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