Line data Source code
1 : //! This module implements the pageserver-global disk-usage-based layer eviction task.
2 : //!
3 : //! # Mechanics
4 : //!
5 : //! Function `launch_disk_usage_global_eviction_task` starts a pageserver-global background
6 : //! loop that evicts layers in response to a shortage of available bytes
7 : //! in the $repo/tenants directory's filesystem.
8 : //!
9 : //! The loop runs periodically at a configurable `period`.
10 : //!
11 : //! Each loop iteration uses `statvfs` to determine filesystem-level space usage.
12 : //! It compares the returned usage data against two different types of thresholds.
13 : //! The iteration tries to evict layers until app-internal accounting says we should be below the thresholds.
14 : //! We cross-check this internal accounting with the real world by making another `statvfs` at the end of the iteration.
15 : //! We're good if that second statvfs shows that we're _actually_ below the configured thresholds.
16 : //! If we're still above one or more thresholds, we emit a warning log message, leaving it to the operator to investigate further.
17 : //!
18 : //! # Eviction Policy
19 : //!
20 : //! There are two thresholds:
21 : //! `max_usage_pct` is the relative available space, expressed in percent of the total filesystem space.
22 : //! If the actual usage is higher, the threshold is exceeded.
23 : //! `min_avail_bytes` is the absolute available space in bytes.
24 : //! If the actual usage is lower, the threshold is exceeded.
25 : //! If either of these thresholds is exceeded, the system is considered to have "disk pressure", and eviction
26 : //! is performed on the next iteration, to release disk space and bring the usage below the thresholds again.
27 : //! The iteration evicts layers in LRU fashion, but, with a weak reservation per tenant.
28 : //! The reservation is to keep the most recently accessed X bytes per tenant resident.
29 : //! If we cannot relieve pressure by evicting layers outside of the reservation, we
30 : //! start evicting layers that are part of the reservation, LRU first.
31 : //!
32 : //! The value for the per-tenant reservation is referred to as `tenant_min_resident_size`
33 : //! throughout the code, but, no actual variable carries that name.
34 : //! The per-tenant default value is the `max(tenant's layer file sizes, regardless of local or remote)`.
35 : //! The idea is to allow at least one layer to be resident per tenant, to ensure it can make forward progress
36 : //! during page reconstruction.
37 : //! An alternative default for all tenants can be specified in the `tenant_config` section of the config.
38 : //! Lastly, each tenant can have an override in their respective tenant config (`min_resident_size_override`).
39 :
40 : // Implementation notes:
41 : // - The `#[allow(dead_code)]` above various structs are to suppress warnings about only the Debug impl
42 : // reading these fields. We use the Debug impl for semi-structured logging, though.
43 :
44 : use std::{
45 : sync::Arc,
46 : time::{Duration, SystemTime},
47 : };
48 :
49 : use anyhow::Context;
50 : use pageserver_api::shard::TenantShardId;
51 : use remote_storage::GenericRemoteStorage;
52 : use serde::{Deserialize, Serialize};
53 : use tokio::time::Instant;
54 : use tokio_util::sync::CancellationToken;
55 : use tracing::{debug, error, info, instrument, warn, Instrument};
56 : use utils::serde_percent::Percent;
57 : use utils::{completion, id::TimelineId};
58 :
59 : use crate::{
60 : config::PageServerConf,
61 : task_mgr::{self, TaskKind, BACKGROUND_RUNTIME},
62 : tenant::{
63 : self,
64 : mgr::TenantManager,
65 : remote_timeline_client::LayerFileMetadata,
66 : secondary::SecondaryTenant,
67 : storage_layer::{AsLayerDesc, EvictionError, Layer, LayerFileName},
68 : Timeline,
69 : },
70 : };
71 :
72 16 : #[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
73 : pub struct DiskUsageEvictionTaskConfig {
74 : pub max_usage_pct: Percent,
75 : pub min_avail_bytes: u64,
76 : #[serde(with = "humantime_serde")]
77 : pub period: Duration,
78 : #[cfg(feature = "testing")]
79 : pub mock_statvfs: Option<crate::statvfs::mock::Behavior>,
80 : /// Select sorting for evicted layers
81 : #[serde(default)]
82 : pub eviction_order: EvictionOrder,
83 : }
84 :
85 : /// Selects the sort order for eviction candidates *after* per tenant `min_resident_size`
86 : /// partitioning.
87 4 : #[derive(Default, Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
88 : #[serde(tag = "type", content = "args")]
89 : pub enum EvictionOrder {
90 : /// Order the layers to be evicted by how recently they have been accessed in absolute
91 : /// time.
92 : ///
93 : /// This strategy is unfair when some tenants grow faster than others towards the slower
94 : /// growing.
95 : #[default]
96 : AbsoluteAccessed,
97 :
98 : /// Order the layers to be evicted by how recently they have been accessed relatively within
99 : /// the set of resident layers of a tenant.
100 : RelativeAccessed {
101 : /// Determines if the tenant with most layers should lose first.
102 : ///
103 : /// Having this enabled is currently the only reasonable option, because the order in which
104 : /// we read tenants is deterministic. If we find the need to use this as `false`, we need
105 : /// to ensure nondeterminism by adding in a random number to break the
106 : /// `relative_last_activity==0.0` ties.
107 : #[serde(default = "default_highest_layer_count_loses_first")]
108 : highest_layer_count_loses_first: bool,
109 : },
110 : }
111 :
112 0 : fn default_highest_layer_count_loses_first() -> bool {
113 0 : true
114 0 : }
115 :
116 : impl EvictionOrder {
117 0 : fn sort(&self, candidates: &mut [(MinResidentSizePartition, EvictionCandidate)]) {
118 0 : use EvictionOrder::*;
119 0 :
120 0 : match self {
121 0 : AbsoluteAccessed => {
122 0 : candidates.sort_unstable_by_key(|(partition, candidate)| {
123 0 : (*partition, candidate.last_activity_ts)
124 0 : });
125 0 : }
126 0 : RelativeAccessed { .. } => candidates.sort_unstable_by_key(|(partition, candidate)| {
127 0 : (*partition, candidate.relative_last_activity)
128 0 : }),
129 : }
130 0 : }
131 :
132 : /// Called to fill in the [`EvictionCandidate::relative_last_activity`] while iterating tenants
133 : /// layers in **most** recently used order.
134 40 : fn relative_last_activity(&self, total: usize, index: usize) -> finite_f32::FiniteF32 {
135 40 : use EvictionOrder::*;
136 40 :
137 40 : match self {
138 0 : AbsoluteAccessed => finite_f32::FiniteF32::ZERO,
139 : RelativeAccessed {
140 40 : highest_layer_count_loses_first,
141 : } => {
142 : // keeping the -1 or not decides if every tenant should lose their least recently accessed
143 : // layer OR if this should happen in the order of having highest layer count:
144 40 : let fudge = if *highest_layer_count_loses_first {
145 : // relative_last_activity vs. tenant layer count:
146 : // - 0.1..=1.0 (10 layers)
147 : // - 0.01..=1.0 (100 layers)
148 : // - 0.001..=1.0 (1000 layers)
149 : //
150 : // leading to evicting less of the smallest tenants.
151 20 : 0
152 : } else {
153 : // use full 0.0..=1.0 range, which means even the smallest tenants could always lose a
154 : // layer. the actual ordering is unspecified: for 10k tenants on a pageserver it could
155 : // be that less than 10k layer evictions is enough, so we would not need to evict from
156 : // all tenants.
157 : //
158 : // as the tenant ordering is now deterministic this could hit the same tenants
159 : // disproportionetly on multiple invocations. alternative could be to remember how many
160 : // layers did we evict last time from this tenant, and inject that as an additional
161 : // fudge here.
162 20 : 1
163 : };
164 :
165 40 : let total = total.checked_sub(fudge).filter(|&x| x > 1).unwrap_or(1);
166 40 : let divider = total as f32;
167 40 :
168 40 : // most recently used is always (total - 0) / divider == 1.0
169 40 : // least recently used depends on the fudge:
170 40 : // - (total - 1) - (total - 1) / total => 0 / total
171 40 : // - total - (total - 1) / total => 1 / total
172 40 : let distance = (total - index) as f32;
173 40 :
174 40 : finite_f32::FiniteF32::try_from_normalized(distance / divider)
175 40 : .unwrap_or_else(|val| {
176 0 : tracing::warn!(%fudge, "calculated invalid relative_last_activity for i={index}, total={total}: {val}");
177 0 : finite_f32::FiniteF32::ZERO
178 40 : })
179 : }
180 : }
181 40 : }
182 : }
183 :
184 0 : #[derive(Default)]
185 : pub struct State {
186 : /// Exclude http requests and background task from running at the same time.
187 : mutex: tokio::sync::Mutex<()>,
188 : }
189 :
190 0 : pub fn launch_disk_usage_global_eviction_task(
191 0 : conf: &'static PageServerConf,
192 0 : storage: GenericRemoteStorage,
193 0 : state: Arc<State>,
194 0 : tenant_manager: Arc<TenantManager>,
195 0 : background_jobs_barrier: completion::Barrier,
196 0 : ) -> anyhow::Result<()> {
197 0 : let Some(task_config) = &conf.disk_usage_based_eviction else {
198 0 : info!("disk usage based eviction task not configured");
199 0 : return Ok(());
200 : };
201 :
202 0 : info!("launching disk usage based eviction task");
203 :
204 0 : task_mgr::spawn(
205 0 : BACKGROUND_RUNTIME.handle(),
206 0 : TaskKind::DiskUsageEviction,
207 0 : None,
208 0 : None,
209 0 : "disk usage based eviction",
210 0 : false,
211 0 : async move {
212 0 : let cancel = task_mgr::shutdown_token();
213 0 :
214 0 : // wait until initial load is complete, because we cannot evict from loading tenants.
215 0 : tokio::select! {
216 0 : _ = cancel.cancelled() => { return Ok(()); },
217 0 : _ = background_jobs_barrier.wait() => { }
218 0 : };
219 :
220 0 : disk_usage_eviction_task(&state, task_config, &storage, tenant_manager, cancel).await;
221 0 : Ok(())
222 0 : },
223 0 : );
224 0 :
225 0 : Ok(())
226 0 : }
227 :
228 0 : #[instrument(skip_all)]
229 : async fn disk_usage_eviction_task(
230 : state: &State,
231 : task_config: &DiskUsageEvictionTaskConfig,
232 : storage: &GenericRemoteStorage,
233 : tenant_manager: Arc<TenantManager>,
234 : cancel: CancellationToken,
235 : ) {
236 0 : scopeguard::defer! {
237 0 : info!("disk usage based eviction task finishing");
238 : };
239 :
240 : use crate::tenant::tasks::random_init_delay;
241 : {
242 : if random_init_delay(task_config.period, &cancel)
243 : .await
244 : .is_err()
245 : {
246 : return;
247 : }
248 : }
249 :
250 : let mut iteration_no = 0;
251 : loop {
252 : iteration_no += 1;
253 : let start = Instant::now();
254 :
255 0 : async {
256 0 : let res = disk_usage_eviction_task_iteration(
257 0 : state,
258 0 : task_config,
259 0 : storage,
260 0 : &tenant_manager,
261 0 : &cancel,
262 0 : )
263 0 : .await;
264 :
265 0 : match res {
266 0 : Ok(()) => {}
267 0 : Err(e) => {
268 0 : // these stat failures are expected to be very rare
269 0 : warn!("iteration failed, unexpected error: {e:#}");
270 : }
271 : }
272 0 : }
273 : .instrument(tracing::info_span!("iteration", iteration_no))
274 : .await;
275 :
276 : let sleep_until = start + task_config.period;
277 : if tokio::time::timeout_at(sleep_until, cancel.cancelled())
278 : .await
279 : .is_ok()
280 : {
281 : break;
282 : }
283 : }
284 : }
285 :
286 : pub trait Usage: Clone + Copy + std::fmt::Debug {
287 : fn has_pressure(&self) -> bool;
288 : fn add_available_bytes(&mut self, bytes: u64);
289 : }
290 :
291 0 : async fn disk_usage_eviction_task_iteration(
292 0 : state: &State,
293 0 : task_config: &DiskUsageEvictionTaskConfig,
294 0 : storage: &GenericRemoteStorage,
295 0 : tenant_manager: &Arc<TenantManager>,
296 0 : cancel: &CancellationToken,
297 0 : ) -> anyhow::Result<()> {
298 0 : let tenants_dir = tenant_manager.get_conf().tenants_path();
299 0 : let usage_pre = filesystem_level_usage::get(&tenants_dir, task_config)
300 0 : .context("get filesystem-level disk usage before evictions")?;
301 0 : let res = disk_usage_eviction_task_iteration_impl(
302 0 : state,
303 0 : storage,
304 0 : usage_pre,
305 0 : tenant_manager,
306 0 : task_config.eviction_order,
307 0 : cancel,
308 0 : )
309 0 : .await;
310 0 : match res {
311 0 : Ok(outcome) => {
312 0 : debug!(?outcome, "disk_usage_eviction_iteration finished");
313 0 : match outcome {
314 0 : IterationOutcome::NoPressure | IterationOutcome::Cancelled => {
315 0 : // nothing to do, select statement below will handle things
316 0 : }
317 0 : IterationOutcome::Finished(outcome) => {
318 : // Verify with statvfs whether we made any real progress
319 0 : let after = filesystem_level_usage::get(&tenants_dir, task_config)
320 0 : // It's quite unlikely to hit the error here. Keep the code simple and bail out.
321 0 : .context("get filesystem-level disk usage after evictions")?;
322 :
323 0 : debug!(?after, "disk usage");
324 :
325 0 : if after.has_pressure() {
326 : // Don't bother doing an out-of-order iteration here now.
327 : // In practice, the task period is set to a value in the tens-of-seconds range,
328 : // which will cause another iteration to happen soon enough.
329 : // TODO: deltas between the three different usages would be helpful,
330 : // consider MiB, GiB, TiB
331 0 : warn!(?outcome, ?after, "disk usage still high");
332 : } else {
333 0 : info!(?outcome, ?after, "disk usage pressure relieved");
334 : }
335 : }
336 : }
337 : }
338 0 : Err(e) => {
339 0 : error!("disk_usage_eviction_iteration failed: {:#}", e);
340 : }
341 : }
342 :
343 0 : Ok(())
344 0 : }
345 :
346 0 : #[derive(Debug, Serialize)]
347 : #[allow(clippy::large_enum_variant)]
348 : pub enum IterationOutcome<U> {
349 : NoPressure,
350 : Cancelled,
351 : Finished(IterationOutcomeFinished<U>),
352 : }
353 :
354 0 : #[derive(Debug, Serialize)]
355 : pub struct IterationOutcomeFinished<U> {
356 : /// The actual usage observed before we started the iteration.
357 : before: U,
358 : /// The expected value for `after`, according to internal accounting, after phase 1.
359 : planned: PlannedUsage<U>,
360 : /// The outcome of phase 2, where we actually do the evictions.
361 : ///
362 : /// If all layers that phase 1 planned to evict _can_ actually get evicted, this will
363 : /// be the same as `planned`.
364 : assumed: AssumedUsage<U>,
365 : }
366 :
367 0 : #[derive(Debug, Serialize)]
368 : struct AssumedUsage<U> {
369 : /// The expected value for `after`, after phase 2.
370 : projected_after: U,
371 : /// The layers we failed to evict during phase 2.
372 : failed: LayerCount,
373 : }
374 :
375 0 : #[derive(Debug, Serialize)]
376 : struct PlannedUsage<U> {
377 : respecting_tenant_min_resident_size: U,
378 : fallback_to_global_lru: Option<U>,
379 : }
380 :
381 0 : #[derive(Debug, Default, Serialize)]
382 : struct LayerCount {
383 : file_sizes: u64,
384 : count: usize,
385 : }
386 :
387 0 : pub(crate) async fn disk_usage_eviction_task_iteration_impl<U: Usage>(
388 0 : state: &State,
389 0 : _storage: &GenericRemoteStorage,
390 0 : usage_pre: U,
391 0 : tenant_manager: &Arc<TenantManager>,
392 0 : eviction_order: EvictionOrder,
393 0 : cancel: &CancellationToken,
394 0 : ) -> anyhow::Result<IterationOutcome<U>> {
395 : // use tokio's mutex to get a Sync guard (instead of std::sync::Mutex)
396 0 : let _g = state
397 0 : .mutex
398 0 : .try_lock()
399 0 : .map_err(|_| anyhow::anyhow!("iteration is already executing"))?;
400 :
401 0 : debug!(?usage_pre, "disk usage");
402 :
403 0 : if !usage_pre.has_pressure() {
404 0 : return Ok(IterationOutcome::NoPressure);
405 0 : }
406 :
407 0 : warn!(
408 0 : ?usage_pre,
409 0 : "running disk usage based eviction due to pressure"
410 0 : );
411 :
412 0 : let candidates =
413 0 : match collect_eviction_candidates(tenant_manager, eviction_order, cancel).await? {
414 : EvictionCandidates::Cancelled => {
415 0 : return Ok(IterationOutcome::Cancelled);
416 : }
417 0 : EvictionCandidates::Finished(partitioned) => partitioned,
418 0 : };
419 0 :
420 0 : // Debug-log the list of candidates
421 0 : let now = SystemTime::now();
422 0 : for (i, (partition, candidate)) in candidates.iter().enumerate() {
423 0 : let nth = i + 1;
424 0 : let total_candidates = candidates.len();
425 0 : let size = candidate.layer.get_file_size();
426 0 : let rel = candidate.relative_last_activity;
427 0 : debug!(
428 0 : "cand {nth}/{total_candidates}: size={size}, rel_last_activity={rel}, no_access_for={}us, partition={partition:?}, {}/{}/{}",
429 0 : now.duration_since(candidate.last_activity_ts)
430 0 : .unwrap()
431 0 : .as_micros(),
432 0 : candidate.layer.get_tenant_shard_id(),
433 0 : candidate.layer.get_timeline_id(),
434 0 : candidate.layer.get_name(),
435 0 : );
436 : }
437 :
438 : // phase1: select victims to relieve pressure
439 : //
440 : // Walk through the list of candidates, until we have accumulated enough layers to get
441 : // us back under the pressure threshold. 'usage_planned' is updated so that it tracks
442 : // how much disk space would be used after evicting all the layers up to the current
443 : // point in the list.
444 : //
445 : // If we get far enough in the list that we start to evict layers that are below
446 : // the tenant's min-resident-size threshold, print a warning, and memorize the disk
447 : // usage at that point, in 'usage_planned_min_resident_size_respecting'.
448 :
449 0 : let selection = select_victims(&candidates, usage_pre);
450 0 :
451 0 : let (evicted_amount, usage_planned) = selection.into_amount_and_planned();
452 0 :
453 0 : // phase2: evict layers
454 0 :
455 0 : let mut js = tokio::task::JoinSet::new();
456 0 : let limit = 1000;
457 0 :
458 0 : let mut evicted = candidates.into_iter().take(evicted_amount).fuse();
459 0 : let mut consumed_all = false;
460 0 :
461 0 : // After the evictions, `usage_assumed` is the post-eviction usage,
462 0 : // according to internal accounting.
463 0 : let mut usage_assumed = usage_pre;
464 0 : let mut evictions_failed = LayerCount::default();
465 0 :
466 0 : let evict_layers = async move {
467 : loop {
468 0 : let next = if js.len() >= limit || consumed_all {
469 0 : js.join_next().await
470 0 : } else if !js.is_empty() {
471 : // opportunistically consume ready result, one per each new evicted
472 0 : futures::future::FutureExt::now_or_never(js.join_next()).and_then(|x| x)
473 : } else {
474 0 : None
475 : };
476 :
477 0 : if let Some(next) = next {
478 0 : match next {
479 0 : Ok(Ok(file_size)) => {
480 0 : usage_assumed.add_available_bytes(file_size);
481 0 : }
482 0 : Ok(Err((file_size, EvictionError::NotFound | EvictionError::Downloaded))) => {
483 0 : evictions_failed.file_sizes += file_size;
484 0 : evictions_failed.count += 1;
485 0 : }
486 0 : Err(je) if je.is_cancelled() => unreachable!("not used"),
487 0 : Err(je) if je.is_panic() => { /* already logged */ }
488 0 : Err(je) => tracing::error!("unknown JoinError: {je:?}"),
489 : }
490 0 : }
491 :
492 0 : if consumed_all && js.is_empty() {
493 0 : break;
494 0 : }
495 :
496 : // calling again when consumed_all is fine as evicted is fused.
497 0 : let Some((_partition, candidate)) = evicted.next() else {
498 0 : consumed_all = true;
499 0 : continue;
500 : };
501 :
502 0 : match candidate.layer {
503 0 : EvictionLayer::Attached(layer) => {
504 0 : let file_size = layer.layer_desc().file_size;
505 0 : js.spawn(async move {
506 0 : layer
507 0 : .evict_and_wait()
508 0 : .await
509 0 : .map(|()| file_size)
510 0 : .map_err(|e| (file_size, e))
511 0 : });
512 0 : }
513 0 : EvictionLayer::Secondary(layer) => {
514 0 : let file_size = layer.metadata.file_size();
515 0 : let tenant_manager = tenant_manager.clone();
516 0 :
517 0 : js.spawn(async move {
518 0 : layer
519 0 : .secondary_tenant
520 0 : .evict_layer(tenant_manager.get_conf(), layer.timeline_id, layer.name)
521 0 : .await;
522 0 : Ok(file_size)
523 0 : });
524 0 : }
525 : }
526 0 : tokio::task::yield_now().await;
527 : }
528 :
529 0 : (usage_assumed, evictions_failed)
530 0 : };
531 :
532 0 : let (usage_assumed, evictions_failed) = tokio::select! {
533 0 : tuple = evict_layers => { tuple },
534 : _ = cancel.cancelled() => {
535 : // dropping joinset will abort all pending evict_and_waits and that is fine, our
536 : // requests will still stand
537 : return Ok(IterationOutcome::Cancelled);
538 : }
539 : };
540 :
541 0 : Ok(IterationOutcome::Finished(IterationOutcomeFinished {
542 0 : before: usage_pre,
543 0 : planned: usage_planned,
544 0 : assumed: AssumedUsage {
545 0 : projected_after: usage_assumed,
546 0 : failed: evictions_failed,
547 0 : },
548 0 : }))
549 0 : }
550 :
551 0 : #[derive(Clone)]
552 : pub(crate) struct EvictionSecondaryLayer {
553 : pub(crate) secondary_tenant: Arc<SecondaryTenant>,
554 : pub(crate) timeline_id: TimelineId,
555 : pub(crate) name: LayerFileName,
556 : pub(crate) metadata: LayerFileMetadata,
557 : }
558 :
559 : /// Full [`Layer`] objects are specific to tenants in attached mode. This type is a layer
560 : /// of indirection to store either a `Layer`, or a reference to a secondary tenant and a layer name.
561 0 : #[derive(Clone)]
562 : pub(crate) enum EvictionLayer {
563 : Attached(Layer),
564 : Secondary(EvictionSecondaryLayer),
565 : }
566 :
567 : impl From<Layer> for EvictionLayer {
568 0 : fn from(value: Layer) -> Self {
569 0 : Self::Attached(value)
570 0 : }
571 : }
572 :
573 : impl EvictionLayer {
574 0 : pub(crate) fn get_tenant_shard_id(&self) -> &TenantShardId {
575 0 : match self {
576 0 : Self::Attached(l) => &l.layer_desc().tenant_shard_id,
577 0 : Self::Secondary(sl) => sl.secondary_tenant.get_tenant_shard_id(),
578 : }
579 0 : }
580 :
581 0 : pub(crate) fn get_timeline_id(&self) -> &TimelineId {
582 0 : match self {
583 0 : Self::Attached(l) => &l.layer_desc().timeline_id,
584 0 : Self::Secondary(sl) => &sl.timeline_id,
585 : }
586 0 : }
587 :
588 0 : pub(crate) fn get_name(&self) -> LayerFileName {
589 0 : match self {
590 0 : Self::Attached(l) => l.layer_desc().filename(),
591 0 : Self::Secondary(sl) => sl.name.clone(),
592 : }
593 0 : }
594 :
595 0 : pub(crate) fn get_file_size(&self) -> u64 {
596 0 : match self {
597 0 : Self::Attached(l) => l.layer_desc().file_size,
598 0 : Self::Secondary(sl) => sl.metadata.file_size(),
599 : }
600 0 : }
601 : }
602 :
603 0 : #[derive(Clone)]
604 : pub(crate) struct EvictionCandidate {
605 : pub(crate) layer: EvictionLayer,
606 : pub(crate) last_activity_ts: SystemTime,
607 : pub(crate) relative_last_activity: finite_f32::FiniteF32,
608 : }
609 :
610 : impl std::fmt::Display for EvictionLayer {
611 0 : fn fmt(&self, f: &mut std::fmt::Formatter) -> std::fmt::Result {
612 0 : match self {
613 0 : Self::Attached(l) => l.fmt(f),
614 0 : Self::Secondary(sl) => {
615 0 : write!(f, "{}/{}", sl.timeline_id, sl.name)
616 : }
617 : }
618 0 : }
619 : }
620 :
621 0 : #[derive(Default)]
622 : pub(crate) struct DiskUsageEvictionInfo {
623 : /// Timeline's largest layer (remote or resident)
624 : pub max_layer_size: Option<u64>,
625 : /// Timeline's resident layers
626 : pub resident_layers: Vec<EvictionCandidate>,
627 : }
628 :
629 : impl std::fmt::Debug for EvictionCandidate {
630 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
631 0 : // format the tv_sec, tv_nsec into rfc3339 in case someone is looking at it
632 0 : // having to allocate a string to this is bad, but it will rarely be formatted
633 0 : let ts = chrono::DateTime::<chrono::Utc>::from(self.last_activity_ts);
634 0 : let ts = ts.to_rfc3339_opts(chrono::SecondsFormat::Nanos, true);
635 0 : struct DisplayIsDebug<'a, T>(&'a T);
636 0 : impl<'a, T: std::fmt::Display> std::fmt::Debug for DisplayIsDebug<'a, T> {
637 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
638 0 : write!(f, "{}", self.0)
639 0 : }
640 0 : }
641 0 : f.debug_struct("LocalLayerInfoForDiskUsageEviction")
642 0 : .field("layer", &DisplayIsDebug(&self.layer))
643 0 : .field("last_activity", &ts)
644 0 : .finish()
645 0 : }
646 : }
647 :
648 0 : #[derive(Debug, Clone, Copy, PartialEq, Eq, PartialOrd, Ord)]
649 : enum MinResidentSizePartition {
650 : Above,
651 : Below,
652 : }
653 :
654 : enum EvictionCandidates {
655 : Cancelled,
656 : Finished(Vec<(MinResidentSizePartition, EvictionCandidate)>),
657 : }
658 :
659 : /// Gather the eviction candidates.
660 : ///
661 : /// The returned `Ok(EvictionCandidates::Finished(candidates))` is sorted in eviction
662 : /// order. A caller that evicts in that order, until pressure is relieved, implements
663 : /// the eviction policy outlined in the module comment.
664 : ///
665 : /// # Example with EvictionOrder::AbsoluteAccessed
666 : ///
667 : /// Imagine that there are two tenants, A and B, with five layers each, a-e.
668 : /// Each layer has size 100, and both tenant's min_resident_size is 150.
669 : /// The eviction order would be
670 : ///
671 : /// ```text
672 : /// partition last_activity_ts tenant/layer
673 : /// Above 18:30 A/c
674 : /// Above 19:00 A/b
675 : /// Above 18:29 B/c
676 : /// Above 19:05 B/b
677 : /// Above 20:00 B/a
678 : /// Above 20:03 A/a
679 : /// Below 20:30 A/d
680 : /// Below 20:40 B/d
681 : /// Below 20:45 B/e
682 : /// Below 20:58 A/e
683 : /// ```
684 : ///
685 : /// Now, if we need to evict 300 bytes to relieve pressure, we'd evict `A/c, A/b, B/c`.
686 : /// They are all in the `Above` partition, so, we respected each tenant's min_resident_size.
687 : ///
688 : /// But, if we need to evict 900 bytes to relieve pressure, we'd evict
689 : /// `A/c, A/b, B/c, B/b, B/a, A/a, A/d, B/d, B/e`, reaching into the `Below` partition
690 : /// after exhauting the `Above` partition.
691 : /// So, we did not respect each tenant's min_resident_size.
692 : ///
693 : /// # Example with EvictionOrder::RelativeAccessed
694 : ///
695 : /// ```text
696 : /// partition relative_age last_activity_ts tenant/layer
697 : /// Above 0/4 18:30 A/c
698 : /// Above 0/4 18:29 B/c
699 : /// Above 1/4 19:00 A/b
700 : /// Above 1/4 19:05 B/b
701 : /// Above 2/4 20:00 B/a
702 : /// Above 2/4 20:03 A/a
703 : /// Below 3/4 20:30 A/d
704 : /// Below 3/4 20:40 B/d
705 : /// Below 4/4 20:45 B/e
706 : /// Below 4/4 20:58 A/e
707 : /// ```
708 : ///
709 : /// With tenants having the same number of layers the picture does not change much. The same with
710 : /// A having many more layers **resident** (not all of them listed):
711 : ///
712 : /// ```text
713 : /// Above 0/100 18:30 A/c
714 : /// Above 0/4 18:29 B/c
715 : /// Above 1/100 19:00 A/b
716 : /// Above 2/100 20:03 A/a
717 : /// Above 3/100 20:03 A/nth_3
718 : /// Above 4/100 20:03 A/nth_4
719 : /// ...
720 : /// Above 1/4 19:05 B/b
721 : /// Above 25/100 20:04 A/nth_25
722 : /// ...
723 : /// Above 2/4 20:00 B/a
724 : /// Above 50/100 20:10 A/nth_50
725 : /// ...
726 : /// Below 3/4 20:40 B/d
727 : /// Below 99/100 20:30 A/nth_99
728 : /// Below 4/4 20:45 B/e
729 : /// Below 100/100 20:58 A/nth_100
730 : /// ```
731 : ///
732 : /// Now it's easier to see that because A has grown fast it has more layers to get evicted. What is
733 : /// difficult to see is what happens on the next round assuming the evicting 23 from the above list
734 : /// relieves the pressure (22 A layers gone, 1 B layers gone) but a new fast growing tenant C has
735 : /// appeared:
736 : ///
737 : /// ```text
738 : /// Above 0/87 20:04 A/nth_23
739 : /// Above 0/3 19:05 B/b
740 : /// Above 0/50 20:59 C/nth_0
741 : /// Above 1/87 20:04 A/nth_24
742 : /// Above 1/50 21:00 C/nth_1
743 : /// Above 2/87 20:04 A/nth_25
744 : /// ...
745 : /// Above 16/50 21:02 C/nth_16
746 : /// Above 1/3 20:00 B/a
747 : /// Above 27/87 20:10 A/nth_50
748 : /// ...
749 : /// Below 2/3 20:40 B/d
750 : /// Below 49/50 21:05 C/nth_49
751 : /// Below 86/87 20:30 A/nth_99
752 : /// Below 3/3 20:45 B/e
753 : /// Below 50/50 21:05 C/nth_50
754 : /// Below 87/87 20:58 A/nth_100
755 : /// ```
756 : ///
757 : /// Now relieving pressure with 23 layers would cost:
758 : /// - tenant A 14 layers
759 : /// - tenant B 1 layer
760 : /// - tenant C 8 layers
761 0 : async fn collect_eviction_candidates(
762 0 : tenant_manager: &Arc<TenantManager>,
763 0 : eviction_order: EvictionOrder,
764 0 : cancel: &CancellationToken,
765 0 : ) -> anyhow::Result<EvictionCandidates> {
766 : // get a snapshot of the list of tenants
767 0 : let tenants = tenant::mgr::list_tenants()
768 0 : .await
769 0 : .context("get list of tenants")?;
770 :
771 : // TODO: avoid listing every layer in every tenant: this loop can block the executor,
772 : // and the resulting data structure can be huge.
773 : // (https://github.com/neondatabase/neon/issues/6224)
774 0 : let mut candidates = Vec::new();
775 :
776 0 : for (tenant_id, _state, _gen) in tenants {
777 0 : if cancel.is_cancelled() {
778 0 : return Ok(EvictionCandidates::Cancelled);
779 0 : }
780 0 : let tenant = match tenant::mgr::get_tenant(tenant_id, true) {
781 0 : Ok(tenant) => tenant,
782 0 : Err(e) => {
783 : // this can happen if tenant has lifecycle transition after we fetched it
784 0 : debug!("failed to get tenant: {e:#}");
785 0 : continue;
786 : }
787 : };
788 :
789 0 : if tenant.cancel.is_cancelled() {
790 0 : info!(%tenant_id, "Skipping tenant for eviction, it is shutting down");
791 0 : continue;
792 0 : }
793 0 :
794 0 : // collect layers from all timelines in this tenant
795 0 : //
796 0 : // If one of the timelines becomes `!is_active()` during the iteration,
797 0 : // for example because we're shutting down, then `max_layer_size` can be too small.
798 0 : // That's OK. This code only runs under a disk pressure situation, and being
799 0 : // a little unfair to tenants during shutdown in such a situation is tolerable.
800 0 : let mut tenant_candidates = Vec::new();
801 0 : let mut max_layer_size = 0;
802 0 : for tl in tenant.list_timelines() {
803 0 : if !tl.is_active() {
804 0 : continue;
805 0 : }
806 0 : let info = tl.get_local_layers_for_disk_usage_eviction().await;
807 0 : debug!(tenant_id=%tl.tenant_shard_id.tenant_id, shard_id=%tl.tenant_shard_id.shard_slug(), timeline_id=%tl.timeline_id, "timeline resident layers count: {}", info.resident_layers.len());
808 0 : tenant_candidates.extend(info.resident_layers.into_iter());
809 0 : max_layer_size = max_layer_size.max(info.max_layer_size.unwrap_or(0));
810 0 :
811 0 : if cancel.is_cancelled() {
812 0 : return Ok(EvictionCandidates::Cancelled);
813 0 : }
814 : }
815 :
816 : // `min_resident_size` defaults to maximum layer file size of the tenant.
817 : // This ensures that each tenant can have at least one layer resident at a given time,
818 : // ensuring forward progress for a single Timeline::get in that tenant.
819 : // It's a questionable heuristic since, usually, there are many Timeline::get
820 : // requests going on for a tenant, and, at least in Neon prod, the median
821 : // layer file size is much smaller than the compaction target size.
822 : // We could be better here, e.g., sum of all L0 layers + most recent L1 layer.
823 : // That's what's typically used by the various background loops.
824 : //
825 : // The default can be overridden with a fixed value in the tenant conf.
826 : // A default override can be put in the default tenant conf in the pageserver.toml.
827 0 : let min_resident_size = if let Some(s) = tenant.get_min_resident_size_override() {
828 0 : debug!(
829 0 : tenant_id=%tenant.tenant_shard_id().tenant_id,
830 0 : shard_id=%tenant.tenant_shard_id().shard_slug(),
831 0 : overridden_size=s,
832 0 : "using overridden min resident size for tenant"
833 0 : );
834 0 : s
835 : } else {
836 0 : debug!(
837 0 : tenant_id=%tenant.tenant_shard_id().tenant_id,
838 0 : shard_id=%tenant.tenant_shard_id().shard_slug(),
839 0 : max_layer_size,
840 0 : "using max layer size as min_resident_size for tenant",
841 0 : );
842 0 : max_layer_size
843 : };
844 :
845 : // Sort layers most-recently-used first, then partition by
846 : // cumsum above/below min_resident_size.
847 0 : tenant_candidates
848 0 : .sort_unstable_by_key(|layer_info| std::cmp::Reverse(layer_info.last_activity_ts));
849 0 : let mut cumsum: i128 = 0;
850 0 :
851 0 : let total = tenant_candidates.len();
852 0 :
853 0 : let tenant_candidates =
854 0 : tenant_candidates
855 0 : .into_iter()
856 0 : .enumerate()
857 0 : .map(|(i, mut candidate)| {
858 0 : // as we iterate this reverse sorted list, the most recently accessed layer will always
859 0 : // be 1.0; this is for us to evict it last.
860 0 : candidate.relative_last_activity =
861 0 : eviction_order.relative_last_activity(total, i);
862 :
863 0 : let partition = if cumsum > min_resident_size as i128 {
864 0 : MinResidentSizePartition::Above
865 : } else {
866 0 : MinResidentSizePartition::Below
867 : };
868 0 : cumsum += i128::from(candidate.layer.get_file_size());
869 0 :
870 0 : (partition, candidate)
871 0 : });
872 0 :
873 0 : candidates.extend(tenant_candidates);
874 : }
875 :
876 : // Note: the same tenant ID might be hit twice, if it transitions from attached to
877 : // secondary while we run. That is okay: when we eventually try and run the eviction,
878 : // the `Gate` on the object will ensure that whichever one has already been shut down
879 : // will not delete anything.
880 :
881 0 : let mut secondary_tenants = Vec::new();
882 0 : tenant_manager.foreach_secondary_tenants(
883 0 : |_tenant_shard_id: &TenantShardId, state: &Arc<SecondaryTenant>| {
884 0 : secondary_tenants.push(state.clone());
885 0 : },
886 0 : );
887 :
888 0 : for secondary_tenant in secondary_tenants {
889 : // for secondary tenants we use a sum of on_disk layers and already evicted layers. this is
890 : // to prevent repeated disk usage based evictions from completely draining less often
891 : // updating secondaries.
892 0 : let (mut layer_info, total_layers) = secondary_tenant.get_layers_for_eviction();
893 :
894 : debug_assert!(
895 0 : total_layers >= layer_info.resident_layers.len(),
896 0 : "total_layers ({total_layers}) must be at least the resident_layers.len() ({})",
897 0 : layer_info.resident_layers.len()
898 : );
899 :
900 0 : layer_info
901 0 : .resident_layers
902 0 : .sort_unstable_by_key(|layer_info| std::cmp::Reverse(layer_info.last_activity_ts));
903 0 :
904 0 : let tenant_candidates =
905 0 : layer_info
906 0 : .resident_layers
907 0 : .into_iter()
908 0 : .enumerate()
909 0 : .map(|(i, mut candidate)| {
910 0 : candidate.relative_last_activity =
911 0 : eviction_order.relative_last_activity(total_layers, i);
912 0 : (
913 0 : // Secondary locations' layers are always considered above the min resident size,
914 0 : // i.e. secondary locations are permitted to be trimmed to zero layers if all
915 0 : // the layers have sufficiently old access times.
916 0 : MinResidentSizePartition::Above,
917 0 : candidate,
918 0 : )
919 0 : });
920 0 :
921 0 : candidates.extend(tenant_candidates);
922 0 :
923 0 : tokio::task::yield_now().await;
924 : }
925 :
926 0 : debug_assert!(MinResidentSizePartition::Above < MinResidentSizePartition::Below,
927 0 : "as explained in the function's doc comment, layers that aren't in the tenant's min_resident_size are evicted first");
928 :
929 0 : eviction_order.sort(&mut candidates);
930 0 :
931 0 : Ok(EvictionCandidates::Finished(candidates))
932 0 : }
933 :
934 : /// Given a pre-sorted vec of all layers in the system, select the first N which are enough to
935 : /// relieve pressure.
936 : ///
937 : /// Returns the amount of candidates selected, with the planned usage.
938 0 : fn select_victims<U: Usage>(
939 0 : candidates: &[(MinResidentSizePartition, EvictionCandidate)],
940 0 : usage_pre: U,
941 0 : ) -> VictimSelection<U> {
942 0 : let mut usage_when_switched = None;
943 0 : let mut usage_planned = usage_pre;
944 0 : let mut evicted_amount = 0;
945 :
946 0 : for (i, (partition, candidate)) in candidates.iter().enumerate() {
947 0 : if !usage_planned.has_pressure() {
948 0 : break;
949 0 : }
950 0 :
951 0 : if partition == &MinResidentSizePartition::Below && usage_when_switched.is_none() {
952 0 : usage_when_switched = Some((usage_planned, i));
953 0 : }
954 :
955 0 : usage_planned.add_available_bytes(candidate.layer.get_file_size());
956 0 : evicted_amount += 1;
957 : }
958 :
959 0 : VictimSelection {
960 0 : amount: evicted_amount,
961 0 : usage_pre,
962 0 : usage_when_switched,
963 0 : usage_planned,
964 0 : }
965 0 : }
966 :
967 : struct VictimSelection<U> {
968 : amount: usize,
969 : usage_pre: U,
970 : usage_when_switched: Option<(U, usize)>,
971 : usage_planned: U,
972 : }
973 :
974 : impl<U: Usage> VictimSelection<U> {
975 0 : fn into_amount_and_planned(self) -> (usize, PlannedUsage<U>) {
976 0 : debug!(
977 0 : evicted_amount=%self.amount,
978 0 : "took enough candidates for pressure to be relieved"
979 0 : );
980 :
981 0 : if let Some((usage_planned, candidate_no)) = self.usage_when_switched.as_ref() {
982 0 : warn!(usage_pre=?self.usage_pre, ?usage_planned, candidate_no, "tenant_min_resident_size-respecting LRU would not relieve pressure, evicting more following global LRU policy");
983 0 : }
984 :
985 0 : let planned = match self.usage_when_switched {
986 0 : Some((respecting_tenant_min_resident_size, _)) => PlannedUsage {
987 0 : respecting_tenant_min_resident_size,
988 0 : fallback_to_global_lru: Some(self.usage_planned),
989 0 : },
990 0 : None => PlannedUsage {
991 0 : respecting_tenant_min_resident_size: self.usage_planned,
992 0 : fallback_to_global_lru: None,
993 0 : },
994 : };
995 :
996 0 : (self.amount, planned)
997 0 : }
998 : }
999 :
1000 : struct TimelineKey(Arc<Timeline>);
1001 :
1002 : impl PartialEq for TimelineKey {
1003 0 : fn eq(&self, other: &Self) -> bool {
1004 0 : Arc::ptr_eq(&self.0, &other.0)
1005 0 : }
1006 : }
1007 :
1008 : impl Eq for TimelineKey {}
1009 :
1010 : impl std::hash::Hash for TimelineKey {
1011 0 : fn hash<H: std::hash::Hasher>(&self, state: &mut H) {
1012 0 : Arc::as_ptr(&self.0).hash(state);
1013 0 : }
1014 : }
1015 :
1016 : impl std::ops::Deref for TimelineKey {
1017 : type Target = Timeline;
1018 :
1019 0 : fn deref(&self) -> &Self::Target {
1020 0 : self.0.as_ref()
1021 0 : }
1022 : }
1023 :
1024 : /// A totally ordered f32 subset we can use with sorting functions.
1025 : pub(crate) mod finite_f32 {
1026 :
1027 : /// A totally ordered f32 subset we can use with sorting functions.
1028 0 : #[derive(Clone, Copy, PartialEq)]
1029 : pub struct FiniteF32(f32);
1030 :
1031 : impl std::fmt::Debug for FiniteF32 {
1032 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
1033 0 : std::fmt::Debug::fmt(&self.0, f)
1034 0 : }
1035 : }
1036 :
1037 : impl std::fmt::Display for FiniteF32 {
1038 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
1039 0 : std::fmt::Display::fmt(&self.0, f)
1040 0 : }
1041 : }
1042 :
1043 : impl std::cmp::Eq for FiniteF32 {}
1044 :
1045 : impl std::cmp::PartialOrd for FiniteF32 {
1046 0 : fn partial_cmp(&self, other: &Self) -> Option<std::cmp::Ordering> {
1047 0 : Some(self.cmp(other))
1048 0 : }
1049 : }
1050 :
1051 : impl std::cmp::Ord for FiniteF32 {
1052 0 : fn cmp(&self, other: &Self) -> std::cmp::Ordering {
1053 0 : self.0.total_cmp(&other.0)
1054 0 : }
1055 : }
1056 :
1057 : impl TryFrom<f32> for FiniteF32 {
1058 : type Error = f32;
1059 :
1060 0 : fn try_from(value: f32) -> Result<Self, Self::Error> {
1061 0 : if value.is_finite() {
1062 0 : Ok(FiniteF32(value))
1063 : } else {
1064 0 : Err(value)
1065 : }
1066 0 : }
1067 : }
1068 :
1069 : impl From<FiniteF32> for f32 {
1070 40 : fn from(value: FiniteF32) -> f32 {
1071 40 : value.0
1072 40 : }
1073 : }
1074 :
1075 : impl FiniteF32 {
1076 : pub const ZERO: FiniteF32 = FiniteF32(0.0);
1077 :
1078 40 : pub fn try_from_normalized(value: f32) -> Result<Self, f32> {
1079 40 : if (0.0..=1.0).contains(&value) {
1080 : // -0.0 is within the range, make sure it is assumed 0.0..=1.0
1081 40 : let value = value.abs();
1082 40 : Ok(FiniteF32(value))
1083 : } else {
1084 0 : Err(value)
1085 : }
1086 40 : }
1087 :
1088 40 : pub fn into_inner(self) -> f32 {
1089 40 : self.into()
1090 40 : }
1091 : }
1092 : }
1093 :
1094 : mod filesystem_level_usage {
1095 : use anyhow::Context;
1096 : use camino::Utf8Path;
1097 :
1098 : use crate::statvfs::Statvfs;
1099 :
1100 : use super::DiskUsageEvictionTaskConfig;
1101 :
1102 0 : #[derive(Debug, Clone, Copy)]
1103 : pub struct Usage<'a> {
1104 : config: &'a DiskUsageEvictionTaskConfig,
1105 :
1106 : /// Filesystem capacity
1107 : total_bytes: u64,
1108 : /// Free filesystem space
1109 : avail_bytes: u64,
1110 : }
1111 :
1112 : impl super::Usage for Usage<'_> {
1113 14 : fn has_pressure(&self) -> bool {
1114 14 : let usage_pct =
1115 14 : (100.0 * (1.0 - ((self.avail_bytes as f64) / (self.total_bytes as f64)))) as u64;
1116 14 :
1117 14 : let pressures = [
1118 14 : (
1119 14 : "min_avail_bytes",
1120 14 : self.avail_bytes < self.config.min_avail_bytes,
1121 14 : ),
1122 14 : (
1123 14 : "max_usage_pct",
1124 14 : usage_pct >= self.config.max_usage_pct.get() as u64,
1125 14 : ),
1126 14 : ];
1127 14 :
1128 28 : pressures.into_iter().any(|(_, has_pressure)| has_pressure)
1129 14 : }
1130 :
1131 12 : fn add_available_bytes(&mut self, bytes: u64) {
1132 12 : self.avail_bytes += bytes;
1133 12 : }
1134 : }
1135 :
1136 0 : pub fn get<'a>(
1137 0 : tenants_dir: &Utf8Path,
1138 0 : config: &'a DiskUsageEvictionTaskConfig,
1139 0 : ) -> anyhow::Result<Usage<'a>> {
1140 0 : let mock_config = {
1141 0 : #[cfg(feature = "testing")]
1142 0 : {
1143 0 : config.mock_statvfs.as_ref()
1144 : }
1145 : #[cfg(not(feature = "testing"))]
1146 : {
1147 : None
1148 : }
1149 : };
1150 :
1151 0 : let stat = Statvfs::get(tenants_dir, mock_config)
1152 0 : .context("statvfs failed, presumably directory got unlinked")?;
1153 :
1154 : // https://unix.stackexchange.com/a/703650
1155 0 : let blocksize = if stat.fragment_size() > 0 {
1156 0 : stat.fragment_size()
1157 : } else {
1158 0 : stat.block_size()
1159 : };
1160 :
1161 : // use blocks_available (b_avail) since, pageserver runs as unprivileged user
1162 0 : let avail_bytes = stat.blocks_available() * blocksize;
1163 0 : let total_bytes = stat.blocks() * blocksize;
1164 0 :
1165 0 : Ok(Usage {
1166 0 : config,
1167 0 : total_bytes,
1168 0 : avail_bytes,
1169 0 : })
1170 0 : }
1171 :
1172 2 : #[test]
1173 2 : fn max_usage_pct_pressure() {
1174 2 : use super::EvictionOrder;
1175 2 : use super::Usage as _;
1176 2 : use std::time::Duration;
1177 2 : use utils::serde_percent::Percent;
1178 2 :
1179 2 : let mut usage = Usage {
1180 2 : config: &DiskUsageEvictionTaskConfig {
1181 2 : max_usage_pct: Percent::new(85).unwrap(),
1182 2 : min_avail_bytes: 0,
1183 2 : period: Duration::MAX,
1184 2 : #[cfg(feature = "testing")]
1185 2 : mock_statvfs: None,
1186 2 : eviction_order: EvictionOrder::default(),
1187 2 : },
1188 2 : total_bytes: 100_000,
1189 2 : avail_bytes: 0,
1190 2 : };
1191 2 :
1192 2 : assert!(usage.has_pressure(), "expected pressure at 100%");
1193 :
1194 2 : usage.add_available_bytes(14_000);
1195 2 : assert!(usage.has_pressure(), "expected pressure at 86%");
1196 :
1197 2 : usage.add_available_bytes(999);
1198 2 : assert!(usage.has_pressure(), "expected pressure at 85.001%");
1199 :
1200 2 : usage.add_available_bytes(1);
1201 2 : assert!(usage.has_pressure(), "expected pressure at precisely 85%");
1202 :
1203 2 : usage.add_available_bytes(1);
1204 2 : assert!(!usage.has_pressure(), "no pressure at 84.999%");
1205 :
1206 2 : usage.add_available_bytes(999);
1207 2 : assert!(!usage.has_pressure(), "no pressure at 84%");
1208 :
1209 2 : usage.add_available_bytes(16_000);
1210 2 : assert!(!usage.has_pressure());
1211 2 : }
1212 : }
1213 :
1214 : #[cfg(test)]
1215 : mod tests {
1216 : use super::*;
1217 :
1218 2 : #[test]
1219 2 : fn relative_equal_bounds() {
1220 2 : let order = EvictionOrder::RelativeAccessed {
1221 2 : highest_layer_count_loses_first: false,
1222 2 : };
1223 2 :
1224 2 : let len = 10;
1225 2 : let v = (0..len)
1226 20 : .map(|i| order.relative_last_activity(len, i).into_inner())
1227 2 : .collect::<Vec<_>>();
1228 2 :
1229 2 : assert_eq!(v.first(), Some(&1.0));
1230 2 : assert_eq!(v.last(), Some(&0.0));
1231 18 : assert!(v.windows(2).all(|slice| slice[0] > slice[1]));
1232 2 : }
1233 :
1234 2 : #[test]
1235 2 : fn relative_spare_bounds() {
1236 2 : let order = EvictionOrder::RelativeAccessed {
1237 2 : highest_layer_count_loses_first: true,
1238 2 : };
1239 2 :
1240 2 : let len = 10;
1241 2 : let v = (0..len)
1242 20 : .map(|i| order.relative_last_activity(len, i).into_inner())
1243 2 : .collect::<Vec<_>>();
1244 2 :
1245 2 : assert_eq!(v.first(), Some(&1.0));
1246 2 : assert_eq!(v.last(), Some(&0.1));
1247 18 : assert!(v.windows(2).all(|slice| slice[0] > slice[1]));
1248 2 : }
1249 : }
|
