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::sync::Arc;
45 : use std::time::SystemTime;
46 :
47 : use anyhow::Context;
48 : use pageserver_api::config::DiskUsageEvictionTaskConfig;
49 : use pageserver_api::shard::TenantShardId;
50 : use remote_storage::GenericRemoteStorage;
51 : use serde::Serialize;
52 : use tokio::time::Instant;
53 : use tokio_util::sync::CancellationToken;
54 : use tracing::{Instrument, debug, error, info, instrument, warn};
55 : use utils::completion;
56 : use utils::id::TimelineId;
57 :
58 : use crate::config::PageServerConf;
59 : use crate::metrics::disk_usage_based_eviction::METRICS;
60 : use crate::task_mgr::{self, BACKGROUND_RUNTIME};
61 : use crate::tenant::mgr::TenantManager;
62 : use crate::tenant::remote_timeline_client::LayerFileMetadata;
63 : use crate::tenant::secondary::SecondaryTenant;
64 : use crate::tenant::storage_layer::{
65 : AsLayerDesc, EvictionError, Layer, LayerName, LayerVisibilityHint,
66 : };
67 : use crate::tenant::tasks::sleep_random;
68 : use crate::{CancellableTask, DiskUsageEvictionTask};
69 :
70 : /// Selects the sort order for eviction candidates *after* per tenant `min_resident_size`
71 : /// partitioning.
72 : #[derive(Debug, Clone, Copy, PartialEq, Eq)]
73 : pub enum EvictionOrder {
74 : /// Order the layers to be evicted by how recently they have been accessed relatively within
75 : /// the set of resident layers of a tenant.
76 : RelativeAccessed {
77 : /// Determines if the tenant with most layers should lose first.
78 : ///
79 : /// Having this enabled is currently the only reasonable option, because the order in which
80 : /// we read tenants is deterministic. If we find the need to use this as `false`, we need
81 : /// to ensure nondeterminism by adding in a random number to break the
82 : /// `relative_last_activity==0.0` ties.
83 : highest_layer_count_loses_first: bool,
84 : },
85 : }
86 :
87 : impl From<pageserver_api::config::EvictionOrder> for EvictionOrder {
88 0 : fn from(value: pageserver_api::config::EvictionOrder) -> Self {
89 0 : match value {
90 0 : pageserver_api::config::EvictionOrder::RelativeAccessed {
91 0 : highest_layer_count_loses_first,
92 0 : } => Self::RelativeAccessed {
93 0 : highest_layer_count_loses_first,
94 0 : },
95 0 : }
96 0 : }
97 : }
98 :
99 : impl EvictionOrder {
100 0 : fn sort(&self, candidates: &mut [(EvictionPartition, EvictionCandidate)]) {
101 : use EvictionOrder::*;
102 :
103 0 : match self {
104 0 : RelativeAccessed { .. } => candidates.sort_unstable_by_key(|(partition, candidate)| {
105 0 : (*partition, candidate.relative_last_activity)
106 0 : }),
107 0 : }
108 0 : }
109 :
110 : /// Called to fill in the [`EvictionCandidate::relative_last_activity`] while iterating tenants
111 : /// layers in **most** recently used order.
112 80 : fn relative_last_activity(&self, total: usize, index: usize) -> finite_f32::FiniteF32 {
113 : use EvictionOrder::*;
114 :
115 80 : match self {
116 80 : RelativeAccessed {
117 80 : highest_layer_count_loses_first,
118 : } => {
119 : // keeping the -1 or not decides if every tenant should lose their least recently accessed
120 : // layer OR if this should happen in the order of having highest layer count:
121 80 : let fudge = if *highest_layer_count_loses_first {
122 : // relative_last_activity vs. tenant layer count:
123 : // - 0.1..=1.0 (10 layers)
124 : // - 0.01..=1.0 (100 layers)
125 : // - 0.001..=1.0 (1000 layers)
126 : //
127 : // leading to evicting less of the smallest tenants.
128 40 : 0
129 : } else {
130 : // use full 0.0..=1.0 range, which means even the smallest tenants could always lose a
131 : // layer. the actual ordering is unspecified: for 10k tenants on a pageserver it could
132 : // be that less than 10k layer evictions is enough, so we would not need to evict from
133 : // all tenants.
134 : //
135 : // as the tenant ordering is now deterministic this could hit the same tenants
136 : // disproportionetly on multiple invocations. alternative could be to remember how many
137 : // layers did we evict last time from this tenant, and inject that as an additional
138 : // fudge here.
139 40 : 1
140 : };
141 :
142 80 : let total = total.checked_sub(fudge).filter(|&x| x > 1).unwrap_or(1);
143 80 : let divider = total as f32;
144 80 :
145 80 : // most recently used is always (total - 0) / divider == 1.0
146 80 : // least recently used depends on the fudge:
147 80 : // - (total - 1) - (total - 1) / total => 0 / total
148 80 : // - total - (total - 1) / total => 1 / total
149 80 : let distance = (total - index) as f32;
150 80 :
151 80 : finite_f32::FiniteF32::try_from_normalized(distance / divider)
152 80 : .unwrap_or_else(|val| {
153 0 : tracing::warn!(%fudge, "calculated invalid relative_last_activity for i={index}, total={total}: {val}");
154 0 : finite_f32::FiniteF32::ZERO
155 80 : })
156 80 : }
157 80 : }
158 80 : }
159 : }
160 :
161 : #[derive(Default)]
162 : pub struct State {
163 : /// Exclude http requests and background task from running at the same time.
164 : mutex: tokio::sync::Mutex<()>,
165 : }
166 :
167 0 : pub fn launch_disk_usage_global_eviction_task(
168 0 : conf: &'static PageServerConf,
169 0 : storage: GenericRemoteStorage,
170 0 : state: Arc<State>,
171 0 : tenant_manager: Arc<TenantManager>,
172 0 : background_jobs_barrier: completion::Barrier,
173 0 : ) -> Option<DiskUsageEvictionTask> {
174 0 : let Some(task_config) = &conf.disk_usage_based_eviction else {
175 0 : info!("disk usage based eviction task not configured");
176 0 : return None;
177 : };
178 :
179 0 : info!("launching disk usage based eviction task");
180 :
181 0 : let cancel = CancellationToken::new();
182 0 : let task = BACKGROUND_RUNTIME.spawn(task_mgr::exit_on_panic_or_error(
183 0 : "disk usage based eviction",
184 0 : {
185 0 : let cancel = cancel.clone();
186 0 : async move {
187 0 : // wait until initial load is complete, because we cannot evict from loading tenants.
188 0 : tokio::select! {
189 0 : _ = cancel.cancelled() => { return anyhow::Ok(()); },
190 0 : _ = background_jobs_barrier.wait() => { }
191 0 : };
192 0 :
193 0 : disk_usage_eviction_task(&state, task_config, &storage, tenant_manager, cancel)
194 0 : .await;
195 0 : anyhow::Ok(())
196 0 : }
197 0 : },
198 0 : ));
199 0 :
200 0 : Some(DiskUsageEvictionTask(CancellableTask { cancel, task }))
201 0 : }
202 :
203 : #[instrument(skip_all)]
204 : async fn disk_usage_eviction_task(
205 : state: &State,
206 : task_config: &DiskUsageEvictionTaskConfig,
207 : storage: &GenericRemoteStorage,
208 : tenant_manager: Arc<TenantManager>,
209 : cancel: CancellationToken,
210 : ) {
211 : scopeguard::defer! {
212 : info!("disk usage based eviction task finishing");
213 : };
214 :
215 : if sleep_random(task_config.period, &cancel).await.is_err() {
216 : return;
217 : }
218 :
219 : let mut iteration_no = 0;
220 : loop {
221 : iteration_no += 1;
222 : let start = Instant::now();
223 :
224 0 : async {
225 0 : let res = disk_usage_eviction_task_iteration(
226 0 : state,
227 0 : task_config,
228 0 : storage,
229 0 : &tenant_manager,
230 0 : &cancel,
231 0 : )
232 0 : .await;
233 :
234 0 : match res {
235 0 : Ok(()) => {}
236 0 : Err(e) => {
237 0 : // these stat failures are expected to be very rare
238 0 : warn!("iteration failed, unexpected error: {e:#}");
239 : }
240 : }
241 0 : }
242 : .instrument(tracing::info_span!("iteration", iteration_no))
243 : .await;
244 :
245 : let sleep_until = start + task_config.period;
246 : if tokio::time::timeout_at(sleep_until, cancel.cancelled())
247 : .await
248 : .is_ok()
249 : {
250 : break;
251 : }
252 : }
253 : }
254 :
255 : pub trait Usage: Clone + Copy + std::fmt::Debug {
256 : fn has_pressure(&self) -> bool;
257 : fn add_available_bytes(&mut self, bytes: u64);
258 : }
259 :
260 0 : async fn disk_usage_eviction_task_iteration(
261 0 : state: &State,
262 0 : task_config: &DiskUsageEvictionTaskConfig,
263 0 : storage: &GenericRemoteStorage,
264 0 : tenant_manager: &Arc<TenantManager>,
265 0 : cancel: &CancellationToken,
266 0 : ) -> anyhow::Result<()> {
267 0 : let tenants_dir = tenant_manager.get_conf().tenants_path();
268 0 : let usage_pre = filesystem_level_usage::get(&tenants_dir, task_config)
269 0 : .context("get filesystem-level disk usage before evictions")?;
270 0 : let res = disk_usage_eviction_task_iteration_impl(
271 0 : state,
272 0 : storage,
273 0 : usage_pre,
274 0 : tenant_manager,
275 0 : task_config.eviction_order.into(),
276 0 : cancel,
277 0 : )
278 0 : .await;
279 0 : match res {
280 0 : Ok(outcome) => {
281 0 : debug!(?outcome, "disk_usage_eviction_iteration finished");
282 0 : match outcome {
283 0 : IterationOutcome::NoPressure | IterationOutcome::Cancelled => {
284 0 : // nothing to do, select statement below will handle things
285 0 : }
286 0 : IterationOutcome::Finished(outcome) => {
287 : // Verify with statvfs whether we made any real progress
288 0 : let after = filesystem_level_usage::get(&tenants_dir, task_config)
289 0 : // It's quite unlikely to hit the error here. Keep the code simple and bail out.
290 0 : .context("get filesystem-level disk usage after evictions")?;
291 :
292 0 : debug!(?after, "disk usage");
293 :
294 0 : if after.has_pressure() {
295 : // Don't bother doing an out-of-order iteration here now.
296 : // In practice, the task period is set to a value in the tens-of-seconds range,
297 : // which will cause another iteration to happen soon enough.
298 : // TODO: deltas between the three different usages would be helpful,
299 : // consider MiB, GiB, TiB
300 0 : warn!(?outcome, ?after, "disk usage still high");
301 : } else {
302 0 : info!(?outcome, ?after, "disk usage pressure relieved");
303 : }
304 : }
305 : }
306 : }
307 0 : Err(e) => {
308 0 : error!("disk_usage_eviction_iteration failed: {:#}", e);
309 : }
310 : }
311 :
312 0 : Ok(())
313 0 : }
314 :
315 : #[derive(Debug, Serialize)]
316 : #[allow(clippy::large_enum_variant)]
317 : pub enum IterationOutcome<U> {
318 : NoPressure,
319 : Cancelled,
320 : Finished(IterationOutcomeFinished<U>),
321 : }
322 :
323 : #[derive(Debug, Serialize)]
324 : pub struct IterationOutcomeFinished<U> {
325 : /// The actual usage observed before we started the iteration.
326 : before: U,
327 : /// The expected value for `after`, according to internal accounting, after phase 1.
328 : planned: PlannedUsage<U>,
329 : /// The outcome of phase 2, where we actually do the evictions.
330 : ///
331 : /// If all layers that phase 1 planned to evict _can_ actually get evicted, this will
332 : /// be the same as `planned`.
333 : assumed: AssumedUsage<U>,
334 : }
335 :
336 : #[derive(Debug, Serialize)]
337 : struct AssumedUsage<U> {
338 : /// The expected value for `after`, after phase 2.
339 : projected_after: U,
340 : /// The layers we failed to evict during phase 2.
341 : failed: LayerCount,
342 : }
343 :
344 : #[derive(Debug, Serialize)]
345 : struct PlannedUsage<U> {
346 : respecting_tenant_min_resident_size: U,
347 : fallback_to_global_lru: Option<U>,
348 : }
349 :
350 : #[derive(Debug, Default, Serialize)]
351 : struct LayerCount {
352 : file_sizes: u64,
353 : count: usize,
354 : }
355 :
356 0 : pub(crate) async fn disk_usage_eviction_task_iteration_impl<U: Usage>(
357 0 : state: &State,
358 0 : _storage: &GenericRemoteStorage,
359 0 : usage_pre: U,
360 0 : tenant_manager: &Arc<TenantManager>,
361 0 : eviction_order: EvictionOrder,
362 0 : cancel: &CancellationToken,
363 0 : ) -> anyhow::Result<IterationOutcome<U>> {
364 : // use tokio's mutex to get a Sync guard (instead of std::sync::Mutex)
365 0 : let _g = state
366 0 : .mutex
367 0 : .try_lock()
368 0 : .map_err(|_| anyhow::anyhow!("iteration is already executing"))?;
369 :
370 0 : debug!(?usage_pre, "disk usage");
371 :
372 0 : if !usage_pre.has_pressure() {
373 0 : return Ok(IterationOutcome::NoPressure);
374 0 : }
375 0 :
376 0 : warn!(
377 : ?usage_pre,
378 0 : "running disk usage based eviction due to pressure"
379 : );
380 :
381 0 : let (candidates, collection_time) = {
382 0 : let started_at = std::time::Instant::now();
383 0 : match collect_eviction_candidates(tenant_manager, eviction_order, cancel).await? {
384 : EvictionCandidates::Cancelled => {
385 0 : return Ok(IterationOutcome::Cancelled);
386 : }
387 0 : EvictionCandidates::Finished(partitioned) => (partitioned, started_at.elapsed()),
388 0 : }
389 0 : };
390 0 :
391 0 : METRICS.layers_collected.inc_by(candidates.len() as u64);
392 0 :
393 0 : tracing::info!(
394 0 : elapsed_ms = collection_time.as_millis(),
395 0 : total_layers = candidates.len(),
396 0 : "collection completed"
397 : );
398 :
399 : // Debug-log the list of candidates
400 0 : let now = SystemTime::now();
401 0 : for (i, (partition, candidate)) in candidates.iter().enumerate() {
402 0 : let nth = i + 1;
403 0 : let total_candidates = candidates.len();
404 0 : let size = candidate.layer.get_file_size();
405 0 : let rel = candidate.relative_last_activity;
406 0 : debug!(
407 0 : "cand {nth}/{total_candidates}: size={size}, rel_last_activity={rel}, no_access_for={}us, partition={partition:?}, {}/{}/{}",
408 0 : now.duration_since(candidate.last_activity_ts)
409 0 : .unwrap()
410 0 : .as_micros(),
411 0 : candidate.layer.get_tenant_shard_id(),
412 0 : candidate.layer.get_timeline_id(),
413 0 : candidate.layer.get_name(),
414 : );
415 : }
416 :
417 : // phase1: select victims to relieve pressure
418 : //
419 : // Walk through the list of candidates, until we have accumulated enough layers to get
420 : // us back under the pressure threshold. 'usage_planned' is updated so that it tracks
421 : // how much disk space would be used after evicting all the layers up to the current
422 : // point in the list.
423 : //
424 : // If we get far enough in the list that we start to evict layers that are below
425 : // the tenant's min-resident-size threshold, print a warning, and memorize the disk
426 : // usage at that point, in 'usage_planned_min_resident_size_respecting'.
427 :
428 0 : let (evicted_amount, usage_planned) =
429 0 : select_victims(&candidates, usage_pre).into_amount_and_planned();
430 0 :
431 0 : METRICS.layers_selected.inc_by(evicted_amount as u64);
432 0 :
433 0 : // phase2: evict layers
434 0 :
435 0 : let mut js = tokio::task::JoinSet::new();
436 0 : let limit = 1000;
437 0 :
438 0 : let mut evicted = candidates.into_iter().take(evicted_amount).fuse();
439 0 : let mut consumed_all = false;
440 0 :
441 0 : // After the evictions, `usage_assumed` is the post-eviction usage,
442 0 : // according to internal accounting.
443 0 : let mut usage_assumed = usage_pre;
444 0 : let mut evictions_failed = LayerCount::default();
445 0 :
446 0 : let evict_layers = async move {
447 : loop {
448 0 : let next = if js.len() >= limit || consumed_all {
449 0 : js.join_next().await
450 0 : } else if !js.is_empty() {
451 : // opportunistically consume ready result, one per each new evicted
452 0 : futures::future::FutureExt::now_or_never(js.join_next()).and_then(|x| x)
453 : } else {
454 0 : None
455 : };
456 :
457 0 : if let Some(next) = next {
458 0 : match next {
459 0 : Ok(Ok(file_size)) => {
460 0 : METRICS.layers_evicted.inc();
461 0 : usage_assumed.add_available_bytes(file_size);
462 0 : }
463 : Ok(Err((
464 0 : file_size,
465 0 : EvictionError::NotFound
466 0 : | EvictionError::Downloaded
467 0 : | EvictionError::Timeout,
468 0 : ))) => {
469 0 : evictions_failed.file_sizes += file_size;
470 0 : evictions_failed.count += 1;
471 0 : }
472 0 : Err(je) if je.is_cancelled() => unreachable!("not used"),
473 0 : Err(je) if je.is_panic() => { /* already logged */ }
474 0 : Err(je) => tracing::error!("unknown JoinError: {je:?}"),
475 : }
476 0 : }
477 :
478 0 : if consumed_all && js.is_empty() {
479 0 : break;
480 0 : }
481 :
482 : // calling again when consumed_all is fine as evicted is fused.
483 0 : let Some((_partition, candidate)) = evicted.next() else {
484 0 : if !consumed_all {
485 0 : tracing::info!("all evictions started, waiting");
486 0 : consumed_all = true;
487 0 : }
488 0 : continue;
489 : };
490 :
491 0 : match candidate.layer {
492 0 : EvictionLayer::Attached(layer) => {
493 0 : let file_size = layer.layer_desc().file_size;
494 0 : js.spawn(async move {
495 0 : // have a low eviction waiting timeout because our LRU calculations go stale fast;
496 0 : // also individual layer evictions could hang because of bugs and we do not want to
497 0 : // pause disk_usage_based_eviction for such.
498 0 : let timeout = std::time::Duration::from_secs(5);
499 0 :
500 0 : match layer.evict_and_wait(timeout).await {
501 0 : Ok(()) => Ok(file_size),
502 0 : Err(e) => Err((file_size, e)),
503 : }
504 0 : });
505 0 : }
506 0 : EvictionLayer::Secondary(layer) => {
507 0 : let file_size = layer.metadata.file_size;
508 0 :
509 0 : js.spawn(async move {
510 0 : layer
511 0 : .secondary_tenant
512 0 : .evict_layer(layer.timeline_id, layer.name)
513 0 : .await;
514 0 : Ok(file_size)
515 0 : });
516 0 : }
517 : }
518 0 : tokio::task::yield_now().await;
519 : }
520 :
521 0 : (usage_assumed, evictions_failed)
522 0 : };
523 :
524 0 : let started_at = std::time::Instant::now();
525 0 :
526 0 : let evict_layers = async move {
527 0 : let mut evict_layers = std::pin::pin!(evict_layers);
528 0 :
529 0 : let maximum_expected = std::time::Duration::from_secs(10);
530 :
531 0 : let res = tokio::time::timeout(maximum_expected, &mut evict_layers).await;
532 0 : let tuple = if let Ok(tuple) = res {
533 0 : tuple
534 : } else {
535 0 : let elapsed = started_at.elapsed();
536 0 : tracing::info!(elapsed_ms = elapsed.as_millis(), "still ongoing");
537 0 : evict_layers.await
538 : };
539 :
540 0 : let elapsed = started_at.elapsed();
541 0 : tracing::info!(elapsed_ms = elapsed.as_millis(), "completed");
542 0 : tuple
543 0 : };
544 :
545 0 : let evict_layers =
546 0 : evict_layers.instrument(tracing::info_span!("evict_layers", layers=%evicted_amount));
547 :
548 0 : let (usage_assumed, evictions_failed) = tokio::select! {
549 0 : tuple = evict_layers => { tuple },
550 0 : _ = cancel.cancelled() => {
551 : // dropping joinset will abort all pending evict_and_waits and that is fine, our
552 : // requests will still stand
553 0 : return Ok(IterationOutcome::Cancelled);
554 : }
555 : };
556 :
557 0 : Ok(IterationOutcome::Finished(IterationOutcomeFinished {
558 0 : before: usage_pre,
559 0 : planned: usage_planned,
560 0 : assumed: AssumedUsage {
561 0 : projected_after: usage_assumed,
562 0 : failed: evictions_failed,
563 0 : },
564 0 : }))
565 0 : }
566 :
567 : #[derive(Clone)]
568 : pub(crate) struct EvictionSecondaryLayer {
569 : pub(crate) secondary_tenant: Arc<SecondaryTenant>,
570 : pub(crate) timeline_id: TimelineId,
571 : pub(crate) name: LayerName,
572 : pub(crate) metadata: LayerFileMetadata,
573 : }
574 :
575 : /// Full [`Layer`] objects are specific to tenants in attached mode. This type is a layer
576 : /// of indirection to store either a `Layer`, or a reference to a secondary tenant and a layer name.
577 : #[derive(Clone)]
578 : pub(crate) enum EvictionLayer {
579 : Attached(Layer),
580 : Secondary(EvictionSecondaryLayer),
581 : }
582 :
583 : impl From<Layer> for EvictionLayer {
584 0 : fn from(value: Layer) -> Self {
585 0 : Self::Attached(value)
586 0 : }
587 : }
588 :
589 : impl EvictionLayer {
590 0 : pub(crate) fn get_tenant_shard_id(&self) -> &TenantShardId {
591 0 : match self {
592 0 : Self::Attached(l) => &l.layer_desc().tenant_shard_id,
593 0 : Self::Secondary(sl) => sl.secondary_tenant.get_tenant_shard_id(),
594 : }
595 0 : }
596 :
597 0 : pub(crate) fn get_timeline_id(&self) -> &TimelineId {
598 0 : match self {
599 0 : Self::Attached(l) => &l.layer_desc().timeline_id,
600 0 : Self::Secondary(sl) => &sl.timeline_id,
601 : }
602 0 : }
603 :
604 0 : pub(crate) fn get_name(&self) -> LayerName {
605 0 : match self {
606 0 : Self::Attached(l) => l.layer_desc().layer_name(),
607 0 : Self::Secondary(sl) => sl.name.clone(),
608 : }
609 0 : }
610 :
611 0 : pub(crate) fn get_file_size(&self) -> u64 {
612 0 : match self {
613 0 : Self::Attached(l) => l.layer_desc().file_size,
614 0 : Self::Secondary(sl) => sl.metadata.file_size,
615 : }
616 0 : }
617 : }
618 :
619 : #[derive(Clone)]
620 : pub(crate) struct EvictionCandidate {
621 : pub(crate) layer: EvictionLayer,
622 : pub(crate) last_activity_ts: SystemTime,
623 : pub(crate) relative_last_activity: finite_f32::FiniteF32,
624 : pub(crate) visibility: LayerVisibilityHint,
625 : }
626 :
627 : impl std::fmt::Display for EvictionLayer {
628 0 : fn fmt(&self, f: &mut std::fmt::Formatter) -> std::fmt::Result {
629 0 : match self {
630 0 : Self::Attached(l) => l.fmt(f),
631 0 : Self::Secondary(sl) => {
632 0 : write!(f, "{}/{}", sl.timeline_id, sl.name)
633 : }
634 : }
635 0 : }
636 : }
637 :
638 : #[derive(Default)]
639 : pub(crate) struct DiskUsageEvictionInfo {
640 : /// Timeline's largest layer (remote or resident)
641 : pub max_layer_size: Option<u64>,
642 : /// Timeline's resident layers
643 : pub resident_layers: Vec<EvictionCandidate>,
644 : }
645 :
646 : impl std::fmt::Debug for EvictionCandidate {
647 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
648 0 : // format the tv_sec, tv_nsec into rfc3339 in case someone is looking at it
649 0 : // having to allocate a string to this is bad, but it will rarely be formatted
650 0 : let ts = chrono::DateTime::<chrono::Utc>::from(self.last_activity_ts);
651 0 : let ts = ts.to_rfc3339_opts(chrono::SecondsFormat::Nanos, true);
652 : struct DisplayIsDebug<'a, T>(&'a T);
653 : impl<T: std::fmt::Display> std::fmt::Debug for DisplayIsDebug<'_, T> {
654 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
655 0 : write!(f, "{}", self.0)
656 0 : }
657 : }
658 0 : f.debug_struct("LocalLayerInfoForDiskUsageEviction")
659 0 : .field("layer", &DisplayIsDebug(&self.layer))
660 0 : .field("last_activity", &ts)
661 0 : .finish()
662 0 : }
663 : }
664 :
665 : #[derive(Debug, Clone, Copy, PartialEq, Eq, PartialOrd, Ord)]
666 : enum EvictionPartition {
667 : // A layer that is un-wanted by the tenant: evict all these first, before considering
668 : // any other layers
669 : EvictNow,
670 :
671 : // Above the minimum size threshold: this layer is a candidate for eviction.
672 : Above,
673 :
674 : // Below the minimum size threshold: this layer should only be evicted if all the
675 : // tenants' layers above the minimum size threshold have already been considered.
676 : Below,
677 : }
678 :
679 : enum EvictionCandidates {
680 : Cancelled,
681 : Finished(Vec<(EvictionPartition, EvictionCandidate)>),
682 : }
683 :
684 : /// Gather the eviction candidates.
685 : ///
686 : /// The returned `Ok(EvictionCandidates::Finished(candidates))` is sorted in eviction
687 : /// order. A caller that evicts in that order, until pressure is relieved, implements
688 : /// the eviction policy outlined in the module comment.
689 : ///
690 : /// # Example with EvictionOrder::AbsoluteAccessed
691 : ///
692 : /// Imagine that there are two tenants, A and B, with five layers each, a-e.
693 : /// Each layer has size 100, and both tenant's min_resident_size is 150.
694 : /// The eviction order would be
695 : ///
696 : /// ```text
697 : /// partition last_activity_ts tenant/layer
698 : /// Above 18:30 A/c
699 : /// Above 19:00 A/b
700 : /// Above 18:29 B/c
701 : /// Above 19:05 B/b
702 : /// Above 20:00 B/a
703 : /// Above 20:03 A/a
704 : /// Below 20:30 A/d
705 : /// Below 20:40 B/d
706 : /// Below 20:45 B/e
707 : /// Below 20:58 A/e
708 : /// ```
709 : ///
710 : /// Now, if we need to evict 300 bytes to relieve pressure, we'd evict `A/c, A/b, B/c`.
711 : /// They are all in the `Above` partition, so, we respected each tenant's min_resident_size.
712 : ///
713 : /// But, if we need to evict 900 bytes to relieve pressure, we'd evict
714 : /// `A/c, A/b, B/c, B/b, B/a, A/a, A/d, B/d, B/e`, reaching into the `Below` partition
715 : /// after exhauting the `Above` partition.
716 : /// So, we did not respect each tenant's min_resident_size.
717 : ///
718 : /// # Example with EvictionOrder::RelativeAccessed
719 : ///
720 : /// ```text
721 : /// partition relative_age last_activity_ts tenant/layer
722 : /// Above 0/4 18:30 A/c
723 : /// Above 0/4 18:29 B/c
724 : /// Above 1/4 19:00 A/b
725 : /// Above 1/4 19:05 B/b
726 : /// Above 2/4 20:00 B/a
727 : /// Above 2/4 20:03 A/a
728 : /// Below 3/4 20:30 A/d
729 : /// Below 3/4 20:40 B/d
730 : /// Below 4/4 20:45 B/e
731 : /// Below 4/4 20:58 A/e
732 : /// ```
733 : ///
734 : /// With tenants having the same number of layers the picture does not change much. The same with
735 : /// A having many more layers **resident** (not all of them listed):
736 : ///
737 : /// ```text
738 : /// Above 0/100 18:30 A/c
739 : /// Above 0/4 18:29 B/c
740 : /// Above 1/100 19:00 A/b
741 : /// Above 2/100 20:03 A/a
742 : /// Above 3/100 20:03 A/nth_3
743 : /// Above 4/100 20:03 A/nth_4
744 : /// ...
745 : /// Above 1/4 19:05 B/b
746 : /// Above 25/100 20:04 A/nth_25
747 : /// ...
748 : /// Above 2/4 20:00 B/a
749 : /// Above 50/100 20:10 A/nth_50
750 : /// ...
751 : /// Below 3/4 20:40 B/d
752 : /// Below 99/100 20:30 A/nth_99
753 : /// Below 4/4 20:45 B/e
754 : /// Below 100/100 20:58 A/nth_100
755 : /// ```
756 : ///
757 : /// Now it's easier to see that because A has grown fast it has more layers to get evicted. What is
758 : /// difficult to see is what happens on the next round assuming the evicting 23 from the above list
759 : /// relieves the pressure (22 A layers gone, 1 B layers gone) but a new fast growing tenant C has
760 : /// appeared:
761 : ///
762 : /// ```text
763 : /// Above 0/87 20:04 A/nth_23
764 : /// Above 0/3 19:05 B/b
765 : /// Above 0/50 20:59 C/nth_0
766 : /// Above 1/87 20:04 A/nth_24
767 : /// Above 1/50 21:00 C/nth_1
768 : /// Above 2/87 20:04 A/nth_25
769 : /// ...
770 : /// Above 16/50 21:02 C/nth_16
771 : /// Above 1/3 20:00 B/a
772 : /// Above 27/87 20:10 A/nth_50
773 : /// ...
774 : /// Below 2/3 20:40 B/d
775 : /// Below 49/50 21:05 C/nth_49
776 : /// Below 86/87 20:30 A/nth_99
777 : /// Below 3/3 20:45 B/e
778 : /// Below 50/50 21:05 C/nth_50
779 : /// Below 87/87 20:58 A/nth_100
780 : /// ```
781 : ///
782 : /// Now relieving pressure with 23 layers would cost:
783 : /// - tenant A 14 layers
784 : /// - tenant B 1 layer
785 : /// - tenant C 8 layers
786 0 : async fn collect_eviction_candidates(
787 0 : tenant_manager: &Arc<TenantManager>,
788 0 : eviction_order: EvictionOrder,
789 0 : cancel: &CancellationToken,
790 0 : ) -> anyhow::Result<EvictionCandidates> {
791 : const LOG_DURATION_THRESHOLD: std::time::Duration = std::time::Duration::from_secs(10);
792 :
793 : // get a snapshot of the list of tenants
794 0 : let tenants = tenant_manager
795 0 : .list_tenants()
796 0 : .context("get list of tenants")?;
797 :
798 : // TODO: avoid listing every layer in every tenant: this loop can block the executor,
799 : // and the resulting data structure can be huge.
800 : // (https://github.com/neondatabase/neon/issues/6224)
801 0 : let mut candidates = Vec::new();
802 :
803 0 : for (tenant_id, _state, _gen) in tenants {
804 0 : if cancel.is_cancelled() {
805 0 : return Ok(EvictionCandidates::Cancelled);
806 0 : }
807 0 : let tenant = match tenant_manager.get_attached_tenant_shard(tenant_id) {
808 0 : Ok(tenant) if tenant.is_active() => tenant,
809 : Ok(_) => {
810 0 : debug!(tenant_id=%tenant_id.tenant_id, shard_id=%tenant_id.shard_slug(), "Tenant shard is not active");
811 0 : continue;
812 : }
813 0 : Err(e) => {
814 0 : // this can happen if tenant has lifecycle transition after we fetched it
815 0 : debug!("failed to get tenant: {e:#}");
816 0 : continue;
817 : }
818 : };
819 :
820 0 : if tenant.cancel.is_cancelled() {
821 0 : info!(%tenant_id, "Skipping tenant for eviction, it is shutting down");
822 0 : continue;
823 0 : }
824 0 :
825 0 : let started_at = std::time::Instant::now();
826 0 :
827 0 : // collect layers from all timelines in this tenant
828 0 : //
829 0 : // If one of the timelines becomes `!is_active()` during the iteration,
830 0 : // for example because we're shutting down, then `max_layer_size` can be too small.
831 0 : // That's OK. This code only runs under a disk pressure situation, and being
832 0 : // a little unfair to tenants during shutdown in such a situation is tolerable.
833 0 : let mut tenant_candidates = Vec::new();
834 0 : let mut max_layer_size = 0;
835 0 : for tl in tenant.list_timelines() {
836 0 : if !tl.is_active() {
837 0 : continue;
838 0 : }
839 0 : let info = tl.get_local_layers_for_disk_usage_eviction().await;
840 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());
841 :
842 0 : tenant_candidates.extend(info.resident_layers.into_iter());
843 0 : max_layer_size = max_layer_size.max(info.max_layer_size.unwrap_or(0));
844 0 :
845 0 : if cancel.is_cancelled() {
846 0 : return Ok(EvictionCandidates::Cancelled);
847 0 : }
848 : }
849 :
850 : // `min_resident_size` defaults to maximum layer file size of the tenant.
851 : // This ensures that each tenant can have at least one layer resident at a given time,
852 : // ensuring forward progress for a single Timeline::get in that tenant.
853 : // It's a questionable heuristic since, usually, there are many Timeline::get
854 : // requests going on for a tenant, and, at least in Neon prod, the median
855 : // layer file size is much smaller than the compaction target size.
856 : // We could be better here, e.g., sum of all L0 layers + most recent L1 layer.
857 : // That's what's typically used by the various background loops.
858 : //
859 : // The default can be overridden with a fixed value in the tenant conf.
860 : // A default override can be put in the default tenant conf in the pageserver.toml.
861 0 : let min_resident_size = if let Some(s) = tenant.get_min_resident_size_override() {
862 0 : debug!(
863 0 : tenant_id=%tenant.tenant_shard_id().tenant_id,
864 0 : shard_id=%tenant.tenant_shard_id().shard_slug(),
865 0 : overridden_size=s,
866 0 : "using overridden min resident size for tenant"
867 : );
868 0 : s
869 : } else {
870 0 : debug!(
871 0 : tenant_id=%tenant.tenant_shard_id().tenant_id,
872 0 : shard_id=%tenant.tenant_shard_id().shard_slug(),
873 0 : max_layer_size,
874 0 : "using max layer size as min_resident_size for tenant",
875 : );
876 0 : max_layer_size
877 : };
878 :
879 : // Sort layers most-recently-used first, then calculate [`EvictionPartition`] for each layer,
880 : // where the inputs are:
881 : // - whether the layer is visible
882 : // - whether the layer is above/below the min_resident_size cutline
883 0 : tenant_candidates
884 0 : .sort_unstable_by_key(|layer_info| std::cmp::Reverse(layer_info.last_activity_ts));
885 0 : let mut cumsum: i128 = 0;
886 0 :
887 0 : let total = tenant_candidates.len();
888 0 :
889 0 : let tenant_candidates =
890 0 : tenant_candidates
891 0 : .into_iter()
892 0 : .enumerate()
893 0 : .map(|(i, mut candidate)| {
894 0 : // as we iterate this reverse sorted list, the most recently accessed layer will always
895 0 : // be 1.0; this is for us to evict it last.
896 0 : candidate.relative_last_activity =
897 0 : eviction_order.relative_last_activity(total, i);
898 :
899 0 : let partition = match candidate.visibility {
900 : LayerVisibilityHint::Covered => {
901 : // Covered layers are evicted first
902 0 : EvictionPartition::EvictNow
903 : }
904 : LayerVisibilityHint::Visible => {
905 0 : cumsum += i128::from(candidate.layer.get_file_size());
906 0 :
907 0 : if cumsum > min_resident_size as i128 {
908 0 : EvictionPartition::Above
909 : } else {
910 : // The most recent layers below the min_resident_size threshold
911 : // are the last to be evicted.
912 0 : EvictionPartition::Below
913 : }
914 : }
915 : };
916 :
917 0 : (partition, candidate)
918 0 : });
919 0 :
920 0 : METRICS
921 0 : .tenant_layer_count
922 0 : .observe(tenant_candidates.len() as f64);
923 0 :
924 0 : candidates.extend(tenant_candidates);
925 0 :
926 0 : let elapsed = started_at.elapsed();
927 0 : METRICS
928 0 : .tenant_collection_time
929 0 : .observe(elapsed.as_secs_f64());
930 0 :
931 0 : if elapsed > LOG_DURATION_THRESHOLD {
932 0 : tracing::info!(
933 0 : tenant_id=%tenant.tenant_shard_id().tenant_id,
934 0 : shard_id=%tenant.tenant_shard_id().shard_slug(),
935 0 : elapsed_ms = elapsed.as_millis(),
936 0 : "collection took longer than threshold"
937 : );
938 0 : }
939 : }
940 :
941 : // Note: the same tenant ID might be hit twice, if it transitions from attached to
942 : // secondary while we run. That is okay: when we eventually try and run the eviction,
943 : // the `Gate` on the object will ensure that whichever one has already been shut down
944 : // will not delete anything.
945 :
946 0 : let mut secondary_tenants = Vec::new();
947 0 : tenant_manager.foreach_secondary_tenants(
948 0 : |_tenant_shard_id: &TenantShardId, state: &Arc<SecondaryTenant>| {
949 0 : secondary_tenants.push(state.clone());
950 0 : },
951 0 : );
952 :
953 0 : for tenant in secondary_tenants {
954 : // for secondary tenants we use a sum of on_disk layers and already evicted layers. this is
955 : // to prevent repeated disk usage based evictions from completely draining less often
956 : // updating secondaries.
957 0 : let (mut layer_info, total_layers) = tenant.get_layers_for_eviction();
958 0 :
959 0 : debug_assert!(
960 0 : total_layers >= layer_info.resident_layers.len(),
961 0 : "total_layers ({total_layers}) must be at least the resident_layers.len() ({})",
962 0 : layer_info.resident_layers.len()
963 : );
964 :
965 0 : let started_at = std::time::Instant::now();
966 0 :
967 0 : layer_info
968 0 : .resident_layers
969 0 : .sort_unstable_by_key(|layer_info| std::cmp::Reverse(layer_info.last_activity_ts));
970 0 :
971 0 : let tenant_candidates =
972 0 : layer_info
973 0 : .resident_layers
974 0 : .into_iter()
975 0 : .enumerate()
976 0 : .map(|(i, mut candidate)| {
977 0 : candidate.relative_last_activity =
978 0 : eviction_order.relative_last_activity(total_layers, i);
979 0 : (
980 0 : // Secondary locations' layers are always considered above the min resident size,
981 0 : // i.e. secondary locations are permitted to be trimmed to zero layers if all
982 0 : // the layers have sufficiently old access times.
983 0 : EvictionPartition::Above,
984 0 : candidate,
985 0 : )
986 0 : });
987 0 :
988 0 : METRICS
989 0 : .tenant_layer_count
990 0 : .observe(tenant_candidates.len() as f64);
991 0 : candidates.extend(tenant_candidates);
992 0 :
993 0 : tokio::task::yield_now().await;
994 :
995 0 : let elapsed = started_at.elapsed();
996 0 :
997 0 : METRICS
998 0 : .tenant_collection_time
999 0 : .observe(elapsed.as_secs_f64());
1000 0 :
1001 0 : if elapsed > LOG_DURATION_THRESHOLD {
1002 0 : tracing::info!(
1003 0 : tenant_id=%tenant.tenant_shard_id().tenant_id,
1004 0 : shard_id=%tenant.tenant_shard_id().shard_slug(),
1005 0 : elapsed_ms = elapsed.as_millis(),
1006 0 : "collection took longer than threshold"
1007 : );
1008 0 : }
1009 : }
1010 :
1011 0 : debug_assert!(
1012 0 : EvictionPartition::Above < EvictionPartition::Below,
1013 0 : "as explained in the function's doc comment, layers that aren't in the tenant's min_resident_size are evicted first"
1014 : );
1015 0 : debug_assert!(
1016 0 : EvictionPartition::EvictNow < EvictionPartition::Above,
1017 0 : "as explained in the function's doc comment, layers that aren't in the tenant's min_resident_size are evicted first"
1018 : );
1019 :
1020 0 : eviction_order.sort(&mut candidates);
1021 0 :
1022 0 : Ok(EvictionCandidates::Finished(candidates))
1023 0 : }
1024 :
1025 : /// Given a pre-sorted vec of all layers in the system, select the first N which are enough to
1026 : /// relieve pressure.
1027 : ///
1028 : /// Returns the amount of candidates selected, with the planned usage.
1029 0 : fn select_victims<U: Usage>(
1030 0 : candidates: &[(EvictionPartition, EvictionCandidate)],
1031 0 : usage_pre: U,
1032 0 : ) -> VictimSelection<U> {
1033 0 : let mut usage_when_switched = None;
1034 0 : let mut usage_planned = usage_pre;
1035 0 : let mut evicted_amount = 0;
1036 :
1037 0 : for (i, (partition, candidate)) in candidates.iter().enumerate() {
1038 0 : if !usage_planned.has_pressure() {
1039 0 : break;
1040 0 : }
1041 0 :
1042 0 : if partition == &EvictionPartition::Below && usage_when_switched.is_none() {
1043 0 : usage_when_switched = Some((usage_planned, i));
1044 0 : }
1045 :
1046 0 : usage_planned.add_available_bytes(candidate.layer.get_file_size());
1047 0 : evicted_amount += 1;
1048 : }
1049 :
1050 0 : VictimSelection {
1051 0 : amount: evicted_amount,
1052 0 : usage_pre,
1053 0 : usage_when_switched,
1054 0 : usage_planned,
1055 0 : }
1056 0 : }
1057 :
1058 : struct VictimSelection<U> {
1059 : amount: usize,
1060 : usage_pre: U,
1061 : usage_when_switched: Option<(U, usize)>,
1062 : usage_planned: U,
1063 : }
1064 :
1065 : impl<U: Usage> VictimSelection<U> {
1066 0 : fn into_amount_and_planned(self) -> (usize, PlannedUsage<U>) {
1067 0 : debug!(
1068 : evicted_amount=%self.amount,
1069 0 : "took enough candidates for pressure to be relieved"
1070 : );
1071 :
1072 0 : if let Some((usage_planned, candidate_no)) = self.usage_when_switched.as_ref() {
1073 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");
1074 0 : }
1075 :
1076 0 : let planned = match self.usage_when_switched {
1077 0 : Some((respecting_tenant_min_resident_size, _)) => PlannedUsage {
1078 0 : respecting_tenant_min_resident_size,
1079 0 : fallback_to_global_lru: Some(self.usage_planned),
1080 0 : },
1081 0 : None => PlannedUsage {
1082 0 : respecting_tenant_min_resident_size: self.usage_planned,
1083 0 : fallback_to_global_lru: None,
1084 0 : },
1085 : };
1086 :
1087 0 : (self.amount, planned)
1088 0 : }
1089 : }
1090 :
1091 : /// A totally ordered f32 subset we can use with sorting functions.
1092 : pub(crate) mod finite_f32 {
1093 :
1094 : /// A totally ordered f32 subset we can use with sorting functions.
1095 : #[derive(Clone, Copy, PartialEq)]
1096 : pub struct FiniteF32(f32);
1097 :
1098 : impl std::fmt::Debug for FiniteF32 {
1099 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
1100 0 : std::fmt::Debug::fmt(&self.0, f)
1101 0 : }
1102 : }
1103 :
1104 : impl std::fmt::Display for FiniteF32 {
1105 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
1106 0 : std::fmt::Display::fmt(&self.0, f)
1107 0 : }
1108 : }
1109 :
1110 : impl std::cmp::Eq for FiniteF32 {}
1111 :
1112 : impl std::cmp::PartialOrd for FiniteF32 {
1113 0 : fn partial_cmp(&self, other: &Self) -> Option<std::cmp::Ordering> {
1114 0 : Some(self.cmp(other))
1115 0 : }
1116 : }
1117 :
1118 : impl std::cmp::Ord for FiniteF32 {
1119 0 : fn cmp(&self, other: &Self) -> std::cmp::Ordering {
1120 0 : self.0.total_cmp(&other.0)
1121 0 : }
1122 : }
1123 :
1124 : impl TryFrom<f32> for FiniteF32 {
1125 : type Error = f32;
1126 :
1127 0 : fn try_from(value: f32) -> Result<Self, Self::Error> {
1128 0 : if value.is_finite() {
1129 0 : Ok(FiniteF32(value))
1130 : } else {
1131 0 : Err(value)
1132 : }
1133 0 : }
1134 : }
1135 :
1136 : impl From<FiniteF32> for f32 {
1137 80 : fn from(value: FiniteF32) -> f32 {
1138 80 : value.0
1139 80 : }
1140 : }
1141 :
1142 : impl FiniteF32 {
1143 : pub const ZERO: FiniteF32 = FiniteF32(0.0);
1144 :
1145 80 : pub fn try_from_normalized(value: f32) -> Result<Self, f32> {
1146 80 : if (0.0..=1.0).contains(&value) {
1147 : // -0.0 is within the range, make sure it is assumed 0.0..=1.0
1148 80 : let value = value.abs();
1149 80 : Ok(FiniteF32(value))
1150 : } else {
1151 0 : Err(value)
1152 : }
1153 80 : }
1154 :
1155 80 : pub fn into_inner(self) -> f32 {
1156 80 : self.into()
1157 80 : }
1158 : }
1159 : }
1160 :
1161 : mod filesystem_level_usage {
1162 : use anyhow::Context;
1163 : use camino::Utf8Path;
1164 :
1165 : use super::DiskUsageEvictionTaskConfig;
1166 : use crate::statvfs::Statvfs;
1167 :
1168 : #[derive(Debug, Clone, Copy)]
1169 : pub struct Usage<'a> {
1170 : config: &'a DiskUsageEvictionTaskConfig,
1171 :
1172 : /// Filesystem capacity
1173 : total_bytes: u64,
1174 : /// Free filesystem space
1175 : avail_bytes: u64,
1176 : }
1177 :
1178 : impl super::Usage for Usage<'_> {
1179 28 : fn has_pressure(&self) -> bool {
1180 28 : let usage_pct =
1181 28 : (100.0 * (1.0 - ((self.avail_bytes as f64) / (self.total_bytes as f64)))) as u64;
1182 28 :
1183 28 : let pressures = [
1184 28 : (
1185 28 : "min_avail_bytes",
1186 28 : self.avail_bytes < self.config.min_avail_bytes,
1187 28 : ),
1188 28 : (
1189 28 : "max_usage_pct",
1190 28 : usage_pct >= self.config.max_usage_pct.get() as u64,
1191 28 : ),
1192 28 : ];
1193 28 :
1194 56 : pressures.into_iter().any(|(_, has_pressure)| has_pressure)
1195 28 : }
1196 :
1197 24 : fn add_available_bytes(&mut self, bytes: u64) {
1198 24 : self.avail_bytes += bytes;
1199 24 : }
1200 : }
1201 :
1202 0 : pub fn get<'a>(
1203 0 : tenants_dir: &Utf8Path,
1204 0 : config: &'a DiskUsageEvictionTaskConfig,
1205 0 : ) -> anyhow::Result<Usage<'a>> {
1206 0 : let mock_config = {
1207 0 : #[cfg(feature = "testing")]
1208 0 : {
1209 0 : config.mock_statvfs.as_ref()
1210 : }
1211 : #[cfg(not(feature = "testing"))]
1212 : {
1213 : None
1214 : }
1215 : };
1216 :
1217 0 : let stat = Statvfs::get(tenants_dir, mock_config)
1218 0 : .context("statvfs failed, presumably directory got unlinked")?;
1219 :
1220 0 : let (avail_bytes, total_bytes) = stat.get_avail_total_bytes();
1221 0 :
1222 0 : Ok(Usage {
1223 0 : config,
1224 0 : total_bytes,
1225 0 : avail_bytes,
1226 0 : })
1227 0 : }
1228 :
1229 : #[test]
1230 4 : fn max_usage_pct_pressure() {
1231 : use std::time::Duration;
1232 :
1233 : use utils::serde_percent::Percent;
1234 :
1235 : use super::Usage as _;
1236 :
1237 4 : let mut usage = Usage {
1238 4 : config: &DiskUsageEvictionTaskConfig {
1239 4 : max_usage_pct: Percent::new(85).unwrap(),
1240 4 : min_avail_bytes: 0,
1241 4 : period: Duration::MAX,
1242 4 : #[cfg(feature = "testing")]
1243 4 : mock_statvfs: None,
1244 4 : eviction_order: pageserver_api::config::EvictionOrder::default(),
1245 4 : },
1246 4 : total_bytes: 100_000,
1247 4 : avail_bytes: 0,
1248 4 : };
1249 4 :
1250 4 : assert!(usage.has_pressure(), "expected pressure at 100%");
1251 :
1252 4 : usage.add_available_bytes(14_000);
1253 4 : assert!(usage.has_pressure(), "expected pressure at 86%");
1254 :
1255 4 : usage.add_available_bytes(999);
1256 4 : assert!(usage.has_pressure(), "expected pressure at 85.001%");
1257 :
1258 4 : usage.add_available_bytes(1);
1259 4 : assert!(usage.has_pressure(), "expected pressure at precisely 85%");
1260 :
1261 4 : usage.add_available_bytes(1);
1262 4 : assert!(!usage.has_pressure(), "no pressure at 84.999%");
1263 :
1264 4 : usage.add_available_bytes(999);
1265 4 : assert!(!usage.has_pressure(), "no pressure at 84%");
1266 :
1267 4 : usage.add_available_bytes(16_000);
1268 4 : assert!(!usage.has_pressure());
1269 4 : }
1270 : }
1271 :
1272 : #[cfg(test)]
1273 : mod tests {
1274 : use super::*;
1275 :
1276 : #[test]
1277 4 : fn relative_equal_bounds() {
1278 4 : let order = EvictionOrder::RelativeAccessed {
1279 4 : highest_layer_count_loses_first: false,
1280 4 : };
1281 4 :
1282 4 : let len = 10;
1283 4 : let v = (0..len)
1284 40 : .map(|i| order.relative_last_activity(len, i).into_inner())
1285 4 : .collect::<Vec<_>>();
1286 4 :
1287 4 : assert_eq!(v.first(), Some(&1.0));
1288 4 : assert_eq!(v.last(), Some(&0.0));
1289 36 : assert!(v.windows(2).all(|slice| slice[0] > slice[1]));
1290 4 : }
1291 :
1292 : #[test]
1293 4 : fn relative_spare_bounds() {
1294 4 : let order = EvictionOrder::RelativeAccessed {
1295 4 : highest_layer_count_loses_first: true,
1296 4 : };
1297 4 :
1298 4 : let len = 10;
1299 4 : let v = (0..len)
1300 40 : .map(|i| order.relative_last_activity(len, i).into_inner())
1301 4 : .collect::<Vec<_>>();
1302 4 :
1303 4 : assert_eq!(v.first(), Some(&1.0));
1304 4 : assert_eq!(v.last(), Some(&0.1));
1305 36 : assert!(v.windows(2).all(|slice| slice[0] > slice[1]));
1306 4 : }
1307 : }
|
