improve ondemand-download latency observability (#11421)

## Problem

We don't have metrics to exactly quantify the end user impact of
on-demand downloads.

Perf tracing is underway (#11140) to supply us with high-resolution
*samples*.

But it will also be useful to have some aggregate per-timeline and
per-instance metrics that definitively contain all observations.

## Summary of changes

This PR consists of independent commits that should be reviewed
independently.

However, for convenience, we're going to merge them together.

- refactor(metrics): measure_remote_op can use async traits
- impr(pageserver metrics): task_kind dimension for
remote_timeline_client latency histo
  - implements https://github.com/neondatabase/cloud/issues/26800
- refs
https://github.com/neondatabase/cloud/issues/26193#issuecomment-2769705793
- use the opportunity to rename the metric and add a _global suffix;
checked grafana export, it's only used in two personal dashboards, one
of them mine, the other by Heikki
- log on-demand download latency for expensive-to-query but precise
ground truth
- metric for wall clock time spent waiting for on-demand downloads

## Refs

- refs https://github.com/neondatabase/cloud/issues/26800
- a bunch of minor investigations / incidents into latency outliers
This commit is contained in:
Christian Schwarz
2025-04-04 20:04:39 +02:00
committed by GitHub
parent 4f94751b75
commit aad410c8f1
11 changed files with 247 additions and 69 deletions

View File

@@ -0,0 +1,26 @@
use std::time::{Duration, Instant};
#[derive(Default)]
pub struct ElapsedAccum {
accum: Duration,
}
impl ElapsedAccum {
pub fn get(&self) -> Duration {
self.accum
}
pub fn guard(&mut self) -> impl Drop + '_ {
let start = Instant::now();
scopeguard::guard(start, |last_wait_at| {
self.accum += Instant::now() - last_wait_at;
})
}
pub async fn measure<Fut, O>(&mut self, fut: Fut) -> O
where
Fut: Future<Output = O>,
{
let _guard = self.guard();
fut.await
}
}

View File

@@ -93,6 +93,8 @@ pub mod try_rcu;
pub mod guard_arc_swap;
pub mod elapsed_accum;
#[cfg(target_os = "linux")]
pub mod linux_socket_ioctl;

View File

@@ -111,9 +111,17 @@ impl<T> OnceCell<T> {
}
}
/// Like [`Self::get_or_init_detached_measured`], but without out parameter for time spent waiting.
pub async fn get_or_init_detached(&self) -> Result<Guard<'_, T>, InitPermit> {
self.get_or_init_detached_measured(None).await
}
/// Returns a guard to an existing initialized value, or returns an unique initialization
/// permit which can be used to initialize this `OnceCell` using `OnceCell::set`.
pub async fn get_or_init_detached(&self) -> Result<Guard<'_, T>, InitPermit> {
pub async fn get_or_init_detached_measured(
&self,
mut wait_time: Option<&mut crate::elapsed_accum::ElapsedAccum>,
) -> Result<Guard<'_, T>, InitPermit> {
// It looks like OnceCell::get_or_init could be implemented using this method instead of
// duplication. However, that makes the future be !Send due to possibly holding on to the
// MutexGuard over an await point.
@@ -125,12 +133,16 @@ impl<T> OnceCell<T> {
}
guard.init_semaphore.clone()
};
{
let permit = {
// increment the count for the duration of queued
let _guard = CountWaitingInitializers::start(self);
sem.acquire().await
let fut = sem.acquire();
if let Some(wait_time) = wait_time.as_mut() {
wait_time.measure(fut).await
} else {
fut.await
}
};
let Ok(permit) = permit else {

View File

@@ -89,7 +89,7 @@
//! [`RequestContext`] argument. Functions in the middle of the call chain
//! only need to pass it on.
use std::sync::Arc;
use std::{sync::Arc, time::Duration};
use once_cell::sync::Lazy;
use tracing::warn;
@@ -566,6 +566,34 @@ impl RequestContext {
}
}
pub(crate) fn ondemand_download_wait_observe(&self, duration: Duration) {
if duration == Duration::ZERO {
return;
}
match &self.scope {
Scope::Timeline { arc_arc } => arc_arc
.wait_ondemand_download_time
.observe(self.task_kind, duration),
_ => {
use once_cell::sync::Lazy;
use std::sync::Mutex;
use std::time::Duration;
use utils::rate_limit::RateLimit;
static LIMIT: Lazy<Mutex<RateLimit>> =
Lazy::new(|| Mutex::new(RateLimit::new(Duration::from_secs(1))));
let mut guard = LIMIT.lock().unwrap();
guard.call2(|rate_limit_stats| {
warn!(
%rate_limit_stats,
backtrace=%std::backtrace::Backtrace::force_capture(),
"ondemand downloads should always happen within timeline scope",
);
});
}
}
}
pub(crate) fn perf_follows_from(&self, from: &RequestContext) {
if let (Some(span), Some(from_span)) = (&self.perf_span, &from.perf_span) {
span.inner().follows_from(from_span.inner());

View File

@@ -1,10 +1,8 @@
use std::collections::HashMap;
use std::num::NonZeroUsize;
use std::os::fd::RawFd;
use std::pin::Pin;
use std::sync::atomic::AtomicU64;
use std::sync::{Arc, Mutex};
use std::task::{Context, Poll};
use std::time::{Duration, Instant};
use enum_map::{Enum as _, EnumMap};
@@ -23,7 +21,6 @@ use pageserver_api::config::{
};
use pageserver_api::models::InMemoryLayerInfo;
use pageserver_api::shard::TenantShardId;
use pin_project_lite::pin_project;
use postgres_backend::{QueryError, is_expected_io_error};
use pq_proto::framed::ConnectionError;
use strum::{EnumCount, IntoEnumIterator as _, VariantNames};
@@ -500,6 +497,100 @@ pub(crate) static WAIT_LSN_IN_PROGRESS_GLOBAL_MICROS: Lazy<IntCounter> = Lazy::n
.expect("failed to define a metric")
});
pub(crate) mod wait_ondemand_download_time {
use super::*;
const WAIT_ONDEMAND_DOWNLOAD_TIME_BUCKETS: &[f64] = &[
0.01, 0.02, 0.03, 0.04, 0.05, 0.06, 0.07, 0.08, 0.09, // 10 ms - 100ms
0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, // 100ms to 1s
1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0, // 1s to 10s
10.0, 20.0, 30.0, 40.0, 50.0, 60.0, // 10s to 1m
];
/// The task kinds for which we want to track wait times for on-demand downloads.
/// Other task kinds' wait times are accumulated in label value `unknown`.
pub(crate) const WAIT_ONDEMAND_DOWNLOAD_METRIC_TASK_KINDS: [TaskKind; 2] = [
TaskKind::PageRequestHandler,
TaskKind::WalReceiverConnectionHandler,
];
pub(crate) static WAIT_ONDEMAND_DOWNLOAD_TIME_GLOBAL: Lazy<Vec<Histogram>> = Lazy::new(|| {
let histo = register_histogram_vec!(
"pageserver_wait_ondemand_download_seconds_global",
"Observations are individual tasks' wait times for on-demand downloads. \
If N tasks coalesce on an on-demand download, and it takes 10s, than we observe N * 10s.",
&["task_kind"],
WAIT_ONDEMAND_DOWNLOAD_TIME_BUCKETS.into(),
)
.expect("failed to define a metric");
WAIT_ONDEMAND_DOWNLOAD_METRIC_TASK_KINDS
.iter()
.map(|task_kind| histo.with_label_values(&[task_kind.into()]))
.collect::<Vec<_>>()
});
pub(crate) static WAIT_ONDEMAND_DOWNLOAD_TIME_SUM: Lazy<CounterVec> = Lazy::new(|| {
register_counter_vec!(
// use a name that _could_ be evolved into a per-timeline histogram later
"pageserver_wait_ondemand_download_seconds_sum",
"Like `pageserver_wait_ondemand_download_seconds_global` but per timeline",
&["tenant_id", "shard_id", "timeline_id", "task_kind"],
)
.unwrap()
});
pub struct WaitOndemandDownloadTimeSum {
counters: [Counter; WAIT_ONDEMAND_DOWNLOAD_METRIC_TASK_KINDS.len()],
}
impl WaitOndemandDownloadTimeSum {
pub(crate) fn new(tenant_id: &str, shard_id: &str, timeline_id: &str) -> Self {
let counters = WAIT_ONDEMAND_DOWNLOAD_METRIC_TASK_KINDS
.iter()
.map(|task_kind| {
WAIT_ONDEMAND_DOWNLOAD_TIME_SUM
.get_metric_with_label_values(&[
tenant_id,
shard_id,
timeline_id,
task_kind.into(),
])
.unwrap()
})
.collect::<Vec<_>>();
Self {
counters: counters.try_into().unwrap(),
}
}
pub(crate) fn observe(&self, task_kind: TaskKind, duration: Duration) {
let maybe = WAIT_ONDEMAND_DOWNLOAD_METRIC_TASK_KINDS
.iter()
.enumerate()
.find(|(_, kind)| **kind == task_kind);
let Some((idx, _)) = maybe else {
return;
};
WAIT_ONDEMAND_DOWNLOAD_TIME_GLOBAL[idx].observe(duration.as_secs_f64());
let counter = &self.counters[idx];
counter.inc_by(duration.as_secs_f64());
}
}
pub(crate) fn shutdown_timeline(tenant_id: &str, shard_id: &str, timeline_id: &str) {
for task_kind in WAIT_ONDEMAND_DOWNLOAD_METRIC_TASK_KINDS {
let _ = WAIT_ONDEMAND_DOWNLOAD_TIME_SUM.remove_label_values(&[
tenant_id,
shard_id,
timeline_id,
task_kind.into(),
]);
}
}
pub(crate) fn preinitialize_global_metrics() {
Lazy::force(&WAIT_ONDEMAND_DOWNLOAD_TIME_GLOBAL);
}
}
static LAST_RECORD_LSN: Lazy<IntGaugeVec> = Lazy::new(|| {
register_int_gauge_vec!(
"pageserver_last_record_lsn",
@@ -2315,13 +2406,18 @@ impl RemoteOpFileKind {
}
}
pub(crate) static REMOTE_OPERATION_TIME: Lazy<HistogramVec> = Lazy::new(|| {
pub(crate) static REMOTE_TIMELINE_CLIENT_COMPLETION_LATENCY: Lazy<HistogramVec> = Lazy::new(|| {
register_histogram_vec!(
"pageserver_remote_operation_seconds",
"Time spent on remote storage operations. \
Grouped by tenant, timeline, operation_kind and status. \
"pageserver_remote_timeline_client_seconds_global",
"Time spent on remote timeline client operations. \
Grouped by task_kind, file_kind, operation_kind and status. \
The task_kind is \
- for layer downloads, populated from RequestContext (primary objective of having the label) \
- for index downloads, set to 'unknown' \
- for any upload operation, set to 'RemoteUploadTask' \
This keeps dimensionality at bay. \
Does not account for time spent waiting in remote timeline client's queues.",
&["file_kind", "op_kind", "status"]
&["task_kind", "file_kind", "op_kind", "status"]
)
.expect("failed to define a metric")
});
@@ -2883,6 +2979,7 @@ pub(crate) struct TimelineMetrics {
pub storage_io_size: StorageIoSizeMetrics,
pub wait_lsn_in_progress_micros: GlobalAndPerTenantIntCounter,
pub wait_lsn_start_finish_counterpair: IntCounterPair,
pub wait_ondemand_download_time: wait_ondemand_download_time::WaitOndemandDownloadTimeSum,
shutdown: std::sync::atomic::AtomicBool,
}
@@ -3028,6 +3125,13 @@ impl TimelineMetrics {
.get_metric_with_label_values(&[&tenant_id, &shard_id, &timeline_id])
.unwrap();
let wait_ondemand_download_time =
wait_ondemand_download_time::WaitOndemandDownloadTimeSum::new(
&tenant_id,
&shard_id,
&timeline_id,
);
TimelineMetrics {
tenant_id,
shard_id,
@@ -3061,6 +3165,7 @@ impl TimelineMetrics {
wal_records_received,
wait_lsn_in_progress_micros,
wait_lsn_start_finish_counterpair,
wait_ondemand_download_time,
shutdown: std::sync::atomic::AtomicBool::default(),
}
}
@@ -3253,6 +3358,8 @@ impl TimelineMetrics {
.remove_label_values(&mut res, &[tenant_id, shard_id, timeline_id]);
}
wait_ondemand_download_time::shutdown_timeline(tenant_id, shard_id, timeline_id);
let _ = SMGR_QUERY_STARTED_PER_TENANT_TIMELINE.remove_label_values(&[
SmgrQueryType::GetPageAtLsn.into(),
tenant_id,
@@ -3374,13 +3481,18 @@ impl RemoteTimelineClientMetrics {
pub fn remote_operation_time(
&self,
task_kind: Option<TaskKind>,
file_kind: &RemoteOpFileKind,
op_kind: &RemoteOpKind,
status: &'static str,
) -> Histogram {
let key = (file_kind.as_str(), op_kind.as_str(), status);
REMOTE_OPERATION_TIME
.get_metric_with_label_values(&[key.0, key.1, key.2])
REMOTE_TIMELINE_CLIENT_COMPLETION_LATENCY
.get_metric_with_label_values(&[
task_kind.as_ref().map(|tk| tk.into()).unwrap_or("unknown"),
file_kind.as_str(),
op_kind.as_str(),
status,
])
.unwrap()
}
@@ -3625,54 +3737,26 @@ impl Drop for RemoteTimelineClientMetrics {
/// Wrapper future that measures the time spent by a remote storage operation,
/// and records the time and success/failure as a prometheus metric.
pub(crate) trait MeasureRemoteOp: Sized {
fn measure_remote_op(
pub(crate) trait MeasureRemoteOp<O, E>: Sized + Future<Output = Result<O, E>> {
async fn measure_remote_op(
self,
task_kind: Option<TaskKind>, // not all caller contexts have a RequestContext / TaskKind handy
file_kind: RemoteOpFileKind,
op: RemoteOpKind,
metrics: Arc<RemoteTimelineClientMetrics>,
) -> MeasuredRemoteOp<Self> {
) -> Result<O, E> {
let start = Instant::now();
MeasuredRemoteOp {
inner: self,
file_kind,
op,
start,
metrics,
}
let res = self.await;
let duration = start.elapsed();
let status = if res.is_ok() { &"success" } else { &"failure" };
metrics
.remote_operation_time(task_kind, &file_kind, &op, status)
.observe(duration.as_secs_f64());
res
}
}
impl<T: Sized> MeasureRemoteOp for T {}
pin_project! {
pub(crate) struct MeasuredRemoteOp<F>
{
#[pin]
inner: F,
file_kind: RemoteOpFileKind,
op: RemoteOpKind,
start: Instant,
metrics: Arc<RemoteTimelineClientMetrics>,
}
}
impl<F: Future<Output = Result<O, E>>, O, E> Future for MeasuredRemoteOp<F> {
type Output = Result<O, E>;
fn poll(self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<Self::Output> {
let this = self.project();
let poll_result = this.inner.poll(cx);
if let Poll::Ready(ref res) = poll_result {
let duration = this.start.elapsed();
let status = if res.is_ok() { &"success" } else { &"failure" };
this.metrics
.remote_operation_time(this.file_kind, this.op, status)
.observe(duration.as_secs_f64());
}
poll_result
}
}
impl<Fut, O, E> MeasureRemoteOp<O, E> for Fut where Fut: Sized + Future<Output = Result<O, E>> {}
pub mod tokio_epoll_uring {
use std::collections::HashMap;
@@ -4220,4 +4304,5 @@ pub fn preinitialize_metrics(
Lazy::force(&tokio_epoll_uring::THREAD_LOCAL_METRICS_STORAGE);
tenant_throttling::preinitialize_global_metrics();
wait_ondemand_download_time::preinitialize_global_metrics();
}

View File

@@ -642,6 +642,7 @@ impl RemoteTimelineClient {
cancel,
)
.measure_remote_op(
Option::<TaskKind>::None,
RemoteOpFileKind::Index,
RemoteOpKind::Download,
Arc::clone(&self.metrics),
@@ -739,6 +740,7 @@ impl RemoteTimelineClient {
ctx,
)
.measure_remote_op(
Some(ctx.task_kind()),
RemoteOpFileKind::Layer,
RemoteOpKind::Download,
Arc::clone(&self.metrics),
@@ -2175,6 +2177,7 @@ impl RemoteTimelineClient {
&self.cancel,
)
.measure_remote_op(
Some(TaskKind::RemoteUploadTask),
RemoteOpFileKind::Layer,
RemoteOpKind::Upload,
Arc::clone(&self.metrics),
@@ -2191,6 +2194,7 @@ impl RemoteTimelineClient {
&self.cancel,
)
.measure_remote_op(
Some(TaskKind::RemoteUploadTask),
RemoteOpFileKind::Index,
RemoteOpKind::Upload,
Arc::clone(&self.metrics),

View File

@@ -975,6 +975,10 @@ impl LayerInner {
allow_download: bool,
ctx: &RequestContext,
) -> Result<Arc<DownloadedLayer>, DownloadError> {
let mut wait_for_download_recorder =
scopeguard::guard(utils::elapsed_accum::ElapsedAccum::default(), |accum| {
ctx.ondemand_download_wait_observe(accum.get());
});
let (weak, permit) = {
// get_or_init_detached can:
// - be fast (mutex lock) OR uncontested semaphore permit acquire
@@ -983,7 +987,7 @@ impl LayerInner {
let locked = self
.inner
.get_or_init_detached()
.get_or_init_detached_measured(Some(&mut wait_for_download_recorder))
.await
.map(|mut guard| guard.get_and_upgrade().ok_or(guard));
@@ -1013,6 +1017,7 @@ impl LayerInner {
Err(permit) => (None, permit),
}
};
let _guard = wait_for_download_recorder.guard();
if let Some(weak) = weak {
// only drop the weak after dropping the heavier_once_cell guard
@@ -1202,6 +1207,7 @@ impl LayerInner {
permit: heavier_once_cell::InitPermit,
ctx: &RequestContext,
) -> Result<Arc<DownloadedLayer>, remote_storage::DownloadError> {
let start = std::time::Instant::now();
let result = timeline
.remote_client
.download_layer_file(
@@ -1213,7 +1219,8 @@ impl LayerInner {
ctx,
)
.await;
let latency = start.elapsed();
let latency_millis = u64::try_from(latency.as_millis()).unwrap();
match result {
Ok(size) => {
assert_eq!(size, self.desc.file_size);
@@ -1229,9 +1236,8 @@ impl LayerInner {
Err(e) => {
panic!("post-condition failed: needs_download errored: {e:?}");
}
}
tracing::info!(size=%self.desc.file_size, "on-demand download successful");
};
tracing::info!(size=%self.desc.file_size, %latency_millis, "on-demand download successful");
timeline
.metrics
.resident_physical_size_add(self.desc.file_size);
@@ -1260,7 +1266,7 @@ impl LayerInner {
return Err(e);
}
tracing::error!(consecutive_failures, "layer file download failed: {e:#}");
tracing::error!(consecutive_failures, %latency_millis, "layer file download failed: {e:#}");
let backoff = utils::backoff::exponential_backoff_duration_seconds(
consecutive_failures.min(u32::MAX as usize) as u32,

View File

@@ -1,7 +1,7 @@
from __future__ import annotations
from collections import defaultdict
from typing import TYPE_CHECKING
from typing import TYPE_CHECKING, Literal
from prometheus_client.parser import text_string_to_metric_families
@@ -46,14 +46,26 @@ class MetricsGetter:
def get_metrics(self) -> Metrics:
raise NotImplementedError()
def get_metric_value(self, name: str, filter: dict[str, str] | None = None) -> float | None:
def get_metric_value(
self,
name: str,
filter: dict[str, str] | None = None,
aggregate: Literal["sum"] | None = None,
) -> float | None:
metrics = self.get_metrics()
results = metrics.query_all(name, filter=filter)
if not results:
log.info(f'could not find metric "{name}"')
return None
assert len(results) == 1, f"metric {name} with given filters is not unique, got: {results}"
return results[0].value
if aggregate is None:
assert len(results) == 1, (
f"metric {name} with given filters is not unique, got: {results}"
)
return results[0].value
elif aggregate == "sum":
return sum(sample.value for sample in results)
else:
raise RuntimeError(f"unknown aggregate function {aggregate}")
def get_metrics_values(
self, names: list[str], filter: dict[str, str] | None = None, absence_ok: bool = False
@@ -132,7 +144,7 @@ PAGESERVER_GLOBAL_METRICS: tuple[str, ...] = (
*[f"pageserver_basebackup_query_seconds_{x}" for x in ["bucket", "count", "sum"]],
*histogram("pageserver_smgr_query_seconds_global"),
*histogram("pageserver_wait_lsn_seconds"),
*histogram("pageserver_remote_operation_seconds"),
*histogram("pageserver_remote_timeline_client_seconds_global"),
*histogram("pageserver_io_operations_seconds"),
"pageserver_smgr_query_started_global_count_total",
"pageserver_tenant_states_count",
@@ -143,6 +155,7 @@ PAGESERVER_GLOBAL_METRICS: tuple[str, ...] = (
counter("pageserver_tenant_throttling_wait_usecs_sum_global"),
counter("pageserver_tenant_throttling_count_global"),
*histogram("pageserver_tokio_epoll_uring_slots_submission_queue_depth"),
*histogram("pageserver_wait_ondemand_download_seconds_global"),
)
PAGESERVER_PER_TENANT_METRICS: tuple[str, ...] = (
@@ -180,6 +193,7 @@ PAGESERVER_PER_TENANT_METRICS: tuple[str, ...] = (
counter("pageserver_wait_lsn_in_progress_micros"),
counter("pageserver_wait_lsn_started_count"),
counter("pageserver_wait_lsn_finished_count"),
counter("pageserver_wait_ondemand_download_seconds_sum"),
*histogram("pageserver_page_service_batch_size"),
*histogram("pageserver_page_service_pagestream_batch_wait_time_seconds"),
*PAGESERVER_PER_TENANT_REMOTE_TIMELINE_CLIENT_METRICS,

View File

@@ -126,7 +126,7 @@ def test_gc_index_upload(neon_env_builder: NeonEnvBuilder):
ps_metrics = env.pageserver.http_client().get_metrics()
total = 0.0
for sample in ps_metrics.query_all(
name="pageserver_remote_operation_seconds_count",
name="pageserver_remote_timeline_client_seconds_global_count",
filter={
"file_kind": str(file_kind),
"op_kind": str(op_kind),

View File

@@ -38,12 +38,13 @@ def get_num_downloaded_layers(client: PageserverHttpClient):
This assumes that the pageserver only has a single tenant.
"""
value = client.get_metric_value(
"pageserver_remote_operation_seconds_count",
"pageserver_remote_timeline_client_seconds_global_count",
{
"file_kind": "layer",
"op_kind": "download",
"status": "success",
},
"sum",
)
if value is None:
return 0

View File

@@ -107,7 +107,7 @@ def test_metric_collection(
ps_metrics = env.pageserver.http_client().get_metrics()
total = 0.0
for sample in ps_metrics.query_all(
name="pageserver_remote_operation_seconds_count",
name="pageserver_remote_timeline_client_seconds_global_count",
filter={
"file_kind": str(file_kind),
"op_kind": str(op_kind),