Line data Source code
1 : //! This module manages synchronizing local FS with remote storage.
2 : //!
3 : //! # Overview
4 : //!
5 : //! * [`RemoteTimelineClient`] provides functions related to upload/download of a particular timeline.
6 : //! It contains a queue of pending uploads, and manages the queue, performing uploads in parallel
7 : //! when it's safe to do so.
8 : //!
9 : //! * Stand-alone function, [`list_remote_timelines`], to get list of timelines of a tenant.
10 : //!
11 : //! These functions use the low-level remote storage client, [`remote_storage::RemoteStorage`].
12 : //!
13 : //! # APIs & How To Use Them
14 : //!
15 : //! There is a [RemoteTimelineClient] for each [Timeline][`crate::tenant::Timeline`] in the system,
16 : //! unless the pageserver is configured without remote storage.
17 : //!
18 : //! We allocate the client instance in [Timeline][`crate::tenant::Timeline`], i.e.,
19 : //! either in [`crate::tenant::mgr`] during startup or when creating a new
20 : //! timeline.
21 : //! However, the client does not become ready for use until we've initialized its upload queue:
22 : //!
23 : //! - For timelines that already have some state on the remote storage, we use
24 : //! [`RemoteTimelineClient::init_upload_queue`] .
25 : //! - For newly created timelines, we use
26 : //! [`RemoteTimelineClient::init_upload_queue_for_empty_remote`].
27 : //!
28 : //! The former takes the remote's [`IndexPart`] as an argument, possibly retrieved
29 : //! using [`list_remote_timelines`]. We'll elaborate on [`IndexPart`] in the next section.
30 : //!
31 : //! Whenever we've created/updated/deleted a file in a timeline directory, we schedule
32 : //! the corresponding remote operation with the timeline's [`RemoteTimelineClient`]:
33 : //!
34 : //! - [`RemoteTimelineClient::schedule_layer_file_upload`] when we've created a new layer file.
35 : //! - [`RemoteTimelineClient::schedule_index_upload_for_metadata_update`] when we've updated the timeline metadata file.
36 : //! - [`RemoteTimelineClient::schedule_index_upload_for_file_changes`] to upload an updated index file, after we've scheduled file uploads
37 : //! - [`RemoteTimelineClient::schedule_layer_file_deletion`] when we've deleted one or more layer files.
38 : //!
39 : //! Internally, these functions create [`UploadOp`]s and put them in a queue.
40 : //!
41 : //! There are also APIs for downloading files.
42 : //! These are not part of the aforementioned queuing and will not be discussed
43 : //! further here, except in the section covering tenant attach.
44 : //!
45 : //! # Remote Storage Structure & [`IndexPart`] Index File
46 : //!
47 : //! The "directory structure" in the remote storage mirrors the local directory structure, with paths
48 : //! like `tenants/<tenant_id>/timelines/<timeline_id>/<layer filename>`.
49 : //! Yet instead of keeping the `metadata` file remotely, we wrap it with more
50 : //! data in an "index file" aka [`IndexPart`], containing the list of **all** remote
51 : //! files for a given timeline.
52 : //! If a file is not referenced from [`IndexPart`], it's not part of the remote storage state.
53 : //!
54 : //! Having the `IndexPart` also avoids expensive and slow `S3 list` commands.
55 : //!
56 : //! # Consistency
57 : //!
58 : //! To have a consistent remote structure, it's important that uploads and
59 : //! deletions are performed in the right order. For example, the index file
60 : //! contains a list of layer files, so it must not be uploaded until all the
61 : //! layer files that are in its list have been successfully uploaded.
62 : //!
63 : //! The contract between client and its user is that the user is responsible of
64 : //! scheduling operations in an order that keeps the remote consistent as
65 : //! described above.
66 : //! From the user's perspective, the operations are executed sequentially.
67 : //! Internally, the client knows which operations can be performed in parallel,
68 : //! and which operations act like a "barrier" that require preceding operations
69 : //! to finish. The calling code just needs to call the schedule-functions in the
70 : //! correct order, and the client will parallelize the operations in a way that
71 : //! is safe.
72 : //!
73 : //! The caller should be careful with deletion, though. They should not delete
74 : //! local files that have been scheduled for upload but not yet finished uploading.
75 : //! Otherwise the upload will fail. To wait for an upload to finish, use
76 : //! the 'wait_completion' function (more on that later.)
77 : //!
78 : //! All of this relies on the following invariants:
79 : //!
80 : //! - We rely on read-after write consistency in the remote storage.
81 : //! - Layer files are immutable
82 : //!
83 : //! NB: Pageserver assumes that it has exclusive write access to the tenant in remote
84 : //! storage. Different tenants can be attached to different pageservers, but if the
85 : //! same tenant is attached to two pageservers at the same time, they will overwrite
86 : //! each other's index file updates, and confusion will ensue. There's no interlock or
87 : //! mechanism to detect that in the pageserver, we rely on the control plane to ensure
88 : //! that that doesn't happen.
89 : //!
90 : //! ## Implementation Note
91 : //!
92 : //! The *actual* remote state lags behind the *desired* remote state while
93 : //! there are in-flight operations.
94 : //! We keep track of the desired remote state in [`UploadQueueInitialized::dirty`].
95 : //! It is initialized based on the [`IndexPart`] that was passed during init
96 : //! and updated with every `schedule_*` function call.
97 : //! All this is necessary necessary to compute the future [`IndexPart`]s
98 : //! when scheduling an operation while other operations that also affect the
99 : //! remote [`IndexPart`] are in flight.
100 : //!
101 : //! # Retries & Error Handling
102 : //!
103 : //! The client retries operations indefinitely, using exponential back-off.
104 : //! There is no way to force a retry, i.e., interrupt the back-off.
105 : //! This could be built easily.
106 : //!
107 : //! # Cancellation
108 : //!
109 : //! The operations execute as plain [`task_mgr`] tasks, scoped to
110 : //! the client's tenant and timeline.
111 : //! Dropping the client will drop queued operations but not executing operations.
112 : //! These will complete unless the `task_mgr` tasks are cancelled using `task_mgr`
113 : //! APIs, e.g., during pageserver shutdown, timeline delete, or tenant detach.
114 : //!
115 : //! # Completion
116 : //!
117 : //! Once an operation has completed, we update [`UploadQueueInitialized::clean`] immediately,
118 : //! and submit a request through the DeletionQueue to update
119 : //! [`UploadQueueInitialized::visible_remote_consistent_lsn`] after it has
120 : //! validated that our generation is not stale. It is this visible value
121 : //! that is advertized to safekeepers as a signal that that they can
122 : //! delete the WAL up to that LSN.
123 : //!
124 : //! The [`RemoteTimelineClient::wait_completion`] method can be used to wait
125 : //! for all pending operations to complete. It does not prevent more
126 : //! operations from getting scheduled.
127 : //!
128 : //! # Crash Consistency
129 : //!
130 : //! We do not persist the upload queue state.
131 : //! If we drop the client, or crash, all unfinished operations are lost.
132 : //!
133 : //! To recover, the following steps need to be taken:
134 : //! - Retrieve the current remote [`IndexPart`]. This gives us a
135 : //! consistent remote state, assuming the user scheduled the operations in
136 : //! the correct order.
137 : //! - Initiate upload queue with that [`IndexPart`].
138 : //! - Reschedule all lost operations by comparing the local filesystem state
139 : //! and remote state as per [`IndexPart`]. This is done in
140 : //! [`Tenant::timeline_init_and_sync`].
141 : //!
142 : //! Note that if we crash during file deletion between the index update
143 : //! that removes the file from the list of files, and deleting the remote file,
144 : //! the file is leaked in the remote storage. Similarly, if a new file is created
145 : //! and uploaded, but the pageserver dies permanently before updating the
146 : //! remote index file, the new file is leaked in remote storage. We accept and
147 : //! tolerate that for now.
148 : //! Note further that we cannot easily fix this by scheduling deletes for every
149 : //! file that is present only on the remote, because we cannot distinguish the
150 : //! following two cases:
151 : //! - (1) We had the file locally, deleted it locally, scheduled a remote delete,
152 : //! but crashed before it finished remotely.
153 : //! - (2) We never had the file locally because we haven't on-demand downloaded
154 : //! it yet.
155 : //!
156 : //! # Downloads
157 : //!
158 : //! In addition to the upload queue, [`RemoteTimelineClient`] has functions for
159 : //! downloading files from the remote storage. Downloads are performed immediately
160 : //! against the `RemoteStorage`, independently of the upload queue.
161 : //!
162 : //! When we attach a tenant, we perform the following steps:
163 : //! - create `Tenant` object in `TenantState::Attaching` state
164 : //! - List timelines that are present in remote storage, and for each:
165 : //! - download their remote [`IndexPart`]s
166 : //! - create `Timeline` struct and a `RemoteTimelineClient`
167 : //! - initialize the client's upload queue with its `IndexPart`
168 : //! - schedule uploads for layers that are only present locally.
169 : //! - After the above is done for each timeline, open the tenant for business by
170 : //! transitioning it from `TenantState::Attaching` to `TenantState::Active` state.
171 : //! This starts the timelines' WAL-receivers and the tenant's GC & Compaction loops.
172 : //!
173 : //! # Operating Without Remote Storage
174 : //!
175 : //! If no remote storage configuration is provided, the [`RemoteTimelineClient`] is
176 : //! not created and the uploads are skipped.
177 : //!
178 : //! [`Tenant::timeline_init_and_sync`]: super::Tenant::timeline_init_and_sync
179 : //! [`Timeline::load_layer_map`]: super::Timeline::load_layer_map
180 :
181 : pub(crate) mod download;
182 : pub mod index;
183 : pub mod manifest;
184 : pub(crate) mod upload;
185 :
186 : use anyhow::Context;
187 : use camino::Utf8Path;
188 : use chrono::{NaiveDateTime, Utc};
189 :
190 : pub(crate) use download::download_initdb_tar_zst;
191 : use pageserver_api::models::TimelineArchivalState;
192 : use pageserver_api::shard::{ShardIndex, TenantShardId};
193 : use regex::Regex;
194 : use scopeguard::ScopeGuard;
195 : use tokio_util::sync::CancellationToken;
196 : use utils::backoff::{
197 : self, exponential_backoff, DEFAULT_BASE_BACKOFF_SECONDS, DEFAULT_MAX_BACKOFF_SECONDS,
198 : };
199 : use utils::pausable_failpoint;
200 : use utils::shard::ShardNumber;
201 :
202 : use std::collections::{HashMap, HashSet, VecDeque};
203 : use std::sync::atomic::{AtomicU32, Ordering};
204 : use std::sync::{Arc, Mutex, OnceLock};
205 : use std::time::Duration;
206 :
207 : use remote_storage::{
208 : DownloadError, GenericRemoteStorage, ListingMode, RemotePath, TimeoutOrCancel,
209 : };
210 : use std::ops::DerefMut;
211 : use tracing::{debug, error, info, instrument, warn};
212 : use tracing::{info_span, Instrument};
213 : use utils::lsn::Lsn;
214 :
215 : use crate::context::RequestContext;
216 : use crate::deletion_queue::{DeletionQueueClient, DeletionQueueError};
217 : use crate::metrics::{
218 : MeasureRemoteOp, RemoteOpFileKind, RemoteOpKind, RemoteTimelineClientMetrics,
219 : RemoteTimelineClientMetricsCallTrackSize, REMOTE_ONDEMAND_DOWNLOADED_BYTES,
220 : REMOTE_ONDEMAND_DOWNLOADED_LAYERS,
221 : };
222 : use crate::task_mgr::shutdown_token;
223 : use crate::tenant::debug_assert_current_span_has_tenant_and_timeline_id;
224 : use crate::tenant::remote_timeline_client::download::download_retry;
225 : use crate::tenant::storage_layer::AsLayerDesc;
226 : use crate::tenant::upload_queue::{Delete, OpType, UploadQueueStoppedDeletable};
227 : use crate::tenant::TIMELINES_SEGMENT_NAME;
228 : use crate::{
229 : config::PageServerConf,
230 : task_mgr,
231 : task_mgr::TaskKind,
232 : task_mgr::BACKGROUND_RUNTIME,
233 : tenant::metadata::TimelineMetadata,
234 : tenant::upload_queue::{
235 : UploadOp, UploadQueue, UploadQueueInitialized, UploadQueueStopped, UploadTask,
236 : },
237 : TENANT_HEATMAP_BASENAME,
238 : };
239 :
240 : use utils::id::{TenantId, TimelineId};
241 :
242 : use self::index::IndexPart;
243 :
244 : use super::config::AttachedLocationConfig;
245 : use super::metadata::MetadataUpdate;
246 : use super::storage_layer::{Layer, LayerName, ResidentLayer};
247 : use super::timeline::import_pgdata;
248 : use super::upload_queue::{NotInitialized, SetDeletedFlagProgress};
249 : use super::{DeleteTimelineError, Generation};
250 :
251 : pub(crate) use download::{
252 : download_index_part, download_tenant_manifest, is_temp_download_file,
253 : list_remote_tenant_shards, list_remote_timelines,
254 : };
255 : pub(crate) use index::LayerFileMetadata;
256 : pub(crate) use upload::upload_initdb_dir;
257 :
258 : // Occasional network issues and such can cause remote operations to fail, and
259 : // that's expected. If a download fails, we log it at info-level, and retry.
260 : // But after FAILED_DOWNLOAD_WARN_THRESHOLD retries, we start to log it at WARN
261 : // level instead, as repeated failures can mean a more serious problem. If it
262 : // fails more than FAILED_DOWNLOAD_RETRIES times, we give up
263 : pub(crate) const FAILED_DOWNLOAD_WARN_THRESHOLD: u32 = 3;
264 : pub(crate) const FAILED_REMOTE_OP_RETRIES: u32 = 10;
265 :
266 : // Similarly log failed uploads and deletions at WARN level, after this many
267 : // retries. Uploads and deletions are retried forever, though.
268 : pub(crate) const FAILED_UPLOAD_WARN_THRESHOLD: u32 = 3;
269 :
270 : pub(crate) const INITDB_PATH: &str = "initdb.tar.zst";
271 :
272 : pub(crate) const INITDB_PRESERVED_PATH: &str = "initdb-preserved.tar.zst";
273 :
274 : /// Default buffer size when interfacing with [`tokio::fs::File`].
275 : pub(crate) const BUFFER_SIZE: usize = 32 * 1024;
276 :
277 : /// Doing non-essential flushes of deletion queue is subject to this timeout, after
278 : /// which we warn and skip.
279 : const DELETION_QUEUE_FLUSH_TIMEOUT: Duration = Duration::from_secs(10);
280 :
281 : pub enum MaybeDeletedIndexPart {
282 : IndexPart(IndexPart),
283 : Deleted(IndexPart),
284 : }
285 :
286 : #[derive(Debug, thiserror::Error)]
287 : pub enum PersistIndexPartWithDeletedFlagError {
288 : #[error("another task is already setting the deleted_flag, started at {0:?}")]
289 : AlreadyInProgress(NaiveDateTime),
290 : #[error("the deleted_flag was already set, value is {0:?}")]
291 : AlreadyDeleted(NaiveDateTime),
292 : #[error(transparent)]
293 : Other(#[from] anyhow::Error),
294 : }
295 :
296 : #[derive(Debug, thiserror::Error)]
297 : pub enum WaitCompletionError {
298 : #[error(transparent)]
299 : NotInitialized(NotInitialized),
300 : #[error("wait_completion aborted because upload queue was stopped")]
301 : UploadQueueShutDownOrStopped,
302 : }
303 :
304 : #[derive(Debug, thiserror::Error)]
305 : #[error("Upload queue either in unexpected state or hasn't downloaded manifest yet")]
306 : pub struct UploadQueueNotReadyError;
307 :
308 : #[derive(Debug, thiserror::Error)]
309 : pub enum ShutdownIfArchivedError {
310 : #[error(transparent)]
311 : NotInitialized(NotInitialized),
312 : #[error("timeline is not archived")]
313 : NotArchived,
314 : }
315 :
316 : /// Behavioral modes that enable seamless live migration.
317 : ///
318 : /// See docs/rfcs/028-pageserver-migration.md to understand how these fit in.
319 : struct RemoteTimelineClientConfig {
320 : /// If this is false, then update to remote_consistent_lsn are dropped rather
321 : /// than being submitted to DeletionQueue for validation. This behavior is
322 : /// used when a tenant attachment is known to have a stale generation number,
323 : /// such that validation attempts will always fail. This is not necessary
324 : /// for correctness, but avoids spamming error statistics with failed validations
325 : /// when doing migrations of tenants.
326 : process_remote_consistent_lsn_updates: bool,
327 :
328 : /// If this is true, then object deletions are held in a buffer in RemoteTimelineClient
329 : /// rather than being submitted to the DeletionQueue. This behavior is used when a tenant
330 : /// is known to be multi-attached, in order to avoid disrupting other attached tenants
331 : /// whose generations' metadata refers to the deleted objects.
332 : block_deletions: bool,
333 : }
334 :
335 : /// RemoteTimelineClientConfig's state is entirely driven by LocationConf, but we do
336 : /// not carry the entire LocationConf structure: it's much more than we need. The From
337 : /// impl extracts the subset of the LocationConf that is interesting to RemoteTimelineClient.
338 : impl From<&AttachedLocationConfig> for RemoteTimelineClientConfig {
339 432 : fn from(lc: &AttachedLocationConfig) -> Self {
340 432 : Self {
341 432 : block_deletions: !lc.may_delete_layers_hint(),
342 432 : process_remote_consistent_lsn_updates: lc.may_upload_layers_hint(),
343 432 : }
344 432 : }
345 : }
346 :
347 : /// A client for accessing a timeline's data in remote storage.
348 : ///
349 : /// This takes care of managing the number of connections, and balancing them
350 : /// across tenants. This also handles retries of failed uploads.
351 : ///
352 : /// Upload and delete requests are ordered so that before a deletion is
353 : /// performed, we wait for all preceding uploads to finish. This ensures sure
354 : /// that if you perform a compaction operation that reshuffles data in layer
355 : /// files, we don't have a transient state where the old files have already been
356 : /// deleted, but new files have not yet been uploaded.
357 : ///
358 : /// Similarly, this enforces an order between index-file uploads, and layer
359 : /// uploads. Before an index-file upload is performed, all preceding layer
360 : /// uploads must be finished.
361 : ///
362 : /// This also maintains a list of remote files, and automatically includes that
363 : /// in the index part file, whenever timeline metadata is uploaded.
364 : ///
365 : /// Downloads are not queued, they are performed immediately.
366 : pub(crate) struct RemoteTimelineClient {
367 : conf: &'static PageServerConf,
368 :
369 : runtime: tokio::runtime::Handle,
370 :
371 : tenant_shard_id: TenantShardId,
372 : timeline_id: TimelineId,
373 : generation: Generation,
374 :
375 : upload_queue: Mutex<UploadQueue>,
376 :
377 : pub(crate) metrics: Arc<RemoteTimelineClientMetrics>,
378 :
379 : storage_impl: GenericRemoteStorage,
380 :
381 : deletion_queue_client: DeletionQueueClient,
382 :
383 : /// Subset of tenant configuration used to control upload behaviors during migrations
384 : config: std::sync::RwLock<RemoteTimelineClientConfig>,
385 :
386 : cancel: CancellationToken,
387 : }
388 :
389 : impl RemoteTimelineClient {
390 : ///
391 : /// Create a remote storage client for given timeline
392 : ///
393 : /// Note: the caller must initialize the upload queue before any uploads can be scheduled,
394 : /// by calling init_upload_queue.
395 : ///
396 422 : pub(crate) fn new(
397 422 : remote_storage: GenericRemoteStorage,
398 422 : deletion_queue_client: DeletionQueueClient,
399 422 : conf: &'static PageServerConf,
400 422 : tenant_shard_id: TenantShardId,
401 422 : timeline_id: TimelineId,
402 422 : generation: Generation,
403 422 : location_conf: &AttachedLocationConfig,
404 422 : ) -> RemoteTimelineClient {
405 422 : RemoteTimelineClient {
406 422 : conf,
407 422 : runtime: if cfg!(test) {
408 : // remote_timeline_client.rs tests rely on current-thread runtime
409 422 : tokio::runtime::Handle::current()
410 : } else {
411 0 : BACKGROUND_RUNTIME.handle().clone()
412 : },
413 422 : tenant_shard_id,
414 422 : timeline_id,
415 422 : generation,
416 422 : storage_impl: remote_storage,
417 422 : deletion_queue_client,
418 422 : upload_queue: Mutex::new(UploadQueue::Uninitialized),
419 422 : metrics: Arc::new(RemoteTimelineClientMetrics::new(
420 422 : &tenant_shard_id,
421 422 : &timeline_id,
422 422 : )),
423 422 : config: std::sync::RwLock::new(RemoteTimelineClientConfig::from(location_conf)),
424 422 : cancel: CancellationToken::new(),
425 422 : }
426 422 : }
427 :
428 : /// Initialize the upload queue for a remote storage that already received
429 : /// an index file upload, i.e., it's not empty.
430 : /// The given `index_part` must be the one on the remote.
431 6 : pub fn init_upload_queue(&self, index_part: &IndexPart) -> anyhow::Result<()> {
432 6 : let mut upload_queue = self.upload_queue.lock().unwrap();
433 6 : upload_queue.initialize_with_current_remote_index_part(index_part)?;
434 6 : self.update_remote_physical_size_gauge(Some(index_part));
435 6 : info!(
436 0 : "initialized upload queue from remote index with {} layer files",
437 0 : index_part.layer_metadata.len()
438 : );
439 6 : Ok(())
440 6 : }
441 :
442 : /// Initialize the upload queue for the case where the remote storage is empty,
443 : /// i.e., it doesn't have an `IndexPart`.
444 416 : pub fn init_upload_queue_for_empty_remote(
445 416 : &self,
446 416 : local_metadata: &TimelineMetadata,
447 416 : ) -> anyhow::Result<()> {
448 416 : let mut upload_queue = self.upload_queue.lock().unwrap();
449 416 : upload_queue.initialize_empty_remote(local_metadata)?;
450 416 : self.update_remote_physical_size_gauge(None);
451 416 : info!("initialized upload queue as empty");
452 416 : Ok(())
453 416 : }
454 :
455 : /// Initialize the queue in stopped state. Used in startup path
456 : /// to continue deletion operation interrupted by pageserver crash or restart.
457 0 : pub fn init_upload_queue_stopped_to_continue_deletion(
458 0 : &self,
459 0 : index_part: &IndexPart,
460 0 : ) -> anyhow::Result<()> {
461 : // FIXME: consider newtype for DeletedIndexPart.
462 0 : let deleted_at = index_part.deleted_at.ok_or(anyhow::anyhow!(
463 0 : "bug: it is responsibility of the caller to provide index part from MaybeDeletedIndexPart::Deleted"
464 0 : ))?;
465 :
466 0 : let mut upload_queue = self.upload_queue.lock().unwrap();
467 0 : upload_queue.initialize_with_current_remote_index_part(index_part)?;
468 0 : self.update_remote_physical_size_gauge(Some(index_part));
469 0 : self.stop_impl(&mut upload_queue);
470 0 :
471 0 : upload_queue
472 0 : .stopped_mut()
473 0 : .expect("stopped above")
474 0 : .deleted_at = SetDeletedFlagProgress::Successful(deleted_at);
475 0 :
476 0 : Ok(())
477 0 : }
478 :
479 : /// Notify this client of a change to its parent tenant's config, as this may cause us to
480 : /// take action (unblocking deletions when transitioning from AttachedMulti to AttachedSingle)
481 0 : pub(super) fn update_config(&self, location_conf: &AttachedLocationConfig) {
482 0 : let new_conf = RemoteTimelineClientConfig::from(location_conf);
483 0 : let unblocked = !new_conf.block_deletions;
484 0 :
485 0 : // Update config before draining deletions, so that we don't race with more being
486 0 : // inserted. This can result in deletions happening our of order, but that does not
487 0 : // violate any invariants: deletions only need to be ordered relative to upload of the index
488 0 : // that dereferences the deleted objects, and we are not changing that order.
489 0 : *self.config.write().unwrap() = new_conf;
490 0 :
491 0 : if unblocked {
492 : // If we may now delete layers, drain any that were blocked in our old
493 : // configuration state
494 0 : let mut queue_locked = self.upload_queue.lock().unwrap();
495 :
496 0 : if let Ok(queue) = queue_locked.initialized_mut() {
497 0 : let blocked_deletions = std::mem::take(&mut queue.blocked_deletions);
498 0 : for d in blocked_deletions {
499 0 : if let Err(e) = self.deletion_queue_client.push_layers_sync(
500 0 : self.tenant_shard_id,
501 0 : self.timeline_id,
502 0 : self.generation,
503 0 : d.layers,
504 0 : ) {
505 : // This could happen if the pageserver is shut down while a tenant
506 : // is transitioning from a deletion-blocked state: we will leak some
507 : // S3 objects in this case.
508 0 : warn!("Failed to drain blocked deletions: {}", e);
509 0 : break;
510 0 : }
511 : }
512 0 : }
513 0 : }
514 0 : }
515 :
516 : /// Returns `None` if nothing is yet uplodaded, `Some(disk_consistent_lsn)` otherwise.
517 0 : pub fn remote_consistent_lsn_projected(&self) -> Option<Lsn> {
518 0 : match &mut *self.upload_queue.lock().unwrap() {
519 0 : UploadQueue::Uninitialized => None,
520 0 : UploadQueue::Initialized(q) => q.get_last_remote_consistent_lsn_projected(),
521 0 : UploadQueue::Stopped(UploadQueueStopped::Uninitialized) => None,
522 0 : UploadQueue::Stopped(UploadQueueStopped::Deletable(q)) => q
523 0 : .upload_queue_for_deletion
524 0 : .get_last_remote_consistent_lsn_projected(),
525 : }
526 0 : }
527 :
528 0 : pub fn remote_consistent_lsn_visible(&self) -> Option<Lsn> {
529 0 : match &mut *self.upload_queue.lock().unwrap() {
530 0 : UploadQueue::Uninitialized => None,
531 0 : UploadQueue::Initialized(q) => Some(q.get_last_remote_consistent_lsn_visible()),
532 0 : UploadQueue::Stopped(UploadQueueStopped::Uninitialized) => None,
533 0 : UploadQueue::Stopped(UploadQueueStopped::Deletable(q)) => Some(
534 0 : q.upload_queue_for_deletion
535 0 : .get_last_remote_consistent_lsn_visible(),
536 0 : ),
537 : }
538 0 : }
539 :
540 : /// Returns true if this timeline was previously detached at this Lsn and the remote timeline
541 : /// client is currently initialized.
542 0 : pub(crate) fn is_previous_ancestor_lsn(&self, lsn: Lsn) -> bool {
543 0 : self.upload_queue
544 0 : .lock()
545 0 : .unwrap()
546 0 : .initialized_mut()
547 0 : .map(|uq| uq.clean.0.lineage.is_previous_ancestor_lsn(lsn))
548 0 : .unwrap_or(false)
549 0 : }
550 :
551 : /// Returns whether the timeline is archived.
552 : /// Return None if the remote index_part hasn't been downloaded yet.
553 0 : pub(crate) fn is_archived(&self) -> Option<bool> {
554 0 : self.upload_queue
555 0 : .lock()
556 0 : .unwrap()
557 0 : .initialized_mut()
558 0 : .map(|q| q.clean.0.archived_at.is_some())
559 0 : .ok()
560 0 : }
561 :
562 : /// Returns `Ok(Some(timestamp))` if the timeline has been archived, `Ok(None)` if the timeline hasn't been archived.
563 : ///
564 : /// Return Err(_) if the remote index_part hasn't been downloaded yet, or the timeline hasn't been stopped yet.
565 2 : pub(crate) fn archived_at_stopped_queue(
566 2 : &self,
567 2 : ) -> Result<Option<NaiveDateTime>, UploadQueueNotReadyError> {
568 2 : self.upload_queue
569 2 : .lock()
570 2 : .unwrap()
571 2 : .stopped_mut()
572 2 : .map(|q| q.upload_queue_for_deletion.clean.0.archived_at)
573 2 : .map_err(|_| UploadQueueNotReadyError)
574 2 : }
575 :
576 1874 : fn update_remote_physical_size_gauge(&self, current_remote_index_part: Option<&IndexPart>) {
577 1874 : let size: u64 = if let Some(current_remote_index_part) = current_remote_index_part {
578 1458 : current_remote_index_part
579 1458 : .layer_metadata
580 1458 : .values()
581 17590 : .map(|ilmd| ilmd.file_size)
582 1458 : .sum()
583 : } else {
584 416 : 0
585 : };
586 1874 : self.metrics.remote_physical_size_gauge.set(size);
587 1874 : }
588 :
589 2 : pub fn get_remote_physical_size(&self) -> u64 {
590 2 : self.metrics.remote_physical_size_gauge.get()
591 2 : }
592 :
593 : //
594 : // Download operations.
595 : //
596 : // These don't use the per-timeline queue. They do use the global semaphore in
597 : // S3Bucket, to limit the total number of concurrent operations, though.
598 : //
599 :
600 : /// Download index file
601 20 : pub async fn download_index_file(
602 20 : &self,
603 20 : cancel: &CancellationToken,
604 20 : ) -> Result<MaybeDeletedIndexPart, DownloadError> {
605 20 : let _unfinished_gauge_guard = self.metrics.call_begin(
606 20 : &RemoteOpFileKind::Index,
607 20 : &RemoteOpKind::Download,
608 20 : crate::metrics::RemoteTimelineClientMetricsCallTrackSize::DontTrackSize {
609 20 : reason: "no need for a downloads gauge",
610 20 : },
611 20 : );
612 :
613 20 : let (index_part, index_generation, index_last_modified) = download::download_index_part(
614 20 : &self.storage_impl,
615 20 : &self.tenant_shard_id,
616 20 : &self.timeline_id,
617 20 : self.generation,
618 20 : cancel,
619 20 : )
620 20 : .measure_remote_op(
621 20 : RemoteOpFileKind::Index,
622 20 : RemoteOpKind::Download,
623 20 : Arc::clone(&self.metrics),
624 20 : )
625 20 : .await?;
626 :
627 : // Defense in depth: monotonicity of generation numbers is an important correctness guarantee, so when we see a very
628 : // old index, we do extra checks in case this is the result of backward time-travel of the generation number (e.g.
629 : // in case of a bug in the service that issues generation numbers). Indices are allowed to be old, but we expect that
630 : // when we load an old index we are loading the _latest_ index: if we are asked to load an old index and there is
631 : // also a newer index available, that is surprising.
632 : const INDEX_AGE_CHECKS_THRESHOLD: Duration = Duration::from_secs(14 * 24 * 3600);
633 20 : let index_age = index_last_modified.elapsed().unwrap_or_else(|e| {
634 0 : if e.duration() > Duration::from_secs(5) {
635 : // We only warn if the S3 clock and our local clock are >5s out: because this is a low resolution
636 : // timestamp, it is common to be out by at least 1 second.
637 0 : tracing::warn!("Index has modification time in the future: {e}");
638 0 : }
639 0 : Duration::ZERO
640 20 : });
641 20 : if index_age > INDEX_AGE_CHECKS_THRESHOLD {
642 0 : tracing::info!(
643 : ?index_generation,
644 0 : age = index_age.as_secs_f64(),
645 0 : "Loaded an old index, checking for other indices..."
646 : );
647 :
648 : // Find the highest-generation index
649 0 : let (_latest_index_part, latest_index_generation, latest_index_mtime) =
650 0 : download::download_index_part(
651 0 : &self.storage_impl,
652 0 : &self.tenant_shard_id,
653 0 : &self.timeline_id,
654 0 : Generation::MAX,
655 0 : cancel,
656 0 : )
657 0 : .await?;
658 :
659 0 : if latest_index_generation > index_generation {
660 : // Unexpected! Why are we loading such an old index if a more recent one exists?
661 : // We will refuse to proceed, as there is no reasonable scenario where this should happen, but
662 : // there _is_ a clear bug/corruption scenario where it would happen (controller sets the generation
663 : // backwards).
664 0 : tracing::error!(
665 : ?index_generation,
666 : ?latest_index_generation,
667 : ?latest_index_mtime,
668 0 : "Found a newer index while loading an old one"
669 : );
670 0 : return Err(DownloadError::Fatal(
671 0 : "Index age exceeds threshold and a newer index exists".into(),
672 0 : ));
673 0 : }
674 20 : }
675 :
676 20 : if index_part.deleted_at.is_some() {
677 0 : Ok(MaybeDeletedIndexPart::Deleted(index_part))
678 : } else {
679 20 : Ok(MaybeDeletedIndexPart::IndexPart(index_part))
680 : }
681 20 : }
682 :
683 : /// Download a (layer) file from `path`, into local filesystem.
684 : ///
685 : /// 'layer_metadata' is the metadata from the remote index file.
686 : ///
687 : /// On success, returns the size of the downloaded file.
688 6 : pub async fn download_layer_file(
689 6 : &self,
690 6 : layer_file_name: &LayerName,
691 6 : layer_metadata: &LayerFileMetadata,
692 6 : local_path: &Utf8Path,
693 6 : gate: &utils::sync::gate::Gate,
694 6 : cancel: &CancellationToken,
695 6 : ctx: &RequestContext,
696 6 : ) -> Result<u64, DownloadError> {
697 6 : let downloaded_size = {
698 6 : let _unfinished_gauge_guard = self.metrics.call_begin(
699 6 : &RemoteOpFileKind::Layer,
700 6 : &RemoteOpKind::Download,
701 6 : crate::metrics::RemoteTimelineClientMetricsCallTrackSize::DontTrackSize {
702 6 : reason: "no need for a downloads gauge",
703 6 : },
704 6 : );
705 6 : download::download_layer_file(
706 6 : self.conf,
707 6 : &self.storage_impl,
708 6 : self.tenant_shard_id,
709 6 : self.timeline_id,
710 6 : layer_file_name,
711 6 : layer_metadata,
712 6 : local_path,
713 6 : gate,
714 6 : cancel,
715 6 : ctx,
716 6 : )
717 6 : .measure_remote_op(
718 6 : RemoteOpFileKind::Layer,
719 6 : RemoteOpKind::Download,
720 6 : Arc::clone(&self.metrics),
721 6 : )
722 6 : .await?
723 : };
724 :
725 6 : REMOTE_ONDEMAND_DOWNLOADED_LAYERS.inc();
726 6 : REMOTE_ONDEMAND_DOWNLOADED_BYTES.inc_by(downloaded_size);
727 6 :
728 6 : Ok(downloaded_size)
729 6 : }
730 :
731 : //
732 : // Upload operations.
733 : //
734 :
735 : /// Launch an index-file upload operation in the background, with
736 : /// fully updated metadata.
737 : ///
738 : /// This should only be used to upload initial metadata to remote storage.
739 : ///
740 : /// The upload will be added to the queue immediately, but it
741 : /// won't be performed until all previously scheduled layer file
742 : /// upload operations have completed successfully. This is to
743 : /// ensure that when the index file claims that layers X, Y and Z
744 : /// exist in remote storage, they really do. To wait for the upload
745 : /// to complete, use `wait_completion`.
746 : ///
747 : /// If there were any changes to the list of files, i.e. if any
748 : /// layer file uploads were scheduled, since the last index file
749 : /// upload, those will be included too.
750 230 : pub fn schedule_index_upload_for_full_metadata_update(
751 230 : self: &Arc<Self>,
752 230 : metadata: &TimelineMetadata,
753 230 : ) -> anyhow::Result<()> {
754 230 : let mut guard = self.upload_queue.lock().unwrap();
755 230 : let upload_queue = guard.initialized_mut()?;
756 :
757 : // As documented in the struct definition, it's ok for latest_metadata to be
758 : // ahead of what's _actually_ on the remote during index upload.
759 230 : upload_queue.dirty.metadata = metadata.clone();
760 230 :
761 230 : self.schedule_index_upload(upload_queue);
762 230 :
763 230 : Ok(())
764 230 : }
765 :
766 : /// Launch an index-file upload operation in the background, with only parts of the metadata
767 : /// updated.
768 : ///
769 : /// This is the regular way of updating metadata on layer flushes or Gc.
770 : ///
771 : /// Using this lighter update mechanism allows for reparenting and detaching without changes to
772 : /// `index_part.json`, while being more clear on what values update regularly.
773 1198 : pub(crate) fn schedule_index_upload_for_metadata_update(
774 1198 : self: &Arc<Self>,
775 1198 : update: &MetadataUpdate,
776 1198 : ) -> anyhow::Result<()> {
777 1198 : let mut guard = self.upload_queue.lock().unwrap();
778 1198 : let upload_queue = guard.initialized_mut()?;
779 :
780 1198 : upload_queue.dirty.metadata.apply(update);
781 1198 :
782 1198 : self.schedule_index_upload(upload_queue);
783 1198 :
784 1198 : Ok(())
785 1198 : }
786 :
787 : /// Launch an index-file upload operation in the background, with only the `archived_at` field updated.
788 : ///
789 : /// Returns whether it is required to wait for the queue to be empty to ensure that the change is uploaded,
790 : /// so either if the change is already sitting in the queue, but not commited yet, or the change has not
791 : /// been in the queue yet.
792 2 : pub(crate) fn schedule_index_upload_for_timeline_archival_state(
793 2 : self: &Arc<Self>,
794 2 : state: TimelineArchivalState,
795 2 : ) -> anyhow::Result<bool> {
796 2 : let mut guard = self.upload_queue.lock().unwrap();
797 2 : let upload_queue = guard.initialized_mut()?;
798 :
799 : /// Returns Some(_) if a change is needed, and Some(true) if it's a
800 : /// change needed to set archived_at.
801 4 : fn need_change(
802 4 : archived_at: &Option<NaiveDateTime>,
803 4 : state: TimelineArchivalState,
804 4 : ) -> Option<bool> {
805 4 : match (archived_at, state) {
806 : (Some(_), TimelineArchivalState::Archived)
807 : | (None, TimelineArchivalState::Unarchived) => {
808 : // Nothing to do
809 0 : tracing::info!("intended state matches present state");
810 0 : None
811 : }
812 4 : (None, TimelineArchivalState::Archived) => Some(true),
813 0 : (Some(_), TimelineArchivalState::Unarchived) => Some(false),
814 : }
815 4 : }
816 2 : let need_upload_scheduled = need_change(&upload_queue.dirty.archived_at, state);
817 :
818 2 : if let Some(archived_at_set) = need_upload_scheduled {
819 2 : let intended_archived_at = archived_at_set.then(|| Utc::now().naive_utc());
820 2 : upload_queue.dirty.archived_at = intended_archived_at;
821 2 : self.schedule_index_upload(upload_queue);
822 2 : }
823 :
824 2 : let need_wait = need_change(&upload_queue.clean.0.archived_at, state).is_some();
825 2 : Ok(need_wait)
826 2 : }
827 :
828 : /// Shuts the timeline client down, but only if the timeline is archived.
829 : ///
830 : /// This function and [`Self::schedule_index_upload_for_timeline_archival_state`] use the
831 : /// same lock to prevent races between unarchival and offloading: unarchival requires the
832 : /// upload queue to be initialized, and leaves behind an upload queue where either dirty
833 : /// or clean has archived_at of `None`. offloading leaves behind an uninitialized upload
834 : /// queue.
835 2 : pub(crate) async fn shutdown_if_archived(
836 2 : self: &Arc<Self>,
837 2 : ) -> Result<(), ShutdownIfArchivedError> {
838 2 : {
839 2 : let mut guard = self.upload_queue.lock().unwrap();
840 2 : let upload_queue = guard
841 2 : .initialized_mut()
842 2 : .map_err(ShutdownIfArchivedError::NotInitialized)?;
843 :
844 2 : match (
845 2 : upload_queue.dirty.archived_at.is_none(),
846 2 : upload_queue.clean.0.archived_at.is_none(),
847 2 : ) {
848 : // The expected case: the timeline is archived and we don't want to unarchive
849 2 : (false, false) => {}
850 : (true, false) => {
851 0 : tracing::info!("can't shut down timeline: timeline slated for unarchival");
852 0 : return Err(ShutdownIfArchivedError::NotArchived);
853 : }
854 0 : (dirty_archived, true) => {
855 0 : tracing::info!(%dirty_archived, "can't shut down timeline: timeline not archived in remote storage");
856 0 : return Err(ShutdownIfArchivedError::NotArchived);
857 : }
858 : }
859 :
860 : // Set the shutting_down flag while the guard from the archival check is held.
861 : // This prevents a race with unarchival, as initialized_mut will not return
862 : // an upload queue from this point.
863 : // Also launch the queued tasks like shutdown() does.
864 2 : if !upload_queue.shutting_down {
865 2 : upload_queue.shutting_down = true;
866 2 : upload_queue.queued_operations.push_back(UploadOp::Shutdown);
867 2 : // this operation is not counted similar to Barrier
868 2 : self.launch_queued_tasks(upload_queue);
869 2 : }
870 : }
871 :
872 2 : self.shutdown().await;
873 :
874 2 : Ok(())
875 2 : }
876 :
877 : /// Launch an index-file upload operation in the background, setting `import_pgdata` field.
878 0 : pub(crate) fn schedule_index_upload_for_import_pgdata_state_update(
879 0 : self: &Arc<Self>,
880 0 : state: Option<import_pgdata::index_part_format::Root>,
881 0 : ) -> anyhow::Result<()> {
882 0 : let mut guard = self.upload_queue.lock().unwrap();
883 0 : let upload_queue = guard.initialized_mut()?;
884 0 : upload_queue.dirty.import_pgdata = state;
885 0 : self.schedule_index_upload(upload_queue);
886 0 : Ok(())
887 0 : }
888 :
889 : ///
890 : /// Launch an index-file upload operation in the background, if necessary.
891 : ///
892 : /// Use this function to schedule the update of the index file after
893 : /// scheduling file uploads or deletions. If no file uploads or deletions
894 : /// have been scheduled since the last index file upload, this does
895 : /// nothing.
896 : ///
897 : /// Like schedule_index_upload_for_metadata_update(), this merely adds
898 : /// the upload to the upload queue and returns quickly.
899 370 : pub fn schedule_index_upload_for_file_changes(self: &Arc<Self>) -> Result<(), NotInitialized> {
900 370 : let mut guard = self.upload_queue.lock().unwrap();
901 370 : let upload_queue = guard.initialized_mut()?;
902 :
903 370 : if upload_queue.latest_files_changes_since_metadata_upload_scheduled > 0 {
904 14 : self.schedule_index_upload(upload_queue);
905 356 : }
906 :
907 370 : Ok(())
908 370 : }
909 :
910 : /// Launch an index-file upload operation in the background (internal function)
911 1514 : fn schedule_index_upload(self: &Arc<Self>, upload_queue: &mut UploadQueueInitialized) {
912 1514 : let disk_consistent_lsn = upload_queue.dirty.metadata.disk_consistent_lsn();
913 1514 : // fix up the duplicated field
914 1514 : upload_queue.dirty.disk_consistent_lsn = disk_consistent_lsn;
915 1514 :
916 1514 : // make sure it serializes before doing it in perform_upload_task so that it doesn't
917 1514 : // look like a retryable error
918 1514 : let void = std::io::sink();
919 1514 : serde_json::to_writer(void, &upload_queue.dirty).expect("serialize index_part.json");
920 1514 :
921 1514 : let index_part = &upload_queue.dirty;
922 1514 :
923 1514 : info!(
924 0 : "scheduling metadata upload up to consistent LSN {disk_consistent_lsn} with {} files ({} changed)",
925 0 : index_part.layer_metadata.len(),
926 : upload_queue.latest_files_changes_since_metadata_upload_scheduled,
927 : );
928 :
929 1514 : let op = UploadOp::UploadMetadata {
930 1514 : uploaded: Box::new(index_part.clone()),
931 1514 : };
932 1514 : self.metric_begin(&op);
933 1514 : upload_queue.queued_operations.push_back(op);
934 1514 : upload_queue.latest_files_changes_since_metadata_upload_scheduled = 0;
935 1514 :
936 1514 : // Launch the task immediately, if possible
937 1514 : self.launch_queued_tasks(upload_queue);
938 1514 : }
939 :
940 : /// Reparent this timeline to a new parent.
941 : ///
942 : /// A retryable step of timeline ancestor detach.
943 0 : pub(crate) async fn schedule_reparenting_and_wait(
944 0 : self: &Arc<Self>,
945 0 : new_parent: &TimelineId,
946 0 : ) -> anyhow::Result<()> {
947 0 : let receiver = {
948 0 : let mut guard = self.upload_queue.lock().unwrap();
949 0 : let upload_queue = guard.initialized_mut()?;
950 :
951 0 : let Some(prev) = upload_queue.dirty.metadata.ancestor_timeline() else {
952 0 : return Err(anyhow::anyhow!(
953 0 : "cannot reparent without a current ancestor"
954 0 : ));
955 : };
956 :
957 0 : let uploaded = &upload_queue.clean.0.metadata;
958 0 :
959 0 : if uploaded.ancestor_timeline().is_none() && !uploaded.ancestor_lsn().is_valid() {
960 : // nothing to do
961 0 : None
962 : } else {
963 0 : upload_queue.dirty.metadata.reparent(new_parent);
964 0 : upload_queue.dirty.lineage.record_previous_ancestor(&prev);
965 0 :
966 0 : self.schedule_index_upload(upload_queue);
967 0 :
968 0 : Some(self.schedule_barrier0(upload_queue))
969 : }
970 : };
971 :
972 0 : if let Some(receiver) = receiver {
973 0 : Self::wait_completion0(receiver).await?;
974 0 : }
975 0 : Ok(())
976 0 : }
977 :
978 : /// Schedules uploading a new version of `index_part.json` with the given layers added,
979 : /// detaching from ancestor and waits for it to complete.
980 : ///
981 : /// This is used with `Timeline::detach_ancestor` functionality.
982 0 : pub(crate) async fn schedule_adding_existing_layers_to_index_detach_and_wait(
983 0 : self: &Arc<Self>,
984 0 : layers: &[Layer],
985 0 : adopted: (TimelineId, Lsn),
986 0 : ) -> anyhow::Result<()> {
987 0 : let barrier = {
988 0 : let mut guard = self.upload_queue.lock().unwrap();
989 0 : let upload_queue = guard.initialized_mut()?;
990 :
991 0 : if upload_queue.clean.0.lineage.detached_previous_ancestor() == Some(adopted) {
992 0 : None
993 : } else {
994 0 : upload_queue.dirty.metadata.detach_from_ancestor(&adopted);
995 0 : upload_queue.dirty.lineage.record_detaching(&adopted);
996 :
997 0 : for layer in layers {
998 0 : let prev = upload_queue
999 0 : .dirty
1000 0 : .layer_metadata
1001 0 : .insert(layer.layer_desc().layer_name(), layer.metadata());
1002 0 : assert!(prev.is_none(), "copied layer existed already {layer}");
1003 : }
1004 :
1005 0 : self.schedule_index_upload(upload_queue);
1006 0 :
1007 0 : Some(self.schedule_barrier0(upload_queue))
1008 : }
1009 : };
1010 :
1011 0 : if let Some(barrier) = barrier {
1012 0 : Self::wait_completion0(barrier).await?;
1013 0 : }
1014 0 : Ok(())
1015 0 : }
1016 :
1017 : /// Adds a gc blocking reason for this timeline if one does not exist already.
1018 : ///
1019 : /// A retryable step of timeline detach ancestor.
1020 : ///
1021 : /// Returns a future which waits until the completion of the upload.
1022 0 : pub(crate) fn schedule_insert_gc_block_reason(
1023 0 : self: &Arc<Self>,
1024 0 : reason: index::GcBlockingReason,
1025 0 : ) -> Result<impl std::future::Future<Output = Result<(), WaitCompletionError>>, NotInitialized>
1026 0 : {
1027 0 : let maybe_barrier = {
1028 0 : let mut guard = self.upload_queue.lock().unwrap();
1029 0 : let upload_queue = guard.initialized_mut()?;
1030 :
1031 0 : if let index::GcBlockingReason::DetachAncestor = reason {
1032 0 : if upload_queue.dirty.metadata.ancestor_timeline().is_none() {
1033 0 : drop(guard);
1034 0 : panic!("cannot start detach ancestor if there is nothing to detach from");
1035 0 : }
1036 0 : }
1037 :
1038 0 : let wanted = |x: Option<&index::GcBlocking>| x.is_some_and(|x| x.blocked_by(reason));
1039 :
1040 0 : let current = upload_queue.dirty.gc_blocking.as_ref();
1041 0 : let uploaded = upload_queue.clean.0.gc_blocking.as_ref();
1042 0 :
1043 0 : match (current, uploaded) {
1044 0 : (x, y) if wanted(x) && wanted(y) => None,
1045 0 : (x, y) if wanted(x) && !wanted(y) => Some(self.schedule_barrier0(upload_queue)),
1046 : // Usual case: !wanted(x) && !wanted(y)
1047 : //
1048 : // Unusual: !wanted(x) && wanted(y) which means we have two processes waiting to
1049 : // turn on and off some reason.
1050 0 : (x, y) => {
1051 0 : if !wanted(x) && wanted(y) {
1052 : // this could be avoided by having external in-memory synchronization, like
1053 : // timeline detach ancestor
1054 0 : warn!(?reason, op="insert", "unexpected: two racing processes to enable and disable a gc blocking reason");
1055 0 : }
1056 :
1057 : // at this point, the metadata must always show that there is a parent
1058 0 : upload_queue.dirty.gc_blocking = current
1059 0 : .map(|x| x.with_reason(reason))
1060 0 : .or_else(|| Some(index::GcBlocking::started_now_for(reason)));
1061 0 : self.schedule_index_upload(upload_queue);
1062 0 : Some(self.schedule_barrier0(upload_queue))
1063 : }
1064 : }
1065 : };
1066 :
1067 0 : Ok(async move {
1068 0 : if let Some(barrier) = maybe_barrier {
1069 0 : Self::wait_completion0(barrier).await?;
1070 0 : }
1071 0 : Ok(())
1072 0 : })
1073 0 : }
1074 :
1075 : /// Removes a gc blocking reason for this timeline if one exists.
1076 : ///
1077 : /// A retryable step of timeline detach ancestor.
1078 : ///
1079 : /// Returns a future which waits until the completion of the upload.
1080 0 : pub(crate) fn schedule_remove_gc_block_reason(
1081 0 : self: &Arc<Self>,
1082 0 : reason: index::GcBlockingReason,
1083 0 : ) -> Result<impl std::future::Future<Output = Result<(), WaitCompletionError>>, NotInitialized>
1084 0 : {
1085 0 : let maybe_barrier = {
1086 0 : let mut guard = self.upload_queue.lock().unwrap();
1087 0 : let upload_queue = guard.initialized_mut()?;
1088 :
1089 0 : if let index::GcBlockingReason::DetachAncestor = reason {
1090 0 : if !upload_queue.clean.0.lineage.is_detached_from_ancestor() {
1091 0 : drop(guard);
1092 0 : panic!("cannot complete timeline_ancestor_detach while not detached");
1093 0 : }
1094 0 : }
1095 :
1096 0 : let wanted = |x: Option<&index::GcBlocking>| {
1097 0 : x.is_none() || x.is_some_and(|b| !b.blocked_by(reason))
1098 0 : };
1099 :
1100 0 : let current = upload_queue.dirty.gc_blocking.as_ref();
1101 0 : let uploaded = upload_queue.clean.0.gc_blocking.as_ref();
1102 0 :
1103 0 : match (current, uploaded) {
1104 0 : (x, y) if wanted(x) && wanted(y) => None,
1105 0 : (x, y) if wanted(x) && !wanted(y) => Some(self.schedule_barrier0(upload_queue)),
1106 0 : (x, y) => {
1107 0 : if !wanted(x) && wanted(y) {
1108 0 : warn!(?reason, op="remove", "unexpected: two racing processes to enable and disable a gc blocking reason (remove)");
1109 0 : }
1110 :
1111 0 : upload_queue.dirty.gc_blocking =
1112 0 : current.as_ref().and_then(|x| x.without_reason(reason));
1113 0 : assert!(wanted(upload_queue.dirty.gc_blocking.as_ref()));
1114 0 : self.schedule_index_upload(upload_queue);
1115 0 : Some(self.schedule_barrier0(upload_queue))
1116 : }
1117 : }
1118 : };
1119 :
1120 0 : Ok(async move {
1121 0 : if let Some(barrier) = maybe_barrier {
1122 0 : Self::wait_completion0(barrier).await?;
1123 0 : }
1124 0 : Ok(())
1125 0 : })
1126 0 : }
1127 :
1128 : /// Launch an upload operation in the background; the file is added to be included in next
1129 : /// `index_part.json` upload.
1130 1194 : pub(crate) fn schedule_layer_file_upload(
1131 1194 : self: &Arc<Self>,
1132 1194 : layer: ResidentLayer,
1133 1194 : ) -> Result<(), NotInitialized> {
1134 1194 : let mut guard = self.upload_queue.lock().unwrap();
1135 1194 : let upload_queue = guard.initialized_mut()?;
1136 :
1137 1194 : self.schedule_layer_file_upload0(upload_queue, layer);
1138 1194 : self.launch_queued_tasks(upload_queue);
1139 1194 : Ok(())
1140 1194 : }
1141 :
1142 1556 : fn schedule_layer_file_upload0(
1143 1556 : self: &Arc<Self>,
1144 1556 : upload_queue: &mut UploadQueueInitialized,
1145 1556 : layer: ResidentLayer,
1146 1556 : ) {
1147 1556 : let metadata = layer.metadata();
1148 1556 :
1149 1556 : upload_queue
1150 1556 : .dirty
1151 1556 : .layer_metadata
1152 1556 : .insert(layer.layer_desc().layer_name(), metadata.clone());
1153 1556 : upload_queue.latest_files_changes_since_metadata_upload_scheduled += 1;
1154 1556 :
1155 1556 : info!(
1156 : gen=?metadata.generation,
1157 : shard=?metadata.shard,
1158 0 : "scheduled layer file upload {layer}",
1159 : );
1160 :
1161 1556 : let op = UploadOp::UploadLayer(layer, metadata, None);
1162 1556 : self.metric_begin(&op);
1163 1556 : upload_queue.queued_operations.push_back(op);
1164 1556 : }
1165 :
1166 : /// Launch a delete operation in the background.
1167 : ///
1168 : /// The operation does not modify local filesystem state.
1169 : ///
1170 : /// Note: This schedules an index file upload before the deletions. The
1171 : /// deletion won't actually be performed, until all previously scheduled
1172 : /// upload operations, and the index file upload, have completed
1173 : /// successfully.
1174 8 : pub fn schedule_layer_file_deletion(
1175 8 : self: &Arc<Self>,
1176 8 : names: &[LayerName],
1177 8 : ) -> anyhow::Result<()> {
1178 8 : let mut guard = self.upload_queue.lock().unwrap();
1179 8 : let upload_queue = guard.initialized_mut()?;
1180 :
1181 8 : let with_metadata =
1182 8 : self.schedule_unlinking_of_layers_from_index_part0(upload_queue, names.iter().cloned());
1183 8 :
1184 8 : self.schedule_deletion_of_unlinked0(upload_queue, with_metadata);
1185 8 :
1186 8 : // Launch the tasks immediately, if possible
1187 8 : self.launch_queued_tasks(upload_queue);
1188 8 : Ok(())
1189 8 : }
1190 :
1191 : /// Unlinks the layer files from `index_part.json` but does not yet schedule deletion for the
1192 : /// layer files, leaving them dangling.
1193 : ///
1194 : /// The files will be leaked in remote storage unless [`Self::schedule_deletion_of_unlinked`]
1195 : /// is invoked on them.
1196 4 : pub(crate) fn schedule_gc_update(
1197 4 : self: &Arc<Self>,
1198 4 : gc_layers: &[Layer],
1199 4 : ) -> Result<(), NotInitialized> {
1200 4 : let mut guard = self.upload_queue.lock().unwrap();
1201 4 : let upload_queue = guard.initialized_mut()?;
1202 :
1203 : // just forget the return value; after uploading the next index_part.json, we can consider
1204 : // the layer files as "dangling". this is fine, at worst case we create work for the
1205 : // scrubber.
1206 :
1207 4 : let names = gc_layers.iter().map(|x| x.layer_desc().layer_name());
1208 4 :
1209 4 : self.schedule_unlinking_of_layers_from_index_part0(upload_queue, names);
1210 4 :
1211 4 : self.launch_queued_tasks(upload_queue);
1212 4 :
1213 4 : Ok(())
1214 4 : }
1215 :
1216 : /// Update the remote index file, removing the to-be-deleted files from the index,
1217 : /// allowing scheduling of actual deletions later.
1218 88 : fn schedule_unlinking_of_layers_from_index_part0<I>(
1219 88 : self: &Arc<Self>,
1220 88 : upload_queue: &mut UploadQueueInitialized,
1221 88 : names: I,
1222 88 : ) -> Vec<(LayerName, LayerFileMetadata)>
1223 88 : where
1224 88 : I: IntoIterator<Item = LayerName>,
1225 88 : {
1226 88 : // Decorate our list of names with each name's metadata, dropping
1227 88 : // names that are unexpectedly missing from our metadata. This metadata
1228 88 : // is later used when physically deleting layers, to construct key paths.
1229 88 : let with_metadata: Vec<_> = names
1230 88 : .into_iter()
1231 512 : .filter_map(|name| {
1232 512 : let meta = upload_queue.dirty.layer_metadata.remove(&name);
1233 :
1234 512 : if let Some(meta) = meta {
1235 446 : upload_queue.latest_files_changes_since_metadata_upload_scheduled += 1;
1236 446 : Some((name, meta))
1237 : } else {
1238 : // This can only happen if we forgot to to schedule the file upload
1239 : // before scheduling the delete. Log it because it is a rare/strange
1240 : // situation, and in case something is misbehaving, we'd like to know which
1241 : // layers experienced this.
1242 66 : info!("Deleting layer {name} not found in latest_files list, never uploaded?");
1243 66 : None
1244 : }
1245 512 : })
1246 88 : .collect();
1247 :
1248 : #[cfg(feature = "testing")]
1249 534 : for (name, metadata) in &with_metadata {
1250 446 : let gen = metadata.generation;
1251 446 : if let Some(unexpected) = upload_queue.dangling_files.insert(name.to_owned(), gen) {
1252 0 : if unexpected == gen {
1253 0 : tracing::error!("{name} was unlinked twice with same generation");
1254 : } else {
1255 0 : tracing::error!("{name} was unlinked twice with different generations {gen:?} and {unexpected:?}");
1256 : }
1257 446 : }
1258 : }
1259 :
1260 : // after unlinking files from the upload_queue.latest_files we must always schedule an
1261 : // index_part update, because that needs to be uploaded before we can actually delete the
1262 : // files.
1263 88 : if upload_queue.latest_files_changes_since_metadata_upload_scheduled > 0 {
1264 70 : self.schedule_index_upload(upload_queue);
1265 70 : }
1266 :
1267 88 : with_metadata
1268 88 : }
1269 :
1270 : /// Schedules deletion for layer files which have previously been unlinked from the
1271 : /// `index_part.json` with [`Self::schedule_gc_update`] or [`Self::schedule_compaction_update`].
1272 510 : pub(crate) fn schedule_deletion_of_unlinked(
1273 510 : self: &Arc<Self>,
1274 510 : layers: Vec<(LayerName, LayerFileMetadata)>,
1275 510 : ) -> anyhow::Result<()> {
1276 510 : let mut guard = self.upload_queue.lock().unwrap();
1277 510 : let upload_queue = guard.initialized_mut()?;
1278 :
1279 510 : self.schedule_deletion_of_unlinked0(upload_queue, layers);
1280 510 : self.launch_queued_tasks(upload_queue);
1281 510 : Ok(())
1282 510 : }
1283 :
1284 516 : fn schedule_deletion_of_unlinked0(
1285 516 : self: &Arc<Self>,
1286 516 : upload_queue: &mut UploadQueueInitialized,
1287 516 : mut with_metadata: Vec<(LayerName, LayerFileMetadata)>,
1288 516 : ) {
1289 516 : // Filter out any layers which were not created by this tenant shard. These are
1290 516 : // layers that originate from some ancestor shard after a split, and may still
1291 516 : // be referenced by other shards. We are free to delete them locally and remove
1292 516 : // them from our index (and would have already done so when we reach this point
1293 516 : // in the code), but we may not delete them remotely.
1294 516 : with_metadata.retain(|(name, meta)| {
1295 510 : let retain = meta.shard.shard_number == self.tenant_shard_id.shard_number
1296 510 : && meta.shard.shard_count == self.tenant_shard_id.shard_count;
1297 510 : if !retain {
1298 0 : tracing::debug!(
1299 0 : "Skipping deletion of ancestor-shard layer {name}, from shard {}",
1300 : meta.shard
1301 : );
1302 510 : }
1303 510 : retain
1304 516 : });
1305 :
1306 1026 : for (name, meta) in &with_metadata {
1307 510 : info!(
1308 0 : "scheduling deletion of layer {}{} (shard {})",
1309 0 : name,
1310 0 : meta.generation.get_suffix(),
1311 : meta.shard
1312 : );
1313 : }
1314 :
1315 : #[cfg(feature = "testing")]
1316 1026 : for (name, meta) in &with_metadata {
1317 510 : let gen = meta.generation;
1318 510 : match upload_queue.dangling_files.remove(name) {
1319 440 : Some(same) if same == gen => { /* expected */ }
1320 0 : Some(other) => {
1321 0 : tracing::error!("{name} was unlinked with {other:?} but deleted with {gen:?}");
1322 : }
1323 : None => {
1324 70 : tracing::error!("{name} was unlinked but was not dangling");
1325 : }
1326 : }
1327 : }
1328 :
1329 : // schedule the actual deletions
1330 516 : if with_metadata.is_empty() {
1331 : // avoid scheduling the op & bumping the metric
1332 6 : return;
1333 510 : }
1334 510 : let op = UploadOp::Delete(Delete {
1335 510 : layers: with_metadata,
1336 510 : });
1337 510 : self.metric_begin(&op);
1338 510 : upload_queue.queued_operations.push_back(op);
1339 516 : }
1340 :
1341 : /// Schedules a compaction update to the remote `index_part.json`.
1342 : ///
1343 : /// `compacted_from` represent the L0 names which have been `compacted_to` L1 layers.
1344 76 : pub(crate) fn schedule_compaction_update(
1345 76 : self: &Arc<Self>,
1346 76 : compacted_from: &[Layer],
1347 76 : compacted_to: &[ResidentLayer],
1348 76 : ) -> Result<(), NotInitialized> {
1349 76 : let mut guard = self.upload_queue.lock().unwrap();
1350 76 : let upload_queue = guard.initialized_mut()?;
1351 :
1352 438 : for layer in compacted_to {
1353 362 : self.schedule_layer_file_upload0(upload_queue, layer.clone());
1354 362 : }
1355 :
1356 506 : let names = compacted_from.iter().map(|x| x.layer_desc().layer_name());
1357 76 :
1358 76 : self.schedule_unlinking_of_layers_from_index_part0(upload_queue, names);
1359 76 : self.launch_queued_tasks(upload_queue);
1360 76 :
1361 76 : Ok(())
1362 76 : }
1363 :
1364 : /// Wait for all previously scheduled uploads/deletions to complete
1365 1338 : pub(crate) async fn wait_completion(self: &Arc<Self>) -> Result<(), WaitCompletionError> {
1366 1338 : let receiver = {
1367 1338 : let mut guard = self.upload_queue.lock().unwrap();
1368 1338 : let upload_queue = guard
1369 1338 : .initialized_mut()
1370 1338 : .map_err(WaitCompletionError::NotInitialized)?;
1371 1338 : self.schedule_barrier0(upload_queue)
1372 1338 : };
1373 1338 :
1374 1338 : Self::wait_completion0(receiver).await
1375 1338 : }
1376 :
1377 1338 : async fn wait_completion0(
1378 1338 : mut receiver: tokio::sync::watch::Receiver<()>,
1379 1338 : ) -> Result<(), WaitCompletionError> {
1380 1338 : if receiver.changed().await.is_err() {
1381 0 : return Err(WaitCompletionError::UploadQueueShutDownOrStopped);
1382 1338 : }
1383 1338 :
1384 1338 : Ok(())
1385 1338 : }
1386 :
1387 6 : pub(crate) fn schedule_barrier(self: &Arc<Self>) -> anyhow::Result<()> {
1388 6 : let mut guard = self.upload_queue.lock().unwrap();
1389 6 : let upload_queue = guard.initialized_mut()?;
1390 6 : self.schedule_barrier0(upload_queue);
1391 6 : Ok(())
1392 6 : }
1393 :
1394 1344 : fn schedule_barrier0(
1395 1344 : self: &Arc<Self>,
1396 1344 : upload_queue: &mut UploadQueueInitialized,
1397 1344 : ) -> tokio::sync::watch::Receiver<()> {
1398 1344 : let (sender, receiver) = tokio::sync::watch::channel(());
1399 1344 : let barrier_op = UploadOp::Barrier(sender);
1400 1344 :
1401 1344 : upload_queue.queued_operations.push_back(barrier_op);
1402 1344 : // Don't count this kind of operation!
1403 1344 :
1404 1344 : // Launch the task immediately, if possible
1405 1344 : self.launch_queued_tasks(upload_queue);
1406 1344 :
1407 1344 : receiver
1408 1344 : }
1409 :
1410 : /// Wait for all previously scheduled operations to complete, and then stop.
1411 : ///
1412 : /// Not cancellation safe
1413 10 : pub(crate) async fn shutdown(self: &Arc<Self>) {
1414 10 : // On cancellation the queue is left in ackward state of refusing new operations but
1415 10 : // proper stop is yet to be called. On cancel the original or some later task must call
1416 10 : // `stop` or `shutdown`.
1417 10 : let sg = scopeguard::guard((), |_| {
1418 0 : tracing::error!("RemoteTimelineClient::shutdown was cancelled; this should not happen, do not make this into an allowed_error")
1419 10 : });
1420 :
1421 8 : let fut = {
1422 10 : let mut guard = self.upload_queue.lock().unwrap();
1423 10 : let upload_queue = match &mut *guard {
1424 : UploadQueue::Stopped(_) => {
1425 2 : scopeguard::ScopeGuard::into_inner(sg);
1426 2 : return;
1427 : }
1428 : UploadQueue::Uninitialized => {
1429 : // transition into Stopped state
1430 0 : self.stop_impl(&mut guard);
1431 0 : scopeguard::ScopeGuard::into_inner(sg);
1432 0 : return;
1433 : }
1434 8 : UploadQueue::Initialized(ref mut init) => init,
1435 8 : };
1436 8 :
1437 8 : // if the queue is already stuck due to a shutdown operation which was cancelled, then
1438 8 : // just don't add more of these as they would never complete.
1439 8 : //
1440 8 : // TODO: if launch_queued_tasks were to be refactored to accept a &mut UploadQueue
1441 8 : // in every place we would not have to jump through this hoop, and this method could be
1442 8 : // made cancellable.
1443 8 : if !upload_queue.shutting_down {
1444 6 : upload_queue.shutting_down = true;
1445 6 : upload_queue.queued_operations.push_back(UploadOp::Shutdown);
1446 6 : // this operation is not counted similar to Barrier
1447 6 :
1448 6 : self.launch_queued_tasks(upload_queue);
1449 6 : }
1450 :
1451 8 : upload_queue.shutdown_ready.clone().acquire_owned()
1452 : };
1453 :
1454 8 : let res = fut.await;
1455 :
1456 8 : scopeguard::ScopeGuard::into_inner(sg);
1457 8 :
1458 8 : match res {
1459 0 : Ok(_permit) => unreachable!("shutdown_ready should not have been added permits"),
1460 8 : Err(_closed) => {
1461 8 : // expected
1462 8 : }
1463 8 : }
1464 8 :
1465 8 : self.stop();
1466 10 : }
1467 :
1468 : /// Set the deleted_at field in the remote index file.
1469 : ///
1470 : /// This fails if the upload queue has not been `stop()`ed.
1471 : ///
1472 : /// The caller is responsible for calling `stop()` AND for waiting
1473 : /// for any ongoing upload tasks to finish after `stop()` has succeeded.
1474 : /// Check method [`RemoteTimelineClient::stop`] for details.
1475 : #[instrument(skip_all)]
1476 : pub(crate) async fn persist_index_part_with_deleted_flag(
1477 : self: &Arc<Self>,
1478 : ) -> Result<(), PersistIndexPartWithDeletedFlagError> {
1479 : let index_part_with_deleted_at = {
1480 : let mut locked = self.upload_queue.lock().unwrap();
1481 :
1482 : // We must be in stopped state because otherwise
1483 : // we can have inprogress index part upload that can overwrite the file
1484 : // with missing is_deleted flag that we going to set below
1485 : let stopped = locked.stopped_mut()?;
1486 :
1487 : match stopped.deleted_at {
1488 : SetDeletedFlagProgress::NotRunning => (), // proceed
1489 : SetDeletedFlagProgress::InProgress(at) => {
1490 : return Err(PersistIndexPartWithDeletedFlagError::AlreadyInProgress(at));
1491 : }
1492 : SetDeletedFlagProgress::Successful(at) => {
1493 : return Err(PersistIndexPartWithDeletedFlagError::AlreadyDeleted(at));
1494 : }
1495 : };
1496 : let deleted_at = Utc::now().naive_utc();
1497 : stopped.deleted_at = SetDeletedFlagProgress::InProgress(deleted_at);
1498 :
1499 : let mut index_part = stopped.upload_queue_for_deletion.dirty.clone();
1500 : index_part.deleted_at = Some(deleted_at);
1501 : index_part
1502 : };
1503 :
1504 0 : let undo_deleted_at = scopeguard::guard(Arc::clone(self), |self_clone| {
1505 0 : let mut locked = self_clone.upload_queue.lock().unwrap();
1506 0 : let stopped = locked
1507 0 : .stopped_mut()
1508 0 : .expect("there's no way out of Stopping, and we checked it's Stopping above");
1509 0 : stopped.deleted_at = SetDeletedFlagProgress::NotRunning;
1510 0 : });
1511 :
1512 : pausable_failpoint!("persist_deleted_index_part");
1513 :
1514 : backoff::retry(
1515 0 : || {
1516 0 : upload::upload_index_part(
1517 0 : &self.storage_impl,
1518 0 : &self.tenant_shard_id,
1519 0 : &self.timeline_id,
1520 0 : self.generation,
1521 0 : &index_part_with_deleted_at,
1522 0 : &self.cancel,
1523 0 : )
1524 0 : },
1525 0 : |_e| false,
1526 : 1,
1527 : // have just a couple of attempts
1528 : // when executed as part of timeline deletion this happens in context of api call
1529 : // when executed as part of tenant deletion this happens in the background
1530 : 2,
1531 : "persist_index_part_with_deleted_flag",
1532 : &self.cancel,
1533 : )
1534 : .await
1535 0 : .ok_or_else(|| anyhow::Error::new(TimeoutOrCancel::Cancel))
1536 0 : .and_then(|x| x)?;
1537 :
1538 : // all good, disarm the guard and mark as success
1539 : ScopeGuard::into_inner(undo_deleted_at);
1540 : {
1541 : let mut locked = self.upload_queue.lock().unwrap();
1542 :
1543 : let stopped = locked
1544 : .stopped_mut()
1545 : .expect("there's no way out of Stopping, and we checked it's Stopping above");
1546 : stopped.deleted_at = SetDeletedFlagProgress::Successful(
1547 : index_part_with_deleted_at
1548 : .deleted_at
1549 : .expect("we set it above"),
1550 : );
1551 : }
1552 :
1553 : Ok(())
1554 : }
1555 :
1556 0 : pub(crate) fn is_deleting(&self) -> bool {
1557 0 : let mut locked = self.upload_queue.lock().unwrap();
1558 0 : locked.stopped_mut().is_ok()
1559 0 : }
1560 :
1561 0 : pub(crate) async fn preserve_initdb_archive(
1562 0 : self: &Arc<Self>,
1563 0 : tenant_id: &TenantId,
1564 0 : timeline_id: &TimelineId,
1565 0 : cancel: &CancellationToken,
1566 0 : ) -> anyhow::Result<()> {
1567 0 : backoff::retry(
1568 0 : || async {
1569 0 : upload::preserve_initdb_archive(&self.storage_impl, tenant_id, timeline_id, cancel)
1570 0 : .await
1571 0 : },
1572 0 : TimeoutOrCancel::caused_by_cancel,
1573 0 : FAILED_DOWNLOAD_WARN_THRESHOLD,
1574 0 : FAILED_REMOTE_OP_RETRIES,
1575 0 : "preserve_initdb_tar_zst",
1576 0 : &cancel.clone(),
1577 0 : )
1578 0 : .await
1579 0 : .ok_or_else(|| anyhow::Error::new(TimeoutOrCancel::Cancel))
1580 0 : .and_then(|x| x)
1581 0 : .context("backing up initdb archive")?;
1582 0 : Ok(())
1583 0 : }
1584 :
1585 : /// Uploads the given layer **without** adding it to be part of a future `index_part.json` upload.
1586 : ///
1587 : /// This is not normally needed.
1588 0 : pub(crate) async fn upload_layer_file(
1589 0 : self: &Arc<Self>,
1590 0 : uploaded: &ResidentLayer,
1591 0 : cancel: &CancellationToken,
1592 0 : ) -> anyhow::Result<()> {
1593 0 : let remote_path = remote_layer_path(
1594 0 : &self.tenant_shard_id.tenant_id,
1595 0 : &self.timeline_id,
1596 0 : uploaded.metadata().shard,
1597 0 : &uploaded.layer_desc().layer_name(),
1598 0 : uploaded.metadata().generation,
1599 0 : );
1600 0 :
1601 0 : backoff::retry(
1602 0 : || async {
1603 0 : upload::upload_timeline_layer(
1604 0 : &self.storage_impl,
1605 0 : uploaded.local_path(),
1606 0 : &remote_path,
1607 0 : uploaded.metadata().file_size,
1608 0 : cancel,
1609 0 : )
1610 0 : .await
1611 0 : },
1612 0 : TimeoutOrCancel::caused_by_cancel,
1613 0 : FAILED_UPLOAD_WARN_THRESHOLD,
1614 0 : FAILED_REMOTE_OP_RETRIES,
1615 0 : "upload a layer without adding it to latest files",
1616 0 : cancel,
1617 0 : )
1618 0 : .await
1619 0 : .ok_or_else(|| anyhow::Error::new(TimeoutOrCancel::Cancel))
1620 0 : .and_then(|x| x)
1621 0 : .context("upload a layer without adding it to latest files")
1622 0 : }
1623 :
1624 : /// Copies the `adopted` remote existing layer to the remote path of `adopted_as`. The layer is
1625 : /// not added to be part of a future `index_part.json` upload.
1626 0 : pub(crate) async fn copy_timeline_layer(
1627 0 : self: &Arc<Self>,
1628 0 : adopted: &Layer,
1629 0 : adopted_as: &Layer,
1630 0 : cancel: &CancellationToken,
1631 0 : ) -> anyhow::Result<()> {
1632 0 : let source_remote_path = remote_layer_path(
1633 0 : &self.tenant_shard_id.tenant_id,
1634 0 : &adopted
1635 0 : .get_timeline_id()
1636 0 : .expect("Source timeline should be alive"),
1637 0 : adopted.metadata().shard,
1638 0 : &adopted.layer_desc().layer_name(),
1639 0 : adopted.metadata().generation,
1640 0 : );
1641 0 :
1642 0 : let target_remote_path = remote_layer_path(
1643 0 : &self.tenant_shard_id.tenant_id,
1644 0 : &self.timeline_id,
1645 0 : adopted_as.metadata().shard,
1646 0 : &adopted_as.layer_desc().layer_name(),
1647 0 : adopted_as.metadata().generation,
1648 0 : );
1649 0 :
1650 0 : backoff::retry(
1651 0 : || async {
1652 0 : upload::copy_timeline_layer(
1653 0 : &self.storage_impl,
1654 0 : &source_remote_path,
1655 0 : &target_remote_path,
1656 0 : cancel,
1657 0 : )
1658 0 : .await
1659 0 : },
1660 0 : TimeoutOrCancel::caused_by_cancel,
1661 0 : FAILED_UPLOAD_WARN_THRESHOLD,
1662 0 : FAILED_REMOTE_OP_RETRIES,
1663 0 : "copy timeline layer",
1664 0 : cancel,
1665 0 : )
1666 0 : .await
1667 0 : .ok_or_else(|| anyhow::Error::new(TimeoutOrCancel::Cancel))
1668 0 : .and_then(|x| x)
1669 0 : .context("remote copy timeline layer")
1670 0 : }
1671 :
1672 0 : async fn flush_deletion_queue(&self) -> Result<(), DeletionQueueError> {
1673 0 : match tokio::time::timeout(
1674 0 : DELETION_QUEUE_FLUSH_TIMEOUT,
1675 0 : self.deletion_queue_client.flush_immediate(),
1676 0 : )
1677 0 : .await
1678 : {
1679 0 : Ok(result) => result,
1680 0 : Err(_timeout) => {
1681 0 : // Flushing remote deletions is not mandatory: we flush here to make the system easier to test, and
1682 0 : // to ensure that _usually_ objects are really gone after a DELETE is acked. However, in case of deletion
1683 0 : // queue issues (https://github.com/neondatabase/neon/issues/6440), we don't want to wait indefinitely here.
1684 0 : tracing::warn!(
1685 0 : "Timed out waiting for deletion queue flush, acking deletion anyway"
1686 : );
1687 0 : Ok(())
1688 : }
1689 : }
1690 0 : }
1691 :
1692 : /// Prerequisites: UploadQueue should be in stopped state and deleted_at should be successfuly set.
1693 : /// The function deletes layer files one by one, then lists the prefix to see if we leaked something
1694 : /// deletes leaked files if any and proceeds with deletion of index file at the end.
1695 0 : pub(crate) async fn delete_all(self: &Arc<Self>) -> Result<(), DeleteTimelineError> {
1696 0 : debug_assert_current_span_has_tenant_and_timeline_id();
1697 :
1698 0 : let layers: Vec<RemotePath> = {
1699 0 : let mut locked = self.upload_queue.lock().unwrap();
1700 0 : let stopped = locked.stopped_mut().map_err(DeleteTimelineError::Other)?;
1701 :
1702 0 : if !matches!(stopped.deleted_at, SetDeletedFlagProgress::Successful(_)) {
1703 0 : return Err(DeleteTimelineError::Other(anyhow::anyhow!(
1704 0 : "deleted_at is not set"
1705 0 : )));
1706 0 : }
1707 0 :
1708 0 : debug_assert!(stopped.upload_queue_for_deletion.no_pending_work());
1709 :
1710 0 : stopped
1711 0 : .upload_queue_for_deletion
1712 0 : .dirty
1713 0 : .layer_metadata
1714 0 : .drain()
1715 0 : .filter(|(_file_name, meta)| {
1716 0 : // Filter out layers that belonged to an ancestor shard. Since we are deleting the whole timeline from
1717 0 : // all shards anyway, we _could_ delete these, but
1718 0 : // - it creates a potential race if other shards are still
1719 0 : // using the layers while this shard deletes them.
1720 0 : // - it means that if we rolled back the shard split, the ancestor shards would be in a state where
1721 0 : // these timelines are present but corrupt (their index exists but some layers don't)
1722 0 : //
1723 0 : // These layers will eventually be cleaned up by the scrubber when it does physical GC.
1724 0 : meta.shard.shard_number == self.tenant_shard_id.shard_number
1725 0 : && meta.shard.shard_count == self.tenant_shard_id.shard_count
1726 0 : })
1727 0 : .map(|(file_name, meta)| {
1728 0 : remote_layer_path(
1729 0 : &self.tenant_shard_id.tenant_id,
1730 0 : &self.timeline_id,
1731 0 : meta.shard,
1732 0 : &file_name,
1733 0 : meta.generation,
1734 0 : )
1735 0 : })
1736 0 : .collect()
1737 0 : };
1738 0 :
1739 0 : let layer_deletion_count = layers.len();
1740 0 : self.deletion_queue_client
1741 0 : .push_immediate(layers)
1742 0 : .await
1743 0 : .map_err(|_| DeleteTimelineError::Cancelled)?;
1744 :
1745 : // Delete the initdb.tar.zst, which is not always present, but deletion attempts of
1746 : // inexistant objects are not considered errors.
1747 0 : let initdb_path =
1748 0 : remote_initdb_archive_path(&self.tenant_shard_id.tenant_id, &self.timeline_id);
1749 0 : self.deletion_queue_client
1750 0 : .push_immediate(vec![initdb_path])
1751 0 : .await
1752 0 : .map_err(|_| DeleteTimelineError::Cancelled)?;
1753 :
1754 : // Do not delete index part yet, it is needed for possible retry. If we remove it first
1755 : // and retry will arrive to different pageserver there wont be any traces of it on remote storage
1756 0 : let timeline_storage_path = remote_timeline_path(&self.tenant_shard_id, &self.timeline_id);
1757 0 :
1758 0 : // Execute all pending deletions, so that when we proceed to do a listing below, we aren't
1759 0 : // taking the burden of listing all the layers that we already know we should delete.
1760 0 : self.flush_deletion_queue()
1761 0 : .await
1762 0 : .map_err(|_| DeleteTimelineError::Cancelled)?;
1763 :
1764 0 : let cancel = shutdown_token();
1765 :
1766 0 : let remaining = download_retry(
1767 0 : || async {
1768 0 : self.storage_impl
1769 0 : .list(
1770 0 : Some(&timeline_storage_path),
1771 0 : ListingMode::NoDelimiter,
1772 0 : None,
1773 0 : &cancel,
1774 0 : )
1775 0 : .await
1776 0 : },
1777 0 : "list remaining files",
1778 0 : &cancel,
1779 0 : )
1780 0 : .await
1781 0 : .context("list files remaining files")?
1782 : .keys;
1783 :
1784 : // We will delete the current index_part object last, since it acts as a deletion
1785 : // marker via its deleted_at attribute
1786 0 : let latest_index = remaining
1787 0 : .iter()
1788 0 : .filter(|o| {
1789 0 : o.key
1790 0 : .object_name()
1791 0 : .map(|n| n.starts_with(IndexPart::FILE_NAME))
1792 0 : .unwrap_or(false)
1793 0 : })
1794 0 : .filter_map(|o| parse_remote_index_path(o.key.clone()).map(|gen| (o.key.clone(), gen)))
1795 0 : .max_by_key(|i| i.1)
1796 0 : .map(|i| i.0.clone())
1797 0 : .unwrap_or(
1798 0 : // No generation-suffixed indices, assume we are dealing with
1799 0 : // a legacy index.
1800 0 : remote_index_path(&self.tenant_shard_id, &self.timeline_id, Generation::none()),
1801 0 : );
1802 0 :
1803 0 : let remaining_layers: Vec<RemotePath> = remaining
1804 0 : .into_iter()
1805 0 : .filter_map(|o| {
1806 0 : if o.key == latest_index || o.key.object_name() == Some(INITDB_PRESERVED_PATH) {
1807 0 : None
1808 : } else {
1809 0 : Some(o.key)
1810 : }
1811 0 : })
1812 0 : .inspect(|path| {
1813 0 : if let Some(name) = path.object_name() {
1814 0 : info!(%name, "deleting a file not referenced from index_part.json");
1815 : } else {
1816 0 : warn!(%path, "deleting a nameless or non-utf8 object not referenced from index_part.json");
1817 : }
1818 0 : })
1819 0 : .collect();
1820 0 :
1821 0 : let not_referenced_count = remaining_layers.len();
1822 0 : if !remaining_layers.is_empty() {
1823 0 : self.deletion_queue_client
1824 0 : .push_immediate(remaining_layers)
1825 0 : .await
1826 0 : .map_err(|_| DeleteTimelineError::Cancelled)?;
1827 0 : }
1828 :
1829 0 : fail::fail_point!("timeline-delete-before-index-delete", |_| {
1830 0 : Err(DeleteTimelineError::Other(anyhow::anyhow!(
1831 0 : "failpoint: timeline-delete-before-index-delete"
1832 0 : )))?
1833 0 : });
1834 :
1835 0 : debug!("enqueuing index part deletion");
1836 0 : self.deletion_queue_client
1837 0 : .push_immediate([latest_index].to_vec())
1838 0 : .await
1839 0 : .map_err(|_| DeleteTimelineError::Cancelled)?;
1840 :
1841 : // Timeline deletion is rare and we have probably emitted a reasonably number of objects: wait
1842 : // for a flush to a persistent deletion list so that we may be sure deletion will occur.
1843 0 : self.flush_deletion_queue()
1844 0 : .await
1845 0 : .map_err(|_| DeleteTimelineError::Cancelled)?;
1846 :
1847 0 : fail::fail_point!("timeline-delete-after-index-delete", |_| {
1848 0 : Err(DeleteTimelineError::Other(anyhow::anyhow!(
1849 0 : "failpoint: timeline-delete-after-index-delete"
1850 0 : )))?
1851 0 : });
1852 :
1853 0 : info!(prefix=%timeline_storage_path, referenced=layer_deletion_count, not_referenced=%not_referenced_count, "done deleting in timeline prefix, including index_part.json");
1854 :
1855 0 : Ok(())
1856 0 : }
1857 :
1858 : ///
1859 : /// Pick next tasks from the queue, and start as many of them as possible without violating
1860 : /// the ordering constraints.
1861 : ///
1862 : /// The caller needs to already hold the `upload_queue` lock.
1863 7709 : fn launch_queued_tasks(self: &Arc<Self>, upload_queue: &mut UploadQueueInitialized) {
1864 12337 : while let Some(next_op) = upload_queue.queued_operations.front() {
1865 : // Can we run this task now?
1866 8978 : let can_run_now = match next_op {
1867 : UploadOp::UploadLayer(..) => {
1868 : // Can always be scheduled.
1869 1547 : true
1870 : }
1871 : UploadOp::UploadMetadata { .. } => {
1872 : // These can only be performed after all the preceding operations
1873 : // have finished.
1874 4581 : upload_queue.inprogress_tasks.is_empty()
1875 : }
1876 : UploadOp::Delete(..) => {
1877 : // Wait for preceding uploads to finish. Concurrent deletions are OK, though.
1878 344 : upload_queue.num_inprogress_deletions == upload_queue.inprogress_tasks.len()
1879 : }
1880 :
1881 : UploadOp::Barrier(_) | UploadOp::Shutdown => {
1882 2506 : upload_queue.inprogress_tasks.is_empty()
1883 : }
1884 : };
1885 :
1886 : // If we cannot launch this task, don't look any further.
1887 : //
1888 : // In some cases, we could let some non-frontmost tasks to "jump the queue" and launch
1889 : // them now, but we don't try to do that currently. For example, if the frontmost task
1890 : // is an index-file upload that cannot proceed until preceding uploads have finished, we
1891 : // could still start layer uploads that were scheduled later.
1892 8978 : if !can_run_now {
1893 4342 : break;
1894 4636 : }
1895 4636 :
1896 4636 : if let UploadOp::Shutdown = next_op {
1897 : // leave the op in the queue but do not start more tasks; it will be dropped when
1898 : // the stop is called.
1899 8 : upload_queue.shutdown_ready.close();
1900 8 : break;
1901 4628 : }
1902 4628 :
1903 4628 : // We can launch this task. Remove it from the queue first.
1904 4628 : let mut next_op = upload_queue.queued_operations.pop_front().unwrap();
1905 4628 :
1906 4628 : debug!("starting op: {}", next_op);
1907 :
1908 : // Update the counters and prepare
1909 4628 : match &mut next_op {
1910 1547 : UploadOp::UploadLayer(layer, meta, mode) => {
1911 1547 : if upload_queue
1912 1547 : .recently_deleted
1913 1547 : .remove(&(layer.layer_desc().layer_name().clone(), meta.generation))
1914 0 : {
1915 0 : *mode = Some(OpType::FlushDeletion);
1916 0 : } else {
1917 1547 : *mode = Some(OpType::MayReorder)
1918 : }
1919 1547 : upload_queue.num_inprogress_layer_uploads += 1;
1920 : }
1921 1474 : UploadOp::UploadMetadata { .. } => {
1922 1474 : upload_queue.num_inprogress_metadata_uploads += 1;
1923 1474 : }
1924 263 : UploadOp::Delete(Delete { layers }) => {
1925 526 : for (name, meta) in layers {
1926 263 : upload_queue
1927 263 : .recently_deleted
1928 263 : .insert((name.clone(), meta.generation));
1929 263 : }
1930 263 : upload_queue.num_inprogress_deletions += 1;
1931 : }
1932 1344 : UploadOp::Barrier(sender) => {
1933 1344 : sender.send_replace(());
1934 1344 : continue;
1935 : }
1936 0 : UploadOp::Shutdown => unreachable!("shutdown is intentionally never popped off"),
1937 : };
1938 :
1939 : // Assign unique ID to this task
1940 3284 : upload_queue.task_counter += 1;
1941 3284 : let upload_task_id = upload_queue.task_counter;
1942 3284 :
1943 3284 : // Add it to the in-progress map
1944 3284 : let task = Arc::new(UploadTask {
1945 3284 : task_id: upload_task_id,
1946 3284 : op: next_op,
1947 3284 : retries: AtomicU32::new(0),
1948 3284 : });
1949 3284 : upload_queue
1950 3284 : .inprogress_tasks
1951 3284 : .insert(task.task_id, Arc::clone(&task));
1952 3284 :
1953 3284 : // Spawn task to perform the task
1954 3284 : let self_rc = Arc::clone(self);
1955 3284 : let tenant_shard_id = self.tenant_shard_id;
1956 3284 : let timeline_id = self.timeline_id;
1957 3284 : task_mgr::spawn(
1958 3284 : &self.runtime,
1959 3284 : TaskKind::RemoteUploadTask,
1960 3284 : self.tenant_shard_id,
1961 3284 : Some(self.timeline_id),
1962 3284 : "remote upload",
1963 3173 : async move {
1964 3173 : self_rc.perform_upload_task(task).await;
1965 3053 : Ok(())
1966 3053 : }
1967 3284 : .instrument(info_span!(parent: None, "remote_upload", tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), %timeline_id, %upload_task_id)),
1968 : );
1969 :
1970 : // Loop back to process next task
1971 : }
1972 7709 : }
1973 :
1974 : ///
1975 : /// Perform an upload task.
1976 : ///
1977 : /// The task is in the `inprogress_tasks` list. This function will try to
1978 : /// execute it, retrying forever. On successful completion, the task is
1979 : /// removed it from the `inprogress_tasks` list, and any next task(s) in the
1980 : /// queue that were waiting by the completion are launched.
1981 : ///
1982 : /// The task can be shut down, however. That leads to stopping the whole
1983 : /// queue.
1984 : ///
1985 3173 : async fn perform_upload_task(self: &Arc<Self>, task: Arc<UploadTask>) {
1986 3173 : let cancel = shutdown_token();
1987 : // Loop to retry until it completes.
1988 : loop {
1989 : // If we're requested to shut down, close up shop and exit.
1990 : //
1991 : // Note: We only check for the shutdown requests between retries, so
1992 : // if a shutdown request arrives while we're busy uploading, in the
1993 : // upload::upload:*() call below, we will wait not exit until it has
1994 : // finished. We probably could cancel the upload by simply dropping
1995 : // the Future, but we're not 100% sure if the remote storage library
1996 : // is cancellation safe, so we don't dare to do that. Hopefully, the
1997 : // upload finishes or times out soon enough.
1998 3173 : if cancel.is_cancelled() {
1999 0 : info!("upload task cancelled by shutdown request");
2000 0 : self.stop();
2001 0 : return;
2002 3173 : }
2003 3173 :
2004 3173 : // Assert that we don't modify a layer that's referenced by the current index.
2005 3173 : if cfg!(debug_assertions) {
2006 3173 : let modified = match &task.op {
2007 1451 : UploadOp::UploadLayer(layer, layer_metadata, _) => {
2008 1451 : vec![(layer.layer_desc().layer_name(), layer_metadata)]
2009 : }
2010 263 : UploadOp::Delete(delete) => {
2011 263 : delete.layers.iter().map(|(n, m)| (n.clone(), m)).collect()
2012 : }
2013 : // These don't modify layers.
2014 1459 : UploadOp::UploadMetadata { .. } => Vec::new(),
2015 0 : UploadOp::Barrier(_) => Vec::new(),
2016 0 : UploadOp::Shutdown => Vec::new(),
2017 : };
2018 3173 : if let Ok(queue) = self.upload_queue.lock().unwrap().initialized_mut() {
2019 4883 : for (ref name, metadata) in modified {
2020 1714 : debug_assert!(
2021 1714 : !queue.clean.0.references(name, metadata),
2022 4 : "layer {name} modified while referenced by index",
2023 : );
2024 : }
2025 0 : }
2026 0 : }
2027 :
2028 3169 : let upload_result: anyhow::Result<()> = match &task.op {
2029 1451 : UploadOp::UploadLayer(ref layer, ref layer_metadata, mode) => {
2030 1451 : if let Some(OpType::FlushDeletion) = mode {
2031 0 : if self.config.read().unwrap().block_deletions {
2032 : // Of course, this is not efficient... but usually the queue should be empty.
2033 0 : let mut queue_locked = self.upload_queue.lock().unwrap();
2034 0 : let mut detected = false;
2035 0 : if let Ok(queue) = queue_locked.initialized_mut() {
2036 0 : for list in queue.blocked_deletions.iter_mut() {
2037 0 : list.layers.retain(|(name, meta)| {
2038 0 : if name == &layer.layer_desc().layer_name()
2039 0 : && meta.generation == layer_metadata.generation
2040 : {
2041 0 : detected = true;
2042 0 : // remove the layer from deletion queue
2043 0 : false
2044 : } else {
2045 : // keep the layer
2046 0 : true
2047 : }
2048 0 : });
2049 0 : }
2050 0 : }
2051 0 : if detected {
2052 0 : info!(
2053 0 : "cancelled blocked deletion of layer {} at gen {:?}",
2054 0 : layer.layer_desc().layer_name(),
2055 : layer_metadata.generation
2056 : );
2057 0 : }
2058 : } else {
2059 : // TODO: we did not guarantee that upload task starts after deletion task, so there could be possibly race conditions
2060 : // that we still get the layer deleted. But this only happens if someone creates a layer immediately after it's deleted,
2061 : // which is not possible in the current system.
2062 0 : info!(
2063 0 : "waiting for deletion queue flush to complete before uploading layer {} at gen {:?}",
2064 0 : layer.layer_desc().layer_name(),
2065 : layer_metadata.generation
2066 : );
2067 : {
2068 : // We are going to flush, we can clean up the recently deleted list.
2069 0 : let mut queue_locked = self.upload_queue.lock().unwrap();
2070 0 : if let Ok(queue) = queue_locked.initialized_mut() {
2071 0 : queue.recently_deleted.clear();
2072 0 : }
2073 : }
2074 0 : if let Err(e) = self.deletion_queue_client.flush_execute().await {
2075 0 : warn!(
2076 0 : "failed to flush the deletion queue before uploading layer {} at gen {:?}, still proceeding to upload: {e:#} ",
2077 0 : layer.layer_desc().layer_name(),
2078 : layer_metadata.generation
2079 : );
2080 : } else {
2081 0 : info!(
2082 0 : "done flushing deletion queue before uploading layer {} at gen {:?}",
2083 0 : layer.layer_desc().layer_name(),
2084 : layer_metadata.generation
2085 : );
2086 : }
2087 : }
2088 1451 : }
2089 1451 : let local_path = layer.local_path();
2090 1451 :
2091 1451 : // We should only be uploading layers created by this `Tenant`'s lifetime, so
2092 1451 : // the metadata in the upload should always match our current generation.
2093 1451 : assert_eq!(layer_metadata.generation, self.generation);
2094 :
2095 1451 : let remote_path = remote_layer_path(
2096 1451 : &self.tenant_shard_id.tenant_id,
2097 1451 : &self.timeline_id,
2098 1451 : layer_metadata.shard,
2099 1451 : &layer.layer_desc().layer_name(),
2100 1451 : layer_metadata.generation,
2101 1451 : );
2102 1451 :
2103 1451 : upload::upload_timeline_layer(
2104 1451 : &self.storage_impl,
2105 1451 : local_path,
2106 1451 : &remote_path,
2107 1451 : layer_metadata.file_size,
2108 1451 : &self.cancel,
2109 1451 : )
2110 1451 : .measure_remote_op(
2111 1451 : RemoteOpFileKind::Layer,
2112 1451 : RemoteOpKind::Upload,
2113 1451 : Arc::clone(&self.metrics),
2114 1451 : )
2115 1451 : .await
2116 : }
2117 1459 : UploadOp::UploadMetadata { ref uploaded } => {
2118 1459 : let res = upload::upload_index_part(
2119 1459 : &self.storage_impl,
2120 1459 : &self.tenant_shard_id,
2121 1459 : &self.timeline_id,
2122 1459 : self.generation,
2123 1459 : uploaded,
2124 1459 : &self.cancel,
2125 1459 : )
2126 1459 : .measure_remote_op(
2127 1459 : RemoteOpFileKind::Index,
2128 1459 : RemoteOpKind::Upload,
2129 1459 : Arc::clone(&self.metrics),
2130 1459 : )
2131 1459 : .await;
2132 1452 : if res.is_ok() {
2133 1452 : self.update_remote_physical_size_gauge(Some(uploaded));
2134 1452 : let mention_having_future_layers = if cfg!(feature = "testing") {
2135 1452 : uploaded
2136 1452 : .layer_metadata
2137 1452 : .keys()
2138 17534 : .any(|x| x.is_in_future(uploaded.metadata.disk_consistent_lsn()))
2139 : } else {
2140 0 : false
2141 : };
2142 1452 : if mention_having_future_layers {
2143 : // find rationale near crate::tenant::timeline::init::cleanup_future_layer
2144 37 : tracing::info!(
2145 0 : disk_consistent_lsn = %uploaded.metadata.disk_consistent_lsn(),
2146 0 : "uploaded an index_part.json with future layers -- this is ok! if shutdown now, expect future layer cleanup"
2147 : );
2148 1415 : }
2149 0 : }
2150 1452 : res
2151 : }
2152 259 : UploadOp::Delete(delete) => {
2153 259 : if self.config.read().unwrap().block_deletions {
2154 0 : let mut queue_locked = self.upload_queue.lock().unwrap();
2155 0 : if let Ok(queue) = queue_locked.initialized_mut() {
2156 0 : queue.blocked_deletions.push(delete.clone());
2157 0 : }
2158 0 : Ok(())
2159 : } else {
2160 259 : pausable_failpoint!("before-delete-layer-pausable");
2161 256 : self.deletion_queue_client
2162 256 : .push_layers(
2163 256 : self.tenant_shard_id,
2164 256 : self.timeline_id,
2165 256 : self.generation,
2166 256 : delete.layers.clone(),
2167 256 : )
2168 256 : .await
2169 256 : .map_err(|e| anyhow::anyhow!(e))
2170 : }
2171 : }
2172 0 : unexpected @ UploadOp::Barrier(_) | unexpected @ UploadOp::Shutdown => {
2173 : // unreachable. Barrier operations are handled synchronously in
2174 : // launch_queued_tasks
2175 0 : warn!("unexpected {unexpected:?} operation in perform_upload_task");
2176 0 : break;
2177 : }
2178 : };
2179 :
2180 0 : match upload_result {
2181 : Ok(()) => {
2182 3053 : break;
2183 : }
2184 0 : Err(e) if TimeoutOrCancel::caused_by_cancel(&e) => {
2185 0 : // loop around to do the proper stopping
2186 0 : continue;
2187 : }
2188 0 : Err(e) => {
2189 0 : let retries = task.retries.fetch_add(1, Ordering::SeqCst);
2190 0 :
2191 0 : // Uploads can fail due to rate limits (IAM, S3), spurious network problems,
2192 0 : // or other external reasons. Such issues are relatively regular, so log them
2193 0 : // at info level at first, and only WARN if the operation fails repeatedly.
2194 0 : //
2195 0 : // (See similar logic for downloads in `download::download_retry`)
2196 0 : if retries < FAILED_UPLOAD_WARN_THRESHOLD {
2197 0 : info!(
2198 0 : "failed to perform remote task {}, will retry (attempt {}): {:#}",
2199 0 : task.op, retries, e
2200 : );
2201 : } else {
2202 0 : warn!(
2203 0 : "failed to perform remote task {}, will retry (attempt {}): {:?}",
2204 0 : task.op, retries, e
2205 : );
2206 : }
2207 :
2208 : // sleep until it's time to retry, or we're cancelled
2209 0 : exponential_backoff(
2210 0 : retries,
2211 0 : DEFAULT_BASE_BACKOFF_SECONDS,
2212 0 : DEFAULT_MAX_BACKOFF_SECONDS,
2213 0 : &cancel,
2214 0 : )
2215 0 : .await;
2216 : }
2217 : }
2218 : }
2219 :
2220 3053 : let retries = task.retries.load(Ordering::SeqCst);
2221 3053 : if retries > 0 {
2222 0 : info!(
2223 0 : "remote task {} completed successfully after {} retries",
2224 0 : task.op, retries
2225 : );
2226 : } else {
2227 3053 : debug!("remote task {} completed successfully", task.op);
2228 : }
2229 :
2230 : // The task has completed successfully. Remove it from the in-progress list.
2231 3053 : let lsn_update = {
2232 3053 : let mut upload_queue_guard = self.upload_queue.lock().unwrap();
2233 3053 : let upload_queue = match upload_queue_guard.deref_mut() {
2234 0 : UploadQueue::Uninitialized => panic!("callers are responsible for ensuring this is only called on an initialized queue"),
2235 0 : UploadQueue::Stopped(_stopped) => {
2236 0 : None
2237 : },
2238 3053 : UploadQueue::Initialized(qi) => { Some(qi) }
2239 : };
2240 :
2241 3053 : let upload_queue = match upload_queue {
2242 3053 : Some(upload_queue) => upload_queue,
2243 : None => {
2244 0 : info!("another concurrent task already stopped the queue");
2245 0 : return;
2246 : }
2247 : };
2248 :
2249 3053 : upload_queue.inprogress_tasks.remove(&task.task_id);
2250 :
2251 3053 : let lsn_update = match task.op {
2252 : UploadOp::UploadLayer(_, _, _) => {
2253 1345 : upload_queue.num_inprogress_layer_uploads -= 1;
2254 1345 : None
2255 : }
2256 1452 : UploadOp::UploadMetadata { ref uploaded } => {
2257 1452 : upload_queue.num_inprogress_metadata_uploads -= 1;
2258 1452 :
2259 1452 : // the task id is reused as a monotonicity check for storing the "clean"
2260 1452 : // IndexPart.
2261 1452 : let last_updater = upload_queue.clean.1;
2262 1452 : let is_later = last_updater.is_some_and(|task_id| task_id < task.task_id);
2263 1452 : let monotone = is_later || last_updater.is_none();
2264 :
2265 1452 : assert!(monotone, "no two index uploads should be completing at the same time, prev={last_updater:?}, task.task_id={}", task.task_id);
2266 :
2267 : // not taking ownership is wasteful
2268 1452 : upload_queue.clean.0.clone_from(uploaded);
2269 1452 : upload_queue.clean.1 = Some(task.task_id);
2270 1452 :
2271 1452 : let lsn = upload_queue.clean.0.metadata.disk_consistent_lsn();
2272 1452 : self.metrics
2273 1452 : .projected_remote_consistent_lsn_gauge
2274 1452 : .set(lsn.0);
2275 1452 :
2276 1452 : if self.generation.is_none() {
2277 : // Legacy mode: skip validating generation
2278 0 : upload_queue.visible_remote_consistent_lsn.store(lsn);
2279 0 : None
2280 1452 : } else if self
2281 1452 : .config
2282 1452 : .read()
2283 1452 : .unwrap()
2284 1452 : .process_remote_consistent_lsn_updates
2285 : {
2286 1452 : Some((lsn, upload_queue.visible_remote_consistent_lsn.clone()))
2287 : } else {
2288 : // Our config disables remote_consistent_lsn updates: drop it.
2289 0 : None
2290 : }
2291 : }
2292 : UploadOp::Delete(_) => {
2293 256 : upload_queue.num_inprogress_deletions -= 1;
2294 256 : None
2295 : }
2296 0 : UploadOp::Barrier(..) | UploadOp::Shutdown => unreachable!(),
2297 : };
2298 :
2299 : // Launch any queued tasks that were unblocked by this one.
2300 3053 : self.launch_queued_tasks(upload_queue);
2301 3053 : lsn_update
2302 : };
2303 :
2304 3053 : if let Some((lsn, slot)) = lsn_update {
2305 : // Updates to the remote_consistent_lsn we advertise to pageservers
2306 : // are all routed through the DeletionQueue, to enforce important
2307 : // data safety guarantees (see docs/rfcs/025-generation-numbers.md)
2308 1452 : self.deletion_queue_client
2309 1452 : .update_remote_consistent_lsn(
2310 1452 : self.tenant_shard_id,
2311 1452 : self.timeline_id,
2312 1452 : self.generation,
2313 1452 : lsn,
2314 1452 : slot,
2315 1452 : )
2316 1452 : .await;
2317 1601 : }
2318 :
2319 3053 : self.metric_end(&task.op);
2320 3053 : }
2321 :
2322 6641 : fn metric_impl(
2323 6641 : &self,
2324 6641 : op: &UploadOp,
2325 6641 : ) -> Option<(
2326 6641 : RemoteOpFileKind,
2327 6641 : RemoteOpKind,
2328 6641 : RemoteTimelineClientMetricsCallTrackSize,
2329 6641 : )> {
2330 : use RemoteTimelineClientMetricsCallTrackSize::DontTrackSize;
2331 6641 : let res = match op {
2332 2901 : UploadOp::UploadLayer(_, m, _) => (
2333 2901 : RemoteOpFileKind::Layer,
2334 2901 : RemoteOpKind::Upload,
2335 2901 : RemoteTimelineClientMetricsCallTrackSize::Bytes(m.file_size),
2336 2901 : ),
2337 2966 : UploadOp::UploadMetadata { .. } => (
2338 2966 : RemoteOpFileKind::Index,
2339 2966 : RemoteOpKind::Upload,
2340 2966 : DontTrackSize {
2341 2966 : reason: "metadata uploads are tiny",
2342 2966 : },
2343 2966 : ),
2344 766 : UploadOp::Delete(_delete) => (
2345 766 : RemoteOpFileKind::Layer,
2346 766 : RemoteOpKind::Delete,
2347 766 : DontTrackSize {
2348 766 : reason: "should we track deletes? positive or negative sign?",
2349 766 : },
2350 766 : ),
2351 : UploadOp::Barrier(..) | UploadOp::Shutdown => {
2352 : // we do not account these
2353 8 : return None;
2354 : }
2355 : };
2356 6633 : Some(res)
2357 6641 : }
2358 :
2359 3580 : fn metric_begin(&self, op: &UploadOp) {
2360 3580 : let (file_kind, op_kind, track_bytes) = match self.metric_impl(op) {
2361 3580 : Some(x) => x,
2362 0 : None => return,
2363 : };
2364 3580 : let guard = self.metrics.call_begin(&file_kind, &op_kind, track_bytes);
2365 3580 : guard.will_decrement_manually(); // in metric_end(), see right below
2366 3580 : }
2367 :
2368 3061 : fn metric_end(&self, op: &UploadOp) {
2369 3061 : let (file_kind, op_kind, track_bytes) = match self.metric_impl(op) {
2370 3053 : Some(x) => x,
2371 8 : None => return,
2372 : };
2373 3053 : self.metrics.call_end(&file_kind, &op_kind, track_bytes);
2374 3061 : }
2375 :
2376 : /// Close the upload queue for new operations and cancel queued operations.
2377 : ///
2378 : /// Use [`RemoteTimelineClient::shutdown`] for graceful stop.
2379 : ///
2380 : /// In-progress operations will still be running after this function returns.
2381 : /// Use `task_mgr::shutdown_tasks(Some(TaskKind::RemoteUploadTask), Some(self.tenant_shard_id), Some(timeline_id))`
2382 : /// to wait for them to complete, after calling this function.
2383 18 : pub(crate) fn stop(&self) {
2384 18 : // Whichever *task* for this RemoteTimelineClient grabs the mutex first will transition the queue
2385 18 : // into stopped state, thereby dropping all off the queued *ops* which haven't become *tasks* yet.
2386 18 : // The other *tasks* will come here and observe an already shut down queue and hence simply wrap up their business.
2387 18 : let mut guard = self.upload_queue.lock().unwrap();
2388 18 : self.stop_impl(&mut guard);
2389 18 : }
2390 :
2391 18 : fn stop_impl(&self, guard: &mut std::sync::MutexGuard<UploadQueue>) {
2392 18 : match &mut **guard {
2393 : UploadQueue::Uninitialized => {
2394 0 : info!("UploadQueue is in state Uninitialized, nothing to do");
2395 0 : **guard = UploadQueue::Stopped(UploadQueueStopped::Uninitialized);
2396 : }
2397 : UploadQueue::Stopped(_) => {
2398 : // nothing to do
2399 8 : info!("another concurrent task already shut down the queue");
2400 : }
2401 10 : UploadQueue::Initialized(initialized) => {
2402 10 : info!("shutting down upload queue");
2403 :
2404 : // Replace the queue with the Stopped state, taking ownership of the old
2405 : // Initialized queue. We will do some checks on it, and then drop it.
2406 10 : let qi = {
2407 : // Here we preserve working version of the upload queue for possible use during deletions.
2408 : // In-place replace of Initialized to Stopped can be done with the help of https://github.com/Sgeo/take_mut
2409 : // but for this use case it doesnt really makes sense to bring unsafe code only for this usage point.
2410 : // Deletion is not really perf sensitive so there shouldnt be any problems with cloning a fraction of it.
2411 10 : let upload_queue_for_deletion = UploadQueueInitialized {
2412 10 : task_counter: 0,
2413 10 : dirty: initialized.dirty.clone(),
2414 10 : clean: initialized.clean.clone(),
2415 10 : latest_files_changes_since_metadata_upload_scheduled: 0,
2416 10 : visible_remote_consistent_lsn: initialized
2417 10 : .visible_remote_consistent_lsn
2418 10 : .clone(),
2419 10 : num_inprogress_layer_uploads: 0,
2420 10 : num_inprogress_metadata_uploads: 0,
2421 10 : num_inprogress_deletions: 0,
2422 10 : inprogress_tasks: HashMap::default(),
2423 10 : queued_operations: VecDeque::default(),
2424 10 : #[cfg(feature = "testing")]
2425 10 : dangling_files: HashMap::default(),
2426 10 : blocked_deletions: Vec::new(),
2427 10 : shutting_down: false,
2428 10 : shutdown_ready: Arc::new(tokio::sync::Semaphore::new(0)),
2429 10 : recently_deleted: HashSet::new(),
2430 10 : };
2431 10 :
2432 10 : let upload_queue = std::mem::replace(
2433 10 : &mut **guard,
2434 10 : UploadQueue::Stopped(UploadQueueStopped::Deletable(
2435 10 : UploadQueueStoppedDeletable {
2436 10 : upload_queue_for_deletion,
2437 10 : deleted_at: SetDeletedFlagProgress::NotRunning,
2438 10 : },
2439 10 : )),
2440 10 : );
2441 10 : if let UploadQueue::Initialized(qi) = upload_queue {
2442 10 : qi
2443 : } else {
2444 0 : unreachable!("we checked in the match above that it is Initialized");
2445 : }
2446 : };
2447 :
2448 : // consistency check
2449 10 : assert_eq!(
2450 10 : qi.num_inprogress_layer_uploads
2451 10 : + qi.num_inprogress_metadata_uploads
2452 10 : + qi.num_inprogress_deletions,
2453 10 : qi.inprogress_tasks.len()
2454 10 : );
2455 :
2456 : // We don't need to do anything here for in-progress tasks. They will finish
2457 : // on their own, decrement the unfinished-task counter themselves, and observe
2458 : // that the queue is Stopped.
2459 10 : drop(qi.inprogress_tasks);
2460 :
2461 : // Tear down queued ops
2462 10 : for op in qi.queued_operations.into_iter() {
2463 8 : self.metric_end(&op);
2464 8 : // Dropping UploadOp::Barrier() here will make wait_completion() return with an Err()
2465 8 : // which is exactly what we want to happen.
2466 8 : drop(op);
2467 8 : }
2468 : }
2469 : }
2470 18 : }
2471 :
2472 : /// Returns an accessor which will hold the UploadQueue mutex for accessing the upload queue
2473 : /// externally to RemoteTimelineClient.
2474 0 : pub(crate) fn initialized_upload_queue(
2475 0 : &self,
2476 0 : ) -> Result<UploadQueueAccessor<'_>, NotInitialized> {
2477 0 : let mut inner = self.upload_queue.lock().unwrap();
2478 0 : inner.initialized_mut()?;
2479 0 : Ok(UploadQueueAccessor { inner })
2480 0 : }
2481 :
2482 8 : pub(crate) fn no_pending_work(&self) -> bool {
2483 8 : let inner = self.upload_queue.lock().unwrap();
2484 8 : match &*inner {
2485 : UploadQueue::Uninitialized
2486 0 : | UploadQueue::Stopped(UploadQueueStopped::Uninitialized) => true,
2487 8 : UploadQueue::Stopped(UploadQueueStopped::Deletable(x)) => {
2488 8 : x.upload_queue_for_deletion.no_pending_work()
2489 : }
2490 0 : UploadQueue::Initialized(x) => x.no_pending_work(),
2491 : }
2492 8 : }
2493 :
2494 : /// 'foreign' in the sense that it does not belong to this tenant shard. This method
2495 : /// is used during GC for other shards to get the index of shard zero.
2496 0 : pub(crate) async fn download_foreign_index(
2497 0 : &self,
2498 0 : shard_number: ShardNumber,
2499 0 : cancel: &CancellationToken,
2500 0 : ) -> Result<(IndexPart, Generation, std::time::SystemTime), DownloadError> {
2501 0 : let foreign_shard_id = TenantShardId {
2502 0 : shard_number,
2503 0 : shard_count: self.tenant_shard_id.shard_count,
2504 0 : tenant_id: self.tenant_shard_id.tenant_id,
2505 0 : };
2506 0 : download_index_part(
2507 0 : &self.storage_impl,
2508 0 : &foreign_shard_id,
2509 0 : &self.timeline_id,
2510 0 : Generation::MAX,
2511 0 : cancel,
2512 0 : )
2513 0 : .await
2514 0 : }
2515 : }
2516 :
2517 : pub(crate) struct UploadQueueAccessor<'a> {
2518 : inner: std::sync::MutexGuard<'a, UploadQueue>,
2519 : }
2520 :
2521 : impl UploadQueueAccessor<'_> {
2522 0 : pub(crate) fn latest_uploaded_index_part(&self) -> &IndexPart {
2523 0 : match &*self.inner {
2524 0 : UploadQueue::Initialized(x) => &x.clean.0,
2525 : UploadQueue::Uninitialized | UploadQueue::Stopped(_) => {
2526 0 : unreachable!("checked before constructing")
2527 : }
2528 : }
2529 0 : }
2530 : }
2531 :
2532 0 : pub fn remote_tenant_path(tenant_shard_id: &TenantShardId) -> RemotePath {
2533 0 : let path = format!("tenants/{tenant_shard_id}");
2534 0 : RemotePath::from_string(&path).expect("Failed to construct path")
2535 0 : }
2536 :
2537 590 : pub fn remote_tenant_manifest_path(
2538 590 : tenant_shard_id: &TenantShardId,
2539 590 : generation: Generation,
2540 590 : ) -> RemotePath {
2541 590 : let path = format!(
2542 590 : "tenants/{tenant_shard_id}/tenant-manifest{}.json",
2543 590 : generation.get_suffix()
2544 590 : );
2545 590 : RemotePath::from_string(&path).expect("Failed to construct path")
2546 590 : }
2547 :
2548 : /// Prefix to all generations' manifest objects in a tenant shard
2549 196 : pub fn remote_tenant_manifest_prefix(tenant_shard_id: &TenantShardId) -> RemotePath {
2550 196 : let path = format!("tenants/{tenant_shard_id}/tenant-manifest",);
2551 196 : RemotePath::from_string(&path).expect("Failed to construct path")
2552 196 : }
2553 :
2554 226 : pub fn remote_timelines_path(tenant_shard_id: &TenantShardId) -> RemotePath {
2555 226 : let path = format!("tenants/{tenant_shard_id}/{TIMELINES_SEGMENT_NAME}");
2556 226 : RemotePath::from_string(&path).expect("Failed to construct path")
2557 226 : }
2558 :
2559 0 : fn remote_timelines_path_unsharded(tenant_id: &TenantId) -> RemotePath {
2560 0 : let path = format!("tenants/{tenant_id}/{TIMELINES_SEGMENT_NAME}");
2561 0 : RemotePath::from_string(&path).expect("Failed to construct path")
2562 0 : }
2563 :
2564 30 : pub fn remote_timeline_path(
2565 30 : tenant_shard_id: &TenantShardId,
2566 30 : timeline_id: &TimelineId,
2567 30 : ) -> RemotePath {
2568 30 : remote_timelines_path(tenant_shard_id).join(Utf8Path::new(&timeline_id.to_string()))
2569 30 : }
2570 :
2571 : /// Obtains the path of the given Layer in the remote
2572 : ///
2573 : /// Note that the shard component of a remote layer path is _not_ always the same
2574 : /// as in the TenantShardId of the caller: tenants may reference layers from a different
2575 : /// ShardIndex. Use the ShardIndex from the layer's metadata.
2576 1715 : pub fn remote_layer_path(
2577 1715 : tenant_id: &TenantId,
2578 1715 : timeline_id: &TimelineId,
2579 1715 : shard: ShardIndex,
2580 1715 : layer_file_name: &LayerName,
2581 1715 : generation: Generation,
2582 1715 : ) -> RemotePath {
2583 1715 : // Generation-aware key format
2584 1715 : let path = format!(
2585 1715 : "tenants/{tenant_id}{0}/{TIMELINES_SEGMENT_NAME}/{timeline_id}/{1}{2}",
2586 1715 : shard.get_suffix(),
2587 1715 : layer_file_name,
2588 1715 : generation.get_suffix()
2589 1715 : );
2590 1715 :
2591 1715 : RemotePath::from_string(&path).expect("Failed to construct path")
2592 1715 : }
2593 :
2594 : /// Returns true if a and b have the same layer path within a tenant/timeline. This is essentially
2595 : /// remote_layer_path(a) == remote_layer_path(b) without the string allocations.
2596 : ///
2597 : /// TODO: there should be a variant of LayerName for the physical path that contains information
2598 : /// about the shard and generation, such that this could be replaced by a simple comparison.
2599 4 : pub fn is_same_remote_layer_path(
2600 4 : aname: &LayerName,
2601 4 : ameta: &LayerFileMetadata,
2602 4 : bname: &LayerName,
2603 4 : bmeta: &LayerFileMetadata,
2604 4 : ) -> bool {
2605 4 : // NB: don't assert remote_layer_path(a) == remote_layer_path(b); too expensive even for debug.
2606 4 : aname == bname && ameta.shard == bmeta.shard && ameta.generation == bmeta.generation
2607 4 : }
2608 :
2609 4 : pub fn remote_initdb_archive_path(tenant_id: &TenantId, timeline_id: &TimelineId) -> RemotePath {
2610 4 : RemotePath::from_string(&format!(
2611 4 : "tenants/{tenant_id}/{TIMELINES_SEGMENT_NAME}/{timeline_id}/{INITDB_PATH}"
2612 4 : ))
2613 4 : .expect("Failed to construct path")
2614 4 : }
2615 :
2616 2 : pub fn remote_initdb_preserved_archive_path(
2617 2 : tenant_id: &TenantId,
2618 2 : timeline_id: &TimelineId,
2619 2 : ) -> RemotePath {
2620 2 : RemotePath::from_string(&format!(
2621 2 : "tenants/{tenant_id}/{TIMELINES_SEGMENT_NAME}/{timeline_id}/{INITDB_PRESERVED_PATH}"
2622 2 : ))
2623 2 : .expect("Failed to construct path")
2624 2 : }
2625 :
2626 1525 : pub fn remote_index_path(
2627 1525 : tenant_shard_id: &TenantShardId,
2628 1525 : timeline_id: &TimelineId,
2629 1525 : generation: Generation,
2630 1525 : ) -> RemotePath {
2631 1525 : RemotePath::from_string(&format!(
2632 1525 : "tenants/{tenant_shard_id}/{TIMELINES_SEGMENT_NAME}/{timeline_id}/{0}{1}",
2633 1525 : IndexPart::FILE_NAME,
2634 1525 : generation.get_suffix()
2635 1525 : ))
2636 1525 : .expect("Failed to construct path")
2637 1525 : }
2638 :
2639 0 : pub(crate) fn remote_heatmap_path(tenant_shard_id: &TenantShardId) -> RemotePath {
2640 0 : RemotePath::from_string(&format!(
2641 0 : "tenants/{tenant_shard_id}/{TENANT_HEATMAP_BASENAME}"
2642 0 : ))
2643 0 : .expect("Failed to construct path")
2644 0 : }
2645 :
2646 : /// Given the key of an index, parse out the generation part of the name
2647 18 : pub fn parse_remote_index_path(path: RemotePath) -> Option<Generation> {
2648 18 : let file_name = match path.get_path().file_name() {
2649 18 : Some(f) => f,
2650 : None => {
2651 : // Unexpected: we should be seeing index_part.json paths only
2652 0 : tracing::warn!("Malformed index key {}", path);
2653 0 : return None;
2654 : }
2655 : };
2656 :
2657 18 : match file_name.split_once('-') {
2658 12 : Some((_, gen_suffix)) => Generation::parse_suffix(gen_suffix),
2659 6 : None => None,
2660 : }
2661 18 : }
2662 :
2663 : /// Given the key of a tenant manifest, parse out the generation number
2664 0 : pub fn parse_remote_tenant_manifest_path(path: RemotePath) -> Option<Generation> {
2665 : static RE: OnceLock<Regex> = OnceLock::new();
2666 0 : let re = RE.get_or_init(|| Regex::new(r".*tenant-manifest-([0-9a-f]{8}).json").unwrap());
2667 0 : re.captures(path.get_path().as_str())
2668 0 : .and_then(|c| c.get(1))
2669 0 : .and_then(|m| Generation::parse_suffix(m.as_str()))
2670 0 : }
2671 :
2672 : #[cfg(test)]
2673 : mod tests {
2674 : use super::*;
2675 : use crate::{
2676 : context::RequestContext,
2677 : tenant::{
2678 : config::AttachmentMode,
2679 : harness::{TenantHarness, TIMELINE_ID},
2680 : storage_layer::layer::local_layer_path,
2681 : Tenant, Timeline,
2682 : },
2683 : DEFAULT_PG_VERSION,
2684 : };
2685 :
2686 : use std::collections::HashSet;
2687 :
2688 8 : pub(super) fn dummy_contents(name: &str) -> Vec<u8> {
2689 8 : format!("contents for {name}").into()
2690 8 : }
2691 :
2692 2 : pub(super) fn dummy_metadata(disk_consistent_lsn: Lsn) -> TimelineMetadata {
2693 2 : let metadata = TimelineMetadata::new(
2694 2 : disk_consistent_lsn,
2695 2 : None,
2696 2 : None,
2697 2 : Lsn(0),
2698 2 : Lsn(0),
2699 2 : Lsn(0),
2700 2 : // Any version will do
2701 2 : // but it should be consistent with the one in the tests
2702 2 : crate::DEFAULT_PG_VERSION,
2703 2 : );
2704 2 :
2705 2 : // go through serialize + deserialize to fix the header, including checksum
2706 2 : TimelineMetadata::from_bytes(&metadata.to_bytes().unwrap()).unwrap()
2707 2 : }
2708 :
2709 2 : fn assert_file_list(a: &HashSet<LayerName>, b: &[&str]) {
2710 6 : let mut avec: Vec<String> = a.iter().map(|x| x.to_string()).collect();
2711 2 : avec.sort();
2712 2 :
2713 2 : let mut bvec = b.to_vec();
2714 2 : bvec.sort_unstable();
2715 2 :
2716 2 : assert_eq!(avec, bvec);
2717 2 : }
2718 :
2719 4 : fn assert_remote_files(expected: &[&str], remote_path: &Utf8Path, generation: Generation) {
2720 4 : let mut expected: Vec<String> = expected
2721 4 : .iter()
2722 16 : .map(|x| format!("{}{}", x, generation.get_suffix()))
2723 4 : .collect();
2724 4 : expected.sort();
2725 4 :
2726 4 : let mut found: Vec<String> = Vec::new();
2727 16 : for entry in std::fs::read_dir(remote_path).unwrap().flatten() {
2728 16 : let entry_name = entry.file_name();
2729 16 : let fname = entry_name.to_str().unwrap();
2730 16 : found.push(String::from(fname));
2731 16 : }
2732 4 : found.sort();
2733 4 :
2734 4 : assert_eq!(found, expected);
2735 4 : }
2736 :
2737 : struct TestSetup {
2738 : harness: TenantHarness,
2739 : tenant: Arc<Tenant>,
2740 : timeline: Arc<Timeline>,
2741 : tenant_ctx: RequestContext,
2742 : }
2743 :
2744 : impl TestSetup {
2745 8 : async fn new(test_name: &str) -> anyhow::Result<Self> {
2746 8 : let test_name = Box::leak(Box::new(format!("remote_timeline_client__{test_name}")));
2747 8 : let harness = TenantHarness::create(test_name).await?;
2748 8 : let (tenant, ctx) = harness.load().await;
2749 :
2750 8 : let timeline = tenant
2751 8 : .create_test_timeline(TIMELINE_ID, Lsn(8), DEFAULT_PG_VERSION, &ctx)
2752 8 : .await?;
2753 :
2754 8 : Ok(Self {
2755 8 : harness,
2756 8 : tenant,
2757 8 : timeline,
2758 8 : tenant_ctx: ctx,
2759 8 : })
2760 8 : }
2761 :
2762 : /// Construct a RemoteTimelineClient in an arbitrary generation
2763 10 : fn build_client(&self, generation: Generation) -> Arc<RemoteTimelineClient> {
2764 10 : let location_conf = AttachedLocationConfig {
2765 10 : generation,
2766 10 : attach_mode: AttachmentMode::Single,
2767 10 : };
2768 10 : Arc::new(RemoteTimelineClient {
2769 10 : conf: self.harness.conf,
2770 10 : runtime: tokio::runtime::Handle::current(),
2771 10 : tenant_shard_id: self.harness.tenant_shard_id,
2772 10 : timeline_id: TIMELINE_ID,
2773 10 : generation,
2774 10 : storage_impl: self.harness.remote_storage.clone(),
2775 10 : deletion_queue_client: self.harness.deletion_queue.new_client(),
2776 10 : upload_queue: Mutex::new(UploadQueue::Uninitialized),
2777 10 : metrics: Arc::new(RemoteTimelineClientMetrics::new(
2778 10 : &self.harness.tenant_shard_id,
2779 10 : &TIMELINE_ID,
2780 10 : )),
2781 10 : config: std::sync::RwLock::new(RemoteTimelineClientConfig::from(&location_conf)),
2782 10 : cancel: CancellationToken::new(),
2783 10 : })
2784 10 : }
2785 :
2786 : /// A tracing::Span that satisfies remote_timeline_client methods that assert tenant_id
2787 : /// and timeline_id are present.
2788 6 : fn span(&self) -> tracing::Span {
2789 6 : tracing::info_span!(
2790 : "test",
2791 : tenant_id = %self.harness.tenant_shard_id.tenant_id,
2792 0 : shard_id = %self.harness.tenant_shard_id.shard_slug(),
2793 : timeline_id = %TIMELINE_ID
2794 : )
2795 6 : }
2796 : }
2797 :
2798 : // Test scheduling
2799 : #[tokio::test]
2800 2 : async fn upload_scheduling() {
2801 2 : // Test outline:
2802 2 : //
2803 2 : // Schedule upload of a bunch of layers. Check that they are started immediately, not queued
2804 2 : // Schedule upload of index. Check that it is queued
2805 2 : // let the layer file uploads finish. Check that the index-upload is now started
2806 2 : // let the index-upload finish.
2807 2 : //
2808 2 : // Download back the index.json. Check that the list of files is correct
2809 2 : //
2810 2 : // Schedule upload. Schedule deletion. Check that the deletion is queued
2811 2 : // let upload finish. Check that deletion is now started
2812 2 : // Schedule another deletion. Check that it's launched immediately.
2813 2 : // Schedule index upload. Check that it's queued
2814 2 :
2815 2 : let test_setup = TestSetup::new("upload_scheduling").await.unwrap();
2816 2 : let span = test_setup.span();
2817 2 : let _guard = span.enter();
2818 2 :
2819 2 : let TestSetup {
2820 2 : harness,
2821 2 : tenant: _tenant,
2822 2 : timeline,
2823 2 : tenant_ctx: _tenant_ctx,
2824 2 : } = test_setup;
2825 2 :
2826 2 : let client = &timeline.remote_client;
2827 2 :
2828 2 : // Download back the index.json, and check that the list of files is correct
2829 2 : let initial_index_part = match client
2830 2 : .download_index_file(&CancellationToken::new())
2831 2 : .await
2832 2 : .unwrap()
2833 2 : {
2834 2 : MaybeDeletedIndexPart::IndexPart(index_part) => index_part,
2835 2 : MaybeDeletedIndexPart::Deleted(_) => panic!("unexpectedly got deleted index part"),
2836 2 : };
2837 2 : let initial_layers = initial_index_part
2838 2 : .layer_metadata
2839 2 : .keys()
2840 2 : .map(|f| f.to_owned())
2841 2 : .collect::<HashSet<LayerName>>();
2842 2 : let initial_layer = {
2843 2 : assert!(initial_layers.len() == 1);
2844 2 : initial_layers.into_iter().next().unwrap()
2845 2 : };
2846 2 :
2847 2 : let timeline_path = harness.timeline_path(&TIMELINE_ID);
2848 2 :
2849 2 : println!("workdir: {}", harness.conf.workdir);
2850 2 :
2851 2 : let remote_timeline_dir = harness
2852 2 : .remote_fs_dir
2853 2 : .join(timeline_path.strip_prefix(&harness.conf.workdir).unwrap());
2854 2 : println!("remote_timeline_dir: {remote_timeline_dir}");
2855 2 :
2856 2 : let generation = harness.generation;
2857 2 : let shard = harness.shard;
2858 2 :
2859 2 : // Create a couple of dummy files, schedule upload for them
2860 2 :
2861 2 : let layers = [
2862 2 : ("000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__00000000016B59D8-00000000016B5A51".parse().unwrap(), dummy_contents("foo")),
2863 2 : ("000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__00000000016B59D9-00000000016B5A52".parse().unwrap(), dummy_contents("bar")),
2864 2 : ("000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__00000000016B59DA-00000000016B5A53".parse().unwrap(), dummy_contents("baz"))
2865 2 : ]
2866 2 : .into_iter()
2867 6 : .map(|(name, contents): (LayerName, Vec<u8>)| {
2868 6 :
2869 6 : let local_path = local_layer_path(
2870 6 : harness.conf,
2871 6 : &timeline.tenant_shard_id,
2872 6 : &timeline.timeline_id,
2873 6 : &name,
2874 6 : &generation,
2875 6 : );
2876 6 : std::fs::write(&local_path, &contents).unwrap();
2877 6 :
2878 6 : Layer::for_resident(
2879 6 : harness.conf,
2880 6 : &timeline,
2881 6 : local_path,
2882 6 : name,
2883 6 : LayerFileMetadata::new(contents.len() as u64, generation, shard),
2884 6 : )
2885 6 : }).collect::<Vec<_>>();
2886 2 :
2887 2 : client
2888 2 : .schedule_layer_file_upload(layers[0].clone())
2889 2 : .unwrap();
2890 2 : client
2891 2 : .schedule_layer_file_upload(layers[1].clone())
2892 2 : .unwrap();
2893 2 :
2894 2 : // Check that they are started immediately, not queued
2895 2 : //
2896 2 : // this works because we running within block_on, so any futures are now queued up until
2897 2 : // our next await point.
2898 2 : {
2899 2 : let mut guard = client.upload_queue.lock().unwrap();
2900 2 : let upload_queue = guard.initialized_mut().unwrap();
2901 2 : assert!(upload_queue.queued_operations.is_empty());
2902 2 : assert!(upload_queue.inprogress_tasks.len() == 2);
2903 2 : assert!(upload_queue.num_inprogress_layer_uploads == 2);
2904 2 :
2905 2 : // also check that `latest_file_changes` was updated
2906 2 : assert!(upload_queue.latest_files_changes_since_metadata_upload_scheduled == 2);
2907 2 : }
2908 2 :
2909 2 : // Schedule upload of index. Check that it is queued
2910 2 : let metadata = dummy_metadata(Lsn(0x20));
2911 2 : client
2912 2 : .schedule_index_upload_for_full_metadata_update(&metadata)
2913 2 : .unwrap();
2914 2 : {
2915 2 : let mut guard = client.upload_queue.lock().unwrap();
2916 2 : let upload_queue = guard.initialized_mut().unwrap();
2917 2 : assert!(upload_queue.queued_operations.len() == 1);
2918 2 : assert!(upload_queue.latest_files_changes_since_metadata_upload_scheduled == 0);
2919 2 : }
2920 2 :
2921 2 : // Wait for the uploads to finish
2922 2 : client.wait_completion().await.unwrap();
2923 2 : {
2924 2 : let mut guard = client.upload_queue.lock().unwrap();
2925 2 : let upload_queue = guard.initialized_mut().unwrap();
2926 2 :
2927 2 : assert!(upload_queue.queued_operations.is_empty());
2928 2 : assert!(upload_queue.inprogress_tasks.is_empty());
2929 2 : }
2930 2 :
2931 2 : // Download back the index.json, and check that the list of files is correct
2932 2 : let index_part = match client
2933 2 : .download_index_file(&CancellationToken::new())
2934 2 : .await
2935 2 : .unwrap()
2936 2 : {
2937 2 : MaybeDeletedIndexPart::IndexPart(index_part) => index_part,
2938 2 : MaybeDeletedIndexPart::Deleted(_) => panic!("unexpectedly got deleted index part"),
2939 2 : };
2940 2 :
2941 2 : assert_file_list(
2942 2 : &index_part
2943 2 : .layer_metadata
2944 2 : .keys()
2945 6 : .map(|f| f.to_owned())
2946 2 : .collect(),
2947 2 : &[
2948 2 : &initial_layer.to_string(),
2949 2 : &layers[0].layer_desc().layer_name().to_string(),
2950 2 : &layers[1].layer_desc().layer_name().to_string(),
2951 2 : ],
2952 2 : );
2953 2 : assert_eq!(index_part.metadata, metadata);
2954 2 :
2955 2 : // Schedule upload and then a deletion. Check that the deletion is queued
2956 2 : client
2957 2 : .schedule_layer_file_upload(layers[2].clone())
2958 2 : .unwrap();
2959 2 :
2960 2 : // this is no longer consistent with how deletion works with Layer::drop, but in this test
2961 2 : // keep using schedule_layer_file_deletion because we don't have a way to wait for the
2962 2 : // spawn_blocking started by the drop.
2963 2 : client
2964 2 : .schedule_layer_file_deletion(&[layers[0].layer_desc().layer_name()])
2965 2 : .unwrap();
2966 2 : {
2967 2 : let mut guard = client.upload_queue.lock().unwrap();
2968 2 : let upload_queue = guard.initialized_mut().unwrap();
2969 2 :
2970 2 : // Deletion schedules upload of the index file, and the file deletion itself
2971 2 : assert_eq!(upload_queue.queued_operations.len(), 2);
2972 2 : assert_eq!(upload_queue.inprogress_tasks.len(), 1);
2973 2 : assert_eq!(upload_queue.num_inprogress_layer_uploads, 1);
2974 2 : assert_eq!(upload_queue.num_inprogress_deletions, 0);
2975 2 : assert_eq!(
2976 2 : upload_queue.latest_files_changes_since_metadata_upload_scheduled,
2977 2 : 0
2978 2 : );
2979 2 : }
2980 2 : assert_remote_files(
2981 2 : &[
2982 2 : &initial_layer.to_string(),
2983 2 : &layers[0].layer_desc().layer_name().to_string(),
2984 2 : &layers[1].layer_desc().layer_name().to_string(),
2985 2 : "index_part.json",
2986 2 : ],
2987 2 : &remote_timeline_dir,
2988 2 : generation,
2989 2 : );
2990 2 :
2991 2 : // Finish them
2992 2 : client.wait_completion().await.unwrap();
2993 2 : harness.deletion_queue.pump().await;
2994 2 :
2995 2 : assert_remote_files(
2996 2 : &[
2997 2 : &initial_layer.to_string(),
2998 2 : &layers[1].layer_desc().layer_name().to_string(),
2999 2 : &layers[2].layer_desc().layer_name().to_string(),
3000 2 : "index_part.json",
3001 2 : ],
3002 2 : &remote_timeline_dir,
3003 2 : generation,
3004 2 : );
3005 2 : }
3006 :
3007 : #[tokio::test]
3008 2 : async fn bytes_unfinished_gauge_for_layer_file_uploads() {
3009 2 : // Setup
3010 2 :
3011 2 : let TestSetup {
3012 2 : harness,
3013 2 : tenant: _tenant,
3014 2 : timeline,
3015 2 : ..
3016 2 : } = TestSetup::new("metrics").await.unwrap();
3017 2 : let client = &timeline.remote_client;
3018 2 :
3019 2 : let layer_file_name_1: LayerName = "000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__00000000016B59D8-00000000016B5A51".parse().unwrap();
3020 2 : let local_path = local_layer_path(
3021 2 : harness.conf,
3022 2 : &timeline.tenant_shard_id,
3023 2 : &timeline.timeline_id,
3024 2 : &layer_file_name_1,
3025 2 : &harness.generation,
3026 2 : );
3027 2 : let content_1 = dummy_contents("foo");
3028 2 : std::fs::write(&local_path, &content_1).unwrap();
3029 2 :
3030 2 : let layer_file_1 = Layer::for_resident(
3031 2 : harness.conf,
3032 2 : &timeline,
3033 2 : local_path,
3034 2 : layer_file_name_1.clone(),
3035 2 : LayerFileMetadata::new(content_1.len() as u64, harness.generation, harness.shard),
3036 2 : );
3037 2 :
3038 2 : #[derive(Debug, PartialEq, Clone, Copy)]
3039 2 : struct BytesStartedFinished {
3040 2 : started: Option<usize>,
3041 2 : finished: Option<usize>,
3042 2 : }
3043 2 : impl std::ops::Add for BytesStartedFinished {
3044 2 : type Output = Self;
3045 4 : fn add(self, rhs: Self) -> Self::Output {
3046 4 : Self {
3047 4 : started: self.started.map(|v| v + rhs.started.unwrap_or(0)),
3048 4 : finished: self.finished.map(|v| v + rhs.finished.unwrap_or(0)),
3049 4 : }
3050 4 : }
3051 2 : }
3052 6 : let get_bytes_started_stopped = || {
3053 6 : let started = client
3054 6 : .metrics
3055 6 : .get_bytes_started_counter_value(&RemoteOpFileKind::Layer, &RemoteOpKind::Upload)
3056 6 : .map(|v| v.try_into().unwrap());
3057 6 : let stopped = client
3058 6 : .metrics
3059 6 : .get_bytes_finished_counter_value(&RemoteOpFileKind::Layer, &RemoteOpKind::Upload)
3060 6 : .map(|v| v.try_into().unwrap());
3061 6 : BytesStartedFinished {
3062 6 : started,
3063 6 : finished: stopped,
3064 6 : }
3065 6 : };
3066 2 :
3067 2 : // Test
3068 2 : tracing::info!("now doing actual test");
3069 2 :
3070 2 : let actual_a = get_bytes_started_stopped();
3071 2 :
3072 2 : client
3073 2 : .schedule_layer_file_upload(layer_file_1.clone())
3074 2 : .unwrap();
3075 2 :
3076 2 : let actual_b = get_bytes_started_stopped();
3077 2 :
3078 2 : client.wait_completion().await.unwrap();
3079 2 :
3080 2 : let actual_c = get_bytes_started_stopped();
3081 2 :
3082 2 : // Validate
3083 2 :
3084 2 : let expected_b = actual_a
3085 2 : + BytesStartedFinished {
3086 2 : started: Some(content_1.len()),
3087 2 : // assert that the _finished metric is created eagerly so that subtractions work on first sample
3088 2 : finished: Some(0),
3089 2 : };
3090 2 : assert_eq!(actual_b, expected_b);
3091 2 :
3092 2 : let expected_c = actual_a
3093 2 : + BytesStartedFinished {
3094 2 : started: Some(content_1.len()),
3095 2 : finished: Some(content_1.len()),
3096 2 : };
3097 2 : assert_eq!(actual_c, expected_c);
3098 2 : }
3099 :
3100 12 : async fn inject_index_part(test_state: &TestSetup, generation: Generation) -> IndexPart {
3101 12 : // An empty IndexPart, just sufficient to ensure deserialization will succeed
3102 12 : let example_index_part = IndexPart::example();
3103 12 :
3104 12 : let index_part_bytes = serde_json::to_vec(&example_index_part).unwrap();
3105 12 :
3106 12 : let index_path = test_state.harness.remote_fs_dir.join(
3107 12 : remote_index_path(
3108 12 : &test_state.harness.tenant_shard_id,
3109 12 : &TIMELINE_ID,
3110 12 : generation,
3111 12 : )
3112 12 : .get_path(),
3113 12 : );
3114 12 :
3115 12 : std::fs::create_dir_all(index_path.parent().unwrap())
3116 12 : .expect("creating test dir should work");
3117 12 :
3118 12 : eprintln!("Writing {index_path}");
3119 12 : std::fs::write(&index_path, index_part_bytes).unwrap();
3120 12 : example_index_part
3121 12 : }
3122 :
3123 : /// Assert that when a RemoteTimelineclient in generation `get_generation` fetches its
3124 : /// index, the IndexPart returned is equal to `expected`
3125 10 : async fn assert_got_index_part(
3126 10 : test_state: &TestSetup,
3127 10 : get_generation: Generation,
3128 10 : expected: &IndexPart,
3129 10 : ) {
3130 10 : let client = test_state.build_client(get_generation);
3131 :
3132 10 : let download_r = client
3133 10 : .download_index_file(&CancellationToken::new())
3134 10 : .await
3135 10 : .expect("download should always succeed");
3136 10 : assert!(matches!(download_r, MaybeDeletedIndexPart::IndexPart(_)));
3137 10 : match download_r {
3138 10 : MaybeDeletedIndexPart::IndexPart(index_part) => {
3139 10 : assert_eq!(&index_part, expected);
3140 : }
3141 0 : MaybeDeletedIndexPart::Deleted(_index_part) => panic!("Test doesn't set deleted_at"),
3142 : }
3143 10 : }
3144 :
3145 : #[tokio::test]
3146 2 : async fn index_part_download_simple() -> anyhow::Result<()> {
3147 2 : let test_state = TestSetup::new("index_part_download_simple").await.unwrap();
3148 2 : let span = test_state.span();
3149 2 : let _guard = span.enter();
3150 2 :
3151 2 : // Simple case: we are in generation N, load the index from generation N - 1
3152 2 : let generation_n = 5;
3153 2 : let injected = inject_index_part(&test_state, Generation::new(generation_n - 1)).await;
3154 2 :
3155 2 : assert_got_index_part(&test_state, Generation::new(generation_n), &injected).await;
3156 2 :
3157 2 : Ok(())
3158 2 : }
3159 :
3160 : #[tokio::test]
3161 2 : async fn index_part_download_ordering() -> anyhow::Result<()> {
3162 2 : let test_state = TestSetup::new("index_part_download_ordering")
3163 2 : .await
3164 2 : .unwrap();
3165 2 :
3166 2 : let span = test_state.span();
3167 2 : let _guard = span.enter();
3168 2 :
3169 2 : // A generation-less IndexPart exists in the bucket, we should find it
3170 2 : let generation_n = 5;
3171 2 : let injected_none = inject_index_part(&test_state, Generation::none()).await;
3172 2 : assert_got_index_part(&test_state, Generation::new(generation_n), &injected_none).await;
3173 2 :
3174 2 : // If a more recent-than-none generation exists, we should prefer to load that
3175 2 : let injected_1 = inject_index_part(&test_state, Generation::new(1)).await;
3176 2 : assert_got_index_part(&test_state, Generation::new(generation_n), &injected_1).await;
3177 2 :
3178 2 : // If a more-recent-than-me generation exists, we should ignore it.
3179 2 : let _injected_10 = inject_index_part(&test_state, Generation::new(10)).await;
3180 2 : assert_got_index_part(&test_state, Generation::new(generation_n), &injected_1).await;
3181 2 :
3182 2 : // If a directly previous generation exists, _and_ an index exists in my own
3183 2 : // generation, I should prefer my own generation.
3184 2 : let _injected_prev =
3185 2 : inject_index_part(&test_state, Generation::new(generation_n - 1)).await;
3186 2 : let injected_current = inject_index_part(&test_state, Generation::new(generation_n)).await;
3187 2 : assert_got_index_part(
3188 2 : &test_state,
3189 2 : Generation::new(generation_n),
3190 2 : &injected_current,
3191 2 : )
3192 2 : .await;
3193 2 :
3194 2 : Ok(())
3195 2 : }
3196 : }
|
