glean_core/metrics/

timing_distribution.rs

// This Source Code Form is subject to the terms of the Mozilla Public
// License, v. 2.0. If a copy of the MPL was not distributed with this
// file, You can obtain one at https://mozilla.org/MPL/2.0/.

use std::collections::HashMap;
use std::mem;
use std::sync::atomic::{AtomicUsize, Ordering};
use std::sync::{Arc, Mutex};
use std::time::Duration;

use malloc_size_of_derive::MallocSizeOf;

use crate::common_metric_data::{CommonMetricDataInternal, DynamicLabelType};
use crate::error_recording::{record_error, test_get_num_recorded_errors, ErrorType};
use crate::histogram::{Functional, Histogram};
use crate::metrics::time_unit::TimeUnit;
use crate::metrics::{DistributionData, Metric, MetricType};
use crate::storage::StorageManager;
use crate::CommonMetricData;
use crate::Glean;

// The base of the logarithm used to determine bucketing
const LOG_BASE: f64 = 2.0;

// The buckets per each order of magnitude of the logarithm.
const BUCKETS_PER_MAGNITUDE: f64 = 8.0;

// Maximum time, which means we retain a maximum of 316 buckets.
// It is automatically adjusted based on the `time_unit` parameter
// so that:
//
// - `nanosecond` - 10 minutes
// - `microsecond` - ~6.94 days
// - `millisecond` - ~19 years
const MAX_SAMPLE_TIME: u64 = 1000 * 1000 * 1000 * 60 * 10;

/// Identifier for a running timer.
///
/// Its internals are considered private,
/// but due to UniFFI's behavior we expose its field for now.
#[derive(Copy, Clone, Debug, Eq, PartialEq, Hash, MallocSizeOf)]
pub struct TimerId {
    /// This timer's id.
    pub id: u64,
}

impl From<u64> for TimerId {
    fn from(val: u64) -> TimerId {
        TimerId { id: val }
    }
}

impl From<usize> for TimerId {
    fn from(val: usize) -> TimerId {
        TimerId { id: val as u64 }
    }
}

/// A timing distribution metric.
///
/// Timing distributions are used to accumulate and store time measurement, for analyzing distributions of the timing data.
#[derive(Clone, Debug)]
pub struct TimingDistributionMetric {
    meta: Arc<CommonMetricDataInternal>,
    time_unit: TimeUnit,
    next_id: Arc<AtomicUsize>,
    start_times: Arc<Mutex<HashMap<TimerId, u64>>>,
}

impl ::malloc_size_of::MallocSizeOf for TimingDistributionMetric {
    fn size_of(&self, ops: &mut malloc_size_of::MallocSizeOfOps) -> usize {
        // Note: This is behind an `Arc`.
        // `size_of` should only be called on the main thread to avoid double-counting.
        self.meta.size_of(ops)
            + self.time_unit.size_of(ops)
            + self.next_id.size_of(ops)
            + self.start_times.lock().unwrap().size_of(ops)
    }
}

/// Create a snapshot of the histogram with a time unit.
///
/// The snapshot can be serialized into the payload format.
pub(crate) fn snapshot(hist: &Histogram<Functional>) -> DistributionData {
    DistributionData {
        // **Caution**: This cannot use `Histogram::snapshot_values` and needs to use the more
        // specialized snapshot function.
        values: hist
            .snapshot()
            .iter()
            .map(|(&k, &v)| (k as i64, v as i64))
            .collect(),
        sum: hist.sum() as i64,
        count: hist.count() as i64,
    }
}

impl MetricType for TimingDistributionMetric {
    fn meta(&self) -> &CommonMetricDataInternal {
        &self.meta
    }

    fn with_name(&self, name: String) -> Self {
        let mut meta = (*self.meta).clone();
        meta.inner.name = name;
        Self {
            meta: Arc::new(meta),
            time_unit: self.time_unit,
            next_id: Arc::new(AtomicUsize::new(1)),
            start_times: Arc::new(Mutex::new(Default::default())),
        }
    }

    fn with_dynamic_label(&self, label: DynamicLabelType) -> Self {
        let mut meta = (*self.meta).clone();
        meta.inner.dynamic_label = Some(label);
        Self {
            meta: Arc::new(meta),
            time_unit: self.time_unit,
            next_id: Arc::new(AtomicUsize::new(1)),
            start_times: Arc::new(Mutex::new(Default::default())),
        }
    }
}

// IMPORTANT:
//
// When changing this implementation, make sure all the operations are
// also declared in the related trait in `../traits/`.
impl TimingDistributionMetric {
    /// Creates a new timing distribution metric.
    pub fn new(meta: CommonMetricData, time_unit: TimeUnit) -> Self {
        Self {
            meta: Arc::new(meta.into()),
            time_unit,
            next_id: Arc::new(AtomicUsize::new(1)),
            start_times: Arc::new(Mutex::new(Default::default())),
        }
    }

    /// Starts tracking time for the provided metric.
    ///
    /// This records an error if it’s already tracking time (i.e.
    /// [`set_start`](TimingDistributionMetric::set_start) was already called with no
    /// corresponding [`set_stop_and_accumulate`](TimingDistributionMetric::set_stop_and_accumulate)): in
    /// that case the original start time will be preserved.
    ///
    /// # Arguments
    ///
    /// * `start_time` - Timestamp in nanoseconds.
    ///
    /// # Returns
    ///
    /// A unique [`TimerId`] for the new timer.
    pub fn start(&self) -> TimerId {
        let start_time = time::precise_time_ns();
        let id = self.next_id.fetch_add(1, Ordering::SeqCst).into();
        let metric = self.clone();
        crate::launch_with_glean(move |_glean| metric.set_start(id, start_time));
        id
    }

    pub(crate) fn start_sync(&self) -> TimerId {
        let start_time = time::precise_time_ns();
        let id = self.next_id.fetch_add(1, Ordering::SeqCst).into();
        let metric = self.clone();
        metric.set_start(id, start_time);
        id
    }

    /// **Test-only API (exported for testing purposes).**
    ///
    /// Set start time for this metric synchronously.
    ///
    /// Use [`start`](Self::start) instead.
    #[doc(hidden)]
    pub fn set_start(&self, id: TimerId, start_time: u64) {
        let mut map = self.start_times.lock().expect("can't lock timings map");
        map.insert(id, start_time);
    }

    /// Stops tracking time for the provided metric and associated timer id.
    ///
    /// Adds a count to the corresponding bucket in the timing distribution.
    /// This will record an error if no
    /// [`set_start`](TimingDistributionMetric::set_start) was called.
    ///
    /// # Arguments
    ///
    /// * `id` - The [`TimerId`] to associate with this timing. This allows
    ///   for concurrent timing of events associated with different ids to the
    ///   same timespan metric.
    /// * `stop_time` - Timestamp in nanoseconds.
    pub fn stop_and_accumulate(&self, id: TimerId) {
        let stop_time = time::precise_time_ns();
        let metric = self.clone();
        crate::launch_with_glean(move |glean| metric.set_stop_and_accumulate(glean, id, stop_time));
    }

    fn set_stop(&self, id: TimerId, stop_time: u64) -> Result<u64, (ErrorType, &str)> {
        let mut start_times = self.start_times.lock().expect("can't lock timings map");
        let start_time = match start_times.remove(&id) {
            Some(start_time) => start_time,
            None => return Err((ErrorType::InvalidState, "Timing not running")),
        };

        let duration = match stop_time.checked_sub(start_time) {
            Some(duration) => duration,
            None => {
                return Err((
                    ErrorType::InvalidValue,
                    "Timer stopped with negative duration",
                ))
            }
        };

        Ok(duration)
    }

    /// **Test-only API (exported for testing purposes).**
    ///
    /// Set stop time for this metric synchronously.
    ///
    /// Use [`stop_and_accumulate`](Self::stop_and_accumulate) instead.
    #[doc(hidden)]
    pub fn set_stop_and_accumulate(&self, glean: &Glean, id: TimerId, stop_time: u64) {
        if !self.should_record(glean) {
            let mut start_times = self.start_times.lock().expect("can't lock timings map");
            start_times.remove(&id);
            return;
        }

        // Duration is in nanoseconds.
        let mut duration = match self.set_stop(id, stop_time) {
            Err((err_type, err_msg)) => {
                record_error(glean, &self.meta, err_type, err_msg, None);
                return;
            }
            Ok(duration) => duration,
        };

        let min_sample_time = self.time_unit.as_nanos(1);
        let max_sample_time = self.time_unit.as_nanos(MAX_SAMPLE_TIME);

        duration = if duration < min_sample_time {
            // If measurement is less than the minimum, just truncate. This is
            // not recorded as an error.
            min_sample_time
        } else if duration > max_sample_time {
            let msg = format!(
                "Sample is longer than the max for a time_unit of {:?} ({} ns)",
                self.time_unit, max_sample_time
            );
            record_error(glean, &self.meta, ErrorType::InvalidOverflow, msg, None);
            max_sample_time
        } else {
            duration
        };

        if !self.should_record(glean) {
            return;
        }

        // Let's be defensive here:
        // The uploader tries to store some timing distribution metrics,
        // but in tests that storage might be gone already.
        // Let's just ignore those.
        // We do the same for counters.
        // This should never happen in real app usage.
        if let Some(storage) = glean.storage_opt() {
            storage.record_with(glean, &self.meta, |old_value| match old_value {
                Some(Metric::TimingDistribution(mut hist)) => {
                    hist.accumulate(duration);
                    Metric::TimingDistribution(hist)
                }
                _ => {
                    let mut hist = Histogram::functional(LOG_BASE, BUCKETS_PER_MAGNITUDE);
                    hist.accumulate(duration);
                    Metric::TimingDistribution(hist)
                }
            });
        } else {
            log::warn!(
                "Couldn't get storage. Can't record timing distribution '{}'.",
                self.meta.base_identifier()
            );
        }
    }

    /// Aborts a previous [`start`](Self::start) call.
    ///
    /// No error is recorded if no [`start`](Self::start) was called.
    ///
    /// # Arguments
    ///
    /// * `id` - The [`TimerId`] to associate with this timing. This allows
    ///   for concurrent timing of events associated with different ids to the
    ///   same timing distribution metric.
    pub fn cancel(&self, id: TimerId) {
        let metric = self.clone();
        crate::launch_with_glean(move |_glean| metric.cancel_sync(id));
    }

    /// Aborts a previous [`start`](Self::start) call synchronously.
    pub(crate) fn cancel_sync(&self, id: TimerId) {
        let mut map = self.start_times.lock().expect("can't lock timings map");
        map.remove(&id);
    }

    /// Accumulates the provided signed samples in the metric.
    ///
    /// This is required so that the platform-specific code can provide us with
    /// 64 bit signed integers if no `u64` comparable type is available. This
    /// will take care of filtering and reporting errors for any provided negative
    /// sample.
    ///
    /// Please note that this assumes that the provided samples are already in
    /// the "unit" declared by the instance of the metric type (e.g. if the
    /// instance this method was called on is using [`TimeUnit::Second`], then
    /// `samples` are assumed to be in that unit).
    ///
    /// # Arguments
    ///
    /// * `samples` - The vector holding the samples to be recorded by the metric.
    ///
    /// ## Notes
    ///
    /// Discards any negative value in `samples` and report an [`ErrorType::InvalidValue`]
    /// for each of them. Reports an [`ErrorType::InvalidOverflow`] error for samples that
    /// are longer than `MAX_SAMPLE_TIME`.
    pub fn accumulate_samples(&self, samples: Vec<i64>) {
        let metric = self.clone();
        crate::launch_with_glean(move |glean| metric.accumulate_samples_sync(glean, &samples))
    }

    /// Accumulates precisely one signed sample and appends it to the metric.
    ///
    /// Precludes the need for a collection in the most common use case.
    ///
    /// Sign is required so that the platform-specific code can provide us with
    /// a 64 bit signed integer if no `u64` comparable type is available. This
    /// will take care of filtering and reporting errors for any provided negative
    /// sample.
    ///
    /// Please note that this assumes that the provided sample is already in
    /// the "unit" declared by the instance of the metric type (e.g. if the
    /// instance this method was called on is using [`crate::TimeUnit::Second`], then
    /// `sample` is assumed to be in that unit).
    ///
    /// # Arguments
    ///
    /// * `sample` - The singular sample to be recorded by the metric.
    ///
    /// ## Notes
    ///
    /// Discards any negative value and reports an [`ErrorType::InvalidValue`].
    /// Reports an [`ErrorType::InvalidOverflow`] error if the sample is longer than
    /// `MAX_SAMPLE_TIME`.
    pub fn accumulate_single_sample(&self, sample: i64) {
        let metric = self.clone();
        crate::launch_with_glean(move |glean| metric.accumulate_samples_sync(glean, &[sample]))
    }

    /// **Test-only API (exported for testing purposes).**
    /// Accumulates the provided signed samples in the metric.
    ///
    /// Use [`accumulate_samples`](Self::accumulate_samples)
    #[doc(hidden)]
    pub fn accumulate_samples_sync(&self, glean: &Glean, samples: &[i64]) {
        if !self.should_record(glean) {
            return;
        }

        let mut num_negative_samples = 0;
        let mut num_too_long_samples = 0;
        let max_sample_time = self.time_unit.as_nanos(MAX_SAMPLE_TIME);

        glean.storage().record_with(glean, &self.meta, |old_value| {
            let mut hist = match old_value {
                Some(Metric::TimingDistribution(hist)) => hist,
                _ => Histogram::functional(LOG_BASE, BUCKETS_PER_MAGNITUDE),
            };

            for &sample in samples.iter() {
                if sample < 0 {
                    num_negative_samples += 1;
                } else {
                    let mut sample = sample as u64;

                    // Check the range prior to converting the incoming unit to
                    // nanoseconds, so we can compare against the constant
                    // MAX_SAMPLE_TIME.
                    if sample == 0 {
                        sample = 1;
                    } else if sample > MAX_SAMPLE_TIME {
                        num_too_long_samples += 1;
                        sample = MAX_SAMPLE_TIME;
                    }

                    sample = self.time_unit.as_nanos(sample);

                    hist.accumulate(sample);
                }
            }

            Metric::TimingDistribution(hist)
        });

        if num_negative_samples > 0 {
            let msg = format!("Accumulated {} negative samples", num_negative_samples);
            record_error(
                glean,
                &self.meta,
                ErrorType::InvalidValue,
                msg,
                num_negative_samples,
            );
        }

        if num_too_long_samples > 0 {
            let msg = format!(
                "{} samples are longer than the maximum of {}",
                num_too_long_samples, max_sample_time
            );
            record_error(
                glean,
                &self.meta,
                ErrorType::InvalidOverflow,
                msg,
                num_too_long_samples,
            );
        }
    }

    /// Accumulates the provided samples in the metric.
    ///
    /// # Arguments
    ///
    /// * `samples` - A list of samples recorded by the metric.
    ///               Samples must be in nanoseconds.
    /// ## Notes
    ///
    /// Reports an [`ErrorType::InvalidOverflow`] error for samples that
    /// are longer than `MAX_SAMPLE_TIME`.
    pub fn accumulate_raw_samples_nanos(&self, samples: Vec<u64>) {
        let metric = self.clone();
        crate::launch_with_glean(move |glean| {
            metric.accumulate_raw_samples_nanos_sync(glean, &samples)
        })
    }

    /// Accumulates precisely one duration to the metric.
    ///
    /// Like `TimingDistribution::accumulate_single_sample`, but for use when the
    /// duration is:
    ///
    ///  * measured externally, or
    ///  * is in a unit different from the timing_distribution's internal TimeUnit.
    ///
    /// # Arguments
    ///
    /// * `duration` - The single duration to be recorded in the metric.
    ///
    /// ## Notes
    ///
    /// Reports an [`ErrorType::InvalidOverflow`] error if `duration` is longer than
    /// `MAX_SAMPLE_TIME`.
    ///
    /// The API client is responsible for ensuring that `duration` is derived from a
    /// monotonic clock source that behaves consistently over computer sleep across
    /// the application's platforms. Otherwise the resulting data may not share the same
    /// guarantees that other `timing_distribution` metrics' data do.
    pub fn accumulate_raw_duration(&self, duration: Duration) {
        let duration_ns = duration.as_nanos().try_into().unwrap_or(u64::MAX);
        let metric = self.clone();
        crate::launch_with_glean(move |glean| {
            metric.accumulate_raw_samples_nanos_sync(glean, &[duration_ns])
        })
    }

    /// **Test-only API (exported for testing purposes).**
    ///
    /// Accumulates the provided samples in the metric.
    ///
    /// Use [`accumulate_raw_samples_nanos`](Self::accumulate_raw_samples_nanos) instead.
    #[doc(hidden)]
    pub fn accumulate_raw_samples_nanos_sync(&self, glean: &Glean, samples: &[u64]) {
        if !self.should_record(glean) {
            return;
        }

        let mut num_too_long_samples = 0;
        let min_sample_time = self.time_unit.as_nanos(1);
        let max_sample_time = self.time_unit.as_nanos(MAX_SAMPLE_TIME);

        glean.storage().record_with(glean, &self.meta, |old_value| {
            let mut hist = match old_value {
                Some(Metric::TimingDistribution(hist)) => hist,
                _ => Histogram::functional(LOG_BASE, BUCKETS_PER_MAGNITUDE),
            };

            for &sample in samples.iter() {
                let mut sample = sample;

                if sample < min_sample_time {
                    sample = min_sample_time;
                } else if sample > max_sample_time {
                    num_too_long_samples += 1;
                    sample = max_sample_time;
                }

                // `sample` is in nanoseconds.
                hist.accumulate(sample);
            }

            Metric::TimingDistribution(hist)
        });

        if num_too_long_samples > 0 {
            let msg = format!(
                "{} samples are longer than the maximum of {}",
                num_too_long_samples, max_sample_time
            );
            record_error(
                glean,
                &self.meta,
                ErrorType::InvalidOverflow,
                msg,
                num_too_long_samples,
            );
        }
    }

    /// Gets the currently stored value as an integer.
    #[doc(hidden)]
    pub fn get_value<'a, S: Into<Option<&'a str>>>(
        &self,
        glean: &Glean,
        ping_name: S,
    ) -> Option<DistributionData> {
        let queried_ping_name = ping_name
            .into()
            .unwrap_or_else(|| &self.meta().inner.send_in_pings[0]);

        match StorageManager.snapshot_metric_for_test(
            glean.storage(),
            queried_ping_name,
            &self.meta.identifier(glean),
            self.meta.inner.lifetime,
        ) {
            Some(Metric::TimingDistribution(hist)) => Some(snapshot(&hist)),
            _ => None,
        }
    }

    /// **Test-only API (exported for FFI purposes).**
    ///
    /// Gets the currently stored value as an integer.
    ///
    /// This doesn't clear the stored value.
    ///
    /// # Arguments
    ///
    /// * `ping_name` - the optional name of the ping to retrieve the metric
    ///                 for. Defaults to the first value in `send_in_pings`.
    ///
    /// # Returns
    ///
    /// The stored value or `None` if nothing stored.
    pub fn test_get_value(&self, ping_name: Option<String>) -> Option<DistributionData> {
        crate::block_on_dispatcher();
        crate::core::with_glean(|glean| self.get_value(glean, ping_name.as_deref()))
    }

    /// **Exported for test purposes.**
    ///
    /// Gets the number of recorded errors for the given metric and error type.
    ///
    /// # Arguments
    ///
    /// * `error` - The type of error
    ///
    /// # Returns
    ///
    /// The number of errors reported.
    pub fn test_get_num_recorded_errors(&self, error: ErrorType) -> i32 {
        crate::block_on_dispatcher();

        crate::core::with_glean(|glean| {
            test_get_num_recorded_errors(glean, self.meta(), error).unwrap_or(0)
        })
    }

    /// **Experimental:** Start a new histogram buffer associated with this timing distribution metric.
    ///
    /// A histogram buffer accumulates in-memory.
    /// Data is recorded into the metric on drop.
    pub fn start_buffer(&self) -> LocalTimingDistribution<'_> {
        LocalTimingDistribution::new(self)
    }

    fn commit_histogram(&self, histogram: Histogram<Functional>, errors: usize) {
        let metric = self.clone();
        crate::launch_with_glean(move |glean| {
            if errors > 0 {
                let max_sample_time = metric.time_unit.as_nanos(MAX_SAMPLE_TIME);
                let msg = format!(
                    "{} samples are longer than the maximum of {}",
                    errors, max_sample_time
                );
                record_error(
                    glean,
                    &metric.meta,
                    ErrorType::InvalidValue,
                    msg,
                    Some(errors as i32),
                );
            }

            glean
                .storage()
                .record_with(glean, &metric.meta, move |old_value| {
                    let mut hist = match old_value {
                        Some(Metric::TimingDistribution(hist)) => hist,
                        _ => Histogram::functional(LOG_BASE, BUCKETS_PER_MAGNITUDE),
                    };

                    hist.merge(&histogram);
                    Metric::TimingDistribution(hist)
                });
        });
    }
}

/// **Experimental:** A histogram buffer associated with a specific instance of a [`TimingDistributionMetric`].
///
/// Accumulation happens in-memory.
/// Data is merged into the metric on [`Drop::drop`].
#[derive(Debug)]
pub struct LocalTimingDistribution<'a> {
    histogram: Histogram<Functional>,
    metric: &'a TimingDistributionMetric,
    errors: usize,
}

impl<'a> LocalTimingDistribution<'a> {
    /// Create a new histogram buffer referencing the timing distribution it will record into.
    fn new(metric: &'a TimingDistributionMetric) -> Self {
        let histogram = Histogram::functional(LOG_BASE, BUCKETS_PER_MAGNITUDE);
        Self {
            histogram,
            metric,
            errors: 0,
        }
    }

    /// Accumulates one sample into the histogram.
    ///
    /// The provided sample must be in the "unit" declared by the instance of the metric type
    /// (e.g. if the instance this method was called on is using [`crate::TimeUnit::Second`], then
    /// `sample` is assumed to be in seconds).
    ///
    /// Accumulation happens in-memory only.
    pub fn accumulate(&mut self, sample: u64) {
        // Check the range prior to converting the incoming unit to
        // nanoseconds, so we can compare against the constant
        // MAX_SAMPLE_TIME.
        let sample = if sample == 0 {
            1
        } else if sample > MAX_SAMPLE_TIME {
            self.errors += 1;
            MAX_SAMPLE_TIME
        } else {
            sample
        };

        let sample = self.metric.time_unit.as_nanos(sample);
        self.histogram.accumulate(sample)
    }

    /// Abandon this histogram buffer and don't commit accumulated data.
    pub fn abandon(mut self) {
        self.histogram.clear();
    }
}

impl Drop for LocalTimingDistribution<'_> {
    fn drop(&mut self) {
        if self.histogram.is_empty() {
            return;
        }

        // We want to move that value.
        // A `0/0` histogram doesn't allocate.
        let buffer = mem::replace(&mut self.histogram, Histogram::functional(0.0, 0.0));
        self.metric.commit_histogram(buffer, self.errors);
    }
}

#[cfg(test)]
mod test {
    use super::*;

    #[test]
    fn can_snapshot() {
        use serde_json::json;

        let mut hist = Histogram::functional(2.0, 8.0);

        for i in 1..=10 {
            hist.accumulate(i);
        }

        let snap = snapshot(&hist);

        let expected_json = json!({
            "sum": 55,
            "values": {
                "1": 1,
                "2": 1,
                "3": 1,
                "4": 1,
                "5": 1,
                "6": 1,
                "7": 1,
                "8": 1,
                "9": 1,
                "10": 1,
            },
        });

        assert_eq!(expected_json, json!(snap));
    }

    #[test]
    fn can_snapshot_sparse() {
        use serde_json::json;

        let mut hist = Histogram::functional(2.0, 8.0);

        hist.accumulate(1024);
        hist.accumulate(1024);
        hist.accumulate(1116);
        hist.accumulate(1448);

        let snap = snapshot(&hist);

        let expected_json = json!({
            "sum": 4612,
            "values": {
                "1024": 2,
                "1116": 1,
                "1448": 1,
            },
        });

        assert_eq!(expected_json, json!(snap));
    }
}