Line data Source code
1 : //! The per-timeline layer eviction task, which evicts data which has not been accessed for more
2 : //! than a given threshold.
3 : //!
4 : //! Data includes all kinds of caches, namely:
5 : //! - (in-memory layers)
6 : //! - on-demand downloaded layer files on disk
7 : //! - (cached layer file pages)
8 : //! - derived data from layer file contents, namely:
9 : //! - initial logical size
10 : //! - partitioning
11 : //! - (other currently missing unknowns)
12 : //!
13 : //! Items with parentheses are not (yet) touched by this task.
14 : //!
15 : //! See write-up on restart on-demand download spike: <https://gist.github.com/problame/2265bf7b8dc398be834abfead36c76b5>
16 : use std::{
17 : collections::HashMap,
18 : ops::ControlFlow,
19 : sync::Arc,
20 : time::{Duration, SystemTime},
21 : };
22 :
23 : use pageserver_api::models::{EvictionPolicy, EvictionPolicyLayerAccessThreshold};
24 : use tokio::time::Instant;
25 : use tokio_util::sync::CancellationToken;
26 : use tracing::{debug, error, info, info_span, instrument, warn, Instrument};
27 :
28 : use crate::{
29 : context::{DownloadBehavior, RequestContext},
30 : pgdatadir_mapping::CollectKeySpaceError,
31 : task_mgr::{self, TaskKind, BACKGROUND_RUNTIME},
32 : tenant::{
33 : tasks::BackgroundLoopKind, timeline::EvictionError, LogicalSizeCalculationCause, Tenant,
34 : },
35 : };
36 :
37 : use utils::{completion, sync::gate::GateGuard};
38 :
39 : use super::Timeline;
40 :
41 296 : #[derive(Default)]
42 : pub struct EvictionTaskTimelineState {
43 : last_layer_access_imitation: Option<tokio::time::Instant>,
44 : }
45 :
46 88 : #[derive(Default)]
47 : pub struct EvictionTaskTenantState {
48 : last_layer_access_imitation: Option<Instant>,
49 : }
50 :
51 : impl Timeline {
52 0 : pub(super) fn launch_eviction_task(
53 0 : self: &Arc<Self>,
54 0 : background_tasks_can_start: Option<&completion::Barrier>,
55 0 : ) {
56 0 : let self_clone = Arc::clone(self);
57 0 : let background_tasks_can_start = background_tasks_can_start.cloned();
58 0 : task_mgr::spawn(
59 0 : BACKGROUND_RUNTIME.handle(),
60 0 : TaskKind::Eviction,
61 0 : Some(self.tenant_shard_id),
62 0 : Some(self.timeline_id),
63 0 : &format!(
64 0 : "layer eviction for {}/{}",
65 0 : self.tenant_shard_id, self.timeline_id
66 0 : ),
67 0 : false,
68 0 : async move {
69 0 : let cancel = task_mgr::shutdown_token();
70 0 : tokio::select! {
71 0 : _ = cancel.cancelled() => { return Ok(()); }
72 0 : _ = completion::Barrier::maybe_wait(background_tasks_can_start) => {}
73 0 : };
74 :
75 0 : self_clone.eviction_task(cancel).await;
76 0 : Ok(())
77 0 : },
78 0 : );
79 0 : }
80 :
81 0 : #[instrument(skip_all, fields(tenant_id = %self.tenant_shard_id.tenant_id, shard_id = %self.tenant_shard_id.shard_slug(), timeline_id = %self.timeline_id))]
82 : async fn eviction_task(self: Arc<Self>, cancel: CancellationToken) {
83 : use crate::tenant::tasks::random_init_delay;
84 :
85 : // acquire the gate guard only once within a useful span
86 : let Ok(guard) = self.gate.enter() else {
87 : return;
88 : };
89 :
90 : {
91 : let policy = self.get_eviction_policy();
92 : let period = match policy {
93 : EvictionPolicy::LayerAccessThreshold(lat) => lat.period,
94 : EvictionPolicy::OnlyImitiate(lat) => lat.period,
95 : EvictionPolicy::NoEviction => Duration::from_secs(10),
96 : };
97 : if random_init_delay(period, &cancel).await.is_err() {
98 : return;
99 : }
100 : }
101 :
102 : let ctx = RequestContext::new(TaskKind::Eviction, DownloadBehavior::Warn);
103 : loop {
104 : let policy = self.get_eviction_policy();
105 : let cf = self
106 : .eviction_iteration(&policy, &cancel, &guard, &ctx)
107 : .await;
108 :
109 : match cf {
110 : ControlFlow::Break(()) => break,
111 : ControlFlow::Continue(sleep_until) => {
112 : if tokio::time::timeout_at(sleep_until, cancel.cancelled())
113 : .await
114 : .is_ok()
115 : {
116 : break;
117 : }
118 : }
119 : }
120 : }
121 : }
122 :
123 0 : #[instrument(skip_all, fields(policy_kind = policy.discriminant_str()))]
124 : async fn eviction_iteration(
125 : self: &Arc<Self>,
126 : policy: &EvictionPolicy,
127 : cancel: &CancellationToken,
128 : gate: &GateGuard,
129 : ctx: &RequestContext,
130 : ) -> ControlFlow<(), Instant> {
131 0 : debug!("eviction iteration: {policy:?}");
132 : let start = Instant::now();
133 : let (period, threshold) = match policy {
134 : EvictionPolicy::NoEviction => {
135 : // check again in 10 seconds; XXX config watch mechanism
136 : return ControlFlow::Continue(Instant::now() + Duration::from_secs(10));
137 : }
138 : EvictionPolicy::LayerAccessThreshold(p) => {
139 : match self
140 : .eviction_iteration_threshold(p, cancel, gate, ctx)
141 : .await
142 : {
143 : ControlFlow::Break(()) => return ControlFlow::Break(()),
144 : ControlFlow::Continue(()) => (),
145 : }
146 : (p.period, p.threshold)
147 : }
148 : EvictionPolicy::OnlyImitiate(p) => {
149 : if self.imitiate_only(p, cancel, gate, ctx).await.is_break() {
150 : return ControlFlow::Break(());
151 : }
152 : (p.period, p.threshold)
153 : }
154 : };
155 :
156 : let elapsed = start.elapsed();
157 : crate::tenant::tasks::warn_when_period_overrun(
158 : elapsed,
159 : period,
160 : BackgroundLoopKind::Eviction,
161 : );
162 : // FIXME: if we were to mix policies on a pageserver, we would have no way to sense this. I
163 : // don't think that is a relevant fear however, and regardless the imitation should be the
164 : // most costly part.
165 : crate::metrics::EVICTION_ITERATION_DURATION
166 : .get_metric_with_label_values(&[
167 : &format!("{}", period.as_secs()),
168 : &format!("{}", threshold.as_secs()),
169 : ])
170 : .unwrap()
171 : .observe(elapsed.as_secs_f64());
172 :
173 : ControlFlow::Continue(start + period)
174 : }
175 :
176 0 : async fn eviction_iteration_threshold(
177 0 : self: &Arc<Self>,
178 0 : p: &EvictionPolicyLayerAccessThreshold,
179 0 : cancel: &CancellationToken,
180 0 : gate: &GateGuard,
181 0 : ctx: &RequestContext,
182 0 : ) -> ControlFlow<()> {
183 0 : let now = SystemTime::now();
184 0 :
185 0 : let acquire_permit = crate::tenant::tasks::concurrent_background_tasks_rate_limit_permit(
186 0 : BackgroundLoopKind::Eviction,
187 0 : ctx,
188 0 : );
189 :
190 0 : let _permit = tokio::select! {
191 0 : permit = acquire_permit => permit,
192 : _ = cancel.cancelled() => return ControlFlow::Break(()),
193 : _ = self.cancel.cancelled() => return ControlFlow::Break(()),
194 : };
195 :
196 0 : match self.imitate_layer_accesses(p, cancel, gate, ctx).await {
197 0 : ControlFlow::Break(()) => return ControlFlow::Break(()),
198 0 : ControlFlow::Continue(()) => (),
199 0 : }
200 0 :
201 0 : #[derive(Debug, Default)]
202 0 : struct EvictionStats {
203 0 : candidates: usize,
204 0 : evicted: usize,
205 0 : errors: usize,
206 0 : not_evictable: usize,
207 0 : #[allow(dead_code)]
208 0 : skipped_for_shutdown: usize,
209 0 : }
210 0 :
211 0 : let mut stats = EvictionStats::default();
212 0 : // Gather layers for eviction.
213 0 : // NB: all the checks can be invalidated as soon as we release the layer map lock.
214 0 : // We don't want to hold the layer map lock during eviction.
215 0 :
216 0 : // So, we just need to deal with this.
217 0 :
218 0 : if self.remote_client.is_none() {
219 0 : error!("no remote storage configured, cannot evict layers");
220 0 : return ControlFlow::Continue(());
221 0 : }
222 0 :
223 0 : let mut js = tokio::task::JoinSet::new();
224 : {
225 0 : let guard = self.layers.read().await;
226 0 : let layers = guard.layer_map();
227 0 : for hist_layer in layers.iter_historic_layers() {
228 0 : let hist_layer = guard.get_from_desc(&hist_layer);
229 :
230 : // guard against eviction while we inspect it; it might be that eviction_task and
231 : // disk_usage_eviction_task both select the same layers to be evicted, and
232 : // seemingly free up double the space. both succeeding is of no consequence.
233 0 : let guard = match hist_layer.keep_resident().await {
234 0 : Ok(Some(l)) => l,
235 0 : Ok(None) => continue,
236 0 : Err(e) => {
237 : // these should not happen, but we cannot make them statically impossible right
238 : // now.
239 0 : tracing::warn!(layer=%hist_layer, "failed to keep the layer resident: {e:#}");
240 0 : continue;
241 : }
242 : };
243 :
244 0 : let last_activity_ts = hist_layer.access_stats().latest_activity_or_now();
245 :
246 0 : let no_activity_for = match now.duration_since(last_activity_ts) {
247 0 : Ok(d) => d,
248 0 : Err(_e) => {
249 0 : // We reach here if `now` < `last_activity_ts`, which can legitimately
250 0 : // happen if there is an access between us getting `now`, and us getting
251 0 : // the access stats from the layer.
252 0 : //
253 0 : // The other reason why it can happen is system clock skew because
254 0 : // SystemTime::now() is not monotonic, so, even if there is no access
255 0 : // to the layer after we get `now` at the beginning of this function,
256 0 : // it could be that `now` < `last_activity_ts`.
257 0 : //
258 0 : // To distinguish the cases, we would need to record `Instant`s in the
259 0 : // access stats (i.e., monotonic timestamps), but then, the timestamps
260 0 : // values in the access stats would need to be `Instant`'s, and hence
261 0 : // they would be meaningless outside of the pageserver process.
262 0 : // At the time of writing, the trade-off is that access stats are more
263 0 : // valuable than detecting clock skew.
264 0 : continue;
265 : }
266 : };
267 0 : let layer = guard.drop_eviction_guard();
268 0 : if no_activity_for > p.threshold {
269 0 : // this could cause a lot of allocations in some cases
270 0 : js.spawn(async move { layer.evict_and_wait().await });
271 0 : stats.candidates += 1;
272 0 : }
273 : }
274 : };
275 :
276 0 : let join_all = async move {
277 0 : while let Some(next) = js.join_next().await {
278 0 : match next {
279 0 : Ok(Ok(())) => stats.evicted += 1,
280 0 : Ok(Err(EvictionError::NotFound | EvictionError::Downloaded)) => {
281 0 : stats.not_evictable += 1;
282 0 : }
283 0 : Err(je) if je.is_cancelled() => unreachable!("not used"),
284 0 : Err(je) if je.is_panic() => {
285 0 : /* already logged */
286 0 : stats.errors += 1;
287 0 : }
288 0 : Err(je) => tracing::error!("unknown JoinError: {je:?}"),
289 : }
290 : }
291 0 : stats
292 0 : };
293 :
294 0 : tokio::select! {
295 0 : stats = join_all => {
296 : if stats.candidates == stats.not_evictable {
297 0 : debug!(stats=?stats, "eviction iteration complete");
298 : } else if stats.errors > 0 || stats.not_evictable > 0 {
299 0 : warn!(stats=?stats, "eviction iteration complete");
300 : } else {
301 0 : info!(stats=?stats, "eviction iteration complete");
302 : }
303 : }
304 : _ = cancel.cancelled() => {
305 : // just drop the joinset to "abort"
306 : }
307 : }
308 :
309 0 : ControlFlow::Continue(())
310 0 : }
311 :
312 : /// Like `eviction_iteration_threshold`, but without any eviction. Eviction will be done by
313 : /// disk usage based eviction task.
314 0 : async fn imitiate_only(
315 0 : self: &Arc<Self>,
316 0 : p: &EvictionPolicyLayerAccessThreshold,
317 0 : cancel: &CancellationToken,
318 0 : gate: &GateGuard,
319 0 : ctx: &RequestContext,
320 0 : ) -> ControlFlow<()> {
321 0 : let acquire_permit = crate::tenant::tasks::concurrent_background_tasks_rate_limit_permit(
322 0 : BackgroundLoopKind::Eviction,
323 0 : ctx,
324 0 : );
325 :
326 0 : let _permit = tokio::select! {
327 0 : permit = acquire_permit => permit,
328 : _ = cancel.cancelled() => return ControlFlow::Break(()),
329 : _ = self.cancel.cancelled() => return ControlFlow::Break(()),
330 : };
331 :
332 0 : self.imitate_layer_accesses(p, cancel, gate, ctx).await
333 0 : }
334 :
335 : /// If we evict layers but keep cached values derived from those layers, then
336 : /// we face a storm of on-demand downloads after pageserver restart.
337 : /// The reason is that the restart empties the caches, and so, the values
338 : /// need to be re-computed by accessing layers, which we evicted while the
339 : /// caches were filled.
340 : ///
341 : /// Solutions here would be one of the following:
342 : /// 1. Have a persistent cache.
343 : /// 2. Count every access to a cached value to the access stats of all layers
344 : /// that were accessed to compute the value in the first place.
345 : /// 3. Invalidate the caches at a period of < p.threshold/2, so that the values
346 : /// get re-computed from layers, thereby counting towards layer access stats.
347 : /// 4. Make the eviction task imitate the layer accesses that typically hit caches.
348 : ///
349 : /// We follow approach (4) here because in Neon prod deployment:
350 : /// - page cache is quite small => high churn => low hit rate
351 : /// => eviction gets correct access stats
352 : /// - value-level caches such as logical size & repatition have a high hit rate,
353 : /// especially for inactive tenants
354 : /// => eviction sees zero accesses for these
355 : /// => they cause the on-demand download storm on pageserver restart
356 : ///
357 : /// We should probably move to persistent caches in the future, or avoid
358 : /// having inactive tenants attached to pageserver in the first place.
359 0 : #[instrument(skip_all)]
360 : async fn imitate_layer_accesses(
361 : &self,
362 : p: &EvictionPolicyLayerAccessThreshold,
363 : cancel: &CancellationToken,
364 : gate: &GateGuard,
365 : ctx: &RequestContext,
366 : ) -> ControlFlow<()> {
367 : if !self.tenant_shard_id.is_zero() {
368 : // Shards !=0 do not maintain accurate relation sizes, and do not need to calculate logical size
369 : // for consumption metrics (consumption metrics are only sent from shard 0). We may therefore
370 : // skip imitating logical size accesses for eviction purposes.
371 : return ControlFlow::Continue(());
372 : }
373 :
374 : let mut state = self.eviction_task_timeline_state.lock().await;
375 :
376 : // Only do the imitate_layer accesses approximately as often as the threshold. A little
377 : // more frequently, to avoid this period racing with the threshold/period-th eviction iteration.
378 : let inter_imitate_period = p.threshold.checked_sub(p.period).unwrap_or(p.threshold);
379 :
380 : match state.last_layer_access_imitation {
381 : Some(ts) if ts.elapsed() < inter_imitate_period => { /* no need to run */ }
382 : _ => {
383 : self.imitate_timeline_cached_layer_accesses(gate, ctx).await;
384 : state.last_layer_access_imitation = Some(tokio::time::Instant::now())
385 : }
386 : }
387 : drop(state);
388 :
389 : if cancel.is_cancelled() {
390 : return ControlFlow::Break(());
391 : }
392 :
393 : // This task is timeline-scoped, but the synthetic size calculation is tenant-scoped.
394 : // Make one of the tenant's timelines draw the short straw and run the calculation.
395 : // The others wait until the calculation is done so that they take into account the
396 : // imitated accesses that the winner made.
397 : let tenant = match crate::tenant::mgr::get_tenant(self.tenant_shard_id, true) {
398 : Ok(t) => t,
399 : Err(_) => {
400 : return ControlFlow::Break(());
401 : }
402 : };
403 : let mut state = tenant.eviction_task_tenant_state.lock().await;
404 : match state.last_layer_access_imitation {
405 : Some(ts) if ts.elapsed() < inter_imitate_period => { /* no need to run */ }
406 : _ => {
407 : self.imitate_synthetic_size_calculation_worker(&tenant, cancel, ctx)
408 : .await;
409 : state.last_layer_access_imitation = Some(tokio::time::Instant::now());
410 : }
411 : }
412 : drop(state);
413 :
414 : if cancel.is_cancelled() {
415 : return ControlFlow::Break(());
416 : }
417 :
418 : ControlFlow::Continue(())
419 : }
420 :
421 : /// Recompute the values which would cause on-demand downloads during restart.
422 0 : #[instrument(skip_all)]
423 : async fn imitate_timeline_cached_layer_accesses(
424 : &self,
425 : guard: &GateGuard,
426 : ctx: &RequestContext,
427 : ) {
428 : let lsn = self.get_last_record_lsn();
429 :
430 : // imitiate on-restart initial logical size
431 : let size = self
432 : .calculate_logical_size(
433 : lsn,
434 : LogicalSizeCalculationCause::EvictionTaskImitation,
435 : guard,
436 : ctx,
437 : )
438 : .instrument(info_span!("calculate_logical_size"))
439 : .await;
440 :
441 : match &size {
442 : Ok(_size) => {
443 : // good, don't log it to avoid confusion
444 : }
445 : Err(_) => {
446 : // we have known issues for which we already log this on consumption metrics,
447 : // gc, and compaction. leave logging out for now.
448 : //
449 : // https://github.com/neondatabase/neon/issues/2539
450 : }
451 : }
452 :
453 : // imitiate repartiting on first compactation
454 : if let Err(e) = self
455 : .collect_keyspace(lsn, ctx)
456 : .instrument(info_span!("collect_keyspace"))
457 : .await
458 : {
459 : // if this failed, we probably failed logical size because these use the same keys
460 : if size.is_err() {
461 : // ignore, see above comment
462 : } else {
463 : match e {
464 : CollectKeySpaceError::Cancelled => {
465 : // Shutting down, ignore
466 : }
467 : err => {
468 0 : warn!(
469 0 : "failed to collect keyspace but succeeded in calculating logical size: {err:#}"
470 0 : );
471 : }
472 : }
473 : }
474 : }
475 : }
476 :
477 : // Imitate the synthetic size calculation done by the consumption_metrics module.
478 0 : #[instrument(skip_all)]
479 : async fn imitate_synthetic_size_calculation_worker(
480 : &self,
481 : tenant: &Arc<Tenant>,
482 : cancel: &CancellationToken,
483 : ctx: &RequestContext,
484 : ) {
485 : if self.conf.metric_collection_endpoint.is_none() {
486 : // We don't start the consumption metrics task if this is not set in the config.
487 : // So, no need to imitate the accesses in that case.
488 : return;
489 : }
490 :
491 : // The consumption metrics are collected on a per-tenant basis, by a single
492 : // global background loop.
493 : // It limits the number of synthetic size calculations using the global
494 : // `concurrent_tenant_size_logical_size_queries` semaphore to not overload
495 : // the pageserver. (size calculation is somewhat expensive in terms of CPU and IOs).
496 : //
497 : // If we used that same semaphore here, then we'd compete for the
498 : // same permits, which may impact timeliness of consumption metrics.
499 : // That is a no-go, as consumption metrics are much more important
500 : // than what we do here.
501 : //
502 : // So, we have a separate semaphore, initialized to the same
503 : // number of permits as the `concurrent_tenant_size_logical_size_queries`.
504 : // In the worst, we would have twice the amount of concurrenct size calculations.
505 : // But in practice, the `p.threshold` >> `consumption metric interval`, and
506 : // we spread out the eviction task using `random_init_delay`.
507 : // So, the chance of the worst case is quite low in practice.
508 : // It runs as a per-tenant task, but the eviction_task.rs is per-timeline.
509 : // So, we must coordinate with other with other eviction tasks of this tenant.
510 : let limit = self
511 : .conf
512 : .eviction_task_immitated_concurrent_logical_size_queries
513 : .inner();
514 :
515 : let mut throwaway_cache = HashMap::new();
516 : let gather = crate::tenant::size::gather_inputs(
517 : tenant,
518 : limit,
519 : None,
520 : &mut throwaway_cache,
521 : LogicalSizeCalculationCause::EvictionTaskImitation,
522 : cancel,
523 : ctx,
524 : )
525 : .instrument(info_span!("gather_inputs"));
526 :
527 0 : tokio::select! {
528 : _ = cancel.cancelled() => {}
529 0 : gather_result = gather => {
530 : match gather_result {
531 : Ok(_) => {},
532 : Err(e) => {
533 : // We don't care about the result, but, if it failed, we should log it,
534 : // since consumption metric might be hitting the cached value and
535 : // thus not encountering this error.
536 0 : warn!("failed to imitate synthetic size calculation accesses: {e:#}")
537 : }
538 : }
539 : }
540 : }
541 : }
542 : }
|
