    /// Remaining seconds until replica expiration.
    pub(crate) replica_expiration_remaining_seconds: raw::Gauge,
    /// Heap capacity of the shared row.
    shared_row_heap_capacity_bytes: UIntGauge,
}

impl WorkerMetrics {
    pub fn for_history(&self) -> HistoryMetrics<UIntGauge> {
        let command_counts = CommandMetrics::build(|typ| {
            self.metrics
                .history_command_count
                .with_label_values(&[&self.worker_label, typ])
        });
        let dataflow_count = self
            .metrics
            .history_dataflow_count
            .with_label_values(&[&self.worker_label]);

        HistoryMetrics {
            command_counts,
            dataflow_count,
        }
    }

    /// Record the reconciliation result for a single dataflow.
    ///
    /// Reconciliation is recorded as successful if the given properties all hold. Otherwise it is
    /// recorded as unsuccessful, with a reason based on the first property that does not hold.
    ///
    /// The properties are:
    ///  * compatible: The old and new dataflow descriptions are compatible.
    ///  * uncompacted: Collections currently installed for the dataflow exports have not been
    ///                 allowed to compact beyond that new dataflow as-of.
    ///  * subscribe_free: The dataflow does not export a subscribe sink.
    ///  * copy_to_free: The dataflow does not export a copy-to sink.
    ///  * dependencies_retained: All local inputs to the dataflow were retained by compute
    ///                           reconciliation.
    pub fn record_dataflow_reconciliation(
        &self,
        compatible: bool,
        uncompacted: bool,
        subscribe_free: bool,
        copy_to_free: bool,
        dependencies_retained: bool,
    ) {
        if !compatible {
            self.metrics
                .reconciliation_replaced_dataflows_count_total
                .with_label_values(&[&self.worker_label, "incompatible"])
                .inc();
        } else if !uncompacted {
            self.metrics
                .reconciliation_replaced_dataflows_count_total
                .with_label_values(&[&self.worker_label, "compacted"])
                .inc();
        } else if !subscribe_free {
            self.metrics
                .reconciliation_replaced_dataflows_count_total
                .with_label_values(&[&self.worker_label, "subscribe"])
                .inc();
        } else if !copy_to_free {
            self.metrics
                .reconciliation_replaced_dataflows_count_total
                .with_label_values(&[&self.worker_label, "copy-to"])
                .inc();
        } else if !dependencies_retained {
            self.metrics
                .reconciliation_replaced_dataflows_count_total
                .with_label_values(&[&self.worker_label, "dependency"])
                .inc();
        } else {
            self.metrics
                .reconciliation_reused_dataflows_count_total
                .with_label_values(&[&self.worker_label])
                .inc();
        }
    }

    /// Record the heap capacity of the shared row.
    pub fn record_shared_row_metrics(&self) {
        let binding = SharedRow::get();
        self.shared_row_heap_capacity_bytes
            .set(u64::cast_from(binding.byte_capacity()));
    }

    /// Increase the count of maintained collections.
    fn inc_collection_count(&self, collection_type: &str, hydrated: bool) {
        let hydrated = if hydrated { "1" } else { "0" };
        self.metrics
            .collection_count
            .with_label_values(&[&self.worker_label, collection_type, hydrated])
            .inc();
    }

    /// Decrease the count of maintained collections.
    fn dec_collection_count(&self, collection_type: &str, hydrated: bool) {
        let hydrated = if hydrated { "1" } else { "0" };
        self.metrics
            .collection_count
            .with_label_values(&[&self.worker_label, collection_type, hydrated])
            .dec();
    }

    /// Sets the workload class for the compute metrics.
    pub fn set_workload_class(&self, workload_class: Option<String>) {
        self.metrics.set_workload_class(workload_class);
    }

    pub fn for_collection(&self, id: GlobalId) -> CollectionMetrics {
        CollectionMetrics::new(id, self.clone())
    }
}

/// Collection metrics.
///
/// Note that these metrics do _not_ have a `collection_id` label. We avoid introducing
/// per-collection, per-worker metrics because the number of resulting time series would
/// potentially be huge. Instead we count classes of collections, such as hydrated collections.
#[derive(Clone, Debug)]
pub struct CollectionMetrics {
    metrics: WorkerMetrics,
    collection_type: &'static str,
    collection_hydrated: bool,
}

impl CollectionMetrics {
    pub fn new(collection_id: GlobalId, metrics: WorkerMetrics) -> Self {
        let collection_type = match collection_id {
            GlobalId::System(_) => "system",
            GlobalId::IntrospectionSourceIndex(_) => "log",
            GlobalId::User(_) => "user",
            GlobalId::Transient(_) => "transient",
            GlobalId::Explain => "explain",
        };
        let collection_hydrated = false;

        metrics.inc_collection_count(collection_type, collection_hydrated);

        Self {
            metrics,
            collection_type,
            collection_hydrated,
        }
    }

    /// Record this collection as hydration.
    pub fn record_collection_hydrated(&mut self) {
        if self.collection_hydrated {
            return;
        }

        self.metrics
            .dec_collection_count(self.collection_type, false);
        self.metrics
            .inc_collection_count(self.collection_type, true);
        self.collection_hydrated = true;
    }
}

impl Drop for CollectionMetrics {
    fn drop(&mut self) {
        self.metrics
            .dec_collection_count(self.collection_type, self.collection_hydrated);
    }
}