Line data Source code
1 : mod deleter;
2 : mod list_writer;
3 : mod validator;
4 :
5 : use std::collections::HashMap;
6 : use std::sync::Arc;
7 : use std::time::Duration;
8 :
9 : use anyhow::Context;
10 : use camino::Utf8PathBuf;
11 : use deleter::DeleterMessage;
12 : use list_writer::ListWriterQueueMessage;
13 : use pageserver_api::shard::TenantShardId;
14 : use remote_storage::{GenericRemoteStorage, RemotePath};
15 : use serde::{Deserialize, Serialize};
16 : use thiserror::Error;
17 : use tokio_util::sync::CancellationToken;
18 : use tracing::{Instrument, debug, error};
19 : use utils::crashsafe::path_with_suffix_extension;
20 : use utils::generation::Generation;
21 : use utils::id::TimelineId;
22 : use utils::lsn::{AtomicLsn, Lsn};
23 : use validator::ValidatorQueueMessage;
24 :
25 : use self::deleter::Deleter;
26 : use self::list_writer::{DeletionOp, ListWriter, RecoverOp};
27 : use self::validator::Validator;
28 : use crate::config::PageServerConf;
29 : use crate::controller_upcall_client::StorageControllerUpcallApi;
30 : use crate::metrics;
31 : use crate::tenant::remote_timeline_client::{LayerFileMetadata, remote_timeline_path};
32 : use crate::tenant::storage_layer::LayerName;
33 : use crate::virtual_file::{MaybeFatalIo, VirtualFile};
34 :
35 : // TODO: configurable for how long to wait before executing deletions
36 :
37 : /// We aggregate object deletions from many tenants in one place, for several reasons:
38 : /// - Coalesce deletions into fewer DeleteObjects calls
39 : /// - Enable Tenant/Timeline lifetimes to be shorter than the time it takes
40 : /// to flush any outstanding deletions.
41 : /// - Globally control throughput of deletions, as these are a low priority task: do
42 : /// not compete with the same S3 clients/connections used for higher priority uploads.
43 : /// - Enable gating deletions on validation of a tenant's generation number, to make
44 : /// it safe to multi-attach tenants (see docs/rfcs/025-generation-numbers.md)
45 : ///
46 : /// There are two kinds of deletion: deferred and immediate. A deferred deletion
47 : /// may be intentionally delayed to protect passive readers of S3 data, and is
48 : /// subject to a generation number validation step. An immediate deletion is
49 : /// ready to execute immediately, and is only queued up so that it can be coalesced
50 : /// with other deletions in flight.
51 : ///
52 : /// Deferred deletions pass through three steps:
53 : /// - ListWriter: accumulate deletion requests from Timelines, and batch them up into
54 : /// DeletionLists, which are persisted to disk.
55 : /// - Validator: accumulate deletion lists, and validate them en-masse prior to passing
56 : /// the keys in the list onward for actual deletion. Also validate remote_consistent_lsn
57 : /// updates for running timelines.
58 : /// - Deleter: accumulate object keys that the validator has validated, and execute them in
59 : /// batches of 1000 keys via DeleteObjects.
60 : ///
61 : /// Non-deferred deletions, such as during timeline deletion, bypass the first
62 : /// two stages and are passed straight into the Deleter.
63 : ///
64 : /// Internally, each stage is joined by a channel to the next. On disk, there is only
65 : /// one queue (of DeletionLists), which is written by the frontend and consumed
66 : /// by the backend.
67 : #[derive(Clone)]
68 : pub struct DeletionQueue {
69 : client: DeletionQueueClient,
70 :
71 : // Parent cancellation token for the tokens passed into background workers
72 : cancel: CancellationToken,
73 : }
74 :
75 : /// Opaque wrapper around individual worker tasks, to avoid making the
76 : /// worker objects themselves public
77 : pub struct DeletionQueueWorkers<C>
78 : where
79 : C: StorageControllerUpcallApi + Send + Sync,
80 : {
81 : frontend: ListWriter,
82 : backend: Validator<C>,
83 : executor: Deleter,
84 : }
85 :
86 : impl<C> DeletionQueueWorkers<C>
87 : where
88 : C: StorageControllerUpcallApi + Send + Sync + 'static,
89 : {
90 4 : pub fn spawn_with(mut self, runtime: &tokio::runtime::Handle) -> tokio::task::JoinHandle<()> {
91 4 : let jh_frontend = runtime.spawn(async move {
92 4 : self.frontend
93 4 : .background()
94 4 : .instrument(tracing::info_span!(parent:None, "deletion frontend"))
95 4 : .await
96 1 : });
97 4 : let jh_backend = runtime.spawn(async move {
98 4 : self.backend
99 4 : .background()
100 4 : .instrument(tracing::info_span!(parent:None, "deletion backend"))
101 4 : .await
102 1 : });
103 4 : let jh_executor = runtime.spawn(async move {
104 4 : self.executor
105 4 : .background()
106 4 : .instrument(tracing::info_span!(parent:None, "deletion executor"))
107 4 : .await
108 1 : });
109 :
110 4 : runtime.spawn({
111 4 : async move {
112 4 : jh_frontend.await.expect("error joining frontend worker");
113 1 : jh_backend.await.expect("error joining backend worker");
114 1 : drop(jh_executor.await.expect("error joining executor worker"));
115 1 : }
116 : })
117 4 : }
118 : }
119 :
120 : /// A FlushOp is just a oneshot channel, where we send the transmit side down
121 : /// another channel, and the receive side will receive a message when the channel
122 : /// we're flushing has reached the FlushOp we sent into it.
123 : ///
124 : /// The only extra behavior beyond the channel is that the notify() method does not
125 : /// return an error when the receive side has been dropped, because in this use case
126 : /// it is harmless (the code that initiated the flush no longer cares about the result).
127 : #[derive(Debug)]
128 : struct FlushOp {
129 : tx: tokio::sync::oneshot::Sender<()>,
130 : }
131 :
132 : impl FlushOp {
133 21 : fn new() -> (Self, tokio::sync::oneshot::Receiver<()>) {
134 21 : let (tx, rx) = tokio::sync::oneshot::channel::<()>();
135 21 : (Self { tx }, rx)
136 21 : }
137 :
138 22 : fn notify(self) {
139 22 : if self.tx.send(()).is_err() {
140 : // oneshot channel closed. This is legal: a client could be destroyed while waiting for a flush.
141 0 : debug!("deletion queue flush from dropped client");
142 22 : };
143 22 : }
144 : }
145 :
146 : #[derive(Clone, Debug)]
147 : pub struct DeletionQueueClient {
148 : tx: tokio::sync::mpsc::UnboundedSender<ListWriterQueueMessage>,
149 : executor_tx: tokio::sync::mpsc::Sender<DeleterMessage>,
150 :
151 : lsn_table: Arc<std::sync::RwLock<VisibleLsnUpdates>>,
152 : }
153 :
154 0 : #[derive(Debug, Serialize, Deserialize, PartialEq, Eq)]
155 : struct TenantDeletionList {
156 : /// For each Timeline, a list of key fragments to append to the timeline remote path
157 : /// when reconstructing a full key
158 : timelines: HashMap<TimelineId, Vec<String>>,
159 :
160 : /// The generation in which this deletion was emitted: note that this may not be the
161 : /// same as the generation of any layers being deleted. The generation of the layer
162 : /// has already been absorbed into the keys in `objects`
163 : generation: Generation,
164 : }
165 :
166 : impl TenantDeletionList {
167 5 : pub(crate) fn len(&self) -> usize {
168 5 : self.timelines.values().map(|v| v.len()).sum()
169 5 : }
170 : }
171 :
172 : /// Files ending with this suffix will be ignored and erased
173 : /// during recovery as startup.
174 : const TEMP_SUFFIX: &str = "tmp";
175 :
176 0 : #[derive(Debug, Serialize, Deserialize, PartialEq, Eq)]
177 : struct DeletionList {
178 : /// Serialization version, for future use
179 : version: u8,
180 :
181 : /// Used for constructing a unique key for each deletion list we write out.
182 : sequence: u64,
183 :
184 : /// To avoid repeating tenant/timeline IDs in every key, we store keys in
185 : /// nested HashMaps by TenantTimelineID. Each Tenant only appears once
186 : /// with one unique generation ID: if someone tries to push a second generation
187 : /// ID for the same tenant, we will start a new DeletionList.
188 : tenants: HashMap<TenantShardId, TenantDeletionList>,
189 :
190 : /// Avoid having to walk `tenants` to calculate the number of keys in
191 : /// the nested deletion lists
192 : size: usize,
193 :
194 : /// Set to true when the list has undergone validation with the control
195 : /// plane and the remaining contents of `tenants` are valid. A list may
196 : /// also be implicitly marked valid by DeletionHeader.validated_sequence
197 : /// advancing to >= DeletionList.sequence
198 : #[serde(default)]
199 : #[serde(skip_serializing_if = "std::ops::Not::not")]
200 : validated: bool,
201 : }
202 :
203 0 : #[derive(Debug, Serialize, Deserialize)]
204 : struct DeletionHeader {
205 : /// Serialization version, for future use
206 : version: u8,
207 :
208 : /// The highest sequence number (inclusive) that has been validated. All deletion
209 : /// lists on disk with a sequence <= this value are safe to execute.
210 : validated_sequence: u64,
211 : }
212 :
213 : impl DeletionHeader {
214 : const VERSION_LATEST: u8 = 1;
215 :
216 4 : fn new(validated_sequence: u64) -> Self {
217 4 : Self {
218 4 : version: Self::VERSION_LATEST,
219 4 : validated_sequence,
220 4 : }
221 4 : }
222 :
223 4 : async fn save(&self, conf: &'static PageServerConf) -> anyhow::Result<()> {
224 4 : debug!("Saving deletion list header {:?}", self);
225 4 : let header_bytes = serde_json::to_vec(self).context("serialize deletion header")?;
226 4 : let header_path = conf.deletion_header_path();
227 4 : let temp_path = path_with_suffix_extension(&header_path, TEMP_SUFFIX);
228 4 : VirtualFile::crashsafe_overwrite(header_path, temp_path, header_bytes)
229 4 : .await
230 4 : .maybe_fatal_err("save deletion header")?;
231 :
232 4 : Ok(())
233 4 : }
234 : }
235 :
236 : impl DeletionList {
237 : const VERSION_LATEST: u8 = 1;
238 10 : fn new(sequence: u64) -> Self {
239 10 : Self {
240 10 : version: Self::VERSION_LATEST,
241 10 : sequence,
242 10 : tenants: HashMap::new(),
243 10 : size: 0,
244 10 : validated: false,
245 10 : }
246 10 : }
247 :
248 13 : fn is_empty(&self) -> bool {
249 13 : self.tenants.is_empty()
250 13 : }
251 :
252 30 : fn len(&self) -> usize {
253 30 : self.size
254 30 : }
255 :
256 : /// Returns true if the push was accepted, false if the caller must start a new
257 : /// deletion list.
258 7 : fn push(
259 7 : &mut self,
260 7 : tenant: &TenantShardId,
261 7 : timeline: &TimelineId,
262 7 : generation: Generation,
263 7 : objects: &mut Vec<RemotePath>,
264 7 : ) -> bool {
265 7 : if objects.is_empty() {
266 : // Avoid inserting an empty TimelineDeletionList: this preserves the property
267 : // that if we have no keys, then self.objects is empty (used in Self::is_empty)
268 0 : return true;
269 7 : }
270 :
271 7 : let tenant_entry = self
272 7 : .tenants
273 7 : .entry(*tenant)
274 7 : .or_insert_with(|| TenantDeletionList {
275 6 : timelines: HashMap::new(),
276 6 : generation,
277 6 : });
278 :
279 7 : if tenant_entry.generation != generation {
280 : // Only one generation per tenant per list: signal to
281 : // caller to start a new list.
282 1 : return false;
283 6 : }
284 :
285 6 : let timeline_entry = tenant_entry.timelines.entry(*timeline).or_default();
286 :
287 6 : let timeline_remote_path = remote_timeline_path(tenant, timeline);
288 :
289 6 : self.size += objects.len();
290 6 : timeline_entry.extend(objects.drain(..).map(|p| {
291 6 : p.strip_prefix(&timeline_remote_path)
292 6 : .expect("Timeline paths always start with the timeline prefix")
293 6 : .to_string()
294 6 : }));
295 6 : true
296 7 : }
297 :
298 5 : fn into_remote_paths(self) -> Vec<RemotePath> {
299 5 : let mut result = Vec::new();
300 5 : for (tenant, tenant_deletions) in self.tenants.into_iter() {
301 3 : for (timeline, timeline_layers) in tenant_deletions.timelines.into_iter() {
302 3 : let timeline_remote_path = remote_timeline_path(&tenant, &timeline);
303 3 : result.extend(
304 3 : timeline_layers
305 3 : .into_iter()
306 3 : .map(|l| timeline_remote_path.join(Utf8PathBuf::from(l))),
307 : );
308 : }
309 : }
310 :
311 5 : result
312 5 : }
313 :
314 7 : async fn save(&self, conf: &'static PageServerConf) -> anyhow::Result<()> {
315 7 : let path = conf.deletion_list_path(self.sequence);
316 7 : let temp_path = path_with_suffix_extension(&path, TEMP_SUFFIX);
317 :
318 7 : let bytes = serde_json::to_vec(self).expect("Failed to serialize deletion list");
319 :
320 7 : VirtualFile::crashsafe_overwrite(path, temp_path, bytes)
321 7 : .await
322 7 : .maybe_fatal_err("save deletion list")
323 7 : .map_err(Into::into)
324 7 : }
325 : }
326 :
327 : impl std::fmt::Display for DeletionList {
328 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
329 0 : write!(
330 0 : f,
331 0 : "DeletionList<seq={}, tenants={}, keys={}>",
332 : self.sequence,
333 0 : self.tenants.len(),
334 : self.size
335 : )
336 0 : }
337 : }
338 :
339 : struct PendingLsn {
340 : projected: Lsn,
341 : result_slot: Arc<AtomicLsn>,
342 : }
343 :
344 : struct TenantLsnState {
345 : timelines: HashMap<TimelineId, PendingLsn>,
346 :
347 : // In what generation was the most recent update proposed?
348 : generation: Generation,
349 : }
350 :
351 : #[derive(Default)]
352 : struct VisibleLsnUpdates {
353 : tenants: HashMap<TenantShardId, TenantLsnState>,
354 : }
355 :
356 : impl VisibleLsnUpdates {
357 123 : fn new() -> Self {
358 123 : Self {
359 123 : tenants: HashMap::new(),
360 123 : }
361 123 : }
362 : }
363 :
364 : impl std::fmt::Debug for VisibleLsnUpdates {
365 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
366 0 : write!(f, "VisibleLsnUpdates({} tenants)", self.tenants.len())
367 0 : }
368 : }
369 :
370 : #[derive(Error, Debug)]
371 : pub enum DeletionQueueError {
372 : #[error("Deletion queue unavailable during shutdown")]
373 : ShuttingDown,
374 : }
375 :
376 : impl DeletionQueueClient {
377 : /// This is cancel-safe. If you drop the future before it completes, the message
378 : /// is not pushed, although in the context of the deletion queue it doesn't matter: once
379 : /// we decide to do a deletion the decision is always final.
380 228 : fn do_push<T>(
381 228 : &self,
382 228 : queue: &tokio::sync::mpsc::UnboundedSender<T>,
383 228 : msg: T,
384 228 : ) -> Result<(), DeletionQueueError> {
385 228 : match queue.send(msg) {
386 228 : Ok(_) => Ok(()),
387 0 : Err(e) => {
388 : // This shouldn't happen, we should shut down all tenants before
389 : // we shut down the global delete queue. If we encounter a bug like this,
390 : // we may leak objects as deletions won't be processed.
391 0 : error!("Deletion queue closed while pushing, shutting down? ({e})");
392 0 : Err(DeletionQueueError::ShuttingDown)
393 : }
394 : }
395 228 : }
396 :
397 4 : pub(crate) fn recover(
398 4 : &self,
399 4 : attached_tenants: HashMap<TenantShardId, Generation>,
400 4 : ) -> Result<(), DeletionQueueError> {
401 4 : self.do_push(
402 4 : &self.tx,
403 4 : ListWriterQueueMessage::Recover(RecoverOp { attached_tenants }),
404 : )
405 4 : }
406 :
407 : /// When a Timeline wishes to update the remote_consistent_lsn that it exposes to the outside
408 : /// world, it must validate its generation number before doing so. Rather than do this synchronously,
409 : /// we allow the timeline to publish updates at will via this API, and then read back what LSN was most
410 : /// recently validated separately.
411 : ///
412 : /// In this function we publish the LSN to the `projected` field of the timeline's entry in the VisibleLsnUpdates. The
413 : /// backend will later wake up and notice that the tenant's generation requires validation.
414 770 : pub(crate) async fn update_remote_consistent_lsn(
415 770 : &self,
416 770 : tenant_shard_id: TenantShardId,
417 770 : timeline_id: TimelineId,
418 770 : current_generation: Generation,
419 770 : lsn: Lsn,
420 770 : result_slot: Arc<AtomicLsn>,
421 770 : ) {
422 770 : let mut locked = self
423 770 : .lsn_table
424 770 : .write()
425 770 : .expect("Lock should never be poisoned");
426 :
427 770 : let tenant_entry = locked
428 770 : .tenants
429 770 : .entry(tenant_shard_id)
430 770 : .or_insert(TenantLsnState {
431 770 : timelines: HashMap::new(),
432 770 : generation: current_generation,
433 770 : });
434 :
435 770 : if tenant_entry.generation != current_generation {
436 0 : // Generation might have changed if we were detached and then re-attached: in this case,
437 0 : // state from the previous generation cannot be trusted.
438 0 : tenant_entry.timelines.clear();
439 0 : tenant_entry.generation = current_generation;
440 770 : }
441 :
442 770 : tenant_entry.timelines.insert(
443 770 : timeline_id,
444 770 : PendingLsn {
445 770 : projected: lsn,
446 770 : result_slot,
447 770 : },
448 : );
449 770 : }
450 :
451 : /// Submit a list of layers for deletion: this function will return before the deletion is
452 : /// persistent, but it may be executed at any time after this function enters: do not push
453 : /// layers until you're sure they can be deleted safely (i.e. remote metadata no longer
454 : /// references them).
455 : ///
456 : /// The `current_generation` is the generation of this pageserver's current attachment. The
457 : /// generations in `layers` are the generations in which those layers were written.
458 212 : pub(crate) fn push_layers(
459 212 : &self,
460 212 : tenant_shard_id: TenantShardId,
461 212 : timeline_id: TimelineId,
462 212 : current_generation: Generation,
463 212 : layers: Vec<(LayerName, LayerFileMetadata)>,
464 212 : ) -> Result<(), DeletionQueueError> {
465 : // None generations are not valid for attached tenants: they must always be attached in
466 : // a known generation. None generations are still permitted for layers in the index because
467 : // they may be historical.
468 212 : assert!(!current_generation.is_none());
469 :
470 212 : metrics::DELETION_QUEUE
471 212 : .keys_submitted
472 212 : .inc_by(layers.len() as u64);
473 212 : self.do_push(
474 212 : &self.tx,
475 212 : ListWriterQueueMessage::Delete(DeletionOp {
476 212 : tenant_shard_id,
477 212 : timeline_id,
478 212 : layers,
479 212 : generation: current_generation,
480 212 : objects: Vec::new(),
481 212 : }),
482 : )
483 212 : }
484 :
485 : /// This is cancel-safe. If you drop the future the flush may still happen in the background.
486 12 : async fn do_flush<T>(
487 12 : &self,
488 12 : queue: &tokio::sync::mpsc::UnboundedSender<T>,
489 12 : msg: T,
490 12 : rx: tokio::sync::oneshot::Receiver<()>,
491 12 : ) -> Result<(), DeletionQueueError> {
492 12 : self.do_push(queue, msg)?;
493 12 : if rx.await.is_err() {
494 : // This shouldn't happen if tenants are shut down before deletion queue. If we
495 : // encounter a bug like this, then a flusher will incorrectly believe it has flushed
496 : // when it hasn't, possibly leading to leaking objects.
497 0 : error!("Deletion queue dropped flush op while client was still waiting");
498 0 : Err(DeletionQueueError::ShuttingDown)
499 : } else {
500 12 : Ok(())
501 : }
502 12 : }
503 :
504 : /// Wait until all previous deletions are persistent (either executed, or written to a DeletionList)
505 : ///
506 : /// This is cancel-safe. If you drop the future the flush may still happen in the background.
507 7 : pub async fn flush(&self) -> Result<(), DeletionQueueError> {
508 7 : let (flush_op, rx) = FlushOp::new();
509 7 : self.do_flush(&self.tx, ListWriterQueueMessage::Flush(flush_op), rx)
510 7 : .await
511 7 : }
512 :
513 : /// Issue a flush without waiting for it to complete. This is useful on advisory flushes where
514 : /// the caller wants to avoid the risk of waiting for lots of enqueued work, such as on tenant
515 : /// detach where flushing is nice but not necessary.
516 : ///
517 : /// This function provides no guarantees of work being done.
518 0 : pub fn flush_advisory(&self) {
519 0 : let (flush_op, _) = FlushOp::new();
520 :
521 : // Transmit the flush message, ignoring any result (such as a closed channel during shutdown).
522 0 : drop(self.tx.send(ListWriterQueueMessage::FlushExecute(flush_op)));
523 0 : }
524 :
525 : // Wait until all previous deletions are executed
526 5 : pub(crate) async fn flush_execute(&self) -> Result<(), DeletionQueueError> {
527 5 : debug!("flush_execute: flushing to deletion lists...");
528 : // Flush any buffered work to deletion lists
529 5 : self.flush().await?;
530 :
531 : // Flush the backend into the executor of deletion lists
532 5 : let (flush_op, rx) = FlushOp::new();
533 5 : debug!("flush_execute: flushing backend...");
534 5 : self.do_flush(&self.tx, ListWriterQueueMessage::FlushExecute(flush_op), rx)
535 5 : .await?;
536 5 : debug!("flush_execute: finished flushing backend...");
537 :
538 : // Flush any immediate-mode deletions (the above backend flush will only flush
539 : // the executor if deletions had flowed through the backend)
540 5 : debug!("flush_execute: flushing execution...");
541 5 : self.flush_immediate().await?;
542 5 : debug!("flush_execute: finished flushing execution...");
543 5 : Ok(())
544 5 : }
545 :
546 : /// This interface bypasses the persistent deletion queue, and any validation
547 : /// that this pageserver is still elegible to execute the deletions. It is for
548 : /// use in timeline deletions, where the control plane is telling us we may
549 : /// delete everything in the timeline.
550 : ///
551 : /// DO NOT USE THIS FROM GC OR COMPACTION CODE. Use the regular `push_layers`.
552 0 : pub(crate) async fn push_immediate(
553 0 : &self,
554 0 : objects: Vec<RemotePath>,
555 0 : ) -> Result<(), DeletionQueueError> {
556 0 : metrics::DELETION_QUEUE
557 0 : .keys_submitted
558 0 : .inc_by(objects.len() as u64);
559 0 : self.executor_tx
560 0 : .send(DeleterMessage::Delete(objects))
561 0 : .await
562 0 : .map_err(|_| DeletionQueueError::ShuttingDown)
563 0 : }
564 :
565 : /// Companion to push_immediate. When this returns Ok, all prior objects sent
566 : /// into push_immediate have been deleted from remote storage.
567 5 : pub(crate) async fn flush_immediate(&self) -> Result<(), DeletionQueueError> {
568 5 : let (flush_op, rx) = FlushOp::new();
569 5 : self.executor_tx
570 5 : .send(DeleterMessage::Flush(flush_op))
571 5 : .await
572 5 : .map_err(|_| DeletionQueueError::ShuttingDown)?;
573 :
574 5 : rx.await.map_err(|_| DeletionQueueError::ShuttingDown)
575 5 : }
576 : }
577 :
578 : impl DeletionQueue {
579 4 : pub fn new_client(&self) -> DeletionQueueClient {
580 4 : self.client.clone()
581 4 : }
582 :
583 : /// Caller may use the returned object to construct clients with new_client.
584 : /// Caller should tokio::spawn the background() members of the two worker objects returned:
585 : /// we don't spawn those inside new() so that the caller can use their runtime/spans of choice.
586 4 : pub fn new<C>(
587 4 : remote_storage: GenericRemoteStorage,
588 4 : controller_upcall_client: C,
589 4 : conf: &'static PageServerConf,
590 4 : ) -> (Self, DeletionQueueWorkers<C>)
591 4 : where
592 4 : C: StorageControllerUpcallApi + Send + Sync,
593 : {
594 : // Unbounded channel: enables non-async functions to submit deletions. The actual length is
595 : // constrained by how promptly the ListWriter wakes up and drains it, which should be frequent
596 : // enough to avoid this taking pathologically large amount of memory.
597 4 : let (tx, rx) = tokio::sync::mpsc::unbounded_channel();
598 :
599 : // Shallow channel: it carries DeletionLists which each contain up to thousands of deletions
600 4 : let (backend_tx, backend_rx) = tokio::sync::mpsc::channel(16);
601 :
602 : // Shallow channel: it carries lists of paths, and we expect the main queueing to
603 : // happen in the backend (persistent), not in this queue.
604 4 : let (executor_tx, executor_rx) = tokio::sync::mpsc::channel(16);
605 :
606 4 : let lsn_table = Arc::new(std::sync::RwLock::new(VisibleLsnUpdates::new()));
607 :
608 : // The deletion queue has an independent cancellation token to
609 : // the general pageserver shutdown token, because it stays alive a bit
610 : // longer to flush after Tenants have all been torn down.
611 4 : let cancel = CancellationToken::new();
612 :
613 4 : (
614 4 : Self {
615 4 : client: DeletionQueueClient {
616 4 : tx,
617 4 : executor_tx: executor_tx.clone(),
618 4 : lsn_table: lsn_table.clone(),
619 4 : },
620 4 : cancel: cancel.clone(),
621 4 : },
622 4 : DeletionQueueWorkers {
623 4 : frontend: ListWriter::new(conf, rx, backend_tx, cancel.clone()),
624 4 : backend: Validator::new(
625 4 : conf,
626 4 : backend_rx,
627 4 : executor_tx,
628 4 : controller_upcall_client,
629 4 : lsn_table.clone(),
630 4 : cancel.clone(),
631 4 : ),
632 4 : executor: Deleter::new(remote_storage, executor_rx, cancel.clone()),
633 4 : },
634 4 : )
635 4 : }
636 :
637 0 : pub async fn shutdown(&mut self, timeout: Duration) {
638 0 : match tokio::time::timeout(timeout, self.client.flush()).await {
639 : Ok(Ok(())) => {
640 0 : tracing::info!("Deletion queue flushed successfully on shutdown")
641 : }
642 : Ok(Err(DeletionQueueError::ShuttingDown)) => {
643 : // This is not harmful for correctness, but is unexpected: the deletion
644 : // queue's workers should stay alive as long as there are any client handles instantiated.
645 0 : tracing::warn!("Deletion queue stopped prematurely");
646 : }
647 0 : Err(_timeout) => {
648 0 : tracing::warn!("Timed out flushing deletion queue on shutdown")
649 : }
650 : }
651 :
652 : // We only cancel _after_ flushing: otherwise we would be shutting down the
653 : // components that do the flush.
654 0 : self.cancel.cancel();
655 0 : }
656 : }
657 :
658 : #[cfg(test)]
659 : mod test {
660 : use std::io::ErrorKind;
661 : use std::time::Duration;
662 :
663 : use camino::Utf8Path;
664 : use hex_literal::hex;
665 : use pageserver_api::key::Key;
666 : use pageserver_api::models::ShardImportStatus;
667 : use pageserver_api::shard::ShardIndex;
668 : use pageserver_api::upcall_api::ReAttachResponseTenant;
669 : use remote_storage::{RemoteStorageConfig, RemoteStorageKind};
670 : use tokio::task::JoinHandle;
671 : use tracing::info;
672 :
673 : use super::*;
674 : use crate::controller_upcall_client::RetryForeverError;
675 : use crate::tenant::harness::TenantHarness;
676 : use crate::tenant::storage_layer::DeltaLayerName;
677 : pub const TIMELINE_ID: TimelineId =
678 : TimelineId::from_array(hex!("11223344556677881122334455667788"));
679 :
680 : pub const EXAMPLE_LAYER_NAME: LayerName = LayerName::Delta(DeltaLayerName {
681 : key_range: Key::from_i128(0x0)..Key::from_i128(0xFFFFFFFFFFFFFFFF),
682 : lsn_range: Lsn(0x00000000016B59D8)..Lsn(0x00000000016B5A51),
683 : });
684 :
685 : // When you need a second layer in a test.
686 : pub const EXAMPLE_LAYER_NAME_ALT: LayerName = LayerName::Delta(DeltaLayerName {
687 : key_range: Key::from_i128(0x0)..Key::from_i128(0xFFFFFFFFFFFFFFFF),
688 : lsn_range: Lsn(0x00000000016B5A51)..Lsn(0x00000000016B5A61),
689 : });
690 :
691 : struct TestSetup {
692 : harness: TenantHarness,
693 : remote_fs_dir: Utf8PathBuf,
694 : storage: GenericRemoteStorage,
695 : mock_control_plane: MockStorageController,
696 : deletion_queue: DeletionQueue,
697 : worker_join: JoinHandle<()>,
698 : }
699 :
700 : impl TestSetup {
701 : /// Simulate a pageserver restart by destroying and recreating the deletion queue
702 1 : async fn restart(&mut self) {
703 1 : let (deletion_queue, workers) = DeletionQueue::new(
704 1 : self.storage.clone(),
705 1 : self.mock_control_plane.clone(),
706 1 : self.harness.conf,
707 1 : );
708 :
709 1 : tracing::debug!("Spawning worker for new queue queue");
710 1 : let worker_join = workers.spawn_with(&tokio::runtime::Handle::current());
711 :
712 1 : let old_worker_join = std::mem::replace(&mut self.worker_join, worker_join);
713 1 : let old_deletion_queue = std::mem::replace(&mut self.deletion_queue, deletion_queue);
714 :
715 1 : tracing::debug!("Joining worker from previous queue");
716 1 : old_deletion_queue.cancel.cancel();
717 1 : old_worker_join
718 1 : .await
719 1 : .expect("Failed to join workers for previous deletion queue");
720 1 : }
721 :
722 3 : fn set_latest_generation(&self, gen_: Generation) {
723 3 : let tenant_shard_id = self.harness.tenant_shard_id;
724 3 : self.mock_control_plane
725 3 : .latest_generation
726 3 : .lock()
727 3 : .unwrap()
728 3 : .insert(tenant_shard_id, gen_);
729 3 : }
730 :
731 : /// Returns remote layer file name, suitable for use in assert_remote_files
732 3 : fn write_remote_layer(
733 3 : &self,
734 3 : file_name: LayerName,
735 3 : gen_: Generation,
736 3 : ) -> anyhow::Result<String> {
737 3 : let tenant_shard_id = self.harness.tenant_shard_id;
738 3 : let relative_remote_path = remote_timeline_path(&tenant_shard_id, &TIMELINE_ID);
739 3 : let remote_timeline_path = self.remote_fs_dir.join(relative_remote_path.get_path());
740 3 : std::fs::create_dir_all(&remote_timeline_path)?;
741 3 : let remote_layer_file_name = format!("{}{}", file_name, gen_.get_suffix());
742 :
743 3 : let content: Vec<u8> = format!("placeholder contents of {file_name}").into();
744 :
745 3 : std::fs::write(
746 3 : remote_timeline_path.join(remote_layer_file_name.clone()),
747 3 : content,
748 0 : )?;
749 :
750 3 : Ok(remote_layer_file_name)
751 3 : }
752 : }
753 :
754 : #[derive(Debug, Clone)]
755 : struct MockStorageController {
756 : pub latest_generation: std::sync::Arc<std::sync::Mutex<HashMap<TenantShardId, Generation>>>,
757 : }
758 :
759 : impl MockStorageController {
760 3 : fn new() -> Self {
761 3 : Self {
762 3 : latest_generation: Arc::default(),
763 3 : }
764 3 : }
765 : }
766 :
767 : impl StorageControllerUpcallApi for MockStorageController {
768 0 : async fn re_attach(
769 0 : &self,
770 0 : _conf: &PageServerConf,
771 0 : ) -> Result<HashMap<TenantShardId, ReAttachResponseTenant>, RetryForeverError> {
772 0 : unimplemented!()
773 : }
774 :
775 4 : async fn validate(
776 4 : &self,
777 4 : tenants: Vec<(TenantShardId, Generation)>,
778 4 : ) -> Result<HashMap<TenantShardId, bool>, RetryForeverError> {
779 4 : let mut result = HashMap::new();
780 :
781 4 : let latest_generation = self.latest_generation.lock().unwrap();
782 :
783 8 : for (tenant_shard_id, generation) in tenants {
784 4 : if let Some(latest) = latest_generation.get(&tenant_shard_id) {
785 4 : result.insert(tenant_shard_id, *latest == generation);
786 4 : }
787 : }
788 :
789 4 : Ok(result)
790 4 : }
791 :
792 0 : async fn put_timeline_import_status(
793 0 : &self,
794 0 : _tenant_shard_id: TenantShardId,
795 0 : _timeline_id: TimelineId,
796 0 : _generation: Generation,
797 0 : _status: pageserver_api::models::ShardImportStatus,
798 0 : ) -> Result<(), RetryForeverError> {
799 0 : unimplemented!()
800 : }
801 :
802 0 : async fn get_timeline_import_status(
803 0 : &self,
804 0 : _tenant_shard_id: TenantShardId,
805 0 : _timeline_id: TimelineId,
806 0 : _generation: Generation,
807 0 : ) -> Result<ShardImportStatus, RetryForeverError> {
808 0 : unimplemented!()
809 : }
810 : }
811 :
812 3 : async fn setup(test_name: &str) -> anyhow::Result<TestSetup> {
813 3 : let test_name = Box::leak(Box::new(format!("deletion_queue__{test_name}")));
814 3 : let harness = TenantHarness::create(test_name).await?;
815 :
816 : // We do not load() the harness: we only need its config and remote_storage
817 :
818 : // Set up a GenericRemoteStorage targetting a directory
819 3 : let remote_fs_dir = harness.conf.workdir.join("remote_fs");
820 3 : std::fs::create_dir_all(remote_fs_dir)?;
821 3 : let remote_fs_dir = harness.conf.workdir.join("remote_fs").canonicalize_utf8()?;
822 3 : let storage_config = RemoteStorageConfig {
823 3 : storage: RemoteStorageKind::LocalFs {
824 3 : local_path: remote_fs_dir.clone(),
825 3 : },
826 3 : timeout: RemoteStorageConfig::DEFAULT_TIMEOUT,
827 3 : small_timeout: RemoteStorageConfig::DEFAULT_SMALL_TIMEOUT,
828 3 : };
829 3 : let storage = GenericRemoteStorage::from_config(&storage_config)
830 3 : .await
831 3 : .unwrap();
832 :
833 3 : let mock_control_plane = MockStorageController::new();
834 :
835 3 : let (deletion_queue, worker) =
836 3 : DeletionQueue::new(storage.clone(), mock_control_plane.clone(), harness.conf);
837 :
838 3 : let worker_join = worker.spawn_with(&tokio::runtime::Handle::current());
839 :
840 3 : Ok(TestSetup {
841 3 : harness,
842 3 : remote_fs_dir,
843 3 : storage,
844 3 : mock_control_plane,
845 3 : deletion_queue,
846 3 : worker_join,
847 3 : })
848 3 : }
849 :
850 : // TODO: put this in a common location so that we can share with remote_timeline_client's tests
851 9 : fn assert_remote_files(expected: &[&str], remote_path: &Utf8Path) {
852 9 : let mut expected: Vec<String> = expected.iter().map(|x| String::from(*x)).collect();
853 9 : expected.sort();
854 :
855 9 : let mut found: Vec<String> = Vec::new();
856 9 : let dir = match std::fs::read_dir(remote_path) {
857 9 : Ok(d) => d,
858 0 : Err(e) => {
859 0 : if e.kind() == ErrorKind::NotFound {
860 0 : if expected.is_empty() {
861 : // We are asserting prefix is empty: it is expected that the dir is missing
862 0 : return;
863 : } else {
864 0 : assert_eq!(expected, Vec::<String>::new());
865 0 : unreachable!();
866 : }
867 : } else {
868 0 : panic!("Unexpected error listing {remote_path}: {e}");
869 : }
870 : }
871 : };
872 :
873 9 : for entry in dir.flatten() {
874 8 : let entry_name = entry.file_name();
875 8 : let fname = entry_name.to_str().unwrap();
876 8 : found.push(String::from(fname));
877 8 : }
878 9 : found.sort();
879 :
880 9 : assert_eq!(expected, found);
881 9 : }
882 :
883 5 : fn assert_local_files(expected: &[&str], directory: &Utf8Path) {
884 5 : let dir = match std::fs::read_dir(directory) {
885 4 : Ok(d) => d,
886 : Err(_) => {
887 1 : assert_eq!(expected, &Vec::<String>::new());
888 1 : return;
889 : }
890 : };
891 4 : let mut found = Vec::new();
892 9 : for dentry in dir {
893 5 : let dentry = dentry.unwrap();
894 5 : let file_name = dentry.file_name();
895 5 : let file_name_str = file_name.to_string_lossy();
896 5 : found.push(file_name_str.to_string());
897 5 : }
898 4 : found.sort();
899 4 : assert_eq!(expected, found);
900 5 : }
901 :
902 : #[tokio::test]
903 1 : async fn deletion_queue_smoke() -> anyhow::Result<()> {
904 : // Basic test that the deletion queue processes the deletions we pass into it
905 1 : let ctx = setup("deletion_queue_smoke")
906 1 : .await
907 1 : .expect("Failed test setup");
908 1 : let client = ctx.deletion_queue.new_client();
909 1 : client.recover(HashMap::new())?;
910 :
911 1 : let layer_file_name_1: LayerName = "000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__00000000016B59D8-00000000016B5A51".parse().unwrap();
912 1 : let tenant_shard_id = ctx.harness.tenant_shard_id;
913 :
914 1 : let content: Vec<u8> = "victim1 contents".into();
915 1 : let relative_remote_path = remote_timeline_path(&tenant_shard_id, &TIMELINE_ID);
916 1 : let remote_timeline_path = ctx.remote_fs_dir.join(relative_remote_path.get_path());
917 1 : let deletion_prefix = ctx.harness.conf.deletion_prefix();
918 :
919 : // Exercise the distinction between the generation of the layers
920 : // we delete, and the generation of the running Tenant.
921 1 : let layer_generation = Generation::new(0xdeadbeef);
922 1 : let now_generation = Generation::new(0xfeedbeef);
923 1 : let layer_metadata =
924 1 : LayerFileMetadata::new(0xf00, layer_generation, ShardIndex::unsharded());
925 :
926 1 : let remote_layer_file_name_1 =
927 1 : format!("{}{}", layer_file_name_1, layer_generation.get_suffix());
928 :
929 : // Set mock control plane state to valid for our generation
930 1 : ctx.set_latest_generation(now_generation);
931 :
932 : // Inject a victim file to remote storage
933 1 : info!("Writing");
934 1 : std::fs::create_dir_all(&remote_timeline_path)?;
935 1 : std::fs::write(
936 1 : remote_timeline_path.join(remote_layer_file_name_1.clone()),
937 1 : content,
938 0 : )?;
939 1 : assert_remote_files(&[&remote_layer_file_name_1], &remote_timeline_path);
940 :
941 : // File should still be there after we push it to the queue (we haven't pushed enough to flush anything)
942 1 : info!("Pushing");
943 1 : client.push_layers(
944 1 : tenant_shard_id,
945 : TIMELINE_ID,
946 1 : now_generation,
947 1 : [(layer_file_name_1.clone(), layer_metadata)].to_vec(),
948 0 : )?;
949 1 : assert_remote_files(&[&remote_layer_file_name_1], &remote_timeline_path);
950 :
951 1 : assert_local_files(&[], &deletion_prefix);
952 :
953 : // File should still be there after we write a deletion list (we haven't pushed enough to execute anything)
954 1 : info!("Flushing");
955 1 : client.flush().await?;
956 1 : assert_remote_files(&[&remote_layer_file_name_1], &remote_timeline_path);
957 1 : assert_local_files(&["0000000000000001-01.list"], &deletion_prefix);
958 :
959 : // File should go away when we execute
960 1 : info!("Flush-executing");
961 1 : client.flush_execute().await?;
962 1 : assert_remote_files(&[], &remote_timeline_path);
963 1 : assert_local_files(&["header-01"], &deletion_prefix);
964 :
965 : // Flushing on an empty queue should succeed immediately, and not write any lists
966 1 : info!("Flush-executing on empty");
967 1 : client.flush_execute().await?;
968 1 : assert_local_files(&["header-01"], &deletion_prefix);
969 :
970 2 : Ok(())
971 1 : }
972 :
973 : #[tokio::test]
974 1 : async fn deletion_queue_validation() -> anyhow::Result<()> {
975 1 : let ctx = setup("deletion_queue_validation")
976 1 : .await
977 1 : .expect("Failed test setup");
978 1 : let client = ctx.deletion_queue.new_client();
979 1 : client.recover(HashMap::new())?;
980 :
981 : // Generation that the control plane thinks is current
982 1 : let latest_generation = Generation::new(0xdeadbeef);
983 : // Generation that our DeletionQueue thinks the tenant is running with
984 1 : let stale_generation = latest_generation.previous();
985 : // Generation that our example layer file was written with
986 1 : let layer_generation = stale_generation.previous();
987 1 : let layer_metadata =
988 1 : LayerFileMetadata::new(0xf00, layer_generation, ShardIndex::unsharded());
989 :
990 1 : ctx.set_latest_generation(latest_generation);
991 :
992 1 : let tenant_shard_id = ctx.harness.tenant_shard_id;
993 1 : let relative_remote_path = remote_timeline_path(&tenant_shard_id, &TIMELINE_ID);
994 1 : let remote_timeline_path = ctx.remote_fs_dir.join(relative_remote_path.get_path());
995 :
996 : // Initial state: a remote layer exists
997 1 : let remote_layer_name = ctx.write_remote_layer(EXAMPLE_LAYER_NAME, layer_generation)?;
998 1 : assert_remote_files(&[&remote_layer_name], &remote_timeline_path);
999 :
1000 1 : tracing::debug!("Pushing...");
1001 1 : client.push_layers(
1002 1 : tenant_shard_id,
1003 : TIMELINE_ID,
1004 1 : stale_generation,
1005 1 : [(EXAMPLE_LAYER_NAME.clone(), layer_metadata.clone())].to_vec(),
1006 0 : )?;
1007 :
1008 : // We enqueued the operation in a stale generation: it should have failed validation
1009 1 : tracing::debug!("Flushing...");
1010 1 : tokio::time::timeout(Duration::from_secs(5), client.flush_execute()).await??;
1011 1 : assert_remote_files(&[&remote_layer_name], &remote_timeline_path);
1012 :
1013 1 : tracing::debug!("Pushing...");
1014 1 : client.push_layers(
1015 1 : tenant_shard_id,
1016 : TIMELINE_ID,
1017 1 : latest_generation,
1018 1 : [(EXAMPLE_LAYER_NAME.clone(), layer_metadata.clone())].to_vec(),
1019 0 : )?;
1020 :
1021 : // We enqueued the operation in a fresh generation: it should have passed validation
1022 1 : tracing::debug!("Flushing...");
1023 1 : tokio::time::timeout(Duration::from_secs(5), client.flush_execute()).await??;
1024 1 : assert_remote_files(&[], &remote_timeline_path);
1025 :
1026 2 : Ok(())
1027 1 : }
1028 :
1029 : #[tokio::test]
1030 1 : async fn deletion_queue_recovery() -> anyhow::Result<()> {
1031 : // Basic test that the deletion queue processes the deletions we pass into it
1032 1 : let mut ctx = setup("deletion_queue_recovery")
1033 1 : .await
1034 1 : .expect("Failed test setup");
1035 1 : let client = ctx.deletion_queue.new_client();
1036 1 : client.recover(HashMap::new())?;
1037 :
1038 1 : let tenant_shard_id = ctx.harness.tenant_shard_id;
1039 :
1040 1 : let relative_remote_path = remote_timeline_path(&tenant_shard_id, &TIMELINE_ID);
1041 1 : let remote_timeline_path = ctx.remote_fs_dir.join(relative_remote_path.get_path());
1042 1 : let deletion_prefix = ctx.harness.conf.deletion_prefix();
1043 :
1044 1 : let layer_generation = Generation::new(0xdeadbeef);
1045 1 : let now_generation = Generation::new(0xfeedbeef);
1046 1 : let layer_metadata =
1047 1 : LayerFileMetadata::new(0xf00, layer_generation, ShardIndex::unsharded());
1048 :
1049 : // Inject a deletion in the generation before generation_now: after restart,
1050 : // this deletion should _not_ get executed (only the immediately previous
1051 : // generation gets that treatment)
1052 1 : let remote_layer_file_name_historical =
1053 1 : ctx.write_remote_layer(EXAMPLE_LAYER_NAME, layer_generation)?;
1054 1 : client.push_layers(
1055 1 : tenant_shard_id,
1056 : TIMELINE_ID,
1057 1 : now_generation.previous(),
1058 1 : [(EXAMPLE_LAYER_NAME.clone(), layer_metadata.clone())].to_vec(),
1059 0 : )?;
1060 :
1061 : // Inject a deletion in the generation before generation_now: after restart,
1062 : // this deletion should get executed, because we execute deletions in the
1063 : // immediately previous generation on the same node.
1064 1 : let remote_layer_file_name_previous =
1065 1 : ctx.write_remote_layer(EXAMPLE_LAYER_NAME_ALT, layer_generation)?;
1066 1 : client.push_layers(
1067 1 : tenant_shard_id,
1068 : TIMELINE_ID,
1069 1 : now_generation,
1070 1 : [(EXAMPLE_LAYER_NAME_ALT.clone(), layer_metadata.clone())].to_vec(),
1071 0 : )?;
1072 :
1073 1 : client.flush().await?;
1074 1 : assert_remote_files(
1075 1 : &[
1076 1 : &remote_layer_file_name_historical,
1077 1 : &remote_layer_file_name_previous,
1078 1 : ],
1079 1 : &remote_timeline_path,
1080 : );
1081 :
1082 : // Different generatinos for the same tenant will cause two separate
1083 : // deletion lists to be emitted.
1084 1 : assert_local_files(
1085 1 : &["0000000000000001-01.list", "0000000000000002-01.list"],
1086 1 : &deletion_prefix,
1087 : );
1088 :
1089 : // Simulate a node restart: the latest generation advances
1090 1 : let now_generation = now_generation.next();
1091 1 : ctx.set_latest_generation(now_generation);
1092 :
1093 : // Restart the deletion queue
1094 1 : drop(client);
1095 1 : ctx.restart().await;
1096 1 : let client = ctx.deletion_queue.new_client();
1097 1 : client.recover(HashMap::from([(tenant_shard_id, now_generation)]))?;
1098 :
1099 1 : info!("Flush-executing");
1100 1 : client.flush_execute().await?;
1101 : // The deletion from immediately prior generation was executed, the one from
1102 : // an older generation was not.
1103 1 : assert_remote_files(&[&remote_layer_file_name_historical], &remote_timeline_path);
1104 2 : Ok(())
1105 1 : }
1106 : }
1107 :
1108 : /// A lightweight queue which can issue ordinary DeletionQueueClient objects, but doesn't do any persistence
1109 : /// or coalescing, and doesn't actually execute any deletions unless you call pump() to kick it.
1110 : #[cfg(test)]
1111 : pub(crate) mod mock {
1112 : use std::sync::atomic::{AtomicUsize, Ordering};
1113 :
1114 : use tracing::info;
1115 :
1116 : use super::*;
1117 : use crate::tenant::remote_timeline_client::remote_layer_path;
1118 :
1119 : pub struct ConsumerState {
1120 : rx: tokio::sync::mpsc::UnboundedReceiver<ListWriterQueueMessage>,
1121 : executor_rx: tokio::sync::mpsc::Receiver<DeleterMessage>,
1122 : cancel: CancellationToken,
1123 : executed: Arc<AtomicUsize>,
1124 : }
1125 :
1126 : impl ConsumerState {
1127 119 : async fn consume(&mut self, remote_storage: &GenericRemoteStorage) {
1128 119 : info!("Executing all pending deletions");
1129 :
1130 : // Transform all executor messages to generic frontend messages
1131 : loop {
1132 : use either::Either;
1133 293 : let msg = tokio::select! {
1134 293 : left = self.executor_rx.recv() => Either::Left(left),
1135 293 : right = self.rx.recv() => Either::Right(right),
1136 : };
1137 1 : match msg {
1138 0 : Either::Left(None) => break,
1139 0 : Either::Right(None) => break,
1140 0 : Either::Left(Some(DeleterMessage::Delete(objects))) => {
1141 0 : for path in objects {
1142 0 : match remote_storage.delete(&path, &self.cancel).await {
1143 : Ok(_) => {
1144 0 : debug!("Deleted {path}");
1145 : }
1146 0 : Err(e) => {
1147 0 : error!("Failed to delete {path}, leaking object! ({e})");
1148 : }
1149 : }
1150 0 : self.executed.fetch_add(1, Ordering::Relaxed);
1151 : }
1152 : }
1153 1 : Either::Left(Some(DeleterMessage::Flush(flush_op))) => {
1154 1 : flush_op.notify();
1155 1 : }
1156 174 : Either::Right(Some(ListWriterQueueMessage::Delete(op))) => {
1157 174 : let mut objects = op.objects;
1158 348 : for (layer, meta) in op.layers {
1159 174 : objects.push(remote_layer_path(
1160 174 : &op.tenant_shard_id.tenant_id,
1161 174 : &op.timeline_id,
1162 174 : meta.shard,
1163 174 : &layer,
1164 174 : meta.generation,
1165 174 : ));
1166 174 : }
1167 :
1168 347 : for path in objects {
1169 174 : info!("Executing deletion {path}");
1170 174 : match remote_storage.delete(&path, &self.cancel).await {
1171 : Ok(_) => {
1172 173 : debug!("Deleted {path}");
1173 : }
1174 0 : Err(e) => {
1175 0 : error!("Failed to delete {path}, leaking object! ({e})");
1176 : }
1177 : }
1178 173 : self.executed.fetch_add(1, Ordering::Relaxed);
1179 : }
1180 : }
1181 0 : Either::Right(Some(ListWriterQueueMessage::Flush(op))) => {
1182 0 : op.notify();
1183 0 : }
1184 0 : Either::Right(Some(ListWriterQueueMessage::FlushExecute(op))) => {
1185 0 : // We have already executed all prior deletions because mock does them inline
1186 0 : op.notify();
1187 0 : }
1188 0 : Either::Right(Some(ListWriterQueueMessage::Recover(_))) => {
1189 0 : // no-op in mock
1190 0 : }
1191 : }
1192 : }
1193 0 : }
1194 : }
1195 :
1196 : pub struct MockDeletionQueue {
1197 : tx: tokio::sync::mpsc::UnboundedSender<ListWriterQueueMessage>,
1198 : executor_tx: tokio::sync::mpsc::Sender<DeleterMessage>,
1199 : lsn_table: Arc<std::sync::RwLock<VisibleLsnUpdates>>,
1200 : }
1201 :
1202 : impl MockDeletionQueue {
1203 119 : pub fn new(remote_storage: Option<GenericRemoteStorage>) -> Self {
1204 119 : let (tx, rx) = tokio::sync::mpsc::unbounded_channel();
1205 119 : let (executor_tx, executor_rx) = tokio::sync::mpsc::channel(16384);
1206 :
1207 119 : let executed = Arc::new(AtomicUsize::new(0));
1208 :
1209 119 : let mut consumer = ConsumerState {
1210 119 : rx,
1211 119 : executor_rx,
1212 119 : cancel: CancellationToken::new(),
1213 119 : executed: executed.clone(),
1214 119 : };
1215 :
1216 119 : tokio::spawn(async move {
1217 119 : if let Some(remote_storage) = &remote_storage {
1218 119 : consumer.consume(remote_storage).await;
1219 0 : }
1220 0 : });
1221 :
1222 119 : Self {
1223 119 : tx,
1224 119 : executor_tx,
1225 119 : lsn_table: Arc::new(std::sync::RwLock::new(VisibleLsnUpdates::new())),
1226 119 : }
1227 119 : }
1228 :
1229 : #[allow(clippy::await_holding_lock)]
1230 1 : pub async fn pump(&self) {
1231 1 : let (tx, rx) = tokio::sync::oneshot::channel();
1232 1 : self.executor_tx
1233 1 : .send(DeleterMessage::Flush(FlushOp { tx }))
1234 1 : .await
1235 1 : .expect("Failed to send flush message");
1236 1 : rx.await.ok();
1237 1 : }
1238 :
1239 125 : pub(crate) fn new_client(&self) -> DeletionQueueClient {
1240 125 : DeletionQueueClient {
1241 125 : tx: self.tx.clone(),
1242 125 : executor_tx: self.executor_tx.clone(),
1243 125 : lsn_table: self.lsn_table.clone(),
1244 125 : }
1245 125 : }
1246 : }
1247 :
1248 : /// Test round-trip serialization/deserialization, and test stability of the format
1249 : /// vs. a static expected string for the serialized version.
1250 : #[test]
1251 1 : fn deletion_list_serialization() -> anyhow::Result<()> {
1252 1 : let tenant_id = "ad6c1a56f5680419d3a16ff55d97ec3c"
1253 1 : .to_string()
1254 1 : .parse::<TenantShardId>()?;
1255 1 : let timeline_id = "be322c834ed9e709e63b5c9698691910"
1256 1 : .to_string()
1257 1 : .parse::<TimelineId>()?;
1258 1 : let generation = Generation::new(123);
1259 :
1260 1 : let object =
1261 1 : RemotePath::from_string(&format!("tenants/{tenant_id}/timelines/{timeline_id}/foo"))?;
1262 1 : let mut objects = [object].to_vec();
1263 :
1264 1 : let mut example = DeletionList::new(1);
1265 1 : example.push(&tenant_id, &timeline_id, generation, &mut objects);
1266 :
1267 1 : let encoded = serde_json::to_string(&example)?;
1268 :
1269 1 : let expected = "{\"version\":1,\"sequence\":1,\"tenants\":{\"ad6c1a56f5680419d3a16ff55d97ec3c\":{\"timelines\":{\"be322c834ed9e709e63b5c9698691910\":[\"foo\"]},\"generation\":123}},\"size\":1}".to_string();
1270 1 : assert_eq!(encoded, expected);
1271 :
1272 1 : let decoded = serde_json::from_str::<DeletionList>(&encoded)?;
1273 1 : assert_eq!(example, decoded);
1274 :
1275 1 : Ok(())
1276 1 : }
1277 : }
|
