Line data Source code
1 : use std::collections::{BTreeMap, BTreeSet, HashMap};
2 : use std::sync::Arc;
3 : use std::time::Duration;
4 :
5 : use crate::checks::{
6 : list_tenant_manifests, list_timeline_blobs, BlobDataParseResult, ListTenantManifestResult,
7 : RemoteTenantManifestInfo,
8 : };
9 : use crate::metadata_stream::{stream_tenant_timelines, stream_tenants};
10 : use crate::{init_remote, BucketConfig, NodeKind, RootTarget, TenantShardTimelineId, MAX_RETRIES};
11 : use futures_util::{StreamExt, TryStreamExt};
12 : use pageserver::tenant::remote_timeline_client::index::LayerFileMetadata;
13 : use pageserver::tenant::remote_timeline_client::manifest::OffloadedTimelineManifest;
14 : use pageserver::tenant::remote_timeline_client::{
15 : parse_remote_index_path, parse_remote_tenant_manifest_path, remote_layer_path,
16 : };
17 : use pageserver::tenant::storage_layer::LayerName;
18 : use pageserver::tenant::IndexPart;
19 : use pageserver_api::controller_api::TenantDescribeResponse;
20 : use pageserver_api::shard::{ShardIndex, TenantShardId};
21 : use remote_storage::{GenericRemoteStorage, ListingObject, RemotePath};
22 : use reqwest::Method;
23 : use serde::Serialize;
24 : use storage_controller_client::control_api;
25 : use tokio_util::sync::CancellationToken;
26 : use tracing::{info_span, Instrument};
27 : use utils::backoff;
28 : use utils::generation::Generation;
29 : use utils::id::{TenantId, TenantTimelineId};
30 :
31 : #[derive(Serialize, Default)]
32 : pub struct GcSummary {
33 : indices_deleted: usize,
34 : tenant_manifests_deleted: usize,
35 : remote_storage_errors: usize,
36 : controller_api_errors: usize,
37 : ancestor_layers_deleted: usize,
38 : }
39 :
40 : impl GcSummary {
41 0 : fn merge(&mut self, other: Self) {
42 0 : let Self {
43 0 : indices_deleted,
44 0 : tenant_manifests_deleted,
45 0 : remote_storage_errors,
46 0 : ancestor_layers_deleted,
47 0 : controller_api_errors,
48 0 : } = other;
49 0 :
50 0 : self.indices_deleted += indices_deleted;
51 0 : self.tenant_manifests_deleted += tenant_manifests_deleted;
52 0 : self.remote_storage_errors += remote_storage_errors;
53 0 : self.ancestor_layers_deleted += ancestor_layers_deleted;
54 0 : self.controller_api_errors += controller_api_errors;
55 0 : }
56 : }
57 :
58 : #[derive(clap::ValueEnum, Debug, Clone, Copy)]
59 : pub enum GcMode {
60 : // Delete nothing
61 : DryRun,
62 :
63 : // Enable only removing old-generation indices
64 : IndicesOnly,
65 :
66 : // Enable all forms of GC
67 : Full,
68 : }
69 :
70 : impl std::fmt::Display for GcMode {
71 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
72 0 : match self {
73 0 : GcMode::DryRun => write!(f, "dry-run"),
74 0 : GcMode::IndicesOnly => write!(f, "indices-only"),
75 0 : GcMode::Full => write!(f, "full"),
76 : }
77 0 : }
78 : }
79 :
80 : mod refs {
81 : use super::*;
82 : // Map of cross-shard layer references, giving a refcount for each layer in each shard that is referenced by some other
83 : // shard in the same tenant. This is sparse! The vast majority of timelines will have no cross-shard refs, and those that
84 : // do have cross shard refs should eventually drop most of them via compaction.
85 : //
86 : // In our inner map type, the TTID in the key is shard-agnostic, and the ShardIndex in the value refers to the _ancestor
87 : // which is is referenced_.
88 : #[derive(Default)]
89 : pub(super) struct AncestorRefs(
90 : BTreeMap<TenantTimelineId, HashMap<(ShardIndex, LayerName), usize>>,
91 : );
92 :
93 : impl AncestorRefs {
94 : /// Insert references for layers discovered in a particular shard-timeline that refer to an ancestral shard-timeline.
95 0 : pub(super) fn update(
96 0 : &mut self,
97 0 : ttid: TenantShardTimelineId,
98 0 : layers: Vec<(LayerName, LayerFileMetadata)>,
99 0 : ) {
100 0 : let ttid_refs = self.0.entry(ttid.as_tenant_timeline_id()).or_default();
101 0 : for (layer_name, layer_metadata) in layers {
102 0 : // Increment refcount of this layer in the ancestor shard
103 0 : *(ttid_refs
104 0 : .entry((layer_metadata.shard, layer_name))
105 0 : .or_default()) += 1;
106 0 : }
107 0 : }
108 :
109 : /// For a particular TTID, return the map of all ancestor layers referenced by a descendent to their refcount
110 : ///
111 : /// The `ShardIndex` in the result's key is the index of the _ancestor_, not the descendent.
112 0 : pub(super) fn get_ttid_refcounts(
113 0 : &self,
114 0 : ttid: &TenantTimelineId,
115 0 : ) -> Option<&HashMap<(ShardIndex, LayerName), usize>> {
116 0 : self.0.get(ttid)
117 0 : }
118 : }
119 : }
120 :
121 : use refs::AncestorRefs;
122 :
123 : // As we see shards for a tenant, acccumulate knowledge needed for cross-shard GC:
124 : // - Are there any ancestor shards?
125 : // - Are there any refs to ancestor shards' layers?
126 : #[derive(Default)]
127 : struct TenantRefAccumulator {
128 : shards_seen: HashMap<TenantId, BTreeSet<ShardIndex>>,
129 :
130 : // For each shard that has refs to an ancestor's layers, the set of ancestor layers referred to
131 : ancestor_ref_shards: AncestorRefs,
132 : }
133 :
134 : impl TenantRefAccumulator {
135 0 : fn update(&mut self, ttid: TenantShardTimelineId, index_part: &IndexPart) {
136 0 : let this_shard_idx = ttid.tenant_shard_id.to_index();
137 0 : (*self
138 0 : .shards_seen
139 0 : .entry(ttid.tenant_shard_id.tenant_id)
140 0 : .or_default())
141 0 : .insert(this_shard_idx);
142 0 :
143 0 : let mut ancestor_refs = Vec::new();
144 0 : for (layer_name, layer_metadata) in &index_part.layer_metadata {
145 0 : if layer_metadata.shard != this_shard_idx {
146 0 : // This is a reference from this shard to a layer in an ancestor shard: we must track this
147 0 : // as a marker to not GC this layer from the parent.
148 0 : ancestor_refs.push((layer_name.clone(), layer_metadata.clone()));
149 0 : }
150 : }
151 :
152 0 : if !ancestor_refs.is_empty() {
153 0 : tracing::info!(%ttid, "Found {} ancestor refs", ancestor_refs.len());
154 0 : self.ancestor_ref_shards.update(ttid, ancestor_refs);
155 0 : }
156 0 : }
157 :
158 : /// Consume Self and return a vector of ancestor tenant shards that should be GC'd, and map of referenced ancestor layers to preserve
159 0 : async fn into_gc_ancestors(
160 0 : self,
161 0 : controller_client: &control_api::Client,
162 0 : summary: &mut GcSummary,
163 0 : ) -> (Vec<TenantShardId>, AncestorRefs) {
164 0 : let mut ancestors_to_gc = Vec::new();
165 0 : for (tenant_id, shard_indices) in self.shards_seen {
166 : // Find the highest shard count
167 0 : let latest_count = shard_indices
168 0 : .iter()
169 0 : .map(|i| i.shard_count)
170 0 : .max()
171 0 : .expect("Always at least one shard");
172 0 :
173 0 : let mut shard_indices = shard_indices.iter().collect::<Vec<_>>();
174 0 : let (mut latest_shards, ancestor_shards) = {
175 0 : let at =
176 0 : itertools::partition(&mut shard_indices, |i| i.shard_count == latest_count);
177 0 : (shard_indices[0..at].to_owned(), &shard_indices[at..])
178 0 : };
179 0 : // Sort shards, as we will later compare them with a sorted list from the controller
180 0 : latest_shards.sort();
181 0 :
182 0 : // Check that we have a complete view of the latest shard count: this should always be the case unless we happened
183 0 : // to scan the S3 bucket halfway through a shard split.
184 0 : if latest_shards.len() != latest_count.count() as usize {
185 : // This should be extremely rare, so we warn on it.
186 0 : tracing::warn!(%tenant_id, "Missed some shards at count {:?}: {latest_shards:?}", latest_count);
187 0 : continue;
188 0 : }
189 0 :
190 0 : // Check if we have any non-latest-count shards
191 0 : if ancestor_shards.is_empty() {
192 0 : tracing::debug!(%tenant_id, "No ancestor shards to clean up");
193 0 : continue;
194 0 : }
195 0 :
196 0 : // Based on S3 view, this tenant looks like it might have some ancestor shard work to do. We
197 0 : // must only do this work if the tenant is not currently being split: otherwise, it is not safe
198 0 : // to GC ancestors, because if the split fails then the controller will try to attach ancestor
199 0 : // shards again.
200 0 : match controller_client
201 0 : .dispatch::<(), TenantDescribeResponse>(
202 0 : Method::GET,
203 0 : format!("control/v1/tenant/{tenant_id}"),
204 0 : None,
205 0 : )
206 0 : .await
207 : {
208 0 : Err(e) => {
209 0 : // We were not able to learn the latest shard split state from the controller, so we will not
210 0 : // do ancestor GC on this tenant.
211 0 : tracing::warn!(%tenant_id, "Failed to query storage controller, will not do ancestor GC: {e}");
212 0 : summary.controller_api_errors += 1;
213 0 : continue;
214 : }
215 0 : Ok(desc) => {
216 0 : // We expect to see that the latest shard count matches the one we saw in S3, and that none
217 0 : // of the shards indicate splitting in progress.
218 0 :
219 0 : let controller_indices: Vec<ShardIndex> = desc
220 0 : .shards
221 0 : .iter()
222 0 : .map(|s| s.tenant_shard_id.to_index())
223 0 : .collect();
224 0 : if !controller_indices.iter().eq(latest_shards.iter().copied()) {
225 0 : tracing::info!(%tenant_id, "Latest shards seen in S3 ({latest_shards:?}) don't match controller state ({controller_indices:?})");
226 0 : continue;
227 0 : }
228 0 :
229 0 : if desc.shards.iter().any(|s| s.is_splitting) {
230 0 : tracing::info!(%tenant_id, "One or more shards is currently splitting");
231 0 : continue;
232 0 : }
233 0 :
234 0 : // This shouldn't be too noisy, because we only log this for tenants that have some ancestral refs.
235 0 : tracing::info!(%tenant_id, "Validated state with controller: {desc:?}");
236 : }
237 : }
238 :
239 : // GC ancestor shards
240 0 : for ancestor_shard in ancestor_shards.iter().map(|idx| TenantShardId {
241 0 : tenant_id,
242 0 : shard_count: idx.shard_count,
243 0 : shard_number: idx.shard_number,
244 0 : }) {
245 0 : ancestors_to_gc.push(ancestor_shard);
246 0 : }
247 : }
248 :
249 0 : (ancestors_to_gc, self.ancestor_ref_shards)
250 0 : }
251 : }
252 :
253 0 : fn is_old_enough(min_age: &Duration, key: &ListingObject, summary: &mut GcSummary) -> bool {
254 : // Validation: we will only GC indices & layers after a time threshold (e.g. one week) so that during an incident
255 : // it is easier to read old data for analysis, and easier to roll back shard splits without having to un-delete any objects.
256 0 : let age = match key.last_modified.elapsed() {
257 0 : Ok(e) => e,
258 : Err(_) => {
259 0 : tracing::warn!("Bad last_modified time: {:?}", key.last_modified);
260 0 : summary.remote_storage_errors += 1;
261 0 : return false;
262 : }
263 : };
264 0 : let old_enough = &age > min_age;
265 0 :
266 0 : if !old_enough {
267 0 : tracing::info!(
268 0 : "Skipping young object {} < {}",
269 0 : humantime::format_duration(age),
270 0 : humantime::format_duration(*min_age)
271 : );
272 0 : }
273 :
274 0 : old_enough
275 0 : }
276 :
277 : /// Same as [`is_old_enough`], but doesn't require a [`ListingObject`] passed to it.
278 0 : async fn check_is_old_enough(
279 0 : remote_client: &GenericRemoteStorage,
280 0 : key: &RemotePath,
281 0 : min_age: &Duration,
282 0 : summary: &mut GcSummary,
283 0 : ) -> Option<bool> {
284 0 : let listing_object = remote_client
285 0 : .head_object(key, &CancellationToken::new())
286 0 : .await
287 0 : .ok()?;
288 0 : Some(is_old_enough(min_age, &listing_object, summary))
289 0 : }
290 :
291 0 : async fn maybe_delete_index(
292 0 : remote_client: &GenericRemoteStorage,
293 0 : min_age: &Duration,
294 0 : latest_gen: Generation,
295 0 : obj: &ListingObject,
296 0 : mode: GcMode,
297 0 : summary: &mut GcSummary,
298 0 : ) {
299 0 : // Validation: we will only delete things that parse cleanly
300 0 : let basename = obj.key.get_path().file_name().unwrap();
301 0 : let candidate_generation =
302 0 : match parse_remote_index_path(RemotePath::from_string(basename).unwrap()) {
303 0 : Some(g) => g,
304 : None => {
305 0 : if basename == IndexPart::FILE_NAME {
306 : // A legacy pre-generation index
307 0 : Generation::none()
308 : } else {
309 : // A strange key: we will not delete this because we don't understand it.
310 0 : tracing::warn!("Bad index key");
311 0 : return;
312 : }
313 : }
314 : };
315 :
316 : // Validation: we will only delete indices more than one generation old, to avoid interfering
317 : // in typical migrations, even if they are very long running.
318 0 : if candidate_generation >= latest_gen {
319 : // This shouldn't happen: when we loaded metadata, it should have selected the latest
320 : // generation already, and only populated [`S3TimelineBlobData::unused_index_keys`]
321 : // with older generations.
322 0 : tracing::warn!("Deletion candidate is >= latest generation, this is a bug!");
323 0 : return;
324 0 : } else if candidate_generation.next() == latest_gen {
325 : // Skip deleting the latest-1th generation's index.
326 0 : return;
327 0 : }
328 0 :
329 0 : if !is_old_enough(min_age, obj, summary) {
330 0 : return;
331 0 : }
332 :
333 0 : if matches!(mode, GcMode::DryRun) {
334 0 : tracing::info!("Dry run: would delete this key");
335 0 : return;
336 0 : }
337 0 :
338 0 : // All validations passed: erase the object
339 0 : let cancel = CancellationToken::new();
340 0 : match backoff::retry(
341 0 : || remote_client.delete(&obj.key, &cancel),
342 0 : |_| false,
343 0 : 3,
344 0 : MAX_RETRIES as u32,
345 0 : "maybe_delete_index",
346 0 : &cancel,
347 0 : )
348 0 : .await
349 : {
350 : None => {
351 0 : unreachable!("Using a dummy cancellation token");
352 : }
353 : Some(Ok(_)) => {
354 0 : tracing::info!("Successfully deleted index");
355 0 : summary.indices_deleted += 1;
356 : }
357 0 : Some(Err(e)) => {
358 0 : tracing::warn!("Failed to delete index: {e}");
359 0 : summary.remote_storage_errors += 1;
360 : }
361 : }
362 0 : }
363 :
364 0 : async fn maybe_delete_tenant_manifest(
365 0 : remote_client: &GenericRemoteStorage,
366 0 : min_age: &Duration,
367 0 : latest_gen: Generation,
368 0 : obj: &ListingObject,
369 0 : mode: GcMode,
370 0 : summary: &mut GcSummary,
371 0 : ) {
372 0 : // Validation: we will only delete things that parse cleanly
373 0 : let basename = obj.key.get_path().file_name().unwrap();
374 0 : let Some(candidate_generation) =
375 0 : parse_remote_tenant_manifest_path(RemotePath::from_string(basename).unwrap())
376 : else {
377 : // A strange key: we will not delete this because we don't understand it.
378 0 : tracing::warn!("Bad index key");
379 0 : return;
380 : };
381 :
382 : // Validation: we will only delete manifests more than one generation old, and in fact we
383 : // should never be called with such recent generations.
384 0 : if candidate_generation >= latest_gen {
385 0 : tracing::warn!("Deletion candidate is >= latest generation, this is a bug!");
386 0 : return;
387 0 : } else if candidate_generation.next() == latest_gen {
388 0 : tracing::warn!("Deletion candidate is >= latest generation - 1, this is a bug!");
389 0 : return;
390 0 : }
391 0 :
392 0 : if !is_old_enough(min_age, obj, summary) {
393 0 : return;
394 0 : }
395 :
396 0 : if matches!(mode, GcMode::DryRun) {
397 0 : tracing::info!("Dry run: would delete this key");
398 0 : return;
399 0 : }
400 0 :
401 0 : // All validations passed: erase the object
402 0 : let cancel = CancellationToken::new();
403 0 : match backoff::retry(
404 0 : || remote_client.delete(&obj.key, &cancel),
405 0 : |_| false,
406 0 : 3,
407 0 : MAX_RETRIES as u32,
408 0 : "maybe_delete_tenant_manifest",
409 0 : &cancel,
410 0 : )
411 0 : .await
412 : {
413 : None => {
414 0 : unreachable!("Using a dummy cancellation token");
415 : }
416 : Some(Ok(_)) => {
417 0 : tracing::info!("Successfully deleted tenant manifest");
418 0 : summary.tenant_manifests_deleted += 1;
419 : }
420 0 : Some(Err(e)) => {
421 0 : tracing::warn!("Failed to delete tenant manifest: {e}");
422 0 : summary.remote_storage_errors += 1;
423 : }
424 : }
425 0 : }
426 :
427 : #[allow(clippy::too_many_arguments)]
428 0 : async fn gc_ancestor(
429 0 : remote_client: &GenericRemoteStorage,
430 0 : root_target: &RootTarget,
431 0 : min_age: &Duration,
432 0 : ancestor: TenantShardId,
433 0 : refs: &AncestorRefs,
434 0 : mode: GcMode,
435 0 : summary: &mut GcSummary,
436 0 : ) -> anyhow::Result<()> {
437 : // Scan timelines in the ancestor
438 0 : let timelines = stream_tenant_timelines(remote_client, root_target, ancestor).await?;
439 0 : let mut timelines = std::pin::pin!(timelines);
440 :
441 : // Build a list of keys to retain
442 :
443 0 : while let Some(ttid) = timelines.next().await {
444 0 : let ttid = ttid?;
445 :
446 0 : let data = list_timeline_blobs(remote_client, ttid, root_target).await?;
447 :
448 0 : let s3_layers = match data.blob_data {
449 : BlobDataParseResult::Parsed {
450 : index_part: _,
451 : index_part_generation: _,
452 0 : s3_layers,
453 0 : } => s3_layers,
454 : BlobDataParseResult::Relic => {
455 : // Post-deletion tenant location: don't try and GC it.
456 0 : continue;
457 : }
458 : BlobDataParseResult::Incorrect {
459 0 : errors,
460 0 : s3_layers: _, // TODO(yuchen): could still check references to these s3 layers?
461 0 : } => {
462 0 : // Our primary purpose isn't to report on bad data, but log this rather than skipping silently
463 0 : tracing::warn!(
464 0 : "Skipping ancestor GC for timeline {ttid}, bad metadata: {errors:?}"
465 : );
466 0 : continue;
467 : }
468 : };
469 :
470 0 : let ttid_refs = refs.get_ttid_refcounts(&ttid.as_tenant_timeline_id());
471 0 : let ancestor_shard_index = ttid.tenant_shard_id.to_index();
472 :
473 0 : for (layer_name, layer_gen) in s3_layers {
474 0 : let ref_count = ttid_refs
475 0 : .and_then(|m| m.get(&(ancestor_shard_index, layer_name.clone())))
476 0 : .copied()
477 0 : .unwrap_or(0);
478 0 :
479 0 : if ref_count > 0 {
480 0 : tracing::debug!(%ttid, "Ancestor layer {layer_name} has {ref_count} refs");
481 0 : continue;
482 0 : }
483 0 :
484 0 : tracing::info!(%ttid, "Ancestor layer {layer_name} is not referenced");
485 :
486 : // Build the key for the layer we are considering deleting
487 0 : let key = root_target.absolute_key(&remote_layer_path(
488 0 : &ttid.tenant_shard_id.tenant_id,
489 0 : &ttid.timeline_id,
490 0 : ancestor_shard_index,
491 0 : &layer_name,
492 0 : layer_gen,
493 0 : ));
494 0 :
495 0 : // We apply a time threshold to GCing objects that are un-referenced: this preserves our ability
496 0 : // to roll back a shard split if we have to, by avoiding deleting ancestor layers right away
497 0 : let path = RemotePath::from_string(key.strip_prefix("/").unwrap_or(&key)).unwrap();
498 0 : if check_is_old_enough(remote_client, &path, min_age, summary).await != Some(true) {
499 0 : continue;
500 0 : }
501 :
502 0 : if !matches!(mode, GcMode::Full) {
503 0 : tracing::info!("Dry run: would delete key {key}");
504 0 : continue;
505 0 : }
506 0 :
507 0 : // All validations passed: erase the object
508 0 : match remote_client.delete(&path, &CancellationToken::new()).await {
509 : Ok(_) => {
510 0 : tracing::info!("Successfully deleted unreferenced ancestor layer {key}");
511 0 : summary.ancestor_layers_deleted += 1;
512 : }
513 0 : Err(e) => {
514 0 : tracing::warn!("Failed to delete layer {key}: {e}");
515 0 : summary.remote_storage_errors += 1;
516 : }
517 : }
518 : }
519 :
520 : // TODO: if all the layers are gone, clean up the whole timeline dir (remove index)
521 : }
522 :
523 0 : Ok(())
524 0 : }
525 :
526 0 : async fn gc_tenant_manifests(
527 0 : remote_client: &GenericRemoteStorage,
528 0 : min_age: Duration,
529 0 : target: &RootTarget,
530 0 : mode: GcMode,
531 0 : tenant_shard_id: TenantShardId,
532 0 : ) -> anyhow::Result<(GcSummary, Option<RemoteTenantManifestInfo>)> {
533 0 : let mut gc_summary = GcSummary::default();
534 0 : match list_tenant_manifests(remote_client, tenant_shard_id, target).await? {
535 : ListTenantManifestResult::WithErrors {
536 0 : errors,
537 : unknown_keys: _,
538 : } => {
539 0 : for (_key, error) in errors {
540 0 : tracing::warn!(%tenant_shard_id, "list_tenant_manifests: {error}");
541 : }
542 0 : Ok((gc_summary, None))
543 : }
544 : ListTenantManifestResult::NoErrors {
545 0 : latest_generation,
546 0 : mut manifests,
547 : } => {
548 0 : let Some(latest_generation) = latest_generation else {
549 0 : return Ok((gc_summary, None));
550 : };
551 0 : manifests.sort_by_key(|(generation, _obj)| *generation);
552 0 : // skip the two latest generations (they don't neccessarily have to be 1 apart from each other)
553 0 : let candidates = manifests.iter().rev().skip(2);
554 0 : for (_generation, key) in candidates {
555 0 : maybe_delete_tenant_manifest(
556 0 : remote_client,
557 0 : &min_age,
558 0 : latest_generation.generation,
559 0 : key,
560 0 : mode,
561 0 : &mut gc_summary,
562 0 : )
563 0 : .instrument(
564 0 : info_span!("maybe_delete_tenant_manifest", %tenant_shard_id, ?latest_generation.generation, %key.key),
565 : )
566 0 : .await;
567 : }
568 0 : Ok((gc_summary, Some(latest_generation)))
569 : }
570 : }
571 0 : }
572 :
573 0 : async fn gc_timeline(
574 0 : remote_client: &GenericRemoteStorage,
575 0 : min_age: &Duration,
576 0 : target: &RootTarget,
577 0 : mode: GcMode,
578 0 : ttid: TenantShardTimelineId,
579 0 : accumulator: &Arc<std::sync::Mutex<TenantRefAccumulator>>,
580 0 : tenant_manifest_info: Arc<Option<RemoteTenantManifestInfo>>,
581 0 : ) -> anyhow::Result<GcSummary> {
582 0 : let mut summary = GcSummary::default();
583 0 : let data = list_timeline_blobs(remote_client, ttid, target).await?;
584 :
585 0 : let (index_part, latest_gen, candidates) = match &data.blob_data {
586 : BlobDataParseResult::Parsed {
587 0 : index_part,
588 0 : index_part_generation,
589 0 : s3_layers: _s3_layers,
590 0 : } => (index_part, *index_part_generation, data.unused_index_keys),
591 : BlobDataParseResult::Relic => {
592 : // Post-deletion tenant location: don't try and GC it.
593 0 : return Ok(summary);
594 : }
595 : BlobDataParseResult::Incorrect {
596 0 : errors,
597 0 : s3_layers: _,
598 0 : } => {
599 0 : // Our primary purpose isn't to report on bad data, but log this rather than skipping silently
600 0 : tracing::warn!("Skipping timeline {ttid}, bad metadata: {errors:?}");
601 0 : return Ok(summary);
602 : }
603 : };
604 :
605 0 : if let Some(tenant_manifest_info) = &*tenant_manifest_info {
606 : // TODO: this is O(n^2) in the number of offloaded timelines. Do a hashmap lookup instead.
607 0 : let maybe_offloaded = tenant_manifest_info
608 0 : .manifest
609 0 : .offloaded_timelines
610 0 : .iter()
611 0 : .find(|offloaded_timeline| offloaded_timeline.timeline_id == ttid.timeline_id);
612 0 : if let Some(offloaded) = maybe_offloaded {
613 0 : let warnings = validate_index_part_with_offloaded(index_part, offloaded);
614 0 : let warn = if warnings.is_empty() {
615 0 : false
616 : } else {
617 : // Verify that the manifest hasn't changed. If it has, a potential racing change could have been cause for our troubles.
618 0 : match list_tenant_manifests(remote_client, ttid.tenant_shard_id, target).await? {
619 : ListTenantManifestResult::WithErrors {
620 0 : errors,
621 : unknown_keys: _,
622 : } => {
623 0 : for (_key, error) in errors {
624 0 : tracing::warn!(%ttid, "list_tenant_manifests in gc_timeline: {error}");
625 : }
626 0 : true
627 : }
628 : ListTenantManifestResult::NoErrors {
629 0 : latest_generation,
630 : manifests: _,
631 : } => {
632 0 : if let Some(new_latest_gen) = latest_generation {
633 0 : let manifest_changed = (
634 0 : new_latest_gen.generation,
635 0 : new_latest_gen.listing_object.last_modified,
636 0 : ) == (
637 0 : tenant_manifest_info.generation,
638 0 : tenant_manifest_info.listing_object.last_modified,
639 0 : );
640 0 : if manifest_changed {
641 0 : tracing::debug!(%ttid, "tenant manifest changed since it was loaded, suppressing {} warnings", warnings.len());
642 0 : }
643 0 : manifest_changed
644 : } else {
645 : // The latest generation is gone. This timeline is in the progress of being deleted?
646 0 : false
647 : }
648 : }
649 : }
650 : };
651 0 : if warn {
652 0 : for warning in warnings {
653 0 : tracing::warn!(%ttid, "{}", warning);
654 : }
655 0 : }
656 0 : }
657 0 : }
658 :
659 0 : accumulator.lock().unwrap().update(ttid, index_part);
660 :
661 0 : for key in candidates {
662 0 : maybe_delete_index(remote_client, min_age, latest_gen, &key, mode, &mut summary)
663 0 : .instrument(info_span!("maybe_delete_index", %ttid, ?latest_gen, %key.key))
664 0 : .await;
665 : }
666 :
667 0 : Ok(summary)
668 0 : }
669 :
670 0 : fn validate_index_part_with_offloaded(
671 0 : index_part: &IndexPart,
672 0 : offloaded: &OffloadedTimelineManifest,
673 0 : ) -> Vec<String> {
674 0 : let mut warnings = Vec::new();
675 0 : if let Some(archived_at_index_part) = index_part.archived_at {
676 0 : if archived_at_index_part
677 0 : .signed_duration_since(offloaded.archived_at)
678 0 : .num_seconds()
679 0 : != 0
680 0 : {
681 0 : warnings.push(format!(
682 0 : "index-part archived_at={} differs from manifest archived_at={}",
683 0 : archived_at_index_part, offloaded.archived_at
684 0 : ));
685 0 : }
686 0 : } else {
687 0 : warnings.push("Timeline offloaded in manifest but not archived in index-part".to_string());
688 0 : }
689 0 : if index_part.metadata.ancestor_timeline() != offloaded.ancestor_timeline_id {
690 0 : warnings.push(format!(
691 0 : "index-part anestor={:?} differs from manifest ancestor={:?}",
692 0 : index_part.metadata.ancestor_timeline(),
693 0 : offloaded.ancestor_timeline_id
694 0 : ));
695 0 : }
696 0 : warnings
697 0 : }
698 :
699 : /// Physical garbage collection: removing unused S3 objects.
700 : ///
701 : /// This is distinct from the garbage collection done inside the pageserver, which operates at a higher level
702 : /// (keys, layers). This type of garbage collection is about removing:
703 : /// - Objects that were uploaded but never referenced in the remote index (e.g. because of a shutdown between
704 : /// uploading a layer and uploading an index)
705 : /// - Index objects and tenant manifests from historic generations
706 : ///
707 : /// This type of GC is not necessary for correctness: rather it serves to reduce wasted storage capacity, and
708 : /// make sure that object listings don't get slowed down by large numbers of garbage objects.
709 0 : pub async fn pageserver_physical_gc(
710 0 : bucket_config: &BucketConfig,
711 0 : controller_client: Option<&control_api::Client>,
712 0 : tenant_shard_ids: Vec<TenantShardId>,
713 0 : min_age: Duration,
714 0 : mode: GcMode,
715 0 : ) -> anyhow::Result<GcSummary> {
716 0 : let (remote_client, target) = init_remote(bucket_config.clone(), NodeKind::Pageserver).await?;
717 :
718 0 : let remote_client = Arc::new(remote_client);
719 0 : let tenants = if tenant_shard_ids.is_empty() {
720 0 : futures::future::Either::Left(stream_tenants(&remote_client, &target))
721 : } else {
722 0 : futures::future::Either::Right(futures::stream::iter(tenant_shard_ids.into_iter().map(Ok)))
723 : };
724 :
725 : // How many tenants to process in parallel. We need to be mindful of pageservers
726 : // accessing the same per tenant prefixes, so use a lower setting than pageservers.
727 : const CONCURRENCY: usize = 32;
728 :
729 : // Accumulate information about each tenant for cross-shard GC step we'll do at the end
730 0 : let accumulator = Arc::new(std::sync::Mutex::new(TenantRefAccumulator::default()));
731 :
732 : // Generate a stream of TenantTimelineId
733 : enum GcSummaryOrContent<T> {
734 : Content(T),
735 : GcSummary(GcSummary),
736 : }
737 0 : let timelines = tenants.map_ok(|tenant_shard_id| {
738 0 : let target_ref = ⌖
739 0 : let remote_client_ref = &remote_client;
740 0 : async move {
741 0 : let gc_manifest_result = gc_tenant_manifests(
742 0 : remote_client_ref,
743 0 : min_age,
744 0 : target_ref,
745 0 : mode,
746 0 : tenant_shard_id,
747 0 : )
748 0 : .await;
749 0 : let (summary_from_manifest, tenant_manifest_opt) = match gc_manifest_result {
750 0 : Ok((gc_summary, tenant_manifest)) => (gc_summary, tenant_manifest),
751 0 : Err(e) => {
752 0 : tracing::warn!(%tenant_shard_id, "Error in gc_tenant_manifests: {e}");
753 0 : (GcSummary::default(), None)
754 : }
755 : };
756 0 : let tenant_manifest_arc = Arc::new(tenant_manifest_opt);
757 0 : let summary_from_manifest = Ok(GcSummaryOrContent::<(_, _)>::GcSummary(
758 0 : summary_from_manifest,
759 0 : ));
760 0 : stream_tenant_timelines(remote_client_ref, target_ref, tenant_shard_id)
761 0 : .await
762 0 : .map(|stream| {
763 0 : stream
764 0 : .zip(futures::stream::iter(std::iter::repeat(
765 0 : tenant_manifest_arc,
766 0 : )))
767 0 : .map(|(ttid_res, tenant_manifest_arc)| {
768 0 : ttid_res.map(move |ttid| {
769 0 : GcSummaryOrContent::Content((ttid, tenant_manifest_arc))
770 0 : })
771 0 : })
772 0 : .chain(futures::stream::iter([summary_from_manifest].into_iter()))
773 0 : })
774 0 : }
775 0 : });
776 0 : let timelines = std::pin::pin!(timelines.try_buffered(CONCURRENCY));
777 0 : let timelines = timelines.try_flatten();
778 0 :
779 0 : let mut summary = GcSummary::default();
780 0 :
781 0 : // Drain futures for per-shard GC, populating accumulator as a side effect
782 0 : {
783 0 : let timelines = timelines.map_ok(|summary_or_ttid| match summary_or_ttid {
784 0 : GcSummaryOrContent::Content((ttid, tenant_manifest_arc)) => {
785 0 : futures::future::Either::Left(gc_timeline(
786 0 : &remote_client,
787 0 : &min_age,
788 0 : &target,
789 0 : mode,
790 0 : ttid,
791 0 : &accumulator,
792 0 : tenant_manifest_arc,
793 0 : ))
794 : }
795 0 : GcSummaryOrContent::GcSummary(gc_summary) => {
796 0 : futures::future::Either::Right(futures::future::ok(gc_summary))
797 : }
798 0 : });
799 0 : let mut timelines = std::pin::pin!(timelines.try_buffered(CONCURRENCY));
800 :
801 0 : while let Some(i) = timelines.next().await {
802 0 : summary.merge(i?);
803 : }
804 : }
805 :
806 : // Execute cross-shard GC, using the accumulator's full view of all the shards built in the per-shard GC
807 0 : let Some(client) = controller_client else {
808 0 : tracing::info!("Skipping ancestor layer GC, because no `--controller-api` was specified");
809 0 : return Ok(summary);
810 : };
811 :
812 0 : let (ancestor_shards, ancestor_refs) = Arc::into_inner(accumulator)
813 0 : .unwrap()
814 0 : .into_inner()
815 0 : .unwrap()
816 0 : .into_gc_ancestors(client, &mut summary)
817 0 : .await;
818 :
819 0 : for ancestor_shard in ancestor_shards {
820 0 : gc_ancestor(
821 0 : &remote_client,
822 0 : &target,
823 0 : &min_age,
824 0 : ancestor_shard,
825 0 : &ancestor_refs,
826 0 : mode,
827 0 : &mut summary,
828 0 : )
829 0 : .instrument(info_span!("gc_ancestor", %ancestor_shard))
830 0 : .await?;
831 : }
832 :
833 0 : Ok(summary)
834 0 : }
|
