Line data Source code
1 : use enum_map::EnumMap;
2 : use metrics::{
3 : register_counter_vec, register_gauge_vec, register_histogram, register_histogram_vec,
4 : register_int_counter, register_int_counter_pair_vec, register_int_counter_vec,
5 : register_int_gauge, register_int_gauge_vec, register_uint_gauge, register_uint_gauge_vec,
6 : Counter, CounterVec, GaugeVec, Histogram, HistogramVec, IntCounter, IntCounterPair,
7 : IntCounterPairVec, IntCounterVec, IntGauge, IntGaugeVec, UIntGauge, UIntGaugeVec,
8 : };
9 : use once_cell::sync::Lazy;
10 : use pageserver_api::shard::TenantShardId;
11 : use strum::{EnumCount, VariantNames};
12 : use strum_macros::{IntoStaticStr, VariantNames};
13 : use tracing::warn;
14 : use utils::id::TimelineId;
15 :
16 : /// Prometheus histogram buckets (in seconds) for operations in the critical
17 : /// path. In other words, operations that directly affect that latency of user
18 : /// queries.
19 : ///
20 : /// The buckets capture the majority of latencies in the microsecond and
21 : /// millisecond range but also extend far enough up to distinguish "bad" from
22 : /// "really bad".
23 : const CRITICAL_OP_BUCKETS: &[f64] = &[
24 : 0.000_001, 0.000_010, 0.000_100, // 1 us, 10 us, 100 us
25 : 0.001_000, 0.010_000, 0.100_000, // 1 ms, 10 ms, 100 ms
26 : 1.0, 10.0, 100.0, // 1 s, 10 s, 100 s
27 : ];
28 :
29 : // Metrics collected on operations on the storage repository.
30 9936 : #[derive(Debug, VariantNames, IntoStaticStr)]
31 : #[strum(serialize_all = "kebab_case")]
32 : pub(crate) enum StorageTimeOperation {
33 : #[strum(serialize = "layer flush")]
34 : LayerFlush,
35 :
36 : #[strum(serialize = "compact")]
37 : Compact,
38 :
39 : #[strum(serialize = "create images")]
40 : CreateImages,
41 :
42 : #[strum(serialize = "logical size")]
43 : LogicalSize,
44 :
45 : #[strum(serialize = "imitate logical size")]
46 : ImitateLogicalSize,
47 :
48 : #[strum(serialize = "load layer map")]
49 : LoadLayerMap,
50 :
51 : #[strum(serialize = "gc")]
52 : Gc,
53 :
54 : #[strum(serialize = "find gc cutoffs")]
55 : FindGcCutoffs,
56 : }
57 :
58 510 : pub(crate) static STORAGE_TIME_SUM_PER_TIMELINE: Lazy<CounterVec> = Lazy::new(|| {
59 510 : register_counter_vec!(
60 510 : "pageserver_storage_operations_seconds_sum",
61 510 : "Total time spent on storage operations with operation, tenant and timeline dimensions",
62 510 : &["operation", "tenant_id", "shard_id", "timeline_id"],
63 510 : )
64 510 : .expect("failed to define a metric")
65 510 : });
66 :
67 510 : pub(crate) static STORAGE_TIME_COUNT_PER_TIMELINE: Lazy<IntCounterVec> = Lazy::new(|| {
68 510 : register_int_counter_vec!(
69 510 : "pageserver_storage_operations_seconds_count",
70 510 : "Count of storage operations with operation, tenant and timeline dimensions",
71 510 : &["operation", "tenant_id", "shard_id", "timeline_id"],
72 510 : )
73 510 : .expect("failed to define a metric")
74 510 : });
75 :
76 : // Buckets for background operations like compaction, GC, size calculation
77 : const STORAGE_OP_BUCKETS: &[f64] = &[0.010, 0.100, 1.0, 10.0, 100.0, 1000.0];
78 :
79 510 : pub(crate) static STORAGE_TIME_GLOBAL: Lazy<HistogramVec> = Lazy::new(|| {
80 510 : register_histogram_vec!(
81 510 : "pageserver_storage_operations_seconds_global",
82 510 : "Time spent on storage operations",
83 510 : &["operation"],
84 510 : STORAGE_OP_BUCKETS.into(),
85 510 : )
86 510 : .expect("failed to define a metric")
87 510 : });
88 :
89 0 : pub(crate) static READ_NUM_LAYERS_VISITED: Lazy<Histogram> = Lazy::new(|| {
90 0 : register_histogram!(
91 0 : "pageserver_layers_visited_per_read_global",
92 0 : "Number of layers visited to reconstruct one key",
93 0 : vec![1.0, 4.0, 8.0, 16.0, 32.0, 64.0, 128.0, 256.0, 512.0, 1024.0],
94 0 : )
95 0 : .expect("failed to define a metric")
96 0 : });
97 :
98 498 : pub(crate) static VEC_READ_NUM_LAYERS_VISITED: Lazy<Histogram> = Lazy::new(|| {
99 498 : register_histogram!(
100 498 : "pageserver_layers_visited_per_vectored_read_global",
101 498 : "Average number of layers visited to reconstruct one key",
102 498 : vec![1.0, 4.0, 8.0, 16.0, 32.0, 64.0, 128.0, 256.0, 512.0, 1024.0],
103 498 : )
104 498 : .expect("failed to define a metric")
105 498 : });
106 :
107 : // Metrics collected on operations on the storage repository.
108 : #[derive(
109 2016 : Clone, Copy, enum_map::Enum, strum_macros::EnumString, strum_macros::Display, IntoStaticStr,
110 : )]
111 : pub(crate) enum GetKind {
112 : Singular,
113 : Vectored,
114 : }
115 :
116 : pub(crate) struct ReconstructTimeMetrics {
117 : singular: Histogram,
118 : vectored: Histogram,
119 : }
120 :
121 504 : pub(crate) static RECONSTRUCT_TIME: Lazy<ReconstructTimeMetrics> = Lazy::new(|| {
122 504 : let inner = register_histogram_vec!(
123 504 : "pageserver_getpage_reconstruct_seconds",
124 504 : "Time spent in reconstruct_value (reconstruct a page from deltas)",
125 504 : &["get_kind"],
126 504 : CRITICAL_OP_BUCKETS.into(),
127 504 : )
128 504 : .expect("failed to define a metric");
129 504 :
130 504 : ReconstructTimeMetrics {
131 504 : singular: inner.with_label_values(&[GetKind::Singular.into()]),
132 504 : vectored: inner.with_label_values(&[GetKind::Vectored.into()]),
133 504 : }
134 504 : });
135 :
136 : impl ReconstructTimeMetrics {
137 1880463 : pub(crate) fn for_get_kind(&self, get_kind: GetKind) -> &Histogram {
138 1880463 : match get_kind {
139 1879131 : GetKind::Singular => &self.singular,
140 1332 : GetKind::Vectored => &self.vectored,
141 : }
142 1880463 : }
143 : }
144 :
145 : pub(crate) struct ReconstructDataTimeMetrics {
146 : singular: Histogram,
147 : vectored: Histogram,
148 : }
149 :
150 : impl ReconstructDataTimeMetrics {
151 1880511 : pub(crate) fn for_get_kind(&self, get_kind: GetKind) -> &Histogram {
152 1880511 : match get_kind {
153 1879179 : GetKind::Singular => &self.singular,
154 1332 : GetKind::Vectored => &self.vectored,
155 : }
156 1880511 : }
157 : }
158 :
159 504 : pub(crate) static GET_RECONSTRUCT_DATA_TIME: Lazy<ReconstructDataTimeMetrics> = Lazy::new(|| {
160 504 : let inner = register_histogram_vec!(
161 504 : "pageserver_getpage_get_reconstruct_data_seconds",
162 504 : "Time spent in get_reconstruct_value_data",
163 504 : &["get_kind"],
164 504 : CRITICAL_OP_BUCKETS.into(),
165 504 : )
166 504 : .expect("failed to define a metric");
167 504 :
168 504 : ReconstructDataTimeMetrics {
169 504 : singular: inner.with_label_values(&[GetKind::Singular.into()]),
170 504 : vectored: inner.with_label_values(&[GetKind::Vectored.into()]),
171 504 : }
172 504 : });
173 :
174 : pub(crate) struct GetVectoredLatency {
175 : map: EnumMap<TaskKind, Option<Histogram>>,
176 : }
177 :
178 : #[allow(dead_code)]
179 : pub(crate) struct ScanLatency {
180 : map: EnumMap<TaskKind, Option<Histogram>>,
181 : }
182 :
183 : impl GetVectoredLatency {
184 : // Only these task types perform vectored gets. Filter all other tasks out to reduce total
185 : // cardinality of the metric.
186 : const TRACKED_TASK_KINDS: [TaskKind; 2] = [TaskKind::Compaction, TaskKind::PageRequestHandler];
187 :
188 3396 : pub(crate) fn for_task_kind(&self, task_kind: TaskKind) -> Option<&Histogram> {
189 3396 : self.map[task_kind].as_ref()
190 3396 : }
191 : }
192 :
193 : impl ScanLatency {
194 : // Only these task types perform vectored gets. Filter all other tasks out to reduce total
195 : // cardinality of the metric.
196 : const TRACKED_TASK_KINDS: [TaskKind; 1] = [TaskKind::PageRequestHandler];
197 :
198 72 : pub(crate) fn for_task_kind(&self, task_kind: TaskKind) -> Option<&Histogram> {
199 72 : self.map[task_kind].as_ref()
200 72 : }
201 : }
202 :
203 : pub(crate) struct ScanLatencyOngoingRecording<'a> {
204 : parent: &'a Histogram,
205 : start: std::time::Instant,
206 : }
207 :
208 : impl<'a> ScanLatencyOngoingRecording<'a> {
209 0 : pub(crate) fn start_recording(parent: &'a Histogram) -> ScanLatencyOngoingRecording<'a> {
210 0 : let start = Instant::now();
211 0 : ScanLatencyOngoingRecording { parent, start }
212 0 : }
213 :
214 0 : pub(crate) fn observe(self, throttled: Option<Duration>) {
215 0 : let elapsed = self.start.elapsed();
216 0 : let ex_throttled = if let Some(throttled) = throttled {
217 0 : elapsed.checked_sub(throttled)
218 : } else {
219 0 : Some(elapsed)
220 : };
221 0 : if let Some(ex_throttled) = ex_throttled {
222 0 : self.parent.observe(ex_throttled.as_secs_f64());
223 0 : } else {
224 0 : use utils::rate_limit::RateLimit;
225 0 : static LOGGED: Lazy<Mutex<RateLimit>> =
226 0 : Lazy::new(|| Mutex::new(RateLimit::new(Duration::from_secs(10))));
227 0 : let mut rate_limit = LOGGED.lock().unwrap();
228 0 : rate_limit.call(|| {
229 0 : warn!("error deducting time spent throttled; this message is logged at a global rate limit");
230 0 : });
231 0 : }
232 0 : }
233 : }
234 :
235 486 : pub(crate) static GET_VECTORED_LATENCY: Lazy<GetVectoredLatency> = Lazy::new(|| {
236 486 : let inner = register_histogram_vec!(
237 486 : "pageserver_get_vectored_seconds",
238 486 : "Time spent in get_vectored, excluding time spent in timeline_get_throttle.",
239 486 : &["task_kind"],
240 486 : CRITICAL_OP_BUCKETS.into(),
241 486 : )
242 486 : .expect("failed to define a metric");
243 486 :
244 486 : GetVectoredLatency {
245 14580 : map: EnumMap::from_array(std::array::from_fn(|task_kind_idx| {
246 14580 : let task_kind = <TaskKind as enum_map::Enum>::from_usize(task_kind_idx);
247 14580 :
248 14580 : if GetVectoredLatency::TRACKED_TASK_KINDS.contains(&task_kind) {
249 972 : let task_kind = task_kind.into();
250 972 : Some(inner.with_label_values(&[task_kind]))
251 : } else {
252 13608 : None
253 : }
254 14580 : })),
255 486 : }
256 486 : });
257 :
258 24 : pub(crate) static SCAN_LATENCY: Lazy<ScanLatency> = Lazy::new(|| {
259 24 : let inner = register_histogram_vec!(
260 24 : "pageserver_scan_seconds",
261 24 : "Time spent in scan, excluding time spent in timeline_get_throttle.",
262 24 : &["task_kind"],
263 24 : CRITICAL_OP_BUCKETS.into(),
264 24 : )
265 24 : .expect("failed to define a metric");
266 24 :
267 24 : ScanLatency {
268 720 : map: EnumMap::from_array(std::array::from_fn(|task_kind_idx| {
269 720 : let task_kind = <TaskKind as enum_map::Enum>::from_usize(task_kind_idx);
270 720 :
271 720 : if ScanLatency::TRACKED_TASK_KINDS.contains(&task_kind) {
272 24 : let task_kind = task_kind.into();
273 24 : Some(inner.with_label_values(&[task_kind]))
274 : } else {
275 696 : None
276 : }
277 720 : })),
278 24 : }
279 24 : });
280 :
281 : pub(crate) struct PageCacheMetricsForTaskKind {
282 : pub read_accesses_immutable: IntCounter,
283 : pub read_hits_immutable: IntCounter,
284 : }
285 :
286 : pub(crate) struct PageCacheMetrics {
287 : map: EnumMap<TaskKind, EnumMap<PageContentKind, PageCacheMetricsForTaskKind>>,
288 : }
289 :
290 258 : static PAGE_CACHE_READ_HITS: Lazy<IntCounterVec> = Lazy::new(|| {
291 258 : register_int_counter_vec!(
292 258 : "pageserver_page_cache_read_hits_total",
293 258 : "Number of read accesses to the page cache that hit",
294 258 : &["task_kind", "key_kind", "content_kind", "hit_kind"]
295 258 : )
296 258 : .expect("failed to define a metric")
297 258 : });
298 :
299 258 : static PAGE_CACHE_READ_ACCESSES: Lazy<IntCounterVec> = Lazy::new(|| {
300 258 : register_int_counter_vec!(
301 258 : "pageserver_page_cache_read_accesses_total",
302 258 : "Number of read accesses to the page cache",
303 258 : &["task_kind", "key_kind", "content_kind"]
304 258 : )
305 258 : .expect("failed to define a metric")
306 258 : });
307 :
308 258 : pub(crate) static PAGE_CACHE: Lazy<PageCacheMetrics> = Lazy::new(|| PageCacheMetrics {
309 7740 : map: EnumMap::from_array(std::array::from_fn(|task_kind| {
310 7740 : let task_kind = <TaskKind as enum_map::Enum>::from_usize(task_kind);
311 7740 : let task_kind: &'static str = task_kind.into();
312 61920 : EnumMap::from_array(std::array::from_fn(|content_kind| {
313 61920 : let content_kind = <PageContentKind as enum_map::Enum>::from_usize(content_kind);
314 61920 : let content_kind: &'static str = content_kind.into();
315 61920 : PageCacheMetricsForTaskKind {
316 61920 : read_accesses_immutable: {
317 61920 : PAGE_CACHE_READ_ACCESSES
318 61920 : .get_metric_with_label_values(&[task_kind, "immutable", content_kind])
319 61920 : .unwrap()
320 61920 : },
321 61920 :
322 61920 : read_hits_immutable: {
323 61920 : PAGE_CACHE_READ_HITS
324 61920 : .get_metric_with_label_values(&[task_kind, "immutable", content_kind, "-"])
325 61920 : .unwrap()
326 61920 : },
327 61920 : }
328 61920 : }))
329 7740 : })),
330 258 : });
331 :
332 : impl PageCacheMetrics {
333 3310222 : pub(crate) fn for_ctx(&self, ctx: &RequestContext) -> &PageCacheMetricsForTaskKind {
334 3310222 : &self.map[ctx.task_kind()][ctx.page_content_kind()]
335 3310222 : }
336 : }
337 :
338 : pub(crate) struct PageCacheSizeMetrics {
339 : pub max_bytes: UIntGauge,
340 :
341 : pub current_bytes_immutable: UIntGauge,
342 : }
343 :
344 258 : static PAGE_CACHE_SIZE_CURRENT_BYTES: Lazy<UIntGaugeVec> = Lazy::new(|| {
345 258 : register_uint_gauge_vec!(
346 258 : "pageserver_page_cache_size_current_bytes",
347 258 : "Current size of the page cache in bytes, by key kind",
348 258 : &["key_kind"]
349 258 : )
350 258 : .expect("failed to define a metric")
351 258 : });
352 :
353 : pub(crate) static PAGE_CACHE_SIZE: Lazy<PageCacheSizeMetrics> =
354 258 : Lazy::new(|| PageCacheSizeMetrics {
355 258 : max_bytes: {
356 258 : register_uint_gauge!(
357 258 : "pageserver_page_cache_size_max_bytes",
358 258 : "Maximum size of the page cache in bytes"
359 258 : )
360 258 : .expect("failed to define a metric")
361 258 : },
362 258 : current_bytes_immutable: {
363 258 : PAGE_CACHE_SIZE_CURRENT_BYTES
364 258 : .get_metric_with_label_values(&["immutable"])
365 258 : .unwrap()
366 258 : },
367 258 : });
368 :
369 : pub(crate) mod page_cache_eviction_metrics {
370 : use std::num::NonZeroUsize;
371 :
372 : use metrics::{register_int_counter_vec, IntCounter, IntCounterVec};
373 : use once_cell::sync::Lazy;
374 :
375 : #[derive(Clone, Copy)]
376 : pub(crate) enum Outcome {
377 : FoundSlotUnused { iters: NonZeroUsize },
378 : FoundSlotEvicted { iters: NonZeroUsize },
379 : ItersExceeded { iters: NonZeroUsize },
380 : }
381 :
382 258 : static ITERS_TOTAL_VEC: Lazy<IntCounterVec> = Lazy::new(|| {
383 258 : register_int_counter_vec!(
384 258 : "pageserver_page_cache_find_victim_iters_total",
385 258 : "Counter for the number of iterations in the find_victim loop",
386 258 : &["outcome"],
387 258 : )
388 258 : .expect("failed to define a metric")
389 258 : });
390 :
391 258 : static CALLS_VEC: Lazy<IntCounterVec> = Lazy::new(|| {
392 258 : register_int_counter_vec!(
393 258 : "pageserver_page_cache_find_victim_calls",
394 258 : "Incremented at the end of each find_victim() call.\
395 258 : Filter by outcome to get e.g., eviction rate.",
396 258 : &["outcome"]
397 258 : )
398 258 : .unwrap()
399 258 : });
400 :
401 96463 : pub(crate) fn observe(outcome: Outcome) {
402 : macro_rules! dry {
403 : ($label:literal, $iters:expr) => {{
404 : static LABEL: &'static str = $label;
405 : static ITERS_TOTAL: Lazy<IntCounter> =
406 318 : Lazy::new(|| ITERS_TOTAL_VEC.with_label_values(&[LABEL]));
407 : static CALLS: Lazy<IntCounter> =
408 318 : Lazy::new(|| CALLS_VEC.with_label_values(&[LABEL]));
409 : ITERS_TOTAL.inc_by(($iters.get()) as u64);
410 : CALLS.inc();
411 : }};
412 : }
413 96463 : match outcome {
414 4440 : Outcome::FoundSlotUnused { iters } => dry!("found_empty", iters),
415 92023 : Outcome::FoundSlotEvicted { iters } => {
416 92023 : dry!("found_evicted", iters)
417 : }
418 0 : Outcome::ItersExceeded { iters } => {
419 0 : dry!("err_iters_exceeded", iters);
420 0 : super::page_cache_errors_inc(super::PageCacheErrorKind::EvictIterLimit);
421 0 : }
422 : }
423 96463 : }
424 : }
425 :
426 0 : static PAGE_CACHE_ERRORS: Lazy<IntCounterVec> = Lazy::new(|| {
427 0 : register_int_counter_vec!(
428 0 : "page_cache_errors_total",
429 0 : "Number of timeouts while acquiring a pinned slot in the page cache",
430 0 : &["error_kind"]
431 0 : )
432 0 : .expect("failed to define a metric")
433 0 : });
434 :
435 0 : #[derive(IntoStaticStr)]
436 : #[strum(serialize_all = "kebab_case")]
437 : pub(crate) enum PageCacheErrorKind {
438 : AcquirePinnedSlotTimeout,
439 : EvictIterLimit,
440 : }
441 :
442 0 : pub(crate) fn page_cache_errors_inc(error_kind: PageCacheErrorKind) {
443 0 : PAGE_CACHE_ERRORS
444 0 : .get_metric_with_label_values(&[error_kind.into()])
445 0 : .unwrap()
446 0 : .inc();
447 0 : }
448 :
449 54 : pub(crate) static WAIT_LSN_TIME: Lazy<Histogram> = Lazy::new(|| {
450 54 : register_histogram!(
451 54 : "pageserver_wait_lsn_seconds",
452 54 : "Time spent waiting for WAL to arrive",
453 54 : CRITICAL_OP_BUCKETS.into(),
454 54 : )
455 54 : .expect("failed to define a metric")
456 54 : });
457 :
458 510 : static LAST_RECORD_LSN: Lazy<IntGaugeVec> = Lazy::new(|| {
459 510 : register_int_gauge_vec!(
460 510 : "pageserver_last_record_lsn",
461 510 : "Last record LSN grouped by timeline",
462 510 : &["tenant_id", "shard_id", "timeline_id"]
463 510 : )
464 510 : .expect("failed to define a metric")
465 510 : });
466 :
467 510 : static PITR_HISTORY_SIZE: Lazy<UIntGaugeVec> = Lazy::new(|| {
468 510 : register_uint_gauge_vec!(
469 510 : "pageserver_pitr_history_size",
470 510 : "Data written since PITR cutoff on this timeline",
471 510 : &["tenant_id", "shard_id", "timeline_id"]
472 510 : )
473 510 : .expect("failed to define a metric")
474 510 : });
475 :
476 5064 : #[derive(strum_macros::EnumString, strum_macros::Display, strum_macros::IntoStaticStr)]
477 : #[strum(serialize_all = "kebab_case")]
478 : pub(crate) enum MetricLayerKind {
479 : Delta,
480 : Image,
481 : }
482 :
483 510 : static TIMELINE_LAYER_SIZE: Lazy<UIntGaugeVec> = Lazy::new(|| {
484 510 : register_uint_gauge_vec!(
485 510 : "pageserver_layer_bytes",
486 510 : "Sum of layer physical sizes in bytes",
487 510 : &["tenant_id", "shard_id", "timeline_id", "kind"]
488 510 : )
489 510 : .expect("failed to define a metric")
490 510 : });
491 :
492 510 : static TIMELINE_LAYER_COUNT: Lazy<UIntGaugeVec> = Lazy::new(|| {
493 510 : register_uint_gauge_vec!(
494 510 : "pageserver_layer_count",
495 510 : "Number of layers that exist",
496 510 : &["tenant_id", "shard_id", "timeline_id", "kind"]
497 510 : )
498 510 : .expect("failed to define a metric")
499 510 : });
500 :
501 510 : static TIMELINE_ARCHIVE_SIZE: Lazy<UIntGaugeVec> = Lazy::new(|| {
502 510 : register_uint_gauge_vec!(
503 510 : "pageserver_archive_size",
504 510 : "Timeline's logical size if it is considered eligible for archival (outside PITR window), else zero",
505 510 : &["tenant_id", "shard_id", "timeline_id"]
506 510 : )
507 510 : .expect("failed to define a metric")
508 510 : });
509 :
510 510 : static STANDBY_HORIZON: Lazy<IntGaugeVec> = Lazy::new(|| {
511 510 : register_int_gauge_vec!(
512 510 : "pageserver_standby_horizon",
513 510 : "Standby apply LSN for which GC is hold off, by timeline.",
514 510 : &["tenant_id", "shard_id", "timeline_id"]
515 510 : )
516 510 : .expect("failed to define a metric")
517 510 : });
518 :
519 510 : static RESIDENT_PHYSICAL_SIZE: Lazy<UIntGaugeVec> = Lazy::new(|| {
520 510 : register_uint_gauge_vec!(
521 510 : "pageserver_resident_physical_size",
522 510 : "The size of the layer files present in the pageserver's filesystem, for attached locations.",
523 510 : &["tenant_id", "shard_id", "timeline_id"]
524 510 : )
525 510 : .expect("failed to define a metric")
526 510 : });
527 :
528 510 : static VISIBLE_PHYSICAL_SIZE: Lazy<UIntGaugeVec> = Lazy::new(|| {
529 510 : register_uint_gauge_vec!(
530 510 : "pageserver_visible_physical_size",
531 510 : "The size of the layer files present in the pageserver's filesystem.",
532 510 : &["tenant_id", "shard_id", "timeline_id"]
533 510 : )
534 510 : .expect("failed to define a metric")
535 510 : });
536 :
537 498 : pub(crate) static RESIDENT_PHYSICAL_SIZE_GLOBAL: Lazy<UIntGauge> = Lazy::new(|| {
538 498 : register_uint_gauge!(
539 498 : "pageserver_resident_physical_size_global",
540 498 : "Like `pageserver_resident_physical_size`, but without tenant/timeline dimensions."
541 498 : )
542 498 : .expect("failed to define a metric")
543 498 : });
544 :
545 510 : static REMOTE_PHYSICAL_SIZE: Lazy<UIntGaugeVec> = Lazy::new(|| {
546 510 : register_uint_gauge_vec!(
547 510 : "pageserver_remote_physical_size",
548 510 : "The size of the layer files present in the remote storage that are listed in the remote index_part.json.",
549 510 : // Corollary: If any files are missing from the index part, they won't be included here.
550 510 : &["tenant_id", "shard_id", "timeline_id"]
551 510 : )
552 510 : .expect("failed to define a metric")
553 510 : });
554 :
555 510 : static REMOTE_PHYSICAL_SIZE_GLOBAL: Lazy<UIntGauge> = Lazy::new(|| {
556 510 : register_uint_gauge!(
557 510 : "pageserver_remote_physical_size_global",
558 510 : "Like `pageserver_remote_physical_size`, but without tenant/timeline dimensions."
559 510 : )
560 510 : .expect("failed to define a metric")
561 510 : });
562 :
563 12 : pub(crate) static REMOTE_ONDEMAND_DOWNLOADED_LAYERS: Lazy<IntCounter> = Lazy::new(|| {
564 12 : register_int_counter!(
565 12 : "pageserver_remote_ondemand_downloaded_layers_total",
566 12 : "Total on-demand downloaded layers"
567 12 : )
568 12 : .unwrap()
569 12 : });
570 :
571 12 : pub(crate) static REMOTE_ONDEMAND_DOWNLOADED_BYTES: Lazy<IntCounter> = Lazy::new(|| {
572 12 : register_int_counter!(
573 12 : "pageserver_remote_ondemand_downloaded_bytes_total",
574 12 : "Total bytes of layers on-demand downloaded",
575 12 : )
576 12 : .unwrap()
577 12 : });
578 :
579 510 : static CURRENT_LOGICAL_SIZE: Lazy<UIntGaugeVec> = Lazy::new(|| {
580 510 : register_uint_gauge_vec!(
581 510 : "pageserver_current_logical_size",
582 510 : "Current logical size grouped by timeline",
583 510 : &["tenant_id", "shard_id", "timeline_id"]
584 510 : )
585 510 : .expect("failed to define current logical size metric")
586 510 : });
587 :
588 510 : static AUX_FILE_SIZE: Lazy<IntGaugeVec> = Lazy::new(|| {
589 510 : register_int_gauge_vec!(
590 510 : "pageserver_aux_file_estimated_size",
591 510 : "The size of all aux files for a timeline in aux file v2 store.",
592 510 : &["tenant_id", "shard_id", "timeline_id"]
593 510 : )
594 510 : .expect("failed to define a metric")
595 510 : });
596 :
597 510 : static VALID_LSN_LEASE_COUNT: Lazy<UIntGaugeVec> = Lazy::new(|| {
598 510 : register_uint_gauge_vec!(
599 510 : "pageserver_valid_lsn_lease_count",
600 510 : "The number of valid leases after refreshing gc info.",
601 510 : &["tenant_id", "shard_id", "timeline_id"],
602 510 : )
603 510 : .expect("failed to define a metric")
604 510 : });
605 :
606 0 : pub(crate) static CIRCUIT_BREAKERS_BROKEN: Lazy<IntCounter> = Lazy::new(|| {
607 0 : register_int_counter!(
608 0 : "pageserver_circuit_breaker_broken",
609 0 : "How many times a circuit breaker has broken"
610 0 : )
611 0 : .expect("failed to define a metric")
612 0 : });
613 :
614 0 : pub(crate) static CIRCUIT_BREAKERS_UNBROKEN: Lazy<IntCounter> = Lazy::new(|| {
615 0 : register_int_counter!(
616 0 : "pageserver_circuit_breaker_unbroken",
617 0 : "How many times a circuit breaker has been un-broken (recovered)"
618 0 : )
619 0 : .expect("failed to define a metric")
620 0 : });
621 :
622 486 : pub(crate) static COMPRESSION_IMAGE_INPUT_BYTES: Lazy<IntCounter> = Lazy::new(|| {
623 486 : register_int_counter!(
624 486 : "pageserver_compression_image_in_bytes_total",
625 486 : "Size of data written into image layers before compression"
626 486 : )
627 486 : .expect("failed to define a metric")
628 486 : });
629 :
630 486 : pub(crate) static COMPRESSION_IMAGE_INPUT_BYTES_CONSIDERED: Lazy<IntCounter> = Lazy::new(|| {
631 486 : register_int_counter!(
632 486 : "pageserver_compression_image_in_bytes_considered",
633 486 : "Size of potentially compressible data written into image layers before compression"
634 486 : )
635 486 : .expect("failed to define a metric")
636 486 : });
637 :
638 486 : pub(crate) static COMPRESSION_IMAGE_INPUT_BYTES_CHOSEN: Lazy<IntCounter> = Lazy::new(|| {
639 486 : register_int_counter!(
640 486 : "pageserver_compression_image_in_bytes_chosen",
641 486 : "Size of data whose compressed form was written into image layers"
642 486 : )
643 486 : .expect("failed to define a metric")
644 486 : });
645 :
646 486 : pub(crate) static COMPRESSION_IMAGE_OUTPUT_BYTES: Lazy<IntCounter> = Lazy::new(|| {
647 486 : register_int_counter!(
648 486 : "pageserver_compression_image_out_bytes_total",
649 486 : "Size of compressed image layer written"
650 486 : )
651 486 : .expect("failed to define a metric")
652 486 : });
653 :
654 : pub(crate) mod initial_logical_size {
655 : use metrics::{register_int_counter, register_int_counter_vec, IntCounter, IntCounterVec};
656 : use once_cell::sync::Lazy;
657 :
658 : pub(crate) struct StartCalculation(IntCounterVec);
659 510 : pub(crate) static START_CALCULATION: Lazy<StartCalculation> = Lazy::new(|| {
660 510 : StartCalculation(
661 510 : register_int_counter_vec!(
662 510 : "pageserver_initial_logical_size_start_calculation",
663 510 : "Incremented each time we start an initial logical size calculation attempt. \
664 510 : The `circumstances` label provides some additional details.",
665 510 : &["attempt", "circumstances"]
666 510 : )
667 510 : .unwrap(),
668 510 : )
669 510 : });
670 :
671 : struct DropCalculation {
672 : first: IntCounter,
673 : retry: IntCounter,
674 : }
675 :
676 510 : static DROP_CALCULATION: Lazy<DropCalculation> = Lazy::new(|| {
677 510 : let vec = register_int_counter_vec!(
678 510 : "pageserver_initial_logical_size_drop_calculation",
679 510 : "Incremented each time we abort a started size calculation attmpt.",
680 510 : &["attempt"]
681 510 : )
682 510 : .unwrap();
683 510 : DropCalculation {
684 510 : first: vec.with_label_values(&["first"]),
685 510 : retry: vec.with_label_values(&["retry"]),
686 510 : }
687 510 : });
688 :
689 : pub(crate) struct Calculated {
690 : pub(crate) births: IntCounter,
691 : pub(crate) deaths: IntCounter,
692 : }
693 :
694 510 : pub(crate) static CALCULATED: Lazy<Calculated> = Lazy::new(|| Calculated {
695 510 : births: register_int_counter!(
696 510 : "pageserver_initial_logical_size_finish_calculation",
697 510 : "Incremented every time we finish calculation of initial logical size.\
698 510 : If everything is working well, this should happen at most once per Timeline object."
699 510 : )
700 510 : .unwrap(),
701 510 : deaths: register_int_counter!(
702 510 : "pageserver_initial_logical_size_drop_finished_calculation",
703 510 : "Incremented when we drop a finished initial logical size calculation result.\
704 510 : Mainly useful to turn pageserver_initial_logical_size_finish_calculation into a gauge."
705 510 : )
706 510 : .unwrap(),
707 510 : });
708 :
709 : pub(crate) struct OngoingCalculationGuard {
710 : inc_drop_calculation: Option<IntCounter>,
711 : }
712 :
713 546 : #[derive(strum_macros::IntoStaticStr)]
714 : pub(crate) enum StartCircumstances {
715 : EmptyInitial,
716 : SkippedConcurrencyLimiter,
717 : AfterBackgroundTasksRateLimit,
718 : }
719 :
720 : impl StartCalculation {
721 546 : pub(crate) fn first(&self, circumstances: StartCircumstances) -> OngoingCalculationGuard {
722 546 : let circumstances_label: &'static str = circumstances.into();
723 546 : self.0
724 546 : .with_label_values(&["first", circumstances_label])
725 546 : .inc();
726 546 : OngoingCalculationGuard {
727 546 : inc_drop_calculation: Some(DROP_CALCULATION.first.clone()),
728 546 : }
729 546 : }
730 0 : pub(crate) fn retry(&self, circumstances: StartCircumstances) -> OngoingCalculationGuard {
731 0 : let circumstances_label: &'static str = circumstances.into();
732 0 : self.0
733 0 : .with_label_values(&["retry", circumstances_label])
734 0 : .inc();
735 0 : OngoingCalculationGuard {
736 0 : inc_drop_calculation: Some(DROP_CALCULATION.retry.clone()),
737 0 : }
738 0 : }
739 : }
740 :
741 : impl Drop for OngoingCalculationGuard {
742 546 : fn drop(&mut self) {
743 546 : if let Some(counter) = self.inc_drop_calculation.take() {
744 0 : counter.inc();
745 546 : }
746 546 : }
747 : }
748 :
749 : impl OngoingCalculationGuard {
750 546 : pub(crate) fn calculation_result_saved(mut self) -> FinishedCalculationGuard {
751 546 : drop(self.inc_drop_calculation.take());
752 546 : CALCULATED.births.inc();
753 546 : FinishedCalculationGuard {
754 546 : inc_on_drop: CALCULATED.deaths.clone(),
755 546 : }
756 546 : }
757 : }
758 :
759 : pub(crate) struct FinishedCalculationGuard {
760 : inc_on_drop: IntCounter,
761 : }
762 :
763 : impl Drop for FinishedCalculationGuard {
764 18 : fn drop(&mut self) {
765 18 : self.inc_on_drop.inc();
766 18 : }
767 : }
768 :
769 : // context: https://github.com/neondatabase/neon/issues/5963
770 : pub(crate) static TIMELINES_WHERE_WALRECEIVER_GOT_APPROXIMATE_SIZE: Lazy<IntCounter> =
771 0 : Lazy::new(|| {
772 0 : register_int_counter!(
773 0 : "pageserver_initial_logical_size_timelines_where_walreceiver_got_approximate_size",
774 0 : "Counter for the following event: walreceiver calls\
775 0 : Timeline::get_current_logical_size() and it returns `Approximate` for the first time."
776 0 : )
777 0 : .unwrap()
778 0 : });
779 : }
780 :
781 0 : static DIRECTORY_ENTRIES_COUNT: Lazy<UIntGaugeVec> = Lazy::new(|| {
782 0 : register_uint_gauge_vec!(
783 0 : "pageserver_directory_entries_count",
784 0 : "Sum of the entries in pageserver-stored directory listings",
785 0 : &["tenant_id", "shard_id", "timeline_id"]
786 0 : )
787 0 : .expect("failed to define a metric")
788 0 : });
789 :
790 516 : pub(crate) static TENANT_STATE_METRIC: Lazy<UIntGaugeVec> = Lazy::new(|| {
791 516 : register_uint_gauge_vec!(
792 516 : "pageserver_tenant_states_count",
793 516 : "Count of tenants per state",
794 516 : &["state"]
795 516 : )
796 516 : .expect("Failed to register pageserver_tenant_states_count metric")
797 516 : });
798 :
799 : /// A set of broken tenants.
800 : ///
801 : /// These are expected to be so rare that a set is fine. Set as in a new timeseries per each broken
802 : /// tenant.
803 36 : pub(crate) static BROKEN_TENANTS_SET: Lazy<UIntGaugeVec> = Lazy::new(|| {
804 36 : register_uint_gauge_vec!(
805 36 : "pageserver_broken_tenants_count",
806 36 : "Set of broken tenants",
807 36 : &["tenant_id", "shard_id"]
808 36 : )
809 36 : .expect("Failed to register pageserver_tenant_states_count metric")
810 36 : });
811 :
812 18 : pub(crate) static TENANT_SYNTHETIC_SIZE_METRIC: Lazy<UIntGaugeVec> = Lazy::new(|| {
813 18 : register_uint_gauge_vec!(
814 18 : "pageserver_tenant_synthetic_cached_size_bytes",
815 18 : "Synthetic size of each tenant in bytes",
816 18 : &["tenant_id"]
817 18 : )
818 18 : .expect("Failed to register pageserver_tenant_synthetic_cached_size_bytes metric")
819 18 : });
820 :
821 0 : pub(crate) static EVICTION_ITERATION_DURATION: Lazy<HistogramVec> = Lazy::new(|| {
822 0 : register_histogram_vec!(
823 0 : "pageserver_eviction_iteration_duration_seconds_global",
824 0 : "Time spent on a single eviction iteration",
825 0 : &["period_secs", "threshold_secs"],
826 0 : STORAGE_OP_BUCKETS.into(),
827 0 : )
828 0 : .expect("failed to define a metric")
829 0 : });
830 :
831 510 : static EVICTIONS: Lazy<IntCounterVec> = Lazy::new(|| {
832 510 : register_int_counter_vec!(
833 510 : "pageserver_evictions",
834 510 : "Number of layers evicted from the pageserver",
835 510 : &["tenant_id", "shard_id", "timeline_id"]
836 510 : )
837 510 : .expect("failed to define a metric")
838 510 : });
839 :
840 510 : static EVICTIONS_WITH_LOW_RESIDENCE_DURATION: Lazy<IntCounterVec> = Lazy::new(|| {
841 510 : register_int_counter_vec!(
842 510 : "pageserver_evictions_with_low_residence_duration",
843 510 : "If a layer is evicted that was resident for less than `low_threshold`, it is counted to this counter. \
844 510 : Residence duration is determined using the `residence_duration_data_source`.",
845 510 : &["tenant_id", "shard_id", "timeline_id", "residence_duration_data_source", "low_threshold_secs"]
846 510 : )
847 510 : .expect("failed to define a metric")
848 510 : });
849 :
850 0 : pub(crate) static UNEXPECTED_ONDEMAND_DOWNLOADS: Lazy<IntCounter> = Lazy::new(|| {
851 0 : register_int_counter!(
852 0 : "pageserver_unexpected_ondemand_downloads_count",
853 0 : "Number of unexpected on-demand downloads. \
854 0 : We log more context for each increment, so, forgo any labels in this metric.",
855 0 : )
856 0 : .expect("failed to define a metric")
857 0 : });
858 :
859 : /// How long did we take to start up? Broken down by labels to describe
860 : /// different phases of startup.
861 0 : pub static STARTUP_DURATION: Lazy<GaugeVec> = Lazy::new(|| {
862 0 : register_gauge_vec!(
863 0 : "pageserver_startup_duration_seconds",
864 0 : "Time taken by phases of pageserver startup, in seconds",
865 0 : &["phase"]
866 0 : )
867 0 : .expect("Failed to register pageserver_startup_duration_seconds metric")
868 0 : });
869 :
870 0 : pub static STARTUP_IS_LOADING: Lazy<UIntGauge> = Lazy::new(|| {
871 0 : register_uint_gauge!(
872 0 : "pageserver_startup_is_loading",
873 0 : "1 while in initial startup load of tenants, 0 at other times"
874 0 : )
875 0 : .expect("Failed to register pageserver_startup_is_loading")
876 0 : });
877 :
878 498 : pub(crate) static TIMELINE_EPHEMERAL_BYTES: Lazy<UIntGauge> = Lazy::new(|| {
879 498 : register_uint_gauge!(
880 498 : "pageserver_timeline_ephemeral_bytes",
881 498 : "Total number of bytes in ephemeral layers, summed for all timelines. Approximate, lazily updated."
882 498 : )
883 498 : .expect("Failed to register metric")
884 498 : });
885 :
886 : /// Metrics related to the lifecycle of a [`crate::tenant::Tenant`] object: things
887 : /// like how long it took to load.
888 : ///
889 : /// Note that these are process-global metrics, _not_ per-tenant metrics. Per-tenant
890 : /// metrics are rather expensive, and usually fine grained stuff makes more sense
891 : /// at a timeline level than tenant level.
892 : pub(crate) struct TenantMetrics {
893 : /// How long did tenants take to go from construction to active state?
894 : pub(crate) activation: Histogram,
895 : pub(crate) preload: Histogram,
896 : pub(crate) attach: Histogram,
897 :
898 : /// How many tenants are included in the initial startup of the pagesrever?
899 : pub(crate) startup_scheduled: IntCounter,
900 : pub(crate) startup_complete: IntCounter,
901 : }
902 :
903 0 : pub(crate) static TENANT: Lazy<TenantMetrics> = Lazy::new(|| {
904 0 : TenantMetrics {
905 0 : activation: register_histogram!(
906 0 : "pageserver_tenant_activation_seconds",
907 0 : "Time taken by tenants to activate, in seconds",
908 0 : CRITICAL_OP_BUCKETS.into()
909 0 : )
910 0 : .expect("Failed to register metric"),
911 0 : preload: register_histogram!(
912 0 : "pageserver_tenant_preload_seconds",
913 0 : "Time taken by tenants to load remote metadata on startup/attach, in seconds",
914 0 : CRITICAL_OP_BUCKETS.into()
915 0 : )
916 0 : .expect("Failed to register metric"),
917 0 : attach: register_histogram!(
918 0 : "pageserver_tenant_attach_seconds",
919 0 : "Time taken by tenants to intialize, after remote metadata is already loaded",
920 0 : CRITICAL_OP_BUCKETS.into()
921 0 : )
922 0 : .expect("Failed to register metric"),
923 0 : startup_scheduled: register_int_counter!(
924 0 : "pageserver_tenant_startup_scheduled",
925 0 : "Number of tenants included in pageserver startup (doesn't count tenants attached later)"
926 0 : ).expect("Failed to register metric"),
927 0 : startup_complete: register_int_counter!(
928 0 : "pageserver_tenant_startup_complete",
929 0 : "Number of tenants that have completed warm-up, or activated on-demand during initial startup: \
930 0 : should eventually reach `pageserver_tenant_startup_scheduled_total`. Does not include broken \
931 0 : tenants: such cases will lead to this metric never reaching the scheduled count."
932 0 : ).expect("Failed to register metric"),
933 0 : }
934 0 : });
935 :
936 : /// Each `Timeline`'s [`EVICTIONS_WITH_LOW_RESIDENCE_DURATION`] metric.
937 : #[derive(Debug)]
938 : pub(crate) struct EvictionsWithLowResidenceDuration {
939 : data_source: &'static str,
940 : threshold: Duration,
941 : counter: Option<IntCounter>,
942 : }
943 :
944 : pub(crate) struct EvictionsWithLowResidenceDurationBuilder {
945 : data_source: &'static str,
946 : threshold: Duration,
947 : }
948 :
949 : impl EvictionsWithLowResidenceDurationBuilder {
950 1242 : pub fn new(data_source: &'static str, threshold: Duration) -> Self {
951 1242 : Self {
952 1242 : data_source,
953 1242 : threshold,
954 1242 : }
955 1242 : }
956 :
957 1242 : fn build(
958 1242 : &self,
959 1242 : tenant_id: &str,
960 1242 : shard_id: &str,
961 1242 : timeline_id: &str,
962 1242 : ) -> EvictionsWithLowResidenceDuration {
963 1242 : let counter = EVICTIONS_WITH_LOW_RESIDENCE_DURATION
964 1242 : .get_metric_with_label_values(&[
965 1242 : tenant_id,
966 1242 : shard_id,
967 1242 : timeline_id,
968 1242 : self.data_source,
969 1242 : &EvictionsWithLowResidenceDuration::threshold_label_value(self.threshold),
970 1242 : ])
971 1242 : .unwrap();
972 1242 : EvictionsWithLowResidenceDuration {
973 1242 : data_source: self.data_source,
974 1242 : threshold: self.threshold,
975 1242 : counter: Some(counter),
976 1242 : }
977 1242 : }
978 : }
979 :
980 : impl EvictionsWithLowResidenceDuration {
981 1266 : fn threshold_label_value(threshold: Duration) -> String {
982 1266 : format!("{}", threshold.as_secs())
983 1266 : }
984 :
985 12 : pub fn observe(&self, observed_value: Duration) {
986 12 : if observed_value < self.threshold {
987 12 : self.counter
988 12 : .as_ref()
989 12 : .expect("nobody calls this function after `remove_from_vec`")
990 12 : .inc();
991 12 : }
992 12 : }
993 :
994 24 : pub fn change_threshold(
995 24 : &mut self,
996 24 : tenant_id: &str,
997 24 : shard_id: &str,
998 24 : timeline_id: &str,
999 24 : new_threshold: Duration,
1000 24 : ) {
1001 24 : if new_threshold == self.threshold {
1002 24 : return;
1003 0 : }
1004 0 : let mut with_new = EvictionsWithLowResidenceDurationBuilder::new(
1005 0 : self.data_source,
1006 0 : new_threshold,
1007 0 : )
1008 0 : .build(tenant_id, shard_id, timeline_id);
1009 0 : std::mem::swap(self, &mut with_new);
1010 0 : with_new.remove(tenant_id, shard_id, timeline_id);
1011 24 : }
1012 :
1013 : // This could be a `Drop` impl, but, we need the `tenant_id` and `timeline_id`.
1014 24 : fn remove(&mut self, tenant_id: &str, shard_id: &str, timeline_id: &str) {
1015 24 : let Some(_counter) = self.counter.take() else {
1016 0 : return;
1017 : };
1018 :
1019 24 : let threshold = Self::threshold_label_value(self.threshold);
1020 24 :
1021 24 : let removed = EVICTIONS_WITH_LOW_RESIDENCE_DURATION.remove_label_values(&[
1022 24 : tenant_id,
1023 24 : shard_id,
1024 24 : timeline_id,
1025 24 : self.data_source,
1026 24 : &threshold,
1027 24 : ]);
1028 24 :
1029 24 : match removed {
1030 0 : Err(e) => {
1031 0 : // this has been hit in staging as
1032 0 : // <https://neondatabase.sentry.io/issues/4142396994/>, but we don't know how.
1033 0 : // because we can be in the drop path already, don't risk:
1034 0 : // - "double-panic => illegal instruction" or
1035 0 : // - future "drop panick => abort"
1036 0 : //
1037 0 : // so just nag: (the error has the labels)
1038 0 : tracing::warn!("failed to remove EvictionsWithLowResidenceDuration, it was already removed? {e:#?}");
1039 : }
1040 : Ok(()) => {
1041 : // to help identify cases where we double-remove the same values, let's log all
1042 : // deletions?
1043 24 : tracing::info!("removed EvictionsWithLowResidenceDuration with {tenant_id}, {timeline_id}, {}, {threshold}", self.data_source);
1044 : }
1045 : }
1046 24 : }
1047 : }
1048 :
1049 : // Metrics collected on disk IO operations
1050 : //
1051 : // Roughly logarithmic scale.
1052 : const STORAGE_IO_TIME_BUCKETS: &[f64] = &[
1053 : 0.000030, // 30 usec
1054 : 0.001000, // 1000 usec
1055 : 0.030, // 30 ms
1056 : 1.000, // 1000 ms
1057 : 30.000, // 30000 ms
1058 : ];
1059 :
1060 : /// VirtualFile fs operation variants.
1061 : ///
1062 : /// Operations:
1063 : /// - open ([`std::fs::OpenOptions::open`])
1064 : /// - close (dropping [`crate::virtual_file::VirtualFile`])
1065 : /// - close-by-replace (close by replacement algorithm)
1066 : /// - read (`read_at`)
1067 : /// - write (`write_at`)
1068 : /// - seek (modify internal position or file length query)
1069 : /// - fsync ([`std::fs::File::sync_all`])
1070 : /// - metadata ([`std::fs::File::metadata`])
1071 : #[derive(
1072 5400 : Debug, Clone, Copy, strum_macros::EnumCount, strum_macros::EnumIter, strum_macros::FromRepr,
1073 : )]
1074 : pub(crate) enum StorageIoOperation {
1075 : Open,
1076 : OpenAfterReplace,
1077 : Close,
1078 : CloseByReplace,
1079 : Read,
1080 : Write,
1081 : Seek,
1082 : Fsync,
1083 : Metadata,
1084 : }
1085 :
1086 : impl StorageIoOperation {
1087 5400 : pub fn as_str(&self) -> &'static str {
1088 5400 : match self {
1089 600 : StorageIoOperation::Open => "open",
1090 600 : StorageIoOperation::OpenAfterReplace => "open-after-replace",
1091 600 : StorageIoOperation::Close => "close",
1092 600 : StorageIoOperation::CloseByReplace => "close-by-replace",
1093 600 : StorageIoOperation::Read => "read",
1094 600 : StorageIoOperation::Write => "write",
1095 600 : StorageIoOperation::Seek => "seek",
1096 600 : StorageIoOperation::Fsync => "fsync",
1097 600 : StorageIoOperation::Metadata => "metadata",
1098 : }
1099 5400 : }
1100 : }
1101 :
1102 : /// Tracks time taken by fs operations near VirtualFile.
1103 : #[derive(Debug)]
1104 : pub(crate) struct StorageIoTime {
1105 : metrics: [Histogram; StorageIoOperation::COUNT],
1106 : }
1107 :
1108 : impl StorageIoTime {
1109 600 : fn new() -> Self {
1110 600 : let storage_io_histogram_vec = register_histogram_vec!(
1111 600 : "pageserver_io_operations_seconds",
1112 600 : "Time spent in IO operations",
1113 600 : &["operation"],
1114 600 : STORAGE_IO_TIME_BUCKETS.into()
1115 600 : )
1116 600 : .expect("failed to define a metric");
1117 5400 : let metrics = std::array::from_fn(|i| {
1118 5400 : let op = StorageIoOperation::from_repr(i).unwrap();
1119 5400 : storage_io_histogram_vec
1120 5400 : .get_metric_with_label_values(&[op.as_str()])
1121 5400 : .unwrap()
1122 5400 : });
1123 600 : Self { metrics }
1124 600 : }
1125 :
1126 6886070 : pub(crate) fn get(&self, op: StorageIoOperation) -> &Histogram {
1127 6886070 : &self.metrics[op as usize]
1128 6886070 : }
1129 : }
1130 :
1131 : pub(crate) static STORAGE_IO_TIME_METRIC: Lazy<StorageIoTime> = Lazy::new(StorageIoTime::new);
1132 :
1133 : const STORAGE_IO_SIZE_OPERATIONS: &[&str] = &["read", "write"];
1134 :
1135 : // Needed for the https://neonprod.grafana.net/d/5uK9tHL4k/picking-tenant-for-relocation?orgId=1
1136 588 : pub(crate) static STORAGE_IO_SIZE: Lazy<IntGaugeVec> = Lazy::new(|| {
1137 588 : register_int_gauge_vec!(
1138 588 : "pageserver_io_operations_bytes_total",
1139 588 : "Total amount of bytes read/written in IO operations",
1140 588 : &["operation", "tenant_id", "shard_id", "timeline_id"]
1141 588 : )
1142 588 : .expect("failed to define a metric")
1143 588 : });
1144 :
1145 : #[cfg(not(test))]
1146 : pub(crate) mod virtual_file_descriptor_cache {
1147 : use super::*;
1148 :
1149 0 : pub(crate) static SIZE_MAX: Lazy<UIntGauge> = Lazy::new(|| {
1150 0 : register_uint_gauge!(
1151 0 : "pageserver_virtual_file_descriptor_cache_size_max",
1152 0 : "Maximum number of open file descriptors in the cache."
1153 0 : )
1154 0 : .unwrap()
1155 0 : });
1156 :
1157 : // SIZE_CURRENT: derive it like so:
1158 : // ```
1159 : // sum (pageserver_io_operations_seconds_count{operation=~"^(open|open-after-replace)$")
1160 : // -ignoring(operation)
1161 : // sum(pageserver_io_operations_seconds_count{operation=~"^(close|close-by-replace)$"}
1162 : // ```
1163 : }
1164 :
1165 : #[cfg(not(test))]
1166 : pub(crate) mod virtual_file_io_engine {
1167 : use super::*;
1168 :
1169 0 : pub(crate) static KIND: Lazy<UIntGaugeVec> = Lazy::new(|| {
1170 0 : register_uint_gauge_vec!(
1171 0 : "pageserver_virtual_file_io_engine_kind",
1172 0 : "The configured io engine for VirtualFile",
1173 0 : &["kind"],
1174 0 : )
1175 0 : .unwrap()
1176 0 : });
1177 : }
1178 :
1179 : struct GlobalAndPerTimelineHistogramTimer<'a, 'c> {
1180 : global_metric: &'a Histogram,
1181 :
1182 : // Optional because not all op types are tracked per-timeline
1183 : timeline_metric: Option<&'a Histogram>,
1184 :
1185 : ctx: &'c RequestContext,
1186 : start: std::time::Instant,
1187 : op: SmgrQueryType,
1188 : }
1189 :
1190 : impl<'a, 'c> Drop for GlobalAndPerTimelineHistogramTimer<'a, 'c> {
1191 30 : fn drop(&mut self) {
1192 30 : let elapsed = self.start.elapsed();
1193 30 : let ex_throttled = self
1194 30 : .ctx
1195 30 : .micros_spent_throttled
1196 30 : .close_and_checked_sub_from(elapsed);
1197 30 : let ex_throttled = match ex_throttled {
1198 30 : Ok(res) => res,
1199 0 : Err(error) => {
1200 : use utils::rate_limit::RateLimit;
1201 : static LOGGED: Lazy<Mutex<enum_map::EnumMap<SmgrQueryType, RateLimit>>> =
1202 0 : Lazy::new(|| {
1203 0 : Mutex::new(enum_map::EnumMap::from_array(std::array::from_fn(|_| {
1204 0 : RateLimit::new(Duration::from_secs(10))
1205 0 : })))
1206 0 : });
1207 0 : let mut guard = LOGGED.lock().unwrap();
1208 0 : let rate_limit = &mut guard[self.op];
1209 0 : rate_limit.call(|| {
1210 0 : warn!(op=?self.op, error, "error deducting time spent throttled; this message is logged at a global rate limit");
1211 0 : });
1212 0 : elapsed
1213 : }
1214 : };
1215 30 : self.global_metric.observe(ex_throttled.as_secs_f64());
1216 30 : if let Some(timeline_metric) = self.timeline_metric {
1217 6 : timeline_metric.observe(ex_throttled.as_secs_f64());
1218 24 : }
1219 30 : }
1220 : }
1221 :
1222 : #[derive(
1223 : Debug,
1224 : Clone,
1225 : Copy,
1226 7686 : IntoStaticStr,
1227 : strum_macros::EnumCount,
1228 72 : strum_macros::EnumIter,
1229 6360 : strum_macros::FromRepr,
1230 : enum_map::Enum,
1231 : )]
1232 : #[strum(serialize_all = "snake_case")]
1233 : pub enum SmgrQueryType {
1234 : GetRelExists,
1235 : GetRelSize,
1236 : GetPageAtLsn,
1237 : GetDbSize,
1238 : GetSlruSegment,
1239 : }
1240 :
1241 : #[derive(Debug)]
1242 : pub(crate) struct SmgrQueryTimePerTimeline {
1243 : global_metrics: [Histogram; SmgrQueryType::COUNT],
1244 : per_timeline_getpage: Histogram,
1245 : }
1246 :
1247 516 : static SMGR_QUERY_TIME_PER_TENANT_TIMELINE: Lazy<HistogramVec> = Lazy::new(|| {
1248 516 : register_histogram_vec!(
1249 516 : "pageserver_smgr_query_seconds",
1250 516 : "Time spent on smgr query handling, aggegated by query type and tenant/timeline.",
1251 516 : &["smgr_query_type", "tenant_id", "shard_id", "timeline_id"],
1252 516 : CRITICAL_OP_BUCKETS.into(),
1253 516 : )
1254 516 : .expect("failed to define a metric")
1255 516 : });
1256 :
1257 516 : static SMGR_QUERY_TIME_GLOBAL_BUCKETS: Lazy<Vec<f64>> = Lazy::new(|| {
1258 516 : [
1259 516 : 1,
1260 516 : 10,
1261 516 : 20,
1262 516 : 40,
1263 516 : 60,
1264 516 : 80,
1265 516 : 100,
1266 516 : 200,
1267 516 : 300,
1268 516 : 400,
1269 516 : 500,
1270 516 : 600,
1271 516 : 700,
1272 516 : 800,
1273 516 : 900,
1274 516 : 1_000, // 1ms
1275 516 : 2_000,
1276 516 : 4_000,
1277 516 : 6_000,
1278 516 : 8_000,
1279 516 : 10_000, // 10ms
1280 516 : 20_000,
1281 516 : 40_000,
1282 516 : 60_000,
1283 516 : 80_000,
1284 516 : 100_000,
1285 516 : 200_000,
1286 516 : 400_000,
1287 516 : 600_000,
1288 516 : 800_000,
1289 516 : 1_000_000, // 1s
1290 516 : 2_000_000,
1291 516 : 4_000_000,
1292 516 : 6_000_000,
1293 516 : 8_000_000,
1294 516 : 10_000_000, // 10s
1295 516 : 20_000_000,
1296 516 : 50_000_000,
1297 516 : 100_000_000,
1298 516 : 200_000_000,
1299 516 : 1_000_000_000, // 1000s
1300 516 : ]
1301 516 : .into_iter()
1302 516 : .map(Duration::from_micros)
1303 21156 : .map(|d| d.as_secs_f64())
1304 516 : .collect()
1305 516 : });
1306 :
1307 516 : static SMGR_QUERY_TIME_GLOBAL: Lazy<HistogramVec> = Lazy::new(|| {
1308 516 : register_histogram_vec!(
1309 516 : "pageserver_smgr_query_seconds_global",
1310 516 : "Time spent on smgr query handling, aggregated by query type.",
1311 516 : &["smgr_query_type"],
1312 516 : SMGR_QUERY_TIME_GLOBAL_BUCKETS.clone(),
1313 516 : )
1314 516 : .expect("failed to define a metric")
1315 516 : });
1316 :
1317 : impl SmgrQueryTimePerTimeline {
1318 1272 : pub(crate) fn new(tenant_shard_id: &TenantShardId, timeline_id: &TimelineId) -> Self {
1319 1272 : let tenant_id = tenant_shard_id.tenant_id.to_string();
1320 1272 : let shard_slug = format!("{}", tenant_shard_id.shard_slug());
1321 1272 : let timeline_id = timeline_id.to_string();
1322 6360 : let global_metrics = std::array::from_fn(|i| {
1323 6360 : let op = SmgrQueryType::from_repr(i).unwrap();
1324 6360 : SMGR_QUERY_TIME_GLOBAL
1325 6360 : .get_metric_with_label_values(&[op.into()])
1326 6360 : .unwrap()
1327 6360 : });
1328 1272 :
1329 1272 : let per_timeline_getpage = SMGR_QUERY_TIME_PER_TENANT_TIMELINE
1330 1272 : .get_metric_with_label_values(&[
1331 1272 : SmgrQueryType::GetPageAtLsn.into(),
1332 1272 : &tenant_id,
1333 1272 : &shard_slug,
1334 1272 : &timeline_id,
1335 1272 : ])
1336 1272 : .unwrap();
1337 1272 : Self {
1338 1272 : global_metrics,
1339 1272 : per_timeline_getpage,
1340 1272 : }
1341 1272 : }
1342 30 : pub(crate) fn start_timer<'c: 'a, 'a>(
1343 30 : &'a self,
1344 30 : op: SmgrQueryType,
1345 30 : ctx: &'c RequestContext,
1346 30 : ) -> Option<impl Drop + '_> {
1347 30 : let global_metric = &self.global_metrics[op as usize];
1348 30 : let start = Instant::now();
1349 30 : match ctx.micros_spent_throttled.open() {
1350 30 : Ok(()) => (),
1351 0 : Err(error) => {
1352 0 : use utils::rate_limit::RateLimit;
1353 0 : static LOGGED: Lazy<Mutex<enum_map::EnumMap<SmgrQueryType, RateLimit>>> =
1354 0 : Lazy::new(|| {
1355 0 : Mutex::new(enum_map::EnumMap::from_array(std::array::from_fn(|_| {
1356 0 : RateLimit::new(Duration::from_secs(10))
1357 0 : })))
1358 0 : });
1359 0 : let mut guard = LOGGED.lock().unwrap();
1360 0 : let rate_limit = &mut guard[op];
1361 0 : rate_limit.call(|| {
1362 0 : warn!(?op, error, "error opening micros_spent_throttled; this message is logged at a global rate limit");
1363 0 : });
1364 0 : }
1365 : }
1366 :
1367 30 : let timeline_metric = if matches!(op, SmgrQueryType::GetPageAtLsn) {
1368 6 : Some(&self.per_timeline_getpage)
1369 : } else {
1370 24 : None
1371 : };
1372 :
1373 30 : Some(GlobalAndPerTimelineHistogramTimer {
1374 30 : global_metric,
1375 30 : timeline_metric,
1376 30 : ctx,
1377 30 : start,
1378 30 : op,
1379 30 : })
1380 30 : }
1381 : }
1382 :
1383 : #[cfg(test)]
1384 : mod smgr_query_time_tests {
1385 : use pageserver_api::shard::TenantShardId;
1386 : use strum::IntoEnumIterator;
1387 : use utils::id::{TenantId, TimelineId};
1388 :
1389 : use crate::{
1390 : context::{DownloadBehavior, RequestContext},
1391 : task_mgr::TaskKind,
1392 : };
1393 :
1394 : // Regression test, we used hard-coded string constants before using an enum.
1395 : #[test]
1396 6 : fn op_label_name() {
1397 : use super::SmgrQueryType::*;
1398 6 : let expect: [(super::SmgrQueryType, &'static str); 5] = [
1399 6 : (GetRelExists, "get_rel_exists"),
1400 6 : (GetRelSize, "get_rel_size"),
1401 6 : (GetPageAtLsn, "get_page_at_lsn"),
1402 6 : (GetDbSize, "get_db_size"),
1403 6 : (GetSlruSegment, "get_slru_segment"),
1404 6 : ];
1405 36 : for (op, expect) in expect {
1406 30 : let actual: &'static str = op.into();
1407 30 : assert_eq!(actual, expect);
1408 : }
1409 6 : }
1410 :
1411 : #[test]
1412 6 : fn basic() {
1413 6 : let ops: Vec<_> = super::SmgrQueryType::iter().collect();
1414 :
1415 36 : for op in &ops {
1416 30 : let tenant_id = TenantId::generate();
1417 30 : let timeline_id = TimelineId::generate();
1418 30 : let metrics = super::SmgrQueryTimePerTimeline::new(
1419 30 : &TenantShardId::unsharded(tenant_id),
1420 30 : &timeline_id,
1421 30 : );
1422 30 :
1423 60 : let get_counts = || {
1424 60 : let global: u64 = ops
1425 60 : .iter()
1426 300 : .map(|op| metrics.global_metrics[*op as usize].get_sample_count())
1427 60 : .sum();
1428 60 : (global, metrics.per_timeline_getpage.get_sample_count())
1429 60 : };
1430 :
1431 30 : let (pre_global, pre_per_tenant_timeline) = get_counts();
1432 30 : assert_eq!(pre_per_tenant_timeline, 0);
1433 :
1434 30 : let ctx = RequestContext::new(TaskKind::UnitTest, DownloadBehavior::Download);
1435 30 : let timer = metrics.start_timer(*op, &ctx);
1436 30 : drop(timer);
1437 30 :
1438 30 : let (post_global, post_per_tenant_timeline) = get_counts();
1439 30 : if matches!(op, super::SmgrQueryType::GetPageAtLsn) {
1440 : // getpage ops are tracked per-timeline, others aren't
1441 6 : assert_eq!(post_per_tenant_timeline, 1);
1442 : } else {
1443 24 : assert_eq!(post_per_tenant_timeline, 0);
1444 : }
1445 30 : assert!(post_global > pre_global);
1446 : }
1447 6 : }
1448 : }
1449 :
1450 : // keep in sync with control plane Go code so that we can validate
1451 : // compute's basebackup_ms metric with our perspective in the context of SLI/SLO.
1452 0 : static COMPUTE_STARTUP_BUCKETS: Lazy<[f64; 28]> = Lazy::new(|| {
1453 0 : // Go code uses milliseconds. Variable is called `computeStartupBuckets`
1454 0 : [
1455 0 : 5, 10, 20, 30, 50, 70, 100, 120, 150, 200, 250, 300, 350, 400, 450, 500, 600, 800, 1000,
1456 0 : 1500, 2000, 2500, 3000, 5000, 10000, 20000, 40000, 60000,
1457 0 : ]
1458 0 : .map(|ms| (ms as f64) / 1000.0)
1459 0 : });
1460 :
1461 : pub(crate) struct BasebackupQueryTime {
1462 : ok: Histogram,
1463 : error: Histogram,
1464 : }
1465 :
1466 0 : pub(crate) static BASEBACKUP_QUERY_TIME: Lazy<BasebackupQueryTime> = Lazy::new(|| {
1467 0 : let vec = register_histogram_vec!(
1468 0 : "pageserver_basebackup_query_seconds",
1469 0 : "Histogram of basebackup queries durations, by result type",
1470 0 : &["result"],
1471 0 : COMPUTE_STARTUP_BUCKETS.to_vec(),
1472 0 : )
1473 0 : .expect("failed to define a metric");
1474 0 : BasebackupQueryTime {
1475 0 : ok: vec.get_metric_with_label_values(&["ok"]).unwrap(),
1476 0 : error: vec.get_metric_with_label_values(&["error"]).unwrap(),
1477 0 : }
1478 0 : });
1479 :
1480 : pub(crate) struct BasebackupQueryTimeOngoingRecording<'a, 'c> {
1481 : parent: &'a BasebackupQueryTime,
1482 : ctx: &'c RequestContext,
1483 : start: std::time::Instant,
1484 : }
1485 :
1486 : impl BasebackupQueryTime {
1487 0 : pub(crate) fn start_recording<'c: 'a, 'a>(
1488 0 : &'a self,
1489 0 : ctx: &'c RequestContext,
1490 0 : ) -> BasebackupQueryTimeOngoingRecording<'_, '_> {
1491 0 : let start = Instant::now();
1492 0 : match ctx.micros_spent_throttled.open() {
1493 0 : Ok(()) => (),
1494 0 : Err(error) => {
1495 0 : use utils::rate_limit::RateLimit;
1496 0 : static LOGGED: Lazy<Mutex<RateLimit>> =
1497 0 : Lazy::new(|| Mutex::new(RateLimit::new(Duration::from_secs(10))));
1498 0 : let mut rate_limit = LOGGED.lock().unwrap();
1499 0 : rate_limit.call(|| {
1500 0 : warn!(error, "error opening micros_spent_throttled; this message is logged at a global rate limit");
1501 0 : });
1502 0 : }
1503 : }
1504 0 : BasebackupQueryTimeOngoingRecording {
1505 0 : parent: self,
1506 0 : ctx,
1507 0 : start,
1508 0 : }
1509 0 : }
1510 : }
1511 :
1512 : impl<'a, 'c> BasebackupQueryTimeOngoingRecording<'a, 'c> {
1513 0 : pub(crate) fn observe<T, E>(self, res: &Result<T, E>) {
1514 0 : let elapsed = self.start.elapsed();
1515 0 : let ex_throttled = self
1516 0 : .ctx
1517 0 : .micros_spent_throttled
1518 0 : .close_and_checked_sub_from(elapsed);
1519 0 : let ex_throttled = match ex_throttled {
1520 0 : Ok(ex_throttled) => ex_throttled,
1521 0 : Err(error) => {
1522 : use utils::rate_limit::RateLimit;
1523 : static LOGGED: Lazy<Mutex<RateLimit>> =
1524 0 : Lazy::new(|| Mutex::new(RateLimit::new(Duration::from_secs(10))));
1525 0 : let mut rate_limit = LOGGED.lock().unwrap();
1526 0 : rate_limit.call(|| {
1527 0 : warn!(error, "error deducting time spent throttled; this message is logged at a global rate limit");
1528 0 : });
1529 0 : elapsed
1530 : }
1531 : };
1532 0 : let metric = if res.is_ok() {
1533 0 : &self.parent.ok
1534 : } else {
1535 0 : &self.parent.error
1536 : };
1537 0 : metric.observe(ex_throttled.as_secs_f64());
1538 0 : }
1539 : }
1540 :
1541 0 : pub(crate) static LIVE_CONNECTIONS: Lazy<IntCounterPairVec> = Lazy::new(|| {
1542 0 : register_int_counter_pair_vec!(
1543 0 : "pageserver_live_connections_started",
1544 0 : "Number of network connections that we started handling",
1545 0 : "pageserver_live_connections_finished",
1546 0 : "Number of network connections that we finished handling",
1547 0 : &["pageserver_connection_kind"]
1548 0 : )
1549 0 : .expect("failed to define a metric")
1550 0 : });
1551 :
1552 0 : #[derive(Clone, Copy, enum_map::Enum, IntoStaticStr)]
1553 : pub(crate) enum ComputeCommandKind {
1554 : PageStreamV2,
1555 : Basebackup,
1556 : Fullbackup,
1557 : LeaseLsn,
1558 : }
1559 :
1560 : pub(crate) struct ComputeCommandCounters {
1561 : map: EnumMap<ComputeCommandKind, IntCounter>,
1562 : }
1563 :
1564 0 : pub(crate) static COMPUTE_COMMANDS_COUNTERS: Lazy<ComputeCommandCounters> = Lazy::new(|| {
1565 0 : let inner = register_int_counter_vec!(
1566 0 : "pageserver_compute_commands",
1567 0 : "Number of compute -> pageserver commands processed",
1568 0 : &["command"]
1569 0 : )
1570 0 : .expect("failed to define a metric");
1571 0 :
1572 0 : ComputeCommandCounters {
1573 0 : map: EnumMap::from_array(std::array::from_fn(|i| {
1574 0 : let command = <ComputeCommandKind as enum_map::Enum>::from_usize(i);
1575 0 : let command_str: &'static str = command.into();
1576 0 : inner.with_label_values(&[command_str])
1577 0 : })),
1578 0 : }
1579 0 : });
1580 :
1581 : impl ComputeCommandCounters {
1582 0 : pub(crate) fn for_command(&self, command: ComputeCommandKind) -> &IntCounter {
1583 0 : &self.map[command]
1584 0 : }
1585 : }
1586 :
1587 : // remote storage metrics
1588 :
1589 504 : static REMOTE_TIMELINE_CLIENT_CALLS: Lazy<IntCounterPairVec> = Lazy::new(|| {
1590 504 : register_int_counter_pair_vec!(
1591 504 : "pageserver_remote_timeline_client_calls_started",
1592 504 : "Number of started calls to remote timeline client.",
1593 504 : "pageserver_remote_timeline_client_calls_finished",
1594 504 : "Number of finshed calls to remote timeline client.",
1595 504 : &[
1596 504 : "tenant_id",
1597 504 : "shard_id",
1598 504 : "timeline_id",
1599 504 : "file_kind",
1600 504 : "op_kind"
1601 504 : ],
1602 504 : )
1603 504 : .unwrap()
1604 504 : });
1605 :
1606 : static REMOTE_TIMELINE_CLIENT_BYTES_STARTED_COUNTER: Lazy<IntCounterVec> =
1607 492 : Lazy::new(|| {
1608 492 : register_int_counter_vec!(
1609 492 : "pageserver_remote_timeline_client_bytes_started",
1610 492 : "Incremented by the number of bytes associated with a remote timeline client operation. \
1611 492 : The increment happens when the operation is scheduled.",
1612 492 : &["tenant_id", "shard_id", "timeline_id", "file_kind", "op_kind"],
1613 492 : )
1614 492 : .expect("failed to define a metric")
1615 492 : });
1616 :
1617 492 : static REMOTE_TIMELINE_CLIENT_BYTES_FINISHED_COUNTER: Lazy<IntCounterVec> = Lazy::new(|| {
1618 492 : register_int_counter_vec!(
1619 492 : "pageserver_remote_timeline_client_bytes_finished",
1620 492 : "Incremented by the number of bytes associated with a remote timeline client operation. \
1621 492 : The increment happens when the operation finishes (regardless of success/failure/shutdown).",
1622 492 : &["tenant_id", "shard_id", "timeline_id", "file_kind", "op_kind"],
1623 492 : )
1624 492 : .expect("failed to define a metric")
1625 492 : });
1626 :
1627 : pub(crate) struct TenantManagerMetrics {
1628 : tenant_slots_attached: UIntGauge,
1629 : tenant_slots_secondary: UIntGauge,
1630 : tenant_slots_inprogress: UIntGauge,
1631 : pub(crate) tenant_slot_writes: IntCounter,
1632 : pub(crate) unexpected_errors: IntCounter,
1633 : }
1634 :
1635 : impl TenantManagerMetrics {
1636 : /// Helpers for tracking slots. Note that these do not track the lifetime of TenantSlot objects
1637 : /// exactly: they track the lifetime of the slots _in the tenant map_.
1638 6 : pub(crate) fn slot_inserted(&self, slot: &TenantSlot) {
1639 6 : match slot {
1640 0 : TenantSlot::Attached(_) => {
1641 0 : self.tenant_slots_attached.inc();
1642 0 : }
1643 0 : TenantSlot::Secondary(_) => {
1644 0 : self.tenant_slots_secondary.inc();
1645 0 : }
1646 6 : TenantSlot::InProgress(_) => {
1647 6 : self.tenant_slots_inprogress.inc();
1648 6 : }
1649 : }
1650 6 : }
1651 :
1652 6 : pub(crate) fn slot_removed(&self, slot: &TenantSlot) {
1653 6 : match slot {
1654 6 : TenantSlot::Attached(_) => {
1655 6 : self.tenant_slots_attached.dec();
1656 6 : }
1657 0 : TenantSlot::Secondary(_) => {
1658 0 : self.tenant_slots_secondary.dec();
1659 0 : }
1660 0 : TenantSlot::InProgress(_) => {
1661 0 : self.tenant_slots_inprogress.dec();
1662 0 : }
1663 : }
1664 6 : }
1665 :
1666 : #[cfg(all(debug_assertions, not(test)))]
1667 0 : pub(crate) fn slots_total(&self) -> u64 {
1668 0 : self.tenant_slots_attached.get()
1669 0 : + self.tenant_slots_secondary.get()
1670 0 : + self.tenant_slots_inprogress.get()
1671 0 : }
1672 : }
1673 :
1674 6 : pub(crate) static TENANT_MANAGER: Lazy<TenantManagerMetrics> = Lazy::new(|| {
1675 6 : let tenant_slots = register_uint_gauge_vec!(
1676 6 : "pageserver_tenant_manager_slots",
1677 6 : "How many slots currently exist, including all attached, secondary and in-progress operations",
1678 6 : &["mode"]
1679 6 : )
1680 6 : .expect("failed to define a metric");
1681 6 : TenantManagerMetrics {
1682 6 : tenant_slots_attached: tenant_slots
1683 6 : .get_metric_with_label_values(&["attached"])
1684 6 : .unwrap(),
1685 6 : tenant_slots_secondary: tenant_slots
1686 6 : .get_metric_with_label_values(&["secondary"])
1687 6 : .unwrap(),
1688 6 : tenant_slots_inprogress: tenant_slots
1689 6 : .get_metric_with_label_values(&["inprogress"])
1690 6 : .unwrap(),
1691 6 : tenant_slot_writes: register_int_counter!(
1692 6 : "pageserver_tenant_manager_slot_writes",
1693 6 : "Writes to a tenant slot, including all of create/attach/detach/delete"
1694 6 : )
1695 6 : .expect("failed to define a metric"),
1696 6 : unexpected_errors: register_int_counter!(
1697 6 : "pageserver_tenant_manager_unexpected_errors_total",
1698 6 : "Number of unexpected conditions encountered: nonzero value indicates a non-fatal bug."
1699 6 : )
1700 6 : .expect("failed to define a metric"),
1701 6 : }
1702 6 : });
1703 :
1704 : pub(crate) struct DeletionQueueMetrics {
1705 : pub(crate) keys_submitted: IntCounter,
1706 : pub(crate) keys_dropped: IntCounter,
1707 : pub(crate) keys_executed: IntCounter,
1708 : pub(crate) keys_validated: IntCounter,
1709 : pub(crate) dropped_lsn_updates: IntCounter,
1710 : pub(crate) unexpected_errors: IntCounter,
1711 : pub(crate) remote_errors: IntCounterVec,
1712 : }
1713 83 : pub(crate) static DELETION_QUEUE: Lazy<DeletionQueueMetrics> = Lazy::new(|| {
1714 83 : DeletionQueueMetrics{
1715 83 :
1716 83 : keys_submitted: register_int_counter!(
1717 83 : "pageserver_deletion_queue_submitted_total",
1718 83 : "Number of objects submitted for deletion"
1719 83 : )
1720 83 : .expect("failed to define a metric"),
1721 83 :
1722 83 : keys_dropped: register_int_counter!(
1723 83 : "pageserver_deletion_queue_dropped_total",
1724 83 : "Number of object deletions dropped due to stale generation."
1725 83 : )
1726 83 : .expect("failed to define a metric"),
1727 83 :
1728 83 : keys_executed: register_int_counter!(
1729 83 : "pageserver_deletion_queue_executed_total",
1730 83 : "Number of objects deleted. Only includes objects that we actually deleted, sum with pageserver_deletion_queue_dropped_total for the total number of keys processed to completion"
1731 83 : )
1732 83 : .expect("failed to define a metric"),
1733 83 :
1734 83 : keys_validated: register_int_counter!(
1735 83 : "pageserver_deletion_queue_validated_total",
1736 83 : "Number of keys validated for deletion. Sum with pageserver_deletion_queue_dropped_total for the total number of keys that have passed through the validation stage."
1737 83 : )
1738 83 : .expect("failed to define a metric"),
1739 83 :
1740 83 : dropped_lsn_updates: register_int_counter!(
1741 83 : "pageserver_deletion_queue_dropped_lsn_updates_total",
1742 83 : "Updates to remote_consistent_lsn dropped due to stale generation number."
1743 83 : )
1744 83 : .expect("failed to define a metric"),
1745 83 : unexpected_errors: register_int_counter!(
1746 83 : "pageserver_deletion_queue_unexpected_errors_total",
1747 83 : "Number of unexpected condiions that may stall the queue: any value above zero is unexpected."
1748 83 : )
1749 83 : .expect("failed to define a metric"),
1750 83 : remote_errors: register_int_counter_vec!(
1751 83 : "pageserver_deletion_queue_remote_errors_total",
1752 83 : "Retryable remote I/O errors while executing deletions, for example 503 responses to DeleteObjects",
1753 83 : &["op_kind"],
1754 83 : )
1755 83 : .expect("failed to define a metric")
1756 83 : }
1757 83 : });
1758 :
1759 : pub(crate) struct SecondaryModeMetrics {
1760 : pub(crate) upload_heatmap: IntCounter,
1761 : pub(crate) upload_heatmap_errors: IntCounter,
1762 : pub(crate) upload_heatmap_duration: Histogram,
1763 : pub(crate) download_heatmap: IntCounter,
1764 : pub(crate) download_layer: IntCounter,
1765 : }
1766 0 : pub(crate) static SECONDARY_MODE: Lazy<SecondaryModeMetrics> = Lazy::new(|| {
1767 0 : SecondaryModeMetrics {
1768 0 : upload_heatmap: register_int_counter!(
1769 0 : "pageserver_secondary_upload_heatmap",
1770 0 : "Number of heatmaps written to remote storage by attached tenants"
1771 0 : )
1772 0 : .expect("failed to define a metric"),
1773 0 : upload_heatmap_errors: register_int_counter!(
1774 0 : "pageserver_secondary_upload_heatmap_errors",
1775 0 : "Failures writing heatmap to remote storage"
1776 0 : )
1777 0 : .expect("failed to define a metric"),
1778 0 : upload_heatmap_duration: register_histogram!(
1779 0 : "pageserver_secondary_upload_heatmap_duration",
1780 0 : "Time to build and upload a heatmap, including any waiting inside the remote storage client"
1781 0 : )
1782 0 : .expect("failed to define a metric"),
1783 0 : download_heatmap: register_int_counter!(
1784 0 : "pageserver_secondary_download_heatmap",
1785 0 : "Number of downloads of heatmaps by secondary mode locations, including when it hasn't changed"
1786 0 : )
1787 0 : .expect("failed to define a metric"),
1788 0 : download_layer: register_int_counter!(
1789 0 : "pageserver_secondary_download_layer",
1790 0 : "Number of downloads of layers by secondary mode locations"
1791 0 : )
1792 0 : .expect("failed to define a metric"),
1793 0 : }
1794 0 : });
1795 :
1796 0 : pub(crate) static SECONDARY_RESIDENT_PHYSICAL_SIZE: Lazy<UIntGaugeVec> = Lazy::new(|| {
1797 0 : register_uint_gauge_vec!(
1798 0 : "pageserver_secondary_resident_physical_size",
1799 0 : "The size of the layer files present in the pageserver's filesystem, for secondary locations.",
1800 0 : &["tenant_id", "shard_id"]
1801 0 : )
1802 0 : .expect("failed to define a metric")
1803 0 : });
1804 :
1805 0 : pub(crate) static NODE_UTILIZATION_SCORE: Lazy<UIntGauge> = Lazy::new(|| {
1806 0 : register_uint_gauge!(
1807 0 : "pageserver_utilization_score",
1808 0 : "The utilization score we report to the storage controller for scheduling, where 0 is empty, 1000000 is full, and anything above is considered overloaded",
1809 0 : )
1810 0 : .expect("failed to define a metric")
1811 0 : });
1812 :
1813 0 : pub(crate) static SECONDARY_HEATMAP_TOTAL_SIZE: Lazy<UIntGaugeVec> = Lazy::new(|| {
1814 0 : register_uint_gauge_vec!(
1815 0 : "pageserver_secondary_heatmap_total_size",
1816 0 : "The total size in bytes of all layers in the most recently downloaded heatmap.",
1817 0 : &["tenant_id", "shard_id"]
1818 0 : )
1819 0 : .expect("failed to define a metric")
1820 0 : });
1821 :
1822 : #[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
1823 : pub enum RemoteOpKind {
1824 : Upload,
1825 : Download,
1826 : Delete,
1827 : }
1828 : impl RemoteOpKind {
1829 40296 : pub fn as_str(&self) -> &'static str {
1830 40296 : match self {
1831 38035 : Self::Upload => "upload",
1832 156 : Self::Download => "download",
1833 2105 : Self::Delete => "delete",
1834 : }
1835 40296 : }
1836 : }
1837 :
1838 : #[derive(Debug, Clone, Copy, Hash, PartialEq, Eq)]
1839 : pub enum RemoteOpFileKind {
1840 : Layer,
1841 : Index,
1842 : }
1843 : impl RemoteOpFileKind {
1844 40296 : pub fn as_str(&self) -> &'static str {
1845 40296 : match self {
1846 27374 : Self::Layer => "layer",
1847 12922 : Self::Index => "index",
1848 : }
1849 40296 : }
1850 : }
1851 :
1852 495 : pub(crate) static REMOTE_OPERATION_TIME: Lazy<HistogramVec> = Lazy::new(|| {
1853 495 : register_histogram_vec!(
1854 495 : "pageserver_remote_operation_seconds",
1855 495 : "Time spent on remote storage operations. \
1856 495 : Grouped by tenant, timeline, operation_kind and status. \
1857 495 : Does not account for time spent waiting in remote timeline client's queues.",
1858 495 : &["file_kind", "op_kind", "status"]
1859 495 : )
1860 495 : .expect("failed to define a metric")
1861 495 : });
1862 :
1863 0 : pub(crate) static TENANT_TASK_EVENTS: Lazy<IntCounterVec> = Lazy::new(|| {
1864 0 : register_int_counter_vec!(
1865 0 : "pageserver_tenant_task_events",
1866 0 : "Number of task start/stop/fail events.",
1867 0 : &["event"],
1868 0 : )
1869 0 : .expect("Failed to register tenant_task_events metric")
1870 0 : });
1871 :
1872 : pub struct BackgroundLoopSemaphoreMetrics {
1873 : counters: EnumMap<BackgroundLoopKind, IntCounterPair>,
1874 : durations: EnumMap<BackgroundLoopKind, Counter>,
1875 : }
1876 :
1877 : pub(crate) static BACKGROUND_LOOP_SEMAPHORE: Lazy<BackgroundLoopSemaphoreMetrics> = Lazy::new(
1878 60 : || {
1879 60 : let counters = register_int_counter_pair_vec!(
1880 60 : "pageserver_background_loop_semaphore_wait_start_count",
1881 60 : "Counter for background loop concurrency-limiting semaphore acquire calls started",
1882 60 : "pageserver_background_loop_semaphore_wait_finish_count",
1883 60 : "Counter for background loop concurrency-limiting semaphore acquire calls finished",
1884 60 : &["task"],
1885 60 : )
1886 60 : .unwrap();
1887 60 :
1888 60 : let durations = register_counter_vec!(
1889 60 : "pageserver_background_loop_semaphore_wait_duration_seconds",
1890 60 : "Sum of wall clock time spent waiting on the background loop concurrency-limiting semaphore acquire calls",
1891 60 : &["task"],
1892 60 : )
1893 60 : .unwrap();
1894 60 :
1895 60 : BackgroundLoopSemaphoreMetrics {
1896 540 : counters: enum_map::EnumMap::from_array(std::array::from_fn(|i| {
1897 540 : let kind = <BackgroundLoopKind as enum_map::Enum>::from_usize(i);
1898 540 : counters.with_label_values(&[kind.into()])
1899 540 : })),
1900 540 : durations: enum_map::EnumMap::from_array(std::array::from_fn(|i| {
1901 540 : let kind = <BackgroundLoopKind as enum_map::Enum>::from_usize(i);
1902 540 : durations.with_label_values(&[kind.into()])
1903 540 : })),
1904 60 : }
1905 60 : },
1906 : );
1907 :
1908 : impl BackgroundLoopSemaphoreMetrics {
1909 1092 : pub(crate) fn measure_acquisition(&self, task: BackgroundLoopKind) -> impl Drop + '_ {
1910 : struct Record<'a> {
1911 : metrics: &'a BackgroundLoopSemaphoreMetrics,
1912 : task: BackgroundLoopKind,
1913 : _counter_guard: metrics::IntCounterPairGuard,
1914 : start: Instant,
1915 : }
1916 : impl Drop for Record<'_> {
1917 1092 : fn drop(&mut self) {
1918 1092 : let elapsed = self.start.elapsed().as_secs_f64();
1919 1092 : self.metrics.durations[self.task].inc_by(elapsed);
1920 1092 : }
1921 : }
1922 1092 : Record {
1923 1092 : metrics: self,
1924 1092 : task,
1925 1092 : _counter_guard: self.counters[task].guard(),
1926 1092 : start: Instant::now(),
1927 1092 : }
1928 1092 : }
1929 : }
1930 :
1931 0 : pub(crate) static BACKGROUND_LOOP_PERIOD_OVERRUN_COUNT: Lazy<IntCounterVec> = Lazy::new(|| {
1932 0 : register_int_counter_vec!(
1933 0 : "pageserver_background_loop_period_overrun_count",
1934 0 : "Incremented whenever warn_when_period_overrun() logs a warning.",
1935 0 : &["task", "period"],
1936 0 : )
1937 0 : .expect("failed to define a metric")
1938 0 : });
1939 :
1940 : // walreceiver metrics
1941 :
1942 0 : pub(crate) static WALRECEIVER_STARTED_CONNECTIONS: Lazy<IntCounter> = Lazy::new(|| {
1943 0 : register_int_counter!(
1944 0 : "pageserver_walreceiver_started_connections_total",
1945 0 : "Number of started walreceiver connections"
1946 0 : )
1947 0 : .expect("failed to define a metric")
1948 0 : });
1949 :
1950 0 : pub(crate) static WALRECEIVER_ACTIVE_MANAGERS: Lazy<IntGauge> = Lazy::new(|| {
1951 0 : register_int_gauge!(
1952 0 : "pageserver_walreceiver_active_managers",
1953 0 : "Number of active walreceiver managers"
1954 0 : )
1955 0 : .expect("failed to define a metric")
1956 0 : });
1957 :
1958 0 : pub(crate) static WALRECEIVER_SWITCHES: Lazy<IntCounterVec> = Lazy::new(|| {
1959 0 : register_int_counter_vec!(
1960 0 : "pageserver_walreceiver_switches_total",
1961 0 : "Number of walreceiver manager change_connection calls",
1962 0 : &["reason"]
1963 0 : )
1964 0 : .expect("failed to define a metric")
1965 0 : });
1966 :
1967 0 : pub(crate) static WALRECEIVER_BROKER_UPDATES: Lazy<IntCounter> = Lazy::new(|| {
1968 0 : register_int_counter!(
1969 0 : "pageserver_walreceiver_broker_updates_total",
1970 0 : "Number of received broker updates in walreceiver"
1971 0 : )
1972 0 : .expect("failed to define a metric")
1973 0 : });
1974 :
1975 6 : pub(crate) static WALRECEIVER_CANDIDATES_EVENTS: Lazy<IntCounterVec> = Lazy::new(|| {
1976 6 : register_int_counter_vec!(
1977 6 : "pageserver_walreceiver_candidates_events_total",
1978 6 : "Number of walreceiver candidate events",
1979 6 : &["event"]
1980 6 : )
1981 6 : .expect("failed to define a metric")
1982 6 : });
1983 :
1984 : pub(crate) static WALRECEIVER_CANDIDATES_ADDED: Lazy<IntCounter> =
1985 0 : Lazy::new(|| WALRECEIVER_CANDIDATES_EVENTS.with_label_values(&["add"]));
1986 :
1987 : pub(crate) static WALRECEIVER_CANDIDATES_REMOVED: Lazy<IntCounter> =
1988 6 : Lazy::new(|| WALRECEIVER_CANDIDATES_EVENTS.with_label_values(&["remove"]));
1989 :
1990 : // Metrics collected on WAL redo operations
1991 : //
1992 : // We collect the time spent in actual WAL redo ('redo'), and time waiting
1993 : // for access to the postgres process ('wait') since there is only one for
1994 : // each tenant.
1995 :
1996 : /// Time buckets are small because we want to be able to measure the
1997 : /// smallest redo processing times. These buckets allow us to measure down
1998 : /// to 5us, which equates to 200'000 pages/sec, which equates to 1.6GB/sec.
1999 : /// This is much better than the previous 5ms aka 200 pages/sec aka 1.6MB/sec.
2000 : ///
2001 : /// Values up to 1s are recorded because metrics show that we have redo
2002 : /// durations and lock times larger than 0.250s.
2003 : macro_rules! redo_histogram_time_buckets {
2004 : () => {
2005 : vec![
2006 : 0.000_005, 0.000_010, 0.000_025, 0.000_050, 0.000_100, 0.000_250, 0.000_500, 0.001_000,
2007 : 0.002_500, 0.005_000, 0.010_000, 0.025_000, 0.050_000, 0.100_000, 0.250_000, 0.500_000,
2008 : 1.000_000,
2009 : ]
2010 : };
2011 : }
2012 :
2013 : /// While we're at it, also measure the amount of records replayed in each
2014 : /// operation. We have a global 'total replayed' counter, but that's not
2015 : /// as useful as 'what is the skew for how many records we replay in one
2016 : /// operation'.
2017 : macro_rules! redo_histogram_count_buckets {
2018 : () => {
2019 : vec![0.0, 1.0, 2.0, 5.0, 10.0, 25.0, 50.0, 100.0, 250.0, 500.0]
2020 : };
2021 : }
2022 :
2023 : macro_rules! redo_bytes_histogram_count_buckets {
2024 : () => {
2025 : // powers of (2^.5), from 2^4.5 to 2^15 (22 buckets)
2026 : // rounded up to the next multiple of 8 to capture any MAXALIGNed record of that size, too.
2027 : vec![
2028 : 24.0, 32.0, 48.0, 64.0, 96.0, 128.0, 184.0, 256.0, 368.0, 512.0, 728.0, 1024.0, 1456.0,
2029 : 2048.0, 2904.0, 4096.0, 5800.0, 8192.0, 11592.0, 16384.0, 23176.0, 32768.0,
2030 : ]
2031 : };
2032 : }
2033 :
2034 : pub(crate) struct WalIngestMetrics {
2035 : pub(crate) bytes_received: IntCounter,
2036 : pub(crate) records_received: IntCounter,
2037 : pub(crate) records_committed: IntCounter,
2038 : pub(crate) records_filtered: IntCounter,
2039 : }
2040 :
2041 6 : pub(crate) static WAL_INGEST: Lazy<WalIngestMetrics> = Lazy::new(|| WalIngestMetrics {
2042 6 : bytes_received: register_int_counter!(
2043 6 : "pageserver_wal_ingest_bytes_received",
2044 6 : "Bytes of WAL ingested from safekeepers",
2045 6 : )
2046 6 : .unwrap(),
2047 6 : records_received: register_int_counter!(
2048 6 : "pageserver_wal_ingest_records_received",
2049 6 : "Number of WAL records received from safekeepers"
2050 6 : )
2051 6 : .expect("failed to define a metric"),
2052 6 : records_committed: register_int_counter!(
2053 6 : "pageserver_wal_ingest_records_committed",
2054 6 : "Number of WAL records which resulted in writes to pageserver storage"
2055 6 : )
2056 6 : .expect("failed to define a metric"),
2057 6 : records_filtered: register_int_counter!(
2058 6 : "pageserver_wal_ingest_records_filtered",
2059 6 : "Number of WAL records filtered out due to sharding"
2060 6 : )
2061 6 : .expect("failed to define a metric"),
2062 6 : });
2063 :
2064 18 : pub(crate) static WAL_REDO_TIME: Lazy<Histogram> = Lazy::new(|| {
2065 18 : register_histogram!(
2066 18 : "pageserver_wal_redo_seconds",
2067 18 : "Time spent on WAL redo",
2068 18 : redo_histogram_time_buckets!()
2069 18 : )
2070 18 : .expect("failed to define a metric")
2071 18 : });
2072 :
2073 18 : pub(crate) static WAL_REDO_RECORDS_HISTOGRAM: Lazy<Histogram> = Lazy::new(|| {
2074 18 : register_histogram!(
2075 18 : "pageserver_wal_redo_records_histogram",
2076 18 : "Histogram of number of records replayed per redo in the Postgres WAL redo process",
2077 18 : redo_histogram_count_buckets!(),
2078 18 : )
2079 18 : .expect("failed to define a metric")
2080 18 : });
2081 :
2082 18 : pub(crate) static WAL_REDO_BYTES_HISTOGRAM: Lazy<Histogram> = Lazy::new(|| {
2083 18 : register_histogram!(
2084 18 : "pageserver_wal_redo_bytes_histogram",
2085 18 : "Histogram of number of records replayed per redo sent to Postgres",
2086 18 : redo_bytes_histogram_count_buckets!(),
2087 18 : )
2088 18 : .expect("failed to define a metric")
2089 18 : });
2090 :
2091 : // FIXME: isn't this already included by WAL_REDO_RECORDS_HISTOGRAM which has _count?
2092 18 : pub(crate) static WAL_REDO_RECORD_COUNTER: Lazy<IntCounter> = Lazy::new(|| {
2093 18 : register_int_counter!(
2094 18 : "pageserver_replayed_wal_records_total",
2095 18 : "Number of WAL records replayed in WAL redo process"
2096 18 : )
2097 18 : .unwrap()
2098 18 : });
2099 :
2100 : #[rustfmt::skip]
2101 18 : pub(crate) static WAL_REDO_PROCESS_LAUNCH_DURATION_HISTOGRAM: Lazy<Histogram> = Lazy::new(|| {
2102 18 : register_histogram!(
2103 18 : "pageserver_wal_redo_process_launch_duration",
2104 18 : "Histogram of the duration of successful WalRedoProcess::launch calls",
2105 18 : vec![
2106 18 : 0.0002, 0.0004, 0.0006, 0.0008, 0.0010,
2107 18 : 0.0020, 0.0040, 0.0060, 0.0080, 0.0100,
2108 18 : 0.0200, 0.0400, 0.0600, 0.0800, 0.1000,
2109 18 : 0.2000, 0.4000, 0.6000, 0.8000, 1.0000,
2110 18 : 1.5000, 2.0000, 2.5000, 3.0000, 4.0000, 10.0000
2111 18 : ],
2112 18 : )
2113 18 : .expect("failed to define a metric")
2114 18 : });
2115 :
2116 : pub(crate) struct WalRedoProcessCounters {
2117 : pub(crate) started: IntCounter,
2118 : pub(crate) killed_by_cause: enum_map::EnumMap<WalRedoKillCause, IntCounter>,
2119 : pub(crate) active_stderr_logger_tasks_started: IntCounter,
2120 : pub(crate) active_stderr_logger_tasks_finished: IntCounter,
2121 : }
2122 :
2123 54 : #[derive(Debug, enum_map::Enum, strum_macros::IntoStaticStr)]
2124 : pub(crate) enum WalRedoKillCause {
2125 : WalRedoProcessDrop,
2126 : NoLeakChildDrop,
2127 : Startup,
2128 : }
2129 :
2130 : impl Default for WalRedoProcessCounters {
2131 18 : fn default() -> Self {
2132 18 : let started = register_int_counter!(
2133 18 : "pageserver_wal_redo_process_started_total",
2134 18 : "Number of WAL redo processes started",
2135 18 : )
2136 18 : .unwrap();
2137 18 :
2138 18 : let killed = register_int_counter_vec!(
2139 18 : "pageserver_wal_redo_process_stopped_total",
2140 18 : "Number of WAL redo processes stopped",
2141 18 : &["cause"],
2142 18 : )
2143 18 : .unwrap();
2144 18 :
2145 18 : let active_stderr_logger_tasks_started = register_int_counter!(
2146 18 : "pageserver_walredo_stderr_logger_tasks_started_total",
2147 18 : "Number of active walredo stderr logger tasks that have started",
2148 18 : )
2149 18 : .unwrap();
2150 18 :
2151 18 : let active_stderr_logger_tasks_finished = register_int_counter!(
2152 18 : "pageserver_walredo_stderr_logger_tasks_finished_total",
2153 18 : "Number of active walredo stderr logger tasks that have finished",
2154 18 : )
2155 18 : .unwrap();
2156 18 :
2157 18 : Self {
2158 18 : started,
2159 54 : killed_by_cause: EnumMap::from_array(std::array::from_fn(|i| {
2160 54 : let cause = <WalRedoKillCause as enum_map::Enum>::from_usize(i);
2161 54 : let cause_str: &'static str = cause.into();
2162 54 : killed.with_label_values(&[cause_str])
2163 54 : })),
2164 18 : active_stderr_logger_tasks_started,
2165 18 : active_stderr_logger_tasks_finished,
2166 18 : }
2167 18 : }
2168 : }
2169 :
2170 : pub(crate) static WAL_REDO_PROCESS_COUNTERS: Lazy<WalRedoProcessCounters> =
2171 : Lazy::new(WalRedoProcessCounters::default);
2172 :
2173 : /// Similar to `prometheus::HistogramTimer` but does not record on drop.
2174 : pub(crate) struct StorageTimeMetricsTimer {
2175 : metrics: StorageTimeMetrics,
2176 : start: Instant,
2177 : }
2178 :
2179 : impl StorageTimeMetricsTimer {
2180 14584 : fn new(metrics: StorageTimeMetrics) -> Self {
2181 14584 : Self {
2182 14584 : metrics,
2183 14584 : start: Instant::now(),
2184 14584 : }
2185 14584 : }
2186 :
2187 : /// Record the time from creation to now.
2188 11193 : pub fn stop_and_record(self) {
2189 11193 : let duration = self.start.elapsed().as_secs_f64();
2190 11193 : self.metrics.timeline_sum.inc_by(duration);
2191 11193 : self.metrics.timeline_count.inc();
2192 11193 : self.metrics.global_histogram.observe(duration);
2193 11193 : }
2194 :
2195 : /// Turns this timer into a timer, which will always record -- usually this means recording
2196 : /// regardless an early `?` path was taken in a function.
2197 2262 : pub(crate) fn record_on_drop(self) -> AlwaysRecordingStorageTimeMetricsTimer {
2198 2262 : AlwaysRecordingStorageTimeMetricsTimer(Some(self))
2199 2262 : }
2200 : }
2201 :
2202 : pub(crate) struct AlwaysRecordingStorageTimeMetricsTimer(Option<StorageTimeMetricsTimer>);
2203 :
2204 : impl Drop for AlwaysRecordingStorageTimeMetricsTimer {
2205 2262 : fn drop(&mut self) {
2206 2262 : if let Some(inner) = self.0.take() {
2207 2262 : inner.stop_and_record();
2208 2262 : }
2209 2262 : }
2210 : }
2211 :
2212 : /// Timing facilities for an globally histogrammed metric, which is supported by per tenant and
2213 : /// timeline total sum and count.
2214 : #[derive(Clone, Debug)]
2215 : pub(crate) struct StorageTimeMetrics {
2216 : /// Sum of f64 seconds, per operation, tenant_id and timeline_id
2217 : timeline_sum: Counter,
2218 : /// Number of oeprations, per operation, tenant_id and timeline_id
2219 : timeline_count: IntCounter,
2220 : /// Global histogram having only the "operation" label.
2221 : global_histogram: Histogram,
2222 : }
2223 :
2224 : impl StorageTimeMetrics {
2225 9936 : pub fn new(
2226 9936 : operation: StorageTimeOperation,
2227 9936 : tenant_id: &str,
2228 9936 : shard_id: &str,
2229 9936 : timeline_id: &str,
2230 9936 : ) -> Self {
2231 9936 : let operation: &'static str = operation.into();
2232 9936 :
2233 9936 : let timeline_sum = STORAGE_TIME_SUM_PER_TIMELINE
2234 9936 : .get_metric_with_label_values(&[operation, tenant_id, shard_id, timeline_id])
2235 9936 : .unwrap();
2236 9936 : let timeline_count = STORAGE_TIME_COUNT_PER_TIMELINE
2237 9936 : .get_metric_with_label_values(&[operation, tenant_id, shard_id, timeline_id])
2238 9936 : .unwrap();
2239 9936 : let global_histogram = STORAGE_TIME_GLOBAL
2240 9936 : .get_metric_with_label_values(&[operation])
2241 9936 : .unwrap();
2242 9936 :
2243 9936 : StorageTimeMetrics {
2244 9936 : timeline_sum,
2245 9936 : timeline_count,
2246 9936 : global_histogram,
2247 9936 : }
2248 9936 : }
2249 :
2250 : /// Starts timing a new operation.
2251 : ///
2252 : /// Note: unlike `prometheus::HistogramTimer` the returned timer does not record on drop.
2253 14584 : pub fn start_timer(&self) -> StorageTimeMetricsTimer {
2254 14584 : StorageTimeMetricsTimer::new(self.clone())
2255 14584 : }
2256 : }
2257 :
2258 : #[derive(Debug)]
2259 : pub(crate) struct TimelineMetrics {
2260 : tenant_id: String,
2261 : shard_id: String,
2262 : timeline_id: String,
2263 : pub flush_time_histo: StorageTimeMetrics,
2264 : pub compact_time_histo: StorageTimeMetrics,
2265 : pub create_images_time_histo: StorageTimeMetrics,
2266 : pub logical_size_histo: StorageTimeMetrics,
2267 : pub imitate_logical_size_histo: StorageTimeMetrics,
2268 : pub load_layer_map_histo: StorageTimeMetrics,
2269 : pub garbage_collect_histo: StorageTimeMetrics,
2270 : pub find_gc_cutoffs_histo: StorageTimeMetrics,
2271 : pub last_record_gauge: IntGauge,
2272 : pub pitr_history_size: UIntGauge,
2273 : pub archival_size: UIntGauge,
2274 : pub(crate) layer_size_image: UIntGauge,
2275 : pub(crate) layer_count_image: UIntGauge,
2276 : pub(crate) layer_size_delta: UIntGauge,
2277 : pub(crate) layer_count_delta: UIntGauge,
2278 : pub standby_horizon_gauge: IntGauge,
2279 : pub resident_physical_size_gauge: UIntGauge,
2280 : pub visible_physical_size_gauge: UIntGauge,
2281 : /// copy of LayeredTimeline.current_logical_size
2282 : pub current_logical_size_gauge: UIntGauge,
2283 : pub aux_file_size_gauge: IntGauge,
2284 : pub directory_entries_count_gauge: Lazy<UIntGauge, Box<dyn Send + Fn() -> UIntGauge>>,
2285 : pub evictions: IntCounter,
2286 : pub evictions_with_low_residence_duration: std::sync::RwLock<EvictionsWithLowResidenceDuration>,
2287 : /// Number of valid LSN leases.
2288 : pub valid_lsn_lease_count_gauge: UIntGauge,
2289 : shutdown: std::sync::atomic::AtomicBool,
2290 : }
2291 :
2292 : impl TimelineMetrics {
2293 1242 : pub fn new(
2294 1242 : tenant_shard_id: &TenantShardId,
2295 1242 : timeline_id_raw: &TimelineId,
2296 1242 : evictions_with_low_residence_duration_builder: EvictionsWithLowResidenceDurationBuilder,
2297 1242 : ) -> Self {
2298 1242 : let tenant_id = tenant_shard_id.tenant_id.to_string();
2299 1242 : let shard_id = format!("{}", tenant_shard_id.shard_slug());
2300 1242 : let timeline_id = timeline_id_raw.to_string();
2301 1242 : let flush_time_histo = StorageTimeMetrics::new(
2302 1242 : StorageTimeOperation::LayerFlush,
2303 1242 : &tenant_id,
2304 1242 : &shard_id,
2305 1242 : &timeline_id,
2306 1242 : );
2307 1242 : let compact_time_histo = StorageTimeMetrics::new(
2308 1242 : StorageTimeOperation::Compact,
2309 1242 : &tenant_id,
2310 1242 : &shard_id,
2311 1242 : &timeline_id,
2312 1242 : );
2313 1242 : let create_images_time_histo = StorageTimeMetrics::new(
2314 1242 : StorageTimeOperation::CreateImages,
2315 1242 : &tenant_id,
2316 1242 : &shard_id,
2317 1242 : &timeline_id,
2318 1242 : );
2319 1242 : let logical_size_histo = StorageTimeMetrics::new(
2320 1242 : StorageTimeOperation::LogicalSize,
2321 1242 : &tenant_id,
2322 1242 : &shard_id,
2323 1242 : &timeline_id,
2324 1242 : );
2325 1242 : let imitate_logical_size_histo = StorageTimeMetrics::new(
2326 1242 : StorageTimeOperation::ImitateLogicalSize,
2327 1242 : &tenant_id,
2328 1242 : &shard_id,
2329 1242 : &timeline_id,
2330 1242 : );
2331 1242 : let load_layer_map_histo = StorageTimeMetrics::new(
2332 1242 : StorageTimeOperation::LoadLayerMap,
2333 1242 : &tenant_id,
2334 1242 : &shard_id,
2335 1242 : &timeline_id,
2336 1242 : );
2337 1242 : let garbage_collect_histo = StorageTimeMetrics::new(
2338 1242 : StorageTimeOperation::Gc,
2339 1242 : &tenant_id,
2340 1242 : &shard_id,
2341 1242 : &timeline_id,
2342 1242 : );
2343 1242 : let find_gc_cutoffs_histo = StorageTimeMetrics::new(
2344 1242 : StorageTimeOperation::FindGcCutoffs,
2345 1242 : &tenant_id,
2346 1242 : &shard_id,
2347 1242 : &timeline_id,
2348 1242 : );
2349 1242 : let last_record_gauge = LAST_RECORD_LSN
2350 1242 : .get_metric_with_label_values(&[&tenant_id, &shard_id, &timeline_id])
2351 1242 : .unwrap();
2352 1242 :
2353 1242 : let pitr_history_size = PITR_HISTORY_SIZE
2354 1242 : .get_metric_with_label_values(&[&tenant_id, &shard_id, &timeline_id])
2355 1242 : .unwrap();
2356 1242 :
2357 1242 : let archival_size = TIMELINE_ARCHIVE_SIZE
2358 1242 : .get_metric_with_label_values(&[&tenant_id, &shard_id, &timeline_id])
2359 1242 : .unwrap();
2360 1242 :
2361 1242 : let layer_size_image = TIMELINE_LAYER_SIZE
2362 1242 : .get_metric_with_label_values(&[
2363 1242 : &tenant_id,
2364 1242 : &shard_id,
2365 1242 : &timeline_id,
2366 1242 : MetricLayerKind::Image.into(),
2367 1242 : ])
2368 1242 : .unwrap();
2369 1242 :
2370 1242 : let layer_count_image = TIMELINE_LAYER_COUNT
2371 1242 : .get_metric_with_label_values(&[
2372 1242 : &tenant_id,
2373 1242 : &shard_id,
2374 1242 : &timeline_id,
2375 1242 : MetricLayerKind::Image.into(),
2376 1242 : ])
2377 1242 : .unwrap();
2378 1242 :
2379 1242 : let layer_size_delta = TIMELINE_LAYER_SIZE
2380 1242 : .get_metric_with_label_values(&[
2381 1242 : &tenant_id,
2382 1242 : &shard_id,
2383 1242 : &timeline_id,
2384 1242 : MetricLayerKind::Delta.into(),
2385 1242 : ])
2386 1242 : .unwrap();
2387 1242 :
2388 1242 : let layer_count_delta = TIMELINE_LAYER_COUNT
2389 1242 : .get_metric_with_label_values(&[
2390 1242 : &tenant_id,
2391 1242 : &shard_id,
2392 1242 : &timeline_id,
2393 1242 : MetricLayerKind::Delta.into(),
2394 1242 : ])
2395 1242 : .unwrap();
2396 1242 :
2397 1242 : let standby_horizon_gauge = STANDBY_HORIZON
2398 1242 : .get_metric_with_label_values(&[&tenant_id, &shard_id, &timeline_id])
2399 1242 : .unwrap();
2400 1242 : let resident_physical_size_gauge = RESIDENT_PHYSICAL_SIZE
2401 1242 : .get_metric_with_label_values(&[&tenant_id, &shard_id, &timeline_id])
2402 1242 : .unwrap();
2403 1242 : let visible_physical_size_gauge = VISIBLE_PHYSICAL_SIZE
2404 1242 : .get_metric_with_label_values(&[&tenant_id, &shard_id, &timeline_id])
2405 1242 : .unwrap();
2406 1242 : // TODO: we shouldn't expose this metric
2407 1242 : let current_logical_size_gauge = CURRENT_LOGICAL_SIZE
2408 1242 : .get_metric_with_label_values(&[&tenant_id, &shard_id, &timeline_id])
2409 1242 : .unwrap();
2410 1242 : let aux_file_size_gauge = AUX_FILE_SIZE
2411 1242 : .get_metric_with_label_values(&[&tenant_id, &shard_id, &timeline_id])
2412 1242 : .unwrap();
2413 1242 : // TODO use impl Trait syntax here once we have ability to use it: https://github.com/rust-lang/rust/issues/63065
2414 1242 : let directory_entries_count_gauge_closure = {
2415 1242 : let tenant_shard_id = *tenant_shard_id;
2416 1242 : let timeline_id_raw = *timeline_id_raw;
2417 0 : move || {
2418 0 : let tenant_id = tenant_shard_id.tenant_id.to_string();
2419 0 : let shard_id = format!("{}", tenant_shard_id.shard_slug());
2420 0 : let timeline_id = timeline_id_raw.to_string();
2421 0 : let gauge: UIntGauge = DIRECTORY_ENTRIES_COUNT
2422 0 : .get_metric_with_label_values(&[&tenant_id, &shard_id, &timeline_id])
2423 0 : .unwrap();
2424 0 : gauge
2425 0 : }
2426 : };
2427 1242 : let directory_entries_count_gauge: Lazy<UIntGauge, Box<dyn Send + Fn() -> UIntGauge>> =
2428 1242 : Lazy::new(Box::new(directory_entries_count_gauge_closure));
2429 1242 : let evictions = EVICTIONS
2430 1242 : .get_metric_with_label_values(&[&tenant_id, &shard_id, &timeline_id])
2431 1242 : .unwrap();
2432 1242 : let evictions_with_low_residence_duration = evictions_with_low_residence_duration_builder
2433 1242 : .build(&tenant_id, &shard_id, &timeline_id);
2434 1242 :
2435 1242 : let valid_lsn_lease_count_gauge = VALID_LSN_LEASE_COUNT
2436 1242 : .get_metric_with_label_values(&[&tenant_id, &shard_id, &timeline_id])
2437 1242 : .unwrap();
2438 1242 :
2439 1242 : TimelineMetrics {
2440 1242 : tenant_id,
2441 1242 : shard_id,
2442 1242 : timeline_id,
2443 1242 : flush_time_histo,
2444 1242 : compact_time_histo,
2445 1242 : create_images_time_histo,
2446 1242 : logical_size_histo,
2447 1242 : imitate_logical_size_histo,
2448 1242 : garbage_collect_histo,
2449 1242 : find_gc_cutoffs_histo,
2450 1242 : load_layer_map_histo,
2451 1242 : last_record_gauge,
2452 1242 : pitr_history_size,
2453 1242 : archival_size,
2454 1242 : layer_size_image,
2455 1242 : layer_count_image,
2456 1242 : layer_size_delta,
2457 1242 : layer_count_delta,
2458 1242 : standby_horizon_gauge,
2459 1242 : resident_physical_size_gauge,
2460 1242 : visible_physical_size_gauge,
2461 1242 : current_logical_size_gauge,
2462 1242 : aux_file_size_gauge,
2463 1242 : directory_entries_count_gauge,
2464 1242 : evictions,
2465 1242 : evictions_with_low_residence_duration: std::sync::RwLock::new(
2466 1242 : evictions_with_low_residence_duration,
2467 1242 : ),
2468 1242 : valid_lsn_lease_count_gauge,
2469 1242 : shutdown: std::sync::atomic::AtomicBool::default(),
2470 1242 : }
2471 1242 : }
2472 :
2473 4530 : pub(crate) fn record_new_file_metrics(&self, sz: u64) {
2474 4530 : self.resident_physical_size_add(sz);
2475 4530 : }
2476 :
2477 1430 : pub(crate) fn resident_physical_size_sub(&self, sz: u64) {
2478 1430 : self.resident_physical_size_gauge.sub(sz);
2479 1430 : crate::metrics::RESIDENT_PHYSICAL_SIZE_GLOBAL.sub(sz);
2480 1430 : }
2481 :
2482 4620 : pub(crate) fn resident_physical_size_add(&self, sz: u64) {
2483 4620 : self.resident_physical_size_gauge.add(sz);
2484 4620 : crate::metrics::RESIDENT_PHYSICAL_SIZE_GLOBAL.add(sz);
2485 4620 : }
2486 :
2487 24 : pub(crate) fn resident_physical_size_get(&self) -> u64 {
2488 24 : self.resident_physical_size_gauge.get()
2489 24 : }
2490 :
2491 24 : pub(crate) fn shutdown(&self) {
2492 24 : let was_shutdown = self
2493 24 : .shutdown
2494 24 : .swap(true, std::sync::atomic::Ordering::Relaxed);
2495 24 :
2496 24 : if was_shutdown {
2497 : // this happens on tenant deletion because tenant first shuts down timelines, then
2498 : // invokes timeline deletion which first shuts down the timeline again.
2499 : // TODO: this can be removed once https://github.com/neondatabase/neon/issues/5080
2500 0 : return;
2501 24 : }
2502 24 :
2503 24 : let tenant_id = &self.tenant_id;
2504 24 : let timeline_id = &self.timeline_id;
2505 24 : let shard_id = &self.shard_id;
2506 24 : let _ = LAST_RECORD_LSN.remove_label_values(&[tenant_id, shard_id, timeline_id]);
2507 24 : let _ = STANDBY_HORIZON.remove_label_values(&[tenant_id, shard_id, timeline_id]);
2508 24 : {
2509 24 : RESIDENT_PHYSICAL_SIZE_GLOBAL.sub(self.resident_physical_size_get());
2510 24 : let _ = RESIDENT_PHYSICAL_SIZE.remove_label_values(&[tenant_id, shard_id, timeline_id]);
2511 24 : }
2512 24 : let _ = VISIBLE_PHYSICAL_SIZE.remove_label_values(&[tenant_id, shard_id, timeline_id]);
2513 24 : let _ = CURRENT_LOGICAL_SIZE.remove_label_values(&[tenant_id, shard_id, timeline_id]);
2514 24 : if let Some(metric) = Lazy::get(&DIRECTORY_ENTRIES_COUNT) {
2515 0 : let _ = metric.remove_label_values(&[tenant_id, shard_id, timeline_id]);
2516 24 : }
2517 :
2518 24 : let _ = TIMELINE_ARCHIVE_SIZE.remove_label_values(&[tenant_id, shard_id, timeline_id]);
2519 24 : let _ = PITR_HISTORY_SIZE.remove_label_values(&[tenant_id, shard_id, timeline_id]);
2520 24 :
2521 24 : let _ = TIMELINE_LAYER_SIZE.remove_label_values(&[
2522 24 : tenant_id,
2523 24 : shard_id,
2524 24 : timeline_id,
2525 24 : MetricLayerKind::Image.into(),
2526 24 : ]);
2527 24 : let _ = TIMELINE_LAYER_COUNT.remove_label_values(&[
2528 24 : tenant_id,
2529 24 : shard_id,
2530 24 : timeline_id,
2531 24 : MetricLayerKind::Image.into(),
2532 24 : ]);
2533 24 : let _ = TIMELINE_LAYER_SIZE.remove_label_values(&[
2534 24 : tenant_id,
2535 24 : shard_id,
2536 24 : timeline_id,
2537 24 : MetricLayerKind::Delta.into(),
2538 24 : ]);
2539 24 : let _ = TIMELINE_LAYER_COUNT.remove_label_values(&[
2540 24 : tenant_id,
2541 24 : shard_id,
2542 24 : timeline_id,
2543 24 : MetricLayerKind::Delta.into(),
2544 24 : ]);
2545 24 :
2546 24 : let _ = EVICTIONS.remove_label_values(&[tenant_id, shard_id, timeline_id]);
2547 24 : let _ = AUX_FILE_SIZE.remove_label_values(&[tenant_id, shard_id, timeline_id]);
2548 24 : let _ = VALID_LSN_LEASE_COUNT.remove_label_values(&[tenant_id, shard_id, timeline_id]);
2549 24 :
2550 24 : self.evictions_with_low_residence_duration
2551 24 : .write()
2552 24 : .unwrap()
2553 24 : .remove(tenant_id, shard_id, timeline_id);
2554 :
2555 : // The following metrics are born outside of the TimelineMetrics lifecycle but still
2556 : // removed at the end of it. The idea is to have the metrics outlive the
2557 : // entity during which they're observed, e.g., the smgr metrics shall
2558 : // outlive an individual smgr connection, but not the timeline.
2559 :
2560 216 : for op in StorageTimeOperation::VARIANTS {
2561 192 : let _ = STORAGE_TIME_SUM_PER_TIMELINE.remove_label_values(&[
2562 192 : op,
2563 192 : tenant_id,
2564 192 : shard_id,
2565 192 : timeline_id,
2566 192 : ]);
2567 192 : let _ = STORAGE_TIME_COUNT_PER_TIMELINE.remove_label_values(&[
2568 192 : op,
2569 192 : tenant_id,
2570 192 : shard_id,
2571 192 : timeline_id,
2572 192 : ]);
2573 192 : }
2574 :
2575 72 : for op in STORAGE_IO_SIZE_OPERATIONS {
2576 48 : let _ = STORAGE_IO_SIZE.remove_label_values(&[op, tenant_id, shard_id, timeline_id]);
2577 48 : }
2578 :
2579 24 : let _ = SMGR_QUERY_TIME_PER_TENANT_TIMELINE.remove_label_values(&[
2580 24 : SmgrQueryType::GetPageAtLsn.into(),
2581 24 : tenant_id,
2582 24 : shard_id,
2583 24 : timeline_id,
2584 24 : ]);
2585 24 : }
2586 : }
2587 :
2588 18 : pub(crate) fn remove_tenant_metrics(tenant_shard_id: &TenantShardId) {
2589 18 : // Only shard zero deals in synthetic sizes
2590 18 : if tenant_shard_id.is_shard_zero() {
2591 18 : let tid = tenant_shard_id.tenant_id.to_string();
2592 18 : let _ = TENANT_SYNTHETIC_SIZE_METRIC.remove_label_values(&[&tid]);
2593 18 : }
2594 :
2595 : // we leave the BROKEN_TENANTS_SET entry if any
2596 18 : }
2597 :
2598 : use futures::Future;
2599 : use pin_project_lite::pin_project;
2600 : use std::collections::HashMap;
2601 : use std::num::NonZeroUsize;
2602 : use std::pin::Pin;
2603 : use std::sync::atomic::AtomicU64;
2604 : use std::sync::{Arc, Mutex};
2605 : use std::task::{Context, Poll};
2606 : use std::time::{Duration, Instant};
2607 :
2608 : use crate::context::{PageContentKind, RequestContext};
2609 : use crate::task_mgr::TaskKind;
2610 : use crate::tenant::mgr::TenantSlot;
2611 : use crate::tenant::tasks::BackgroundLoopKind;
2612 :
2613 : /// Maintain a per timeline gauge in addition to the global gauge.
2614 : pub(crate) struct PerTimelineRemotePhysicalSizeGauge {
2615 : last_set: AtomicU64,
2616 : gauge: UIntGauge,
2617 : }
2618 :
2619 : impl PerTimelineRemotePhysicalSizeGauge {
2620 1272 : fn new(per_timeline_gauge: UIntGauge) -> Self {
2621 1272 : Self {
2622 1272 : last_set: AtomicU64::new(0),
2623 1272 : gauge: per_timeline_gauge,
2624 1272 : }
2625 1272 : }
2626 5459 : pub(crate) fn set(&self, sz: u64) {
2627 5459 : self.gauge.set(sz);
2628 5459 : let prev = self.last_set.swap(sz, std::sync::atomic::Ordering::Relaxed);
2629 5459 : if sz < prev {
2630 78 : REMOTE_PHYSICAL_SIZE_GLOBAL.sub(prev - sz);
2631 5381 : } else {
2632 5381 : REMOTE_PHYSICAL_SIZE_GLOBAL.add(sz - prev);
2633 5381 : };
2634 5459 : }
2635 6 : pub(crate) fn get(&self) -> u64 {
2636 6 : self.gauge.get()
2637 6 : }
2638 : }
2639 :
2640 : impl Drop for PerTimelineRemotePhysicalSizeGauge {
2641 54 : fn drop(&mut self) {
2642 54 : REMOTE_PHYSICAL_SIZE_GLOBAL.sub(self.last_set.load(std::sync::atomic::Ordering::Relaxed));
2643 54 : }
2644 : }
2645 :
2646 : pub(crate) struct RemoteTimelineClientMetrics {
2647 : tenant_id: String,
2648 : shard_id: String,
2649 : timeline_id: String,
2650 : pub(crate) remote_physical_size_gauge: PerTimelineRemotePhysicalSizeGauge,
2651 : calls: Mutex<HashMap<(&'static str, &'static str), IntCounterPair>>,
2652 : bytes_started_counter: Mutex<HashMap<(&'static str, &'static str), IntCounter>>,
2653 : bytes_finished_counter: Mutex<HashMap<(&'static str, &'static str), IntCounter>>,
2654 : }
2655 :
2656 : impl RemoteTimelineClientMetrics {
2657 1272 : pub fn new(tenant_shard_id: &TenantShardId, timeline_id: &TimelineId) -> Self {
2658 1272 : let tenant_id_str = tenant_shard_id.tenant_id.to_string();
2659 1272 : let shard_id_str = format!("{}", tenant_shard_id.shard_slug());
2660 1272 : let timeline_id_str = timeline_id.to_string();
2661 1272 :
2662 1272 : let remote_physical_size_gauge = PerTimelineRemotePhysicalSizeGauge::new(
2663 1272 : REMOTE_PHYSICAL_SIZE
2664 1272 : .get_metric_with_label_values(&[&tenant_id_str, &shard_id_str, &timeline_id_str])
2665 1272 : .unwrap(),
2666 1272 : );
2667 1272 :
2668 1272 : RemoteTimelineClientMetrics {
2669 1272 : tenant_id: tenant_id_str,
2670 1272 : shard_id: shard_id_str,
2671 1272 : timeline_id: timeline_id_str,
2672 1272 : calls: Mutex::new(HashMap::default()),
2673 1272 : bytes_started_counter: Mutex::new(HashMap::default()),
2674 1272 : bytes_finished_counter: Mutex::new(HashMap::default()),
2675 1272 : remote_physical_size_gauge,
2676 1272 : }
2677 1272 : }
2678 :
2679 8140 : pub fn remote_operation_time(
2680 8140 : &self,
2681 8140 : file_kind: &RemoteOpFileKind,
2682 8140 : op_kind: &RemoteOpKind,
2683 8140 : status: &'static str,
2684 8140 : ) -> Histogram {
2685 8140 : let key = (file_kind.as_str(), op_kind.as_str(), status);
2686 8140 : REMOTE_OPERATION_TIME
2687 8140 : .get_metric_with_label_values(&[key.0, key.1, key.2])
2688 8140 : .unwrap()
2689 8140 : }
2690 :
2691 19167 : fn calls_counter_pair(
2692 19167 : &self,
2693 19167 : file_kind: &RemoteOpFileKind,
2694 19167 : op_kind: &RemoteOpKind,
2695 19167 : ) -> IntCounterPair {
2696 19167 : let mut guard = self.calls.lock().unwrap();
2697 19167 : let key = (file_kind.as_str(), op_kind.as_str());
2698 19167 : let metric = guard.entry(key).or_insert_with(move || {
2699 2242 : REMOTE_TIMELINE_CLIENT_CALLS
2700 2242 : .get_metric_with_label_values(&[
2701 2242 : &self.tenant_id,
2702 2242 : &self.shard_id,
2703 2242 : &self.timeline_id,
2704 2242 : key.0,
2705 2242 : key.1,
2706 2242 : ])
2707 2242 : .unwrap()
2708 19167 : });
2709 19167 : metric.clone()
2710 19167 : }
2711 :
2712 4554 : fn bytes_started_counter(
2713 4554 : &self,
2714 4554 : file_kind: &RemoteOpFileKind,
2715 4554 : op_kind: &RemoteOpKind,
2716 4554 : ) -> IntCounter {
2717 4554 : let mut guard = self.bytes_started_counter.lock().unwrap();
2718 4554 : let key = (file_kind.as_str(), op_kind.as_str());
2719 4554 : let metric = guard.entry(key).or_insert_with(move || {
2720 864 : REMOTE_TIMELINE_CLIENT_BYTES_STARTED_COUNTER
2721 864 : .get_metric_with_label_values(&[
2722 864 : &self.tenant_id,
2723 864 : &self.shard_id,
2724 864 : &self.timeline_id,
2725 864 : key.0,
2726 864 : key.1,
2727 864 : ])
2728 864 : .unwrap()
2729 4554 : });
2730 4554 : metric.clone()
2731 4554 : }
2732 :
2733 8399 : fn bytes_finished_counter(
2734 8399 : &self,
2735 8399 : file_kind: &RemoteOpFileKind,
2736 8399 : op_kind: &RemoteOpKind,
2737 8399 : ) -> IntCounter {
2738 8399 : let mut guard = self.bytes_finished_counter.lock().unwrap();
2739 8399 : let key = (file_kind.as_str(), op_kind.as_str());
2740 8399 : let metric = guard.entry(key).or_insert_with(move || {
2741 864 : REMOTE_TIMELINE_CLIENT_BYTES_FINISHED_COUNTER
2742 864 : .get_metric_with_label_values(&[
2743 864 : &self.tenant_id,
2744 864 : &self.shard_id,
2745 864 : &self.timeline_id,
2746 864 : key.0,
2747 864 : key.1,
2748 864 : ])
2749 864 : .unwrap()
2750 8399 : });
2751 8399 : metric.clone()
2752 8399 : }
2753 : }
2754 :
2755 : #[cfg(test)]
2756 : impl RemoteTimelineClientMetrics {
2757 18 : pub fn get_bytes_started_counter_value(
2758 18 : &self,
2759 18 : file_kind: &RemoteOpFileKind,
2760 18 : op_kind: &RemoteOpKind,
2761 18 : ) -> Option<u64> {
2762 18 : let guard = self.bytes_started_counter.lock().unwrap();
2763 18 : let key = (file_kind.as_str(), op_kind.as_str());
2764 18 : guard.get(&key).map(|counter| counter.get())
2765 18 : }
2766 :
2767 18 : pub fn get_bytes_finished_counter_value(
2768 18 : &self,
2769 18 : file_kind: &RemoteOpFileKind,
2770 18 : op_kind: &RemoteOpKind,
2771 18 : ) -> Option<u64> {
2772 18 : let guard = self.bytes_finished_counter.lock().unwrap();
2773 18 : let key = (file_kind.as_str(), op_kind.as_str());
2774 18 : guard.get(&key).map(|counter| counter.get())
2775 18 : }
2776 : }
2777 :
2778 : /// See [`RemoteTimelineClientMetrics::call_begin`].
2779 : #[must_use]
2780 : pub(crate) struct RemoteTimelineClientCallMetricGuard {
2781 : /// Decremented on drop.
2782 : calls_counter_pair: Option<IntCounterPair>,
2783 : /// If Some(), this references the bytes_finished metric, and we increment it by the given `u64` on drop.
2784 : bytes_finished: Option<(IntCounter, u64)>,
2785 : }
2786 :
2787 : impl RemoteTimelineClientCallMetricGuard {
2788 : /// Consume this guard object without performing the metric updates it would do on `drop()`.
2789 : /// The caller vouches to do the metric updates manually.
2790 10310 : pub fn will_decrement_manually(mut self) {
2791 10310 : let RemoteTimelineClientCallMetricGuard {
2792 10310 : calls_counter_pair,
2793 10310 : bytes_finished,
2794 10310 : } = &mut self;
2795 10310 : calls_counter_pair.take();
2796 10310 : bytes_finished.take();
2797 10310 : }
2798 : }
2799 :
2800 : impl Drop for RemoteTimelineClientCallMetricGuard {
2801 10388 : fn drop(&mut self) {
2802 10388 : let RemoteTimelineClientCallMetricGuard {
2803 10388 : calls_counter_pair,
2804 10388 : bytes_finished,
2805 10388 : } = self;
2806 10388 : if let Some(guard) = calls_counter_pair.take() {
2807 78 : guard.dec();
2808 10310 : }
2809 10388 : if let Some((bytes_finished_metric, value)) = bytes_finished {
2810 0 : bytes_finished_metric.inc_by(*value);
2811 10388 : }
2812 10388 : }
2813 : }
2814 :
2815 : /// The enum variants communicate to the [`RemoteTimelineClientMetrics`] whether to
2816 : /// track the byte size of this call in applicable metric(s).
2817 : pub(crate) enum RemoteTimelineClientMetricsCallTrackSize {
2818 : /// Do not account for this call's byte size in any metrics.
2819 : /// The `reason` field is there to make the call sites self-documenting
2820 : /// about why they don't need the metric.
2821 : DontTrackSize { reason: &'static str },
2822 : /// Track the byte size of the call in applicable metric(s).
2823 : Bytes(u64),
2824 : }
2825 :
2826 : impl RemoteTimelineClientMetrics {
2827 : /// Update the metrics that change when a call to the remote timeline client instance starts.
2828 : ///
2829 : /// Drop the returned guard object once the operation is finished to updates corresponding metrics that track completions.
2830 : /// Or, use [`RemoteTimelineClientCallMetricGuard::will_decrement_manually`] and [`call_end`](Self::call_end) if that
2831 : /// is more suitable.
2832 : /// Never do both.
2833 10388 : pub(crate) fn call_begin(
2834 10388 : &self,
2835 10388 : file_kind: &RemoteOpFileKind,
2836 10388 : op_kind: &RemoteOpKind,
2837 10388 : size: RemoteTimelineClientMetricsCallTrackSize,
2838 10388 : ) -> RemoteTimelineClientCallMetricGuard {
2839 10388 : let calls_counter_pair = self.calls_counter_pair(file_kind, op_kind);
2840 10388 : calls_counter_pair.inc();
2841 :
2842 10388 : let bytes_finished = match size {
2843 5834 : RemoteTimelineClientMetricsCallTrackSize::DontTrackSize { reason: _reason } => {
2844 5834 : // nothing to do
2845 5834 : None
2846 : }
2847 4554 : RemoteTimelineClientMetricsCallTrackSize::Bytes(size) => {
2848 4554 : self.bytes_started_counter(file_kind, op_kind).inc_by(size);
2849 4554 : let finished_counter = self.bytes_finished_counter(file_kind, op_kind);
2850 4554 : Some((finished_counter, size))
2851 : }
2852 : };
2853 10388 : RemoteTimelineClientCallMetricGuard {
2854 10388 : calls_counter_pair: Some(calls_counter_pair),
2855 10388 : bytes_finished,
2856 10388 : }
2857 10388 : }
2858 :
2859 : /// Manually udpate the metrics that track completions, instead of using the guard object.
2860 : /// Using the guard object is generally preferable.
2861 : /// See [`call_begin`](Self::call_begin) for more context.
2862 8779 : pub(crate) fn call_end(
2863 8779 : &self,
2864 8779 : file_kind: &RemoteOpFileKind,
2865 8779 : op_kind: &RemoteOpKind,
2866 8779 : size: RemoteTimelineClientMetricsCallTrackSize,
2867 8779 : ) {
2868 8779 : let calls_counter_pair = self.calls_counter_pair(file_kind, op_kind);
2869 8779 : calls_counter_pair.dec();
2870 8779 : match size {
2871 4934 : RemoteTimelineClientMetricsCallTrackSize::DontTrackSize { reason: _reason } => {}
2872 3845 : RemoteTimelineClientMetricsCallTrackSize::Bytes(size) => {
2873 3845 : self.bytes_finished_counter(file_kind, op_kind).inc_by(size);
2874 3845 : }
2875 : }
2876 8779 : }
2877 : }
2878 :
2879 : impl Drop for RemoteTimelineClientMetrics {
2880 54 : fn drop(&mut self) {
2881 54 : let RemoteTimelineClientMetrics {
2882 54 : tenant_id,
2883 54 : shard_id,
2884 54 : timeline_id,
2885 54 : remote_physical_size_gauge,
2886 54 : calls,
2887 54 : bytes_started_counter,
2888 54 : bytes_finished_counter,
2889 54 : } = self;
2890 72 : for ((a, b), _) in calls.get_mut().unwrap().drain() {
2891 72 : let mut res = [Ok(()), Ok(())];
2892 72 : REMOTE_TIMELINE_CLIENT_CALLS
2893 72 : .remove_label_values(&mut res, &[tenant_id, shard_id, timeline_id, a, b]);
2894 72 : // don't care about results
2895 72 : }
2896 54 : for ((a, b), _) in bytes_started_counter.get_mut().unwrap().drain() {
2897 18 : let _ = REMOTE_TIMELINE_CLIENT_BYTES_STARTED_COUNTER.remove_label_values(&[
2898 18 : tenant_id,
2899 18 : shard_id,
2900 18 : timeline_id,
2901 18 : a,
2902 18 : b,
2903 18 : ]);
2904 18 : }
2905 54 : for ((a, b), _) in bytes_finished_counter.get_mut().unwrap().drain() {
2906 18 : let _ = REMOTE_TIMELINE_CLIENT_BYTES_FINISHED_COUNTER.remove_label_values(&[
2907 18 : tenant_id,
2908 18 : shard_id,
2909 18 : timeline_id,
2910 18 : a,
2911 18 : b,
2912 18 : ]);
2913 18 : }
2914 54 : {
2915 54 : let _ = remote_physical_size_gauge; // use to avoid 'unused' warning in desctructuring above
2916 54 : let _ = REMOTE_PHYSICAL_SIZE.remove_label_values(&[tenant_id, shard_id, timeline_id]);
2917 54 : }
2918 54 : }
2919 : }
2920 :
2921 : /// Wrapper future that measures the time spent by a remote storage operation,
2922 : /// and records the time and success/failure as a prometheus metric.
2923 : pub(crate) trait MeasureRemoteOp: Sized {
2924 8422 : fn measure_remote_op(
2925 8422 : self,
2926 8422 : file_kind: RemoteOpFileKind,
2927 8422 : op: RemoteOpKind,
2928 8422 : metrics: Arc<RemoteTimelineClientMetrics>,
2929 8422 : ) -> MeasuredRemoteOp<Self> {
2930 8422 : let start = Instant::now();
2931 8422 : MeasuredRemoteOp {
2932 8422 : inner: self,
2933 8422 : file_kind,
2934 8422 : op,
2935 8422 : start,
2936 8422 : metrics,
2937 8422 : }
2938 8422 : }
2939 : }
2940 :
2941 : impl<T: Sized> MeasureRemoteOp for T {}
2942 :
2943 : pin_project! {
2944 : pub(crate) struct MeasuredRemoteOp<F>
2945 : {
2946 : #[pin]
2947 : inner: F,
2948 : file_kind: RemoteOpFileKind,
2949 : op: RemoteOpKind,
2950 : start: Instant,
2951 : metrics: Arc<RemoteTimelineClientMetrics>,
2952 : }
2953 : }
2954 :
2955 : impl<F: Future<Output = Result<O, E>>, O, E> Future for MeasuredRemoteOp<F> {
2956 : type Output = Result<O, E>;
2957 :
2958 130656 : fn poll(self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<Self::Output> {
2959 130656 : let this = self.project();
2960 130656 : let poll_result = this.inner.poll(cx);
2961 130656 : if let Poll::Ready(ref res) = poll_result {
2962 8140 : let duration = this.start.elapsed();
2963 8140 : let status = if res.is_ok() { &"success" } else { &"failure" };
2964 8140 : this.metrics
2965 8140 : .remote_operation_time(this.file_kind, this.op, status)
2966 8140 : .observe(duration.as_secs_f64());
2967 122516 : }
2968 130656 : poll_result
2969 130656 : }
2970 : }
2971 :
2972 : pub mod tokio_epoll_uring {
2973 : use metrics::{register_int_counter, UIntGauge};
2974 : use once_cell::sync::Lazy;
2975 :
2976 : pub struct Collector {
2977 : descs: Vec<metrics::core::Desc>,
2978 : systems_created: UIntGauge,
2979 : systems_destroyed: UIntGauge,
2980 : }
2981 :
2982 : impl metrics::core::Collector for Collector {
2983 0 : fn desc(&self) -> Vec<&metrics::core::Desc> {
2984 0 : self.descs.iter().collect()
2985 0 : }
2986 :
2987 0 : fn collect(&self) -> Vec<metrics::proto::MetricFamily> {
2988 0 : let mut mfs = Vec::with_capacity(Self::NMETRICS);
2989 0 : let tokio_epoll_uring::metrics::Metrics {
2990 0 : systems_created,
2991 0 : systems_destroyed,
2992 0 : } = tokio_epoll_uring::metrics::global();
2993 0 : self.systems_created.set(systems_created);
2994 0 : mfs.extend(self.systems_created.collect());
2995 0 : self.systems_destroyed.set(systems_destroyed);
2996 0 : mfs.extend(self.systems_destroyed.collect());
2997 0 : mfs
2998 0 : }
2999 : }
3000 :
3001 : impl Collector {
3002 : const NMETRICS: usize = 2;
3003 :
3004 : #[allow(clippy::new_without_default)]
3005 0 : pub fn new() -> Self {
3006 0 : let mut descs = Vec::new();
3007 0 :
3008 0 : let systems_created = UIntGauge::new(
3009 0 : "pageserver_tokio_epoll_uring_systems_created",
3010 0 : "counter of tokio-epoll-uring systems that were created",
3011 0 : )
3012 0 : .unwrap();
3013 0 : descs.extend(
3014 0 : metrics::core::Collector::desc(&systems_created)
3015 0 : .into_iter()
3016 0 : .cloned(),
3017 0 : );
3018 0 :
3019 0 : let systems_destroyed = UIntGauge::new(
3020 0 : "pageserver_tokio_epoll_uring_systems_destroyed",
3021 0 : "counter of tokio-epoll-uring systems that were destroyed",
3022 0 : )
3023 0 : .unwrap();
3024 0 : descs.extend(
3025 0 : metrics::core::Collector::desc(&systems_destroyed)
3026 0 : .into_iter()
3027 0 : .cloned(),
3028 0 : );
3029 0 :
3030 0 : Self {
3031 0 : descs,
3032 0 : systems_created,
3033 0 : systems_destroyed,
3034 0 : }
3035 0 : }
3036 : }
3037 :
3038 303 : pub(crate) static THREAD_LOCAL_LAUNCH_SUCCESSES: Lazy<metrics::IntCounter> = Lazy::new(|| {
3039 303 : register_int_counter!(
3040 303 : "pageserver_tokio_epoll_uring_pageserver_thread_local_launch_success_count",
3041 303 : "Number of times where thread_local_system creation spanned multiple executor threads",
3042 303 : )
3043 303 : .unwrap()
3044 303 : });
3045 :
3046 0 : pub(crate) static THREAD_LOCAL_LAUNCH_FAILURES: Lazy<metrics::IntCounter> = Lazy::new(|| {
3047 0 : register_int_counter!(
3048 0 : "pageserver_tokio_epoll_uring_pageserver_thread_local_launch_failures_count",
3049 0 : "Number of times thread_local_system creation failed and was retried after back-off.",
3050 0 : )
3051 0 : .unwrap()
3052 0 : });
3053 : }
3054 :
3055 : pub(crate) mod tenant_throttling {
3056 : use metrics::{register_int_counter_vec, IntCounter};
3057 : use once_cell::sync::Lazy;
3058 :
3059 : use crate::tenant::{self, throttle::Metric};
3060 :
3061 : pub(crate) struct TimelineGet {
3062 : wait_time: IntCounter,
3063 : count: IntCounter,
3064 : }
3065 :
3066 516 : pub(crate) static TIMELINE_GET: Lazy<TimelineGet> = Lazy::new(|| {
3067 516 : static WAIT_USECS: Lazy<metrics::IntCounterVec> = Lazy::new(|| {
3068 516 : register_int_counter_vec!(
3069 516 : "pageserver_tenant_throttling_wait_usecs_sum_global",
3070 516 : "Sum of microseconds that tenants spent waiting for a tenant throttle of a given kind.",
3071 516 : &["kind"]
3072 516 : )
3073 516 : .unwrap()
3074 516 : });
3075 :
3076 516 : static WAIT_COUNT: Lazy<metrics::IntCounterVec> = Lazy::new(|| {
3077 516 : register_int_counter_vec!(
3078 516 : "pageserver_tenant_throttling_count_global",
3079 516 : "Count of tenant throttlings, by kind of throttle.",
3080 516 : &["kind"]
3081 516 : )
3082 516 : .unwrap()
3083 516 : });
3084 :
3085 516 : let kind = "timeline_get";
3086 516 : TimelineGet {
3087 516 : wait_time: WAIT_USECS.with_label_values(&[kind]),
3088 516 : count: WAIT_COUNT.with_label_values(&[kind]),
3089 516 : }
3090 516 : });
3091 :
3092 : impl Metric for &'static TimelineGet {
3093 : #[inline(always)]
3094 0 : fn observe_throttling(
3095 0 : &self,
3096 0 : tenant::throttle::Observation { wait_time }: &tenant::throttle::Observation,
3097 0 : ) {
3098 0 : let val = u64::try_from(wait_time.as_micros()).unwrap();
3099 0 : self.wait_time.inc_by(val);
3100 0 : self.count.inc();
3101 0 : }
3102 : }
3103 : }
3104 :
3105 : pub(crate) mod disk_usage_based_eviction {
3106 : use super::*;
3107 :
3108 : pub(crate) struct Metrics {
3109 : pub(crate) tenant_collection_time: Histogram,
3110 : pub(crate) tenant_layer_count: Histogram,
3111 : pub(crate) layers_collected: IntCounter,
3112 : pub(crate) layers_selected: IntCounter,
3113 : pub(crate) layers_evicted: IntCounter,
3114 : }
3115 :
3116 : impl Default for Metrics {
3117 0 : fn default() -> Self {
3118 0 : let tenant_collection_time = register_histogram!(
3119 0 : "pageserver_disk_usage_based_eviction_tenant_collection_seconds",
3120 0 : "Time spent collecting layers from a tenant -- not normalized by collected layer amount",
3121 0 : vec![0.001, 0.005, 0.01, 0.05, 0.1, 0.5, 1.0, 5.0, 10.0]
3122 0 : )
3123 0 : .unwrap();
3124 0 :
3125 0 : let tenant_layer_count = register_histogram!(
3126 0 : "pageserver_disk_usage_based_eviction_tenant_collected_layers",
3127 0 : "Amount of layers gathered from a tenant",
3128 0 : vec![5.0, 50.0, 500.0, 5000.0, 50000.0]
3129 0 : )
3130 0 : .unwrap();
3131 0 :
3132 0 : let layers_collected = register_int_counter!(
3133 0 : "pageserver_disk_usage_based_eviction_collected_layers_total",
3134 0 : "Amount of layers collected"
3135 0 : )
3136 0 : .unwrap();
3137 0 :
3138 0 : let layers_selected = register_int_counter!(
3139 0 : "pageserver_disk_usage_based_eviction_select_layers_total",
3140 0 : "Amount of layers selected"
3141 0 : )
3142 0 : .unwrap();
3143 0 :
3144 0 : let layers_evicted = register_int_counter!(
3145 0 : "pageserver_disk_usage_based_eviction_evicted_layers_total",
3146 0 : "Amount of layers successfully evicted"
3147 0 : )
3148 0 : .unwrap();
3149 0 :
3150 0 : Self {
3151 0 : tenant_collection_time,
3152 0 : tenant_layer_count,
3153 0 : layers_collected,
3154 0 : layers_selected,
3155 0 : layers_evicted,
3156 0 : }
3157 0 : }
3158 : }
3159 :
3160 : pub(crate) static METRICS: Lazy<Metrics> = Lazy::new(Metrics::default);
3161 : }
3162 :
3163 498 : static TOKIO_EXECUTOR_THREAD_COUNT: Lazy<UIntGaugeVec> = Lazy::new(|| {
3164 498 : register_uint_gauge_vec!(
3165 498 : "pageserver_tokio_executor_thread_configured_count",
3166 498 : "Total number of configued tokio executor threads in the process.
3167 498 : The `setup` label denotes whether we're running with multiple runtimes or a single runtime.",
3168 498 : &["setup"],
3169 498 : )
3170 498 : .unwrap()
3171 498 : });
3172 :
3173 498 : pub(crate) fn set_tokio_runtime_setup(setup: &str, num_threads: NonZeroUsize) {
3174 : static SERIALIZE: std::sync::Mutex<()> = std::sync::Mutex::new(());
3175 498 : let _guard = SERIALIZE.lock().unwrap();
3176 498 : TOKIO_EXECUTOR_THREAD_COUNT.reset();
3177 498 : TOKIO_EXECUTOR_THREAD_COUNT
3178 498 : .get_metric_with_label_values(&[setup])
3179 498 : .unwrap()
3180 498 : .set(u64::try_from(num_threads.get()).unwrap());
3181 498 : }
3182 :
3183 0 : pub fn preinitialize_metrics() {
3184 0 : // Python tests need these and on some we do alerting.
3185 0 : //
3186 0 : // FIXME(4813): make it so that we have no top level metrics as this fn will easily fall out of
3187 0 : // order:
3188 0 : // - global metrics reside in a Lazy<PageserverMetrics>
3189 0 : // - access via crate::metrics::PS_METRICS.some_metric.inc()
3190 0 : // - could move the statics into TimelineMetrics::new()?
3191 0 :
3192 0 : // counters
3193 0 : [
3194 0 : &UNEXPECTED_ONDEMAND_DOWNLOADS,
3195 0 : &WALRECEIVER_STARTED_CONNECTIONS,
3196 0 : &WALRECEIVER_BROKER_UPDATES,
3197 0 : &WALRECEIVER_CANDIDATES_ADDED,
3198 0 : &WALRECEIVER_CANDIDATES_REMOVED,
3199 0 : &tokio_epoll_uring::THREAD_LOCAL_LAUNCH_FAILURES,
3200 0 : &tokio_epoll_uring::THREAD_LOCAL_LAUNCH_SUCCESSES,
3201 0 : &REMOTE_ONDEMAND_DOWNLOADED_LAYERS,
3202 0 : &REMOTE_ONDEMAND_DOWNLOADED_BYTES,
3203 0 : &CIRCUIT_BREAKERS_BROKEN,
3204 0 : &CIRCUIT_BREAKERS_UNBROKEN,
3205 0 : ]
3206 0 : .into_iter()
3207 0 : .for_each(|c| {
3208 0 : Lazy::force(c);
3209 0 : });
3210 0 :
3211 0 : // Deletion queue stats
3212 0 : Lazy::force(&DELETION_QUEUE);
3213 0 :
3214 0 : // Tenant stats
3215 0 : Lazy::force(&TENANT);
3216 0 :
3217 0 : // Tenant manager stats
3218 0 : Lazy::force(&TENANT_MANAGER);
3219 0 :
3220 0 : Lazy::force(&crate::tenant::storage_layer::layer::LAYER_IMPL_METRICS);
3221 0 : Lazy::force(&disk_usage_based_eviction::METRICS);
3222 :
3223 0 : for state_name in pageserver_api::models::TenantState::VARIANTS {
3224 0 : // initialize the metric for all gauges, otherwise the time series might seemingly show
3225 0 : // values from last restart.
3226 0 : TENANT_STATE_METRIC.with_label_values(&[state_name]).set(0);
3227 0 : }
3228 :
3229 : // countervecs
3230 0 : [&BACKGROUND_LOOP_PERIOD_OVERRUN_COUNT]
3231 0 : .into_iter()
3232 0 : .for_each(|c| {
3233 0 : Lazy::force(c);
3234 0 : });
3235 0 :
3236 0 : // gauges
3237 0 : WALRECEIVER_ACTIVE_MANAGERS.get();
3238 0 :
3239 0 : // histograms
3240 0 : [
3241 0 : &READ_NUM_LAYERS_VISITED,
3242 0 : &VEC_READ_NUM_LAYERS_VISITED,
3243 0 : &WAIT_LSN_TIME,
3244 0 : &WAL_REDO_TIME,
3245 0 : &WAL_REDO_RECORDS_HISTOGRAM,
3246 0 : &WAL_REDO_BYTES_HISTOGRAM,
3247 0 : &WAL_REDO_PROCESS_LAUNCH_DURATION_HISTOGRAM,
3248 0 : ]
3249 0 : .into_iter()
3250 0 : .for_each(|h| {
3251 0 : Lazy::force(h);
3252 0 : });
3253 0 :
3254 0 : // Custom
3255 0 : Lazy::force(&RECONSTRUCT_TIME);
3256 0 : Lazy::force(&tenant_throttling::TIMELINE_GET);
3257 0 : Lazy::force(&BASEBACKUP_QUERY_TIME);
3258 0 : Lazy::force(&COMPUTE_COMMANDS_COUNTERS);
3259 0 : }
|
