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 : //!
67 : //! From the user's perspective, the operations are executed sequentially.
68 : //! Internally, the client knows which operations can be performed in parallel,
69 : //! and which operations act like a "barrier" that require preceding operations
70 : //! to finish. The calling code just needs to call the schedule-functions in the
71 : //! correct order, and the client will parallelize the operations in a way that
72 : //! is safe. For more details, see `UploadOp::can_bypass`.
73 : //!
74 : //! All of this relies on the following invariants:
75 : //!
76 : //! - We rely on read-after write consistency in the remote storage.
77 : //! - Layer files are immutable.
78 : //!
79 : //! NB: Pageserver assumes that it has exclusive write access to the tenant in remote
80 : //! storage. Different tenants can be attached to different pageservers, but if the
81 : //! same tenant is attached to two pageservers at the same time, they will overwrite
82 : //! each other's index file updates, and confusion will ensue. There's no interlock or
83 : //! mechanism to detect that in the pageserver, we rely on the control plane to ensure
84 : //! that that doesn't happen.
85 : //!
86 : //! ## Implementation Note
87 : //!
88 : //! The *actual* remote state lags behind the *desired* remote state while
89 : //! there are in-flight operations.
90 : //! We keep track of the desired remote state in [`UploadQueueInitialized::dirty`].
91 : //! It is initialized based on the [`IndexPart`] that was passed during init
92 : //! and updated with every `schedule_*` function call.
93 : //! All this is necessary necessary to compute the future [`IndexPart`]s
94 : //! when scheduling an operation while other operations that also affect the
95 : //! remote [`IndexPart`] are in flight.
96 : //!
97 : //! # Retries & Error Handling
98 : //!
99 : //! The client retries operations indefinitely, using exponential back-off.
100 : //! There is no way to force a retry, i.e., interrupt the back-off.
101 : //! This could be built easily.
102 : //!
103 : //! # Cancellation
104 : //!
105 : //! The operations execute as plain [`task_mgr`] tasks, scoped to
106 : //! the client's tenant and timeline.
107 : //! Dropping the client will drop queued operations but not executing operations.
108 : //! These will complete unless the `task_mgr` tasks are cancelled using `task_mgr`
109 : //! APIs, e.g., during pageserver shutdown, timeline delete, or tenant detach.
110 : //!
111 : //! # Completion
112 : //!
113 : //! Once an operation has completed, we update [`UploadQueueInitialized::clean`] immediately,
114 : //! and submit a request through the DeletionQueue to update
115 : //! [`UploadQueueInitialized::visible_remote_consistent_lsn`] after it has
116 : //! validated that our generation is not stale. It is this visible value
117 : //! that is advertized to safekeepers as a signal that that they can
118 : //! delete the WAL up to that LSN.
119 : //!
120 : //! The [`RemoteTimelineClient::wait_completion`] method can be used to wait
121 : //! for all pending operations to complete. It does not prevent more
122 : //! operations from getting scheduled.
123 : //!
124 : //! # Crash Consistency
125 : //!
126 : //! We do not persist the upload queue state.
127 : //! If we drop the client, or crash, all unfinished operations are lost.
128 : //!
129 : //! To recover, the following steps need to be taken:
130 : //! - Retrieve the current remote [`IndexPart`]. This gives us a
131 : //! consistent remote state, assuming the user scheduled the operations in
132 : //! the correct order.
133 : //! - Initiate upload queue with that [`IndexPart`].
134 : //! - Reschedule all lost operations by comparing the local filesystem state
135 : //! and remote state as per [`IndexPart`]. This is done in
136 : //! [`Tenant::timeline_init_and_sync`].
137 : //!
138 : //! Note that if we crash during file deletion between the index update
139 : //! that removes the file from the list of files, and deleting the remote file,
140 : //! the file is leaked in the remote storage. Similarly, if a new file is created
141 : //! and uploaded, but the pageserver dies permanently before updating the
142 : //! remote index file, the new file is leaked in remote storage. We accept and
143 : //! tolerate that for now.
144 : //! Note further that we cannot easily fix this by scheduling deletes for every
145 : //! file that is present only on the remote, because we cannot distinguish the
146 : //! following two cases:
147 : //! - (1) We had the file locally, deleted it locally, scheduled a remote delete,
148 : //! but crashed before it finished remotely.
149 : //! - (2) We never had the file locally because we haven't on-demand downloaded
150 : //! it yet.
151 : //!
152 : //! # Downloads
153 : //!
154 : //! In addition to the upload queue, [`RemoteTimelineClient`] has functions for
155 : //! downloading files from the remote storage. Downloads are performed immediately
156 : //! against the `RemoteStorage`, independently of the upload queue.
157 : //!
158 : //! When we attach a tenant, we perform the following steps:
159 : //! - create `Tenant` object in `TenantState::Attaching` state
160 : //! - List timelines that are present in remote storage, and for each:
161 : //! - download their remote [`IndexPart`]s
162 : //! - create `Timeline` struct and a `RemoteTimelineClient`
163 : //! - initialize the client's upload queue with its `IndexPart`
164 : //! - schedule uploads for layers that are only present locally.
165 : //! - After the above is done for each timeline, open the tenant for business by
166 : //! transitioning it from `TenantState::Attaching` to `TenantState::Active` state.
167 : //! This starts the timelines' WAL-receivers and the tenant's GC & Compaction loops.
168 : //!
169 : //! # Operating Without Remote Storage
170 : //!
171 : //! If no remote storage configuration is provided, the [`RemoteTimelineClient`] is
172 : //! not created and the uploads are skipped.
173 : //!
174 : //! [`Tenant::timeline_init_and_sync`]: super::Tenant::timeline_init_and_sync
175 : //! [`Timeline::load_layer_map`]: super::Timeline::load_layer_map
176 :
177 : pub(crate) mod download;
178 : pub mod index;
179 : pub mod manifest;
180 : pub(crate) mod upload;
181 :
182 : use anyhow::Context;
183 : use camino::Utf8Path;
184 : use chrono::{NaiveDateTime, Utc};
185 :
186 : pub(crate) use download::download_initdb_tar_zst;
187 : use pageserver_api::models::TimelineArchivalState;
188 : use pageserver_api::shard::{ShardIndex, TenantShardId};
189 : use regex::Regex;
190 : use scopeguard::ScopeGuard;
191 : use tokio_util::sync::CancellationToken;
192 : use utils::backoff::{
193 : self, exponential_backoff, DEFAULT_BASE_BACKOFF_SECONDS, DEFAULT_MAX_BACKOFF_SECONDS,
194 : };
195 : use utils::pausable_failpoint;
196 : use utils::shard::ShardNumber;
197 :
198 : use std::collections::{HashMap, HashSet, VecDeque};
199 : use std::sync::atomic::{AtomicU32, Ordering};
200 : use std::sync::{Arc, Mutex, OnceLock};
201 : use std::time::Duration;
202 :
203 : use remote_storage::{
204 : DownloadError, GenericRemoteStorage, ListingMode, RemotePath, TimeoutOrCancel,
205 : };
206 : use std::ops::DerefMut;
207 : use tracing::{debug, error, info, instrument, warn};
208 : use tracing::{info_span, Instrument};
209 : use utils::lsn::Lsn;
210 :
211 : use crate::context::RequestContext;
212 : use crate::deletion_queue::{DeletionQueueClient, DeletionQueueError};
213 : use crate::metrics::{
214 : MeasureRemoteOp, RemoteOpFileKind, RemoteOpKind, RemoteTimelineClientMetrics,
215 : RemoteTimelineClientMetricsCallTrackSize, REMOTE_ONDEMAND_DOWNLOADED_BYTES,
216 : REMOTE_ONDEMAND_DOWNLOADED_LAYERS,
217 : };
218 : use crate::task_mgr::shutdown_token;
219 : use crate::tenant::debug_assert_current_span_has_tenant_and_timeline_id;
220 : use crate::tenant::remote_timeline_client::download::download_retry;
221 : use crate::tenant::storage_layer::AsLayerDesc;
222 : use crate::tenant::upload_queue::{Delete, OpType, UploadQueueStoppedDeletable};
223 : use crate::tenant::TIMELINES_SEGMENT_NAME;
224 : use crate::{
225 : config::PageServerConf,
226 : task_mgr,
227 : task_mgr::TaskKind,
228 : task_mgr::BACKGROUND_RUNTIME,
229 : tenant::metadata::TimelineMetadata,
230 : tenant::upload_queue::{
231 : UploadOp, UploadQueue, UploadQueueInitialized, UploadQueueStopped, UploadTask,
232 : },
233 : TENANT_HEATMAP_BASENAME,
234 : };
235 :
236 : use utils::id::{TenantId, TimelineId};
237 :
238 : use self::index::IndexPart;
239 :
240 : use super::config::AttachedLocationConfig;
241 : use super::metadata::MetadataUpdate;
242 : use super::storage_layer::{Layer, LayerName, ResidentLayer};
243 : use super::timeline::import_pgdata;
244 : use super::upload_queue::{NotInitialized, SetDeletedFlagProgress};
245 : use super::{DeleteTimelineError, Generation};
246 :
247 : pub(crate) use download::{
248 : download_index_part, download_tenant_manifest, is_temp_download_file,
249 : list_remote_tenant_shards, list_remote_timelines,
250 : };
251 : pub(crate) use index::LayerFileMetadata;
252 : pub(crate) use upload::upload_initdb_dir;
253 :
254 : // Occasional network issues and such can cause remote operations to fail, and
255 : // that's expected. If a download fails, we log it at info-level, and retry.
256 : // But after FAILED_DOWNLOAD_WARN_THRESHOLD retries, we start to log it at WARN
257 : // level instead, as repeated failures can mean a more serious problem. If it
258 : // fails more than FAILED_DOWNLOAD_RETRIES times, we give up
259 : pub(crate) const FAILED_DOWNLOAD_WARN_THRESHOLD: u32 = 3;
260 : pub(crate) const FAILED_REMOTE_OP_RETRIES: u32 = 10;
261 :
262 : // Similarly log failed uploads and deletions at WARN level, after this many
263 : // retries. Uploads and deletions are retried forever, though.
264 : pub(crate) const FAILED_UPLOAD_WARN_THRESHOLD: u32 = 3;
265 :
266 : pub(crate) const INITDB_PATH: &str = "initdb.tar.zst";
267 :
268 : pub(crate) const INITDB_PRESERVED_PATH: &str = "initdb-preserved.tar.zst";
269 :
270 : /// Default buffer size when interfacing with [`tokio::fs::File`].
271 : pub(crate) const BUFFER_SIZE: usize = 32 * 1024;
272 :
273 : /// Doing non-essential flushes of deletion queue is subject to this timeout, after
274 : /// which we warn and skip.
275 : const DELETION_QUEUE_FLUSH_TIMEOUT: Duration = Duration::from_secs(10);
276 :
277 : pub enum MaybeDeletedIndexPart {
278 : IndexPart(IndexPart),
279 : Deleted(IndexPart),
280 : }
281 :
282 : #[derive(Debug, thiserror::Error)]
283 : pub enum PersistIndexPartWithDeletedFlagError {
284 : #[error("another task is already setting the deleted_flag, started at {0:?}")]
285 : AlreadyInProgress(NaiveDateTime),
286 : #[error("the deleted_flag was already set, value is {0:?}")]
287 : AlreadyDeleted(NaiveDateTime),
288 : #[error(transparent)]
289 : Other(#[from] anyhow::Error),
290 : }
291 :
292 : #[derive(Debug, thiserror::Error)]
293 : pub enum WaitCompletionError {
294 : #[error(transparent)]
295 : NotInitialized(NotInitialized),
296 : #[error("wait_completion aborted because upload queue was stopped")]
297 : UploadQueueShutDownOrStopped,
298 : }
299 :
300 : #[derive(Debug, thiserror::Error)]
301 : #[error("Upload queue either in unexpected state or hasn't downloaded manifest yet")]
302 : pub struct UploadQueueNotReadyError;
303 :
304 : #[derive(Debug, thiserror::Error)]
305 : pub enum ShutdownIfArchivedError {
306 : #[error(transparent)]
307 : NotInitialized(NotInitialized),
308 : #[error("timeline is not archived")]
309 : NotArchived,
310 : }
311 :
312 : /// Behavioral modes that enable seamless live migration.
313 : ///
314 : /// See docs/rfcs/028-pageserver-migration.md to understand how these fit in.
315 : struct RemoteTimelineClientConfig {
316 : /// If this is false, then update to remote_consistent_lsn are dropped rather
317 : /// than being submitted to DeletionQueue for validation. This behavior is
318 : /// used when a tenant attachment is known to have a stale generation number,
319 : /// such that validation attempts will always fail. This is not necessary
320 : /// for correctness, but avoids spamming error statistics with failed validations
321 : /// when doing migrations of tenants.
322 : process_remote_consistent_lsn_updates: bool,
323 :
324 : /// If this is true, then object deletions are held in a buffer in RemoteTimelineClient
325 : /// rather than being submitted to the DeletionQueue. This behavior is used when a tenant
326 : /// is known to be multi-attached, in order to avoid disrupting other attached tenants
327 : /// whose generations' metadata refers to the deleted objects.
328 : block_deletions: bool,
329 : }
330 :
331 : /// RemoteTimelineClientConfig's state is entirely driven by LocationConf, but we do
332 : /// not carry the entire LocationConf structure: it's much more than we need. The From
333 : /// impl extracts the subset of the LocationConf that is interesting to RemoteTimelineClient.
334 : impl From<&AttachedLocationConfig> for RemoteTimelineClientConfig {
335 912 : fn from(lc: &AttachedLocationConfig) -> Self {
336 912 : Self {
337 912 : block_deletions: !lc.may_delete_layers_hint(),
338 912 : process_remote_consistent_lsn_updates: lc.may_upload_layers_hint(),
339 912 : }
340 912 : }
341 : }
342 :
343 : /// A client for accessing a timeline's data in remote storage.
344 : ///
345 : /// This takes care of managing the number of connections, and balancing them
346 : /// across tenants. This also handles retries of failed uploads.
347 : ///
348 : /// Upload and delete requests are ordered so that before a deletion is
349 : /// performed, we wait for all preceding uploads to finish. This ensures sure
350 : /// that if you perform a compaction operation that reshuffles data in layer
351 : /// files, we don't have a transient state where the old files have already been
352 : /// deleted, but new files have not yet been uploaded.
353 : ///
354 : /// Similarly, this enforces an order between index-file uploads, and layer
355 : /// uploads. Before an index-file upload is performed, all preceding layer
356 : /// uploads must be finished.
357 : ///
358 : /// This also maintains a list of remote files, and automatically includes that
359 : /// in the index part file, whenever timeline metadata is uploaded.
360 : ///
361 : /// Downloads are not queued, they are performed immediately.
362 : pub(crate) struct RemoteTimelineClient {
363 : conf: &'static PageServerConf,
364 :
365 : runtime: tokio::runtime::Handle,
366 :
367 : tenant_shard_id: TenantShardId,
368 : timeline_id: TimelineId,
369 : generation: Generation,
370 :
371 : upload_queue: Mutex<UploadQueue>,
372 :
373 : pub(crate) metrics: Arc<RemoteTimelineClientMetrics>,
374 :
375 : storage_impl: GenericRemoteStorage,
376 :
377 : deletion_queue_client: DeletionQueueClient,
378 :
379 : /// Subset of tenant configuration used to control upload behaviors during migrations
380 : config: std::sync::RwLock<RemoteTimelineClientConfig>,
381 :
382 : cancel: CancellationToken,
383 : }
384 :
385 : impl Drop for RemoteTimelineClient {
386 40 : fn drop(&mut self) {
387 40 : debug!("dropping RemoteTimelineClient");
388 40 : }
389 : }
390 :
391 : impl RemoteTimelineClient {
392 : ///
393 : /// Create a remote storage client for given timeline
394 : ///
395 : /// Note: the caller must initialize the upload queue before any uploads can be scheduled,
396 : /// by calling init_upload_queue.
397 : ///
398 892 : pub(crate) fn new(
399 892 : remote_storage: GenericRemoteStorage,
400 892 : deletion_queue_client: DeletionQueueClient,
401 892 : conf: &'static PageServerConf,
402 892 : tenant_shard_id: TenantShardId,
403 892 : timeline_id: TimelineId,
404 892 : generation: Generation,
405 892 : location_conf: &AttachedLocationConfig,
406 892 : ) -> RemoteTimelineClient {
407 892 : RemoteTimelineClient {
408 892 : conf,
409 892 : runtime: if cfg!(test) {
410 : // remote_timeline_client.rs tests rely on current-thread runtime
411 892 : tokio::runtime::Handle::current()
412 : } else {
413 0 : BACKGROUND_RUNTIME.handle().clone()
414 : },
415 892 : tenant_shard_id,
416 892 : timeline_id,
417 892 : generation,
418 892 : storage_impl: remote_storage,
419 892 : deletion_queue_client,
420 892 : upload_queue: Mutex::new(UploadQueue::Uninitialized),
421 892 : metrics: Arc::new(RemoteTimelineClientMetrics::new(
422 892 : &tenant_shard_id,
423 892 : &timeline_id,
424 892 : )),
425 892 : config: std::sync::RwLock::new(RemoteTimelineClientConfig::from(location_conf)),
426 892 : cancel: CancellationToken::new(),
427 892 : }
428 892 : }
429 :
430 : /// Initialize the upload queue for a remote storage that already received
431 : /// an index file upload, i.e., it's not empty.
432 : /// The given `index_part` must be the one on the remote.
433 12 : pub fn init_upload_queue(&self, index_part: &IndexPart) -> anyhow::Result<()> {
434 12 : // Set the maximum number of inprogress tasks to the remote storage concurrency. There's
435 12 : // certainly no point in starting more upload tasks than this.
436 12 : let inprogress_limit = self
437 12 : .conf
438 12 : .remote_storage_config
439 12 : .as_ref()
440 12 : .and_then(|r| r.concurrency_limit())
441 12 : .unwrap_or(0);
442 12 : let mut upload_queue = self.upload_queue.lock().unwrap();
443 12 : upload_queue.initialize_with_current_remote_index_part(index_part, inprogress_limit)?;
444 12 : self.update_remote_physical_size_gauge(Some(index_part));
445 12 : info!(
446 0 : "initialized upload queue from remote index with {} layer files",
447 0 : index_part.layer_metadata.len()
448 : );
449 12 : Ok(())
450 12 : }
451 :
452 : /// Initialize the upload queue for the case where the remote storage is empty,
453 : /// i.e., it doesn't have an `IndexPart`.
454 880 : pub fn init_upload_queue_for_empty_remote(
455 880 : &self,
456 880 : local_metadata: &TimelineMetadata,
457 880 : ) -> anyhow::Result<()> {
458 880 : // Set the maximum number of inprogress tasks to the remote storage concurrency. There's
459 880 : // certainly no point in starting more upload tasks than this.
460 880 : let inprogress_limit = self
461 880 : .conf
462 880 : .remote_storage_config
463 880 : .as_ref()
464 880 : .and_then(|r| r.concurrency_limit())
465 880 : .unwrap_or(0);
466 880 : let mut upload_queue = self.upload_queue.lock().unwrap();
467 880 : upload_queue.initialize_empty_remote(local_metadata, inprogress_limit)?;
468 880 : self.update_remote_physical_size_gauge(None);
469 880 : info!("initialized upload queue as empty");
470 880 : Ok(())
471 880 : }
472 :
473 : /// Initialize the queue in stopped state. Used in startup path
474 : /// to continue deletion operation interrupted by pageserver crash or restart.
475 0 : pub fn init_upload_queue_stopped_to_continue_deletion(
476 0 : &self,
477 0 : index_part: &IndexPart,
478 0 : ) -> anyhow::Result<()> {
479 : // FIXME: consider newtype for DeletedIndexPart.
480 0 : let deleted_at = index_part.deleted_at.ok_or(anyhow::anyhow!(
481 0 : "bug: it is responsibility of the caller to provide index part from MaybeDeletedIndexPart::Deleted"
482 0 : ))?;
483 0 : let inprogress_limit = self
484 0 : .conf
485 0 : .remote_storage_config
486 0 : .as_ref()
487 0 : .and_then(|r| r.concurrency_limit())
488 0 : .unwrap_or(0);
489 0 :
490 0 : let mut upload_queue = self.upload_queue.lock().unwrap();
491 0 : upload_queue.initialize_with_current_remote_index_part(index_part, inprogress_limit)?;
492 0 : self.update_remote_physical_size_gauge(Some(index_part));
493 0 : self.stop_impl(&mut upload_queue);
494 0 :
495 0 : upload_queue
496 0 : .stopped_mut()
497 0 : .expect("stopped above")
498 0 : .deleted_at = SetDeletedFlagProgress::Successful(deleted_at);
499 0 :
500 0 : Ok(())
501 0 : }
502 :
503 : /// Notify this client of a change to its parent tenant's config, as this may cause us to
504 : /// take action (unblocking deletions when transitioning from AttachedMulti to AttachedSingle)
505 0 : pub(super) fn update_config(&self, location_conf: &AttachedLocationConfig) {
506 0 : let new_conf = RemoteTimelineClientConfig::from(location_conf);
507 0 : let unblocked = !new_conf.block_deletions;
508 0 :
509 0 : // Update config before draining deletions, so that we don't race with more being
510 0 : // inserted. This can result in deletions happening our of order, but that does not
511 0 : // violate any invariants: deletions only need to be ordered relative to upload of the index
512 0 : // that dereferences the deleted objects, and we are not changing that order.
513 0 : *self.config.write().unwrap() = new_conf;
514 0 :
515 0 : if unblocked {
516 : // If we may now delete layers, drain any that were blocked in our old
517 : // configuration state
518 0 : let mut queue_locked = self.upload_queue.lock().unwrap();
519 :
520 0 : if let Ok(queue) = queue_locked.initialized_mut() {
521 0 : let blocked_deletions = std::mem::take(&mut queue.blocked_deletions);
522 0 : for d in blocked_deletions {
523 0 : if let Err(e) = self.deletion_queue_client.push_layers_sync(
524 0 : self.tenant_shard_id,
525 0 : self.timeline_id,
526 0 : self.generation,
527 0 : d.layers,
528 0 : ) {
529 : // This could happen if the pageserver is shut down while a tenant
530 : // is transitioning from a deletion-blocked state: we will leak some
531 : // S3 objects in this case.
532 0 : warn!("Failed to drain blocked deletions: {}", e);
533 0 : break;
534 0 : }
535 : }
536 0 : }
537 0 : }
538 0 : }
539 :
540 : /// Returns `None` if nothing is yet uplodaded, `Some(disk_consistent_lsn)` otherwise.
541 0 : pub fn remote_consistent_lsn_projected(&self) -> Option<Lsn> {
542 0 : match &mut *self.upload_queue.lock().unwrap() {
543 0 : UploadQueue::Uninitialized => None,
544 0 : UploadQueue::Initialized(q) => q.get_last_remote_consistent_lsn_projected(),
545 0 : UploadQueue::Stopped(UploadQueueStopped::Uninitialized) => None,
546 0 : UploadQueue::Stopped(UploadQueueStopped::Deletable(q)) => q
547 0 : .upload_queue_for_deletion
548 0 : .get_last_remote_consistent_lsn_projected(),
549 : }
550 0 : }
551 :
552 0 : pub fn remote_consistent_lsn_visible(&self) -> Option<Lsn> {
553 0 : match &mut *self.upload_queue.lock().unwrap() {
554 0 : UploadQueue::Uninitialized => None,
555 0 : UploadQueue::Initialized(q) => Some(q.get_last_remote_consistent_lsn_visible()),
556 0 : UploadQueue::Stopped(UploadQueueStopped::Uninitialized) => None,
557 0 : UploadQueue::Stopped(UploadQueueStopped::Deletable(q)) => Some(
558 0 : q.upload_queue_for_deletion
559 0 : .get_last_remote_consistent_lsn_visible(),
560 0 : ),
561 : }
562 0 : }
563 :
564 : /// Returns true if this timeline was previously detached at this Lsn and the remote timeline
565 : /// client is currently initialized.
566 0 : pub(crate) fn is_previous_ancestor_lsn(&self, lsn: Lsn) -> bool {
567 0 : self.upload_queue
568 0 : .lock()
569 0 : .unwrap()
570 0 : .initialized_mut()
571 0 : .map(|uq| uq.clean.0.lineage.is_previous_ancestor_lsn(lsn))
572 0 : .unwrap_or(false)
573 0 : }
574 :
575 : /// Returns whether the timeline is archived.
576 : /// Return None if the remote index_part hasn't been downloaded yet.
577 0 : pub(crate) fn is_archived(&self) -> Option<bool> {
578 0 : self.upload_queue
579 0 : .lock()
580 0 : .unwrap()
581 0 : .initialized_mut()
582 0 : .map(|q| q.clean.0.archived_at.is_some())
583 0 : .ok()
584 0 : }
585 :
586 : /// Returns `Ok(Some(timestamp))` if the timeline has been archived, `Ok(None)` if the timeline hasn't been archived.
587 : ///
588 : /// Return Err(_) if the remote index_part hasn't been downloaded yet, or the timeline hasn't been stopped yet.
589 4 : pub(crate) fn archived_at_stopped_queue(
590 4 : &self,
591 4 : ) -> Result<Option<NaiveDateTime>, UploadQueueNotReadyError> {
592 4 : self.upload_queue
593 4 : .lock()
594 4 : .unwrap()
595 4 : .stopped_mut()
596 4 : .map(|q| q.upload_queue_for_deletion.clean.0.archived_at)
597 4 : .map_err(|_| UploadQueueNotReadyError)
598 4 : }
599 :
600 3854 : fn update_remote_physical_size_gauge(&self, current_remote_index_part: Option<&IndexPart>) {
601 3854 : let size: u64 = if let Some(current_remote_index_part) = current_remote_index_part {
602 2974 : current_remote_index_part
603 2974 : .layer_metadata
604 2974 : .values()
605 35806 : .map(|ilmd| ilmd.file_size)
606 2974 : .sum()
607 : } else {
608 880 : 0
609 : };
610 3854 : self.metrics.remote_physical_size_gauge.set(size);
611 3854 : }
612 :
613 4 : pub fn get_remote_physical_size(&self) -> u64 {
614 4 : self.metrics.remote_physical_size_gauge.get()
615 4 : }
616 :
617 : //
618 : // Download operations.
619 : //
620 : // These don't use the per-timeline queue. They do use the global semaphore in
621 : // S3Bucket, to limit the total number of concurrent operations, though.
622 : //
623 :
624 : /// Download index file
625 40 : pub async fn download_index_file(
626 40 : &self,
627 40 : cancel: &CancellationToken,
628 40 : ) -> Result<MaybeDeletedIndexPart, DownloadError> {
629 40 : let _unfinished_gauge_guard = self.metrics.call_begin(
630 40 : &RemoteOpFileKind::Index,
631 40 : &RemoteOpKind::Download,
632 40 : crate::metrics::RemoteTimelineClientMetricsCallTrackSize::DontTrackSize {
633 40 : reason: "no need for a downloads gauge",
634 40 : },
635 40 : );
636 :
637 40 : let (index_part, index_generation, index_last_modified) = download::download_index_part(
638 40 : &self.storage_impl,
639 40 : &self.tenant_shard_id,
640 40 : &self.timeline_id,
641 40 : self.generation,
642 40 : cancel,
643 40 : )
644 40 : .measure_remote_op(
645 40 : RemoteOpFileKind::Index,
646 40 : RemoteOpKind::Download,
647 40 : Arc::clone(&self.metrics),
648 40 : )
649 40 : .await?;
650 :
651 : // Defense in depth: monotonicity of generation numbers is an important correctness guarantee, so when we see a very
652 : // old index, we do extra checks in case this is the result of backward time-travel of the generation number (e.g.
653 : // in case of a bug in the service that issues generation numbers). Indices are allowed to be old, but we expect that
654 : // when we load an old index we are loading the _latest_ index: if we are asked to load an old index and there is
655 : // also a newer index available, that is surprising.
656 : const INDEX_AGE_CHECKS_THRESHOLD: Duration = Duration::from_secs(14 * 24 * 3600);
657 40 : let index_age = index_last_modified.elapsed().unwrap_or_else(|e| {
658 0 : if e.duration() > Duration::from_secs(5) {
659 : // We only warn if the S3 clock and our local clock are >5s out: because this is a low resolution
660 : // timestamp, it is common to be out by at least 1 second.
661 0 : tracing::warn!("Index has modification time in the future: {e}");
662 0 : }
663 0 : Duration::ZERO
664 40 : });
665 40 : if index_age > INDEX_AGE_CHECKS_THRESHOLD {
666 0 : tracing::info!(
667 : ?index_generation,
668 0 : age = index_age.as_secs_f64(),
669 0 : "Loaded an old index, checking for other indices..."
670 : );
671 :
672 : // Find the highest-generation index
673 0 : let (_latest_index_part, latest_index_generation, latest_index_mtime) =
674 0 : download::download_index_part(
675 0 : &self.storage_impl,
676 0 : &self.tenant_shard_id,
677 0 : &self.timeline_id,
678 0 : Generation::MAX,
679 0 : cancel,
680 0 : )
681 0 : .await?;
682 :
683 0 : if latest_index_generation > index_generation {
684 : // Unexpected! Why are we loading such an old index if a more recent one exists?
685 : // We will refuse to proceed, as there is no reasonable scenario where this should happen, but
686 : // there _is_ a clear bug/corruption scenario where it would happen (controller sets the generation
687 : // backwards).
688 0 : tracing::error!(
689 : ?index_generation,
690 : ?latest_index_generation,
691 : ?latest_index_mtime,
692 0 : "Found a newer index while loading an old one"
693 : );
694 0 : return Err(DownloadError::Fatal(
695 0 : "Index age exceeds threshold and a newer index exists".into(),
696 0 : ));
697 0 : }
698 40 : }
699 :
700 40 : if index_part.deleted_at.is_some() {
701 0 : Ok(MaybeDeletedIndexPart::Deleted(index_part))
702 : } else {
703 40 : Ok(MaybeDeletedIndexPart::IndexPart(index_part))
704 : }
705 40 : }
706 :
707 : /// Download a (layer) file from `path`, into local filesystem.
708 : ///
709 : /// 'layer_metadata' is the metadata from the remote index file.
710 : ///
711 : /// On success, returns the size of the downloaded file.
712 12 : pub async fn download_layer_file(
713 12 : &self,
714 12 : layer_file_name: &LayerName,
715 12 : layer_metadata: &LayerFileMetadata,
716 12 : local_path: &Utf8Path,
717 12 : gate: &utils::sync::gate::Gate,
718 12 : cancel: &CancellationToken,
719 12 : ctx: &RequestContext,
720 12 : ) -> Result<u64, DownloadError> {
721 12 : let downloaded_size = {
722 12 : let _unfinished_gauge_guard = self.metrics.call_begin(
723 12 : &RemoteOpFileKind::Layer,
724 12 : &RemoteOpKind::Download,
725 12 : crate::metrics::RemoteTimelineClientMetricsCallTrackSize::DontTrackSize {
726 12 : reason: "no need for a downloads gauge",
727 12 : },
728 12 : );
729 12 : download::download_layer_file(
730 12 : self.conf,
731 12 : &self.storage_impl,
732 12 : self.tenant_shard_id,
733 12 : self.timeline_id,
734 12 : layer_file_name,
735 12 : layer_metadata,
736 12 : local_path,
737 12 : gate,
738 12 : cancel,
739 12 : ctx,
740 12 : )
741 12 : .measure_remote_op(
742 12 : RemoteOpFileKind::Layer,
743 12 : RemoteOpKind::Download,
744 12 : Arc::clone(&self.metrics),
745 12 : )
746 12 : .await?
747 : };
748 :
749 12 : REMOTE_ONDEMAND_DOWNLOADED_LAYERS.inc();
750 12 : REMOTE_ONDEMAND_DOWNLOADED_BYTES.inc_by(downloaded_size);
751 12 :
752 12 : Ok(downloaded_size)
753 12 : }
754 :
755 : //
756 : // Upload operations.
757 : //
758 :
759 : /// Launch an index-file upload operation in the background, with
760 : /// fully updated metadata.
761 : ///
762 : /// This should only be used to upload initial metadata to remote storage.
763 : ///
764 : /// The upload will be added to the queue immediately, but it
765 : /// won't be performed until all previously scheduled layer file
766 : /// upload operations have completed successfully. This is to
767 : /// ensure that when the index file claims that layers X, Y and Z
768 : /// exist in remote storage, they really do. To wait for the upload
769 : /// to complete, use `wait_completion`.
770 : ///
771 : /// If there were any changes to the list of files, i.e. if any
772 : /// layer file uploads were scheduled, since the last index file
773 : /// upload, those will be included too.
774 460 : pub fn schedule_index_upload_for_full_metadata_update(
775 460 : self: &Arc<Self>,
776 460 : metadata: &TimelineMetadata,
777 460 : ) -> anyhow::Result<()> {
778 460 : let mut guard = self.upload_queue.lock().unwrap();
779 460 : let upload_queue = guard.initialized_mut()?;
780 :
781 : // As documented in the struct definition, it's ok for latest_metadata to be
782 : // ahead of what's _actually_ on the remote during index upload.
783 460 : upload_queue.dirty.metadata = metadata.clone();
784 460 :
785 460 : self.schedule_index_upload(upload_queue);
786 460 :
787 460 : Ok(())
788 460 : }
789 :
790 : /// Launch an index-file upload operation in the background, with only parts of the metadata
791 : /// updated.
792 : ///
793 : /// This is the regular way of updating metadata on layer flushes or Gc.
794 : ///
795 : /// Using this lighter update mechanism allows for reparenting and detaching without changes to
796 : /// `index_part.json`, while being more clear on what values update regularly.
797 2444 : pub(crate) fn schedule_index_upload_for_metadata_update(
798 2444 : self: &Arc<Self>,
799 2444 : update: &MetadataUpdate,
800 2444 : ) -> anyhow::Result<()> {
801 2444 : let mut guard = self.upload_queue.lock().unwrap();
802 2444 : let upload_queue = guard.initialized_mut()?;
803 :
804 2444 : upload_queue.dirty.metadata.apply(update);
805 2444 :
806 2444 : // Defense in depth: if we somehow generated invalid metadata, do not persist it.
807 2444 : upload_queue
808 2444 : .dirty
809 2444 : .validate()
810 2444 : .map_err(|e| anyhow::anyhow!(e))?;
811 :
812 2444 : self.schedule_index_upload(upload_queue);
813 2444 :
814 2444 : Ok(())
815 2444 : }
816 :
817 : /// Launch an index-file upload operation in the background, with only the `archived_at` field updated.
818 : ///
819 : /// Returns whether it is required to wait for the queue to be empty to ensure that the change is uploaded,
820 : /// so either if the change is already sitting in the queue, but not commited yet, or the change has not
821 : /// been in the queue yet.
822 4 : pub(crate) fn schedule_index_upload_for_timeline_archival_state(
823 4 : self: &Arc<Self>,
824 4 : state: TimelineArchivalState,
825 4 : ) -> anyhow::Result<bool> {
826 4 : let mut guard = self.upload_queue.lock().unwrap();
827 4 : let upload_queue = guard.initialized_mut()?;
828 :
829 : /// Returns Some(_) if a change is needed, and Some(true) if it's a
830 : /// change needed to set archived_at.
831 8 : fn need_change(
832 8 : archived_at: &Option<NaiveDateTime>,
833 8 : state: TimelineArchivalState,
834 8 : ) -> Option<bool> {
835 8 : match (archived_at, state) {
836 : (Some(_), TimelineArchivalState::Archived)
837 : | (None, TimelineArchivalState::Unarchived) => {
838 : // Nothing to do
839 0 : tracing::info!("intended state matches present state");
840 0 : None
841 : }
842 8 : (None, TimelineArchivalState::Archived) => Some(true),
843 0 : (Some(_), TimelineArchivalState::Unarchived) => Some(false),
844 : }
845 8 : }
846 4 : let need_upload_scheduled = need_change(&upload_queue.dirty.archived_at, state);
847 :
848 4 : if let Some(archived_at_set) = need_upload_scheduled {
849 4 : let intended_archived_at = archived_at_set.then(|| Utc::now().naive_utc());
850 4 : upload_queue.dirty.archived_at = intended_archived_at;
851 4 : self.schedule_index_upload(upload_queue);
852 4 : }
853 :
854 4 : let need_wait = need_change(&upload_queue.clean.0.archived_at, state).is_some();
855 4 : Ok(need_wait)
856 4 : }
857 :
858 : /// Shuts the timeline client down, but only if the timeline is archived.
859 : ///
860 : /// This function and [`Self::schedule_index_upload_for_timeline_archival_state`] use the
861 : /// same lock to prevent races between unarchival and offloading: unarchival requires the
862 : /// upload queue to be initialized, and leaves behind an upload queue where either dirty
863 : /// or clean has archived_at of `None`. offloading leaves behind an uninitialized upload
864 : /// queue.
865 4 : pub(crate) async fn shutdown_if_archived(
866 4 : self: &Arc<Self>,
867 4 : ) -> Result<(), ShutdownIfArchivedError> {
868 4 : {
869 4 : let mut guard = self.upload_queue.lock().unwrap();
870 4 : let upload_queue = guard
871 4 : .initialized_mut()
872 4 : .map_err(ShutdownIfArchivedError::NotInitialized)?;
873 :
874 4 : match (
875 4 : upload_queue.dirty.archived_at.is_none(),
876 4 : upload_queue.clean.0.archived_at.is_none(),
877 4 : ) {
878 : // The expected case: the timeline is archived and we don't want to unarchive
879 4 : (false, false) => {}
880 : (true, false) => {
881 0 : tracing::info!("can't shut down timeline: timeline slated for unarchival");
882 0 : return Err(ShutdownIfArchivedError::NotArchived);
883 : }
884 0 : (dirty_archived, true) => {
885 0 : tracing::info!(%dirty_archived, "can't shut down timeline: timeline not archived in remote storage");
886 0 : return Err(ShutdownIfArchivedError::NotArchived);
887 : }
888 : }
889 :
890 : // Set the shutting_down flag while the guard from the archival check is held.
891 : // This prevents a race with unarchival, as initialized_mut will not return
892 : // an upload queue from this point.
893 : // Also launch the queued tasks like shutdown() does.
894 4 : if !upload_queue.shutting_down {
895 4 : upload_queue.shutting_down = true;
896 4 : upload_queue.queued_operations.push_back(UploadOp::Shutdown);
897 4 : // this operation is not counted similar to Barrier
898 4 : self.launch_queued_tasks(upload_queue);
899 4 : }
900 : }
901 :
902 4 : self.shutdown().await;
903 :
904 4 : Ok(())
905 4 : }
906 :
907 : /// Launch an index-file upload operation in the background, setting `import_pgdata` field.
908 0 : pub(crate) fn schedule_index_upload_for_import_pgdata_state_update(
909 0 : self: &Arc<Self>,
910 0 : state: Option<import_pgdata::index_part_format::Root>,
911 0 : ) -> anyhow::Result<()> {
912 0 : let mut guard = self.upload_queue.lock().unwrap();
913 0 : let upload_queue = guard.initialized_mut()?;
914 0 : upload_queue.dirty.import_pgdata = state;
915 0 : self.schedule_index_upload(upload_queue);
916 0 : Ok(())
917 0 : }
918 :
919 : ///
920 : /// Launch an index-file upload operation in the background, if necessary.
921 : ///
922 : /// Use this function to schedule the update of the index file after
923 : /// scheduling file uploads or deletions. If no file uploads or deletions
924 : /// have been scheduled since the last index file upload, this does
925 : /// nothing.
926 : ///
927 : /// Like schedule_index_upload_for_metadata_update(), this merely adds
928 : /// the upload to the upload queue and returns quickly.
929 740 : pub fn schedule_index_upload_for_file_changes(self: &Arc<Self>) -> Result<(), NotInitialized> {
930 740 : let mut guard = self.upload_queue.lock().unwrap();
931 740 : let upload_queue = guard.initialized_mut()?;
932 :
933 740 : if upload_queue.latest_files_changes_since_metadata_upload_scheduled > 0 {
934 28 : self.schedule_index_upload(upload_queue);
935 712 : }
936 :
937 740 : Ok(())
938 740 : }
939 :
940 : /// Launch an index-file upload operation in the background (internal function)
941 3080 : fn schedule_index_upload(self: &Arc<Self>, upload_queue: &mut UploadQueueInitialized) {
942 3080 : let disk_consistent_lsn = upload_queue.dirty.metadata.disk_consistent_lsn();
943 3080 : // fix up the duplicated field
944 3080 : upload_queue.dirty.disk_consistent_lsn = disk_consistent_lsn;
945 3080 :
946 3080 : // make sure it serializes before doing it in perform_upload_task so that it doesn't
947 3080 : // look like a retryable error
948 3080 : let void = std::io::sink();
949 3080 : serde_json::to_writer(void, &upload_queue.dirty).expect("serialize index_part.json");
950 3080 :
951 3080 : let index_part = &upload_queue.dirty;
952 3080 :
953 3080 : info!(
954 0 : "scheduling metadata upload up to consistent LSN {disk_consistent_lsn} with {} files ({} changed)",
955 0 : index_part.layer_metadata.len(),
956 : upload_queue.latest_files_changes_since_metadata_upload_scheduled,
957 : );
958 :
959 3080 : let op = UploadOp::UploadMetadata {
960 3080 : uploaded: Box::new(index_part.clone()),
961 3080 : };
962 3080 : self.metric_begin(&op);
963 3080 : upload_queue.queued_operations.push_back(op);
964 3080 : upload_queue.latest_files_changes_since_metadata_upload_scheduled = 0;
965 3080 :
966 3080 : // Launch the task immediately, if possible
967 3080 : self.launch_queued_tasks(upload_queue);
968 3080 : }
969 :
970 : /// Reparent this timeline to a new parent.
971 : ///
972 : /// A retryable step of timeline ancestor detach.
973 0 : pub(crate) async fn schedule_reparenting_and_wait(
974 0 : self: &Arc<Self>,
975 0 : new_parent: &TimelineId,
976 0 : ) -> anyhow::Result<()> {
977 0 : let receiver = {
978 0 : let mut guard = self.upload_queue.lock().unwrap();
979 0 : let upload_queue = guard.initialized_mut()?;
980 :
981 0 : let Some(prev) = upload_queue.dirty.metadata.ancestor_timeline() else {
982 0 : return Err(anyhow::anyhow!(
983 0 : "cannot reparent without a current ancestor"
984 0 : ));
985 : };
986 :
987 0 : let uploaded = &upload_queue.clean.0.metadata;
988 0 :
989 0 : if uploaded.ancestor_timeline().is_none() && !uploaded.ancestor_lsn().is_valid() {
990 : // nothing to do
991 0 : None
992 : } else {
993 0 : upload_queue.dirty.metadata.reparent(new_parent);
994 0 : upload_queue.dirty.lineage.record_previous_ancestor(&prev);
995 0 :
996 0 : self.schedule_index_upload(upload_queue);
997 0 :
998 0 : Some(self.schedule_barrier0(upload_queue))
999 : }
1000 : };
1001 :
1002 0 : if let Some(receiver) = receiver {
1003 0 : Self::wait_completion0(receiver).await?;
1004 0 : }
1005 0 : Ok(())
1006 0 : }
1007 :
1008 : /// Schedules uploading a new version of `index_part.json` with the given layers added,
1009 : /// detaching from ancestor and waits for it to complete.
1010 : ///
1011 : /// This is used with `Timeline::detach_ancestor` functionality.
1012 0 : pub(crate) async fn schedule_adding_existing_layers_to_index_detach_and_wait(
1013 0 : self: &Arc<Self>,
1014 0 : layers: &[Layer],
1015 0 : adopted: (TimelineId, Lsn),
1016 0 : ) -> anyhow::Result<()> {
1017 0 : let barrier = {
1018 0 : let mut guard = self.upload_queue.lock().unwrap();
1019 0 : let upload_queue = guard.initialized_mut()?;
1020 :
1021 0 : if upload_queue.clean.0.lineage.detached_previous_ancestor() == Some(adopted) {
1022 0 : None
1023 : } else {
1024 0 : upload_queue.dirty.metadata.detach_from_ancestor(&adopted);
1025 0 : upload_queue.dirty.lineage.record_detaching(&adopted);
1026 :
1027 0 : for layer in layers {
1028 0 : let prev = upload_queue
1029 0 : .dirty
1030 0 : .layer_metadata
1031 0 : .insert(layer.layer_desc().layer_name(), layer.metadata());
1032 0 : assert!(prev.is_none(), "copied layer existed already {layer}");
1033 : }
1034 :
1035 0 : self.schedule_index_upload(upload_queue);
1036 0 :
1037 0 : Some(self.schedule_barrier0(upload_queue))
1038 : }
1039 : };
1040 :
1041 0 : if let Some(barrier) = barrier {
1042 0 : Self::wait_completion0(barrier).await?;
1043 0 : }
1044 0 : Ok(())
1045 0 : }
1046 :
1047 : /// Adds a gc blocking reason for this timeline if one does not exist already.
1048 : ///
1049 : /// A retryable step of timeline detach ancestor.
1050 : ///
1051 : /// Returns a future which waits until the completion of the upload.
1052 0 : pub(crate) fn schedule_insert_gc_block_reason(
1053 0 : self: &Arc<Self>,
1054 0 : reason: index::GcBlockingReason,
1055 0 : ) -> Result<impl std::future::Future<Output = Result<(), WaitCompletionError>>, NotInitialized>
1056 0 : {
1057 0 : let maybe_barrier = {
1058 0 : let mut guard = self.upload_queue.lock().unwrap();
1059 0 : let upload_queue = guard.initialized_mut()?;
1060 :
1061 0 : if let index::GcBlockingReason::DetachAncestor = reason {
1062 0 : if upload_queue.dirty.metadata.ancestor_timeline().is_none() {
1063 0 : drop(guard);
1064 0 : panic!("cannot start detach ancestor if there is nothing to detach from");
1065 0 : }
1066 0 : }
1067 :
1068 0 : let wanted = |x: Option<&index::GcBlocking>| x.is_some_and(|x| x.blocked_by(reason));
1069 :
1070 0 : let current = upload_queue.dirty.gc_blocking.as_ref();
1071 0 : let uploaded = upload_queue.clean.0.gc_blocking.as_ref();
1072 0 :
1073 0 : match (current, uploaded) {
1074 0 : (x, y) if wanted(x) && wanted(y) => None,
1075 0 : (x, y) if wanted(x) && !wanted(y) => Some(self.schedule_barrier0(upload_queue)),
1076 : // Usual case: !wanted(x) && !wanted(y)
1077 : //
1078 : // Unusual: !wanted(x) && wanted(y) which means we have two processes waiting to
1079 : // turn on and off some reason.
1080 0 : (x, y) => {
1081 0 : if !wanted(x) && wanted(y) {
1082 : // this could be avoided by having external in-memory synchronization, like
1083 : // timeline detach ancestor
1084 0 : warn!(?reason, op="insert", "unexpected: two racing processes to enable and disable a gc blocking reason");
1085 0 : }
1086 :
1087 : // at this point, the metadata must always show that there is a parent
1088 0 : upload_queue.dirty.gc_blocking = current
1089 0 : .map(|x| x.with_reason(reason))
1090 0 : .or_else(|| Some(index::GcBlocking::started_now_for(reason)));
1091 0 : self.schedule_index_upload(upload_queue);
1092 0 : Some(self.schedule_barrier0(upload_queue))
1093 : }
1094 : }
1095 : };
1096 :
1097 0 : Ok(async move {
1098 0 : if let Some(barrier) = maybe_barrier {
1099 0 : Self::wait_completion0(barrier).await?;
1100 0 : }
1101 0 : Ok(())
1102 0 : })
1103 0 : }
1104 :
1105 : /// Removes a gc blocking reason for this timeline if one exists.
1106 : ///
1107 : /// A retryable step of timeline detach ancestor.
1108 : ///
1109 : /// Returns a future which waits until the completion of the upload.
1110 0 : pub(crate) fn schedule_remove_gc_block_reason(
1111 0 : self: &Arc<Self>,
1112 0 : reason: index::GcBlockingReason,
1113 0 : ) -> Result<impl std::future::Future<Output = Result<(), WaitCompletionError>>, NotInitialized>
1114 0 : {
1115 0 : let maybe_barrier = {
1116 0 : let mut guard = self.upload_queue.lock().unwrap();
1117 0 : let upload_queue = guard.initialized_mut()?;
1118 :
1119 0 : if let index::GcBlockingReason::DetachAncestor = reason {
1120 0 : if !upload_queue.clean.0.lineage.is_detached_from_ancestor() {
1121 0 : drop(guard);
1122 0 : panic!("cannot complete timeline_ancestor_detach while not detached");
1123 0 : }
1124 0 : }
1125 :
1126 0 : let wanted = |x: Option<&index::GcBlocking>| {
1127 0 : x.is_none() || x.is_some_and(|b| !b.blocked_by(reason))
1128 0 : };
1129 :
1130 0 : let current = upload_queue.dirty.gc_blocking.as_ref();
1131 0 : let uploaded = upload_queue.clean.0.gc_blocking.as_ref();
1132 0 :
1133 0 : match (current, uploaded) {
1134 0 : (x, y) if wanted(x) && wanted(y) => None,
1135 0 : (x, y) if wanted(x) && !wanted(y) => Some(self.schedule_barrier0(upload_queue)),
1136 0 : (x, y) => {
1137 0 : if !wanted(x) && wanted(y) {
1138 0 : warn!(?reason, op="remove", "unexpected: two racing processes to enable and disable a gc blocking reason (remove)");
1139 0 : }
1140 :
1141 0 : upload_queue.dirty.gc_blocking =
1142 0 : current.as_ref().and_then(|x| x.without_reason(reason));
1143 0 : assert!(wanted(upload_queue.dirty.gc_blocking.as_ref()));
1144 0 : self.schedule_index_upload(upload_queue);
1145 0 : Some(self.schedule_barrier0(upload_queue))
1146 : }
1147 : }
1148 : };
1149 :
1150 0 : Ok(async move {
1151 0 : if let Some(barrier) = maybe_barrier {
1152 0 : Self::wait_completion0(barrier).await?;
1153 0 : }
1154 0 : Ok(())
1155 0 : })
1156 0 : }
1157 :
1158 : /// Launch an upload operation in the background; the file is added to be included in next
1159 : /// `index_part.json` upload.
1160 2732 : pub(crate) fn schedule_layer_file_upload(
1161 2732 : self: &Arc<Self>,
1162 2732 : layer: ResidentLayer,
1163 2732 : ) -> Result<(), NotInitialized> {
1164 2732 : let mut guard = self.upload_queue.lock().unwrap();
1165 2732 : let upload_queue = guard.initialized_mut()?;
1166 :
1167 2732 : self.schedule_layer_file_upload0(upload_queue, layer);
1168 2732 : self.launch_queued_tasks(upload_queue);
1169 2732 : Ok(())
1170 2732 : }
1171 :
1172 3456 : fn schedule_layer_file_upload0(
1173 3456 : self: &Arc<Self>,
1174 3456 : upload_queue: &mut UploadQueueInitialized,
1175 3456 : layer: ResidentLayer,
1176 3456 : ) {
1177 3456 : let metadata = layer.metadata();
1178 3456 :
1179 3456 : upload_queue
1180 3456 : .dirty
1181 3456 : .layer_metadata
1182 3456 : .insert(layer.layer_desc().layer_name(), metadata.clone());
1183 3456 : upload_queue.latest_files_changes_since_metadata_upload_scheduled += 1;
1184 3456 :
1185 3456 : info!(
1186 : gen=?metadata.generation,
1187 : shard=?metadata.shard,
1188 0 : "scheduled layer file upload {layer}",
1189 : );
1190 :
1191 3456 : let op = UploadOp::UploadLayer(layer, metadata, None);
1192 3456 : self.metric_begin(&op);
1193 3456 : upload_queue.queued_operations.push_back(op);
1194 3456 : }
1195 :
1196 : /// Launch a delete operation in the background.
1197 : ///
1198 : /// The operation does not modify local filesystem state.
1199 : ///
1200 : /// Note: This schedules an index file upload before the deletions. The
1201 : /// deletion won't actually be performed, until all previously scheduled
1202 : /// upload operations, and the index file upload, have completed
1203 : /// successfully.
1204 16 : pub fn schedule_layer_file_deletion(
1205 16 : self: &Arc<Self>,
1206 16 : names: &[LayerName],
1207 16 : ) -> anyhow::Result<()> {
1208 16 : let mut guard = self.upload_queue.lock().unwrap();
1209 16 : let upload_queue = guard.initialized_mut()?;
1210 :
1211 16 : let with_metadata =
1212 16 : self.schedule_unlinking_of_layers_from_index_part0(upload_queue, names.iter().cloned());
1213 16 :
1214 16 : self.schedule_deletion_of_unlinked0(upload_queue, with_metadata);
1215 16 :
1216 16 : // Launch the tasks immediately, if possible
1217 16 : self.launch_queued_tasks(upload_queue);
1218 16 : Ok(())
1219 16 : }
1220 :
1221 : /// Unlinks the layer files from `index_part.json` but does not yet schedule deletion for the
1222 : /// layer files, leaving them dangling.
1223 : ///
1224 : /// The files will be leaked in remote storage unless [`Self::schedule_deletion_of_unlinked`]
1225 : /// is invoked on them.
1226 8 : pub(crate) fn schedule_gc_update(
1227 8 : self: &Arc<Self>,
1228 8 : gc_layers: &[Layer],
1229 8 : ) -> Result<(), NotInitialized> {
1230 8 : let mut guard = self.upload_queue.lock().unwrap();
1231 8 : let upload_queue = guard.initialized_mut()?;
1232 :
1233 : // just forget the return value; after uploading the next index_part.json, we can consider
1234 : // the layer files as "dangling". this is fine, at worst case we create work for the
1235 : // scrubber.
1236 :
1237 8 : let names = gc_layers.iter().map(|x| x.layer_desc().layer_name());
1238 8 :
1239 8 : self.schedule_unlinking_of_layers_from_index_part0(upload_queue, names);
1240 8 :
1241 8 : self.launch_queued_tasks(upload_queue);
1242 8 :
1243 8 : Ok(())
1244 8 : }
1245 :
1246 : /// Update the remote index file, removing the to-be-deleted files from the index,
1247 : /// allowing scheduling of actual deletions later.
1248 176 : fn schedule_unlinking_of_layers_from_index_part0<I>(
1249 176 : self: &Arc<Self>,
1250 176 : upload_queue: &mut UploadQueueInitialized,
1251 176 : names: I,
1252 176 : ) -> Vec<(LayerName, LayerFileMetadata)>
1253 176 : where
1254 176 : I: IntoIterator<Item = LayerName>,
1255 176 : {
1256 176 : // Decorate our list of names with each name's metadata, dropping
1257 176 : // names that are unexpectedly missing from our metadata. This metadata
1258 176 : // is later used when physically deleting layers, to construct key paths.
1259 176 : let with_metadata: Vec<_> = names
1260 176 : .into_iter()
1261 1024 : .filter_map(|name| {
1262 1024 : let meta = upload_queue.dirty.layer_metadata.remove(&name);
1263 :
1264 1024 : if let Some(meta) = meta {
1265 1024 : upload_queue.latest_files_changes_since_metadata_upload_scheduled += 1;
1266 1024 : Some((name, meta))
1267 : } else {
1268 : // This can only happen if we forgot to to schedule the file upload
1269 : // before scheduling the delete. Log it because it is a rare/strange
1270 : // situation, and in case something is misbehaving, we'd like to know which
1271 : // layers experienced this.
1272 0 : info!("Deleting layer {name} not found in latest_files list, never uploaded?");
1273 0 : None
1274 : }
1275 1024 : })
1276 176 : .collect();
1277 :
1278 : #[cfg(feature = "testing")]
1279 1200 : for (name, metadata) in &with_metadata {
1280 1024 : let gen = metadata.generation;
1281 1024 : if let Some(unexpected) = upload_queue.dangling_files.insert(name.to_owned(), gen) {
1282 0 : if unexpected == gen {
1283 0 : tracing::error!("{name} was unlinked twice with same generation");
1284 : } else {
1285 0 : tracing::error!("{name} was unlinked twice with different generations {gen:?} and {unexpected:?}");
1286 : }
1287 1024 : }
1288 : }
1289 :
1290 : // after unlinking files from the upload_queue.latest_files we must always schedule an
1291 : // index_part update, because that needs to be uploaded before we can actually delete the
1292 : // files.
1293 176 : if upload_queue.latest_files_changes_since_metadata_upload_scheduled > 0 {
1294 144 : self.schedule_index_upload(upload_queue);
1295 144 : }
1296 :
1297 176 : with_metadata
1298 176 : }
1299 :
1300 : /// Schedules deletion for layer files which have previously been unlinked from the
1301 : /// `index_part.json` with [`Self::schedule_gc_update`] or [`Self::schedule_compaction_update`].
1302 1019 : pub(crate) fn schedule_deletion_of_unlinked(
1303 1019 : self: &Arc<Self>,
1304 1019 : layers: Vec<(LayerName, LayerFileMetadata)>,
1305 1019 : ) -> anyhow::Result<()> {
1306 1019 : let mut guard = self.upload_queue.lock().unwrap();
1307 1019 : let upload_queue = guard.initialized_mut()?;
1308 :
1309 1019 : self.schedule_deletion_of_unlinked0(upload_queue, layers);
1310 1019 : self.launch_queued_tasks(upload_queue);
1311 1019 : Ok(())
1312 1019 : }
1313 :
1314 1031 : fn schedule_deletion_of_unlinked0(
1315 1031 : self: &Arc<Self>,
1316 1031 : upload_queue: &mut UploadQueueInitialized,
1317 1031 : mut with_metadata: Vec<(LayerName, LayerFileMetadata)>,
1318 1031 : ) {
1319 1031 : // Filter out any layers which were not created by this tenant shard. These are
1320 1031 : // layers that originate from some ancestor shard after a split, and may still
1321 1031 : // be referenced by other shards. We are free to delete them locally and remove
1322 1031 : // them from our index (and would have already done so when we reach this point
1323 1031 : // in the code), but we may not delete them remotely.
1324 1031 : with_metadata.retain(|(name, meta)| {
1325 1019 : let retain = meta.shard.shard_number == self.tenant_shard_id.shard_number
1326 1019 : && meta.shard.shard_count == self.tenant_shard_id.shard_count;
1327 1019 : if !retain {
1328 0 : tracing::debug!(
1329 0 : "Skipping deletion of ancestor-shard layer {name}, from shard {}",
1330 : meta.shard
1331 : );
1332 1019 : }
1333 1019 : retain
1334 1031 : });
1335 :
1336 2050 : for (name, meta) in &with_metadata {
1337 1019 : info!(
1338 0 : "scheduling deletion of layer {}{} (shard {})",
1339 0 : name,
1340 0 : meta.generation.get_suffix(),
1341 : meta.shard
1342 : );
1343 : }
1344 :
1345 : #[cfg(feature = "testing")]
1346 2050 : for (name, meta) in &with_metadata {
1347 1019 : let gen = meta.generation;
1348 1019 : match upload_queue.dangling_files.remove(name) {
1349 1011 : Some(same) if same == gen => { /* expected */ }
1350 0 : Some(other) => {
1351 0 : tracing::error!("{name} was unlinked with {other:?} but deleted with {gen:?}");
1352 : }
1353 : None => {
1354 8 : tracing::error!("{name} was unlinked but was not dangling");
1355 : }
1356 : }
1357 : }
1358 :
1359 : // schedule the actual deletions
1360 1031 : if with_metadata.is_empty() {
1361 : // avoid scheduling the op & bumping the metric
1362 12 : return;
1363 1019 : }
1364 1019 : let op = UploadOp::Delete(Delete {
1365 1019 : layers: with_metadata,
1366 1019 : });
1367 1019 : self.metric_begin(&op);
1368 1019 : upload_queue.queued_operations.push_back(op);
1369 1031 : }
1370 :
1371 : /// Schedules a compaction update to the remote `index_part.json`.
1372 : ///
1373 : /// `compacted_from` represent the L0 names which have been `compacted_to` L1 layers.
1374 152 : pub(crate) fn schedule_compaction_update(
1375 152 : self: &Arc<Self>,
1376 152 : compacted_from: &[Layer],
1377 152 : compacted_to: &[ResidentLayer],
1378 152 : ) -> Result<(), NotInitialized> {
1379 152 : let mut guard = self.upload_queue.lock().unwrap();
1380 152 : let upload_queue = guard.initialized_mut()?;
1381 :
1382 876 : for layer in compacted_to {
1383 724 : self.schedule_layer_file_upload0(upload_queue, layer.clone());
1384 724 : }
1385 :
1386 1012 : let names = compacted_from.iter().map(|x| x.layer_desc().layer_name());
1387 152 :
1388 152 : self.schedule_unlinking_of_layers_from_index_part0(upload_queue, names);
1389 152 : self.launch_queued_tasks(upload_queue);
1390 152 :
1391 152 : Ok(())
1392 152 : }
1393 :
1394 : /// Wait for all previously scheduled uploads/deletions to complete
1395 2772 : pub(crate) async fn wait_completion(self: &Arc<Self>) -> Result<(), WaitCompletionError> {
1396 2772 : let receiver = {
1397 2772 : let mut guard = self.upload_queue.lock().unwrap();
1398 2772 : let upload_queue = guard
1399 2772 : .initialized_mut()
1400 2772 : .map_err(WaitCompletionError::NotInitialized)?;
1401 2772 : self.schedule_barrier0(upload_queue)
1402 2772 : };
1403 2772 :
1404 2772 : Self::wait_completion0(receiver).await
1405 2772 : }
1406 :
1407 2772 : async fn wait_completion0(
1408 2772 : mut receiver: tokio::sync::watch::Receiver<()>,
1409 2772 : ) -> Result<(), WaitCompletionError> {
1410 2772 : if receiver.changed().await.is_err() {
1411 0 : return Err(WaitCompletionError::UploadQueueShutDownOrStopped);
1412 2772 : }
1413 2772 :
1414 2772 : Ok(())
1415 2772 : }
1416 :
1417 12 : pub(crate) fn schedule_barrier(self: &Arc<Self>) -> anyhow::Result<()> {
1418 12 : let mut guard = self.upload_queue.lock().unwrap();
1419 12 : let upload_queue = guard.initialized_mut()?;
1420 12 : self.schedule_barrier0(upload_queue);
1421 12 : Ok(())
1422 12 : }
1423 :
1424 2784 : fn schedule_barrier0(
1425 2784 : self: &Arc<Self>,
1426 2784 : upload_queue: &mut UploadQueueInitialized,
1427 2784 : ) -> tokio::sync::watch::Receiver<()> {
1428 2784 : let (sender, receiver) = tokio::sync::watch::channel(());
1429 2784 : let barrier_op = UploadOp::Barrier(sender);
1430 2784 :
1431 2784 : upload_queue.queued_operations.push_back(barrier_op);
1432 2784 : // Don't count this kind of operation!
1433 2784 :
1434 2784 : // Launch the task immediately, if possible
1435 2784 : self.launch_queued_tasks(upload_queue);
1436 2784 :
1437 2784 : receiver
1438 2784 : }
1439 :
1440 : /// Wait for all previously scheduled operations to complete, and then stop.
1441 : ///
1442 : /// Not cancellation safe
1443 20 : pub(crate) async fn shutdown(self: &Arc<Self>) {
1444 20 : // On cancellation the queue is left in ackward state of refusing new operations but
1445 20 : // proper stop is yet to be called. On cancel the original or some later task must call
1446 20 : // `stop` or `shutdown`.
1447 20 : let sg = scopeguard::guard((), |_| {
1448 0 : tracing::error!("RemoteTimelineClient::shutdown was cancelled; this should not happen, do not make this into an allowed_error")
1449 20 : });
1450 :
1451 16 : let fut = {
1452 20 : let mut guard = self.upload_queue.lock().unwrap();
1453 20 : let upload_queue = match &mut *guard {
1454 : UploadQueue::Stopped(_) => {
1455 4 : scopeguard::ScopeGuard::into_inner(sg);
1456 4 : return;
1457 : }
1458 : UploadQueue::Uninitialized => {
1459 : // transition into Stopped state
1460 0 : self.stop_impl(&mut guard);
1461 0 : scopeguard::ScopeGuard::into_inner(sg);
1462 0 : return;
1463 : }
1464 16 : UploadQueue::Initialized(ref mut init) => init,
1465 16 : };
1466 16 :
1467 16 : // if the queue is already stuck due to a shutdown operation which was cancelled, then
1468 16 : // just don't add more of these as they would never complete.
1469 16 : //
1470 16 : // TODO: if launch_queued_tasks were to be refactored to accept a &mut UploadQueue
1471 16 : // in every place we would not have to jump through this hoop, and this method could be
1472 16 : // made cancellable.
1473 16 : if !upload_queue.shutting_down {
1474 12 : upload_queue.shutting_down = true;
1475 12 : upload_queue.queued_operations.push_back(UploadOp::Shutdown);
1476 12 : // this operation is not counted similar to Barrier
1477 12 :
1478 12 : self.launch_queued_tasks(upload_queue);
1479 12 : }
1480 :
1481 16 : upload_queue.shutdown_ready.clone().acquire_owned()
1482 : };
1483 :
1484 16 : let res = fut.await;
1485 :
1486 16 : scopeguard::ScopeGuard::into_inner(sg);
1487 16 :
1488 16 : match res {
1489 0 : Ok(_permit) => unreachable!("shutdown_ready should not have been added permits"),
1490 16 : Err(_closed) => {
1491 16 : // expected
1492 16 : }
1493 16 : }
1494 16 :
1495 16 : self.stop();
1496 20 : }
1497 :
1498 : /// Set the deleted_at field in the remote index file.
1499 : ///
1500 : /// This fails if the upload queue has not been `stop()`ed.
1501 : ///
1502 : /// The caller is responsible for calling `stop()` AND for waiting
1503 : /// for any ongoing upload tasks to finish after `stop()` has succeeded.
1504 : /// Check method [`RemoteTimelineClient::stop`] for details.
1505 : #[instrument(skip_all)]
1506 : pub(crate) async fn persist_index_part_with_deleted_flag(
1507 : self: &Arc<Self>,
1508 : ) -> Result<(), PersistIndexPartWithDeletedFlagError> {
1509 : let index_part_with_deleted_at = {
1510 : let mut locked = self.upload_queue.lock().unwrap();
1511 :
1512 : // We must be in stopped state because otherwise
1513 : // we can have inprogress index part upload that can overwrite the file
1514 : // with missing is_deleted flag that we going to set below
1515 : let stopped = locked.stopped_mut()?;
1516 :
1517 : match stopped.deleted_at {
1518 : SetDeletedFlagProgress::NotRunning => (), // proceed
1519 : SetDeletedFlagProgress::InProgress(at) => {
1520 : return Err(PersistIndexPartWithDeletedFlagError::AlreadyInProgress(at));
1521 : }
1522 : SetDeletedFlagProgress::Successful(at) => {
1523 : return Err(PersistIndexPartWithDeletedFlagError::AlreadyDeleted(at));
1524 : }
1525 : };
1526 : let deleted_at = Utc::now().naive_utc();
1527 : stopped.deleted_at = SetDeletedFlagProgress::InProgress(deleted_at);
1528 :
1529 : let mut index_part = stopped.upload_queue_for_deletion.dirty.clone();
1530 : index_part.deleted_at = Some(deleted_at);
1531 : index_part
1532 : };
1533 :
1534 0 : let undo_deleted_at = scopeguard::guard(Arc::clone(self), |self_clone| {
1535 0 : let mut locked = self_clone.upload_queue.lock().unwrap();
1536 0 : let stopped = locked
1537 0 : .stopped_mut()
1538 0 : .expect("there's no way out of Stopping, and we checked it's Stopping above");
1539 0 : stopped.deleted_at = SetDeletedFlagProgress::NotRunning;
1540 0 : });
1541 :
1542 : pausable_failpoint!("persist_deleted_index_part");
1543 :
1544 : backoff::retry(
1545 0 : || {
1546 0 : upload::upload_index_part(
1547 0 : &self.storage_impl,
1548 0 : &self.tenant_shard_id,
1549 0 : &self.timeline_id,
1550 0 : self.generation,
1551 0 : &index_part_with_deleted_at,
1552 0 : &self.cancel,
1553 0 : )
1554 0 : },
1555 0 : |_e| false,
1556 : 1,
1557 : // have just a couple of attempts
1558 : // when executed as part of timeline deletion this happens in context of api call
1559 : // when executed as part of tenant deletion this happens in the background
1560 : 2,
1561 : "persist_index_part_with_deleted_flag",
1562 : &self.cancel,
1563 : )
1564 : .await
1565 0 : .ok_or_else(|| anyhow::Error::new(TimeoutOrCancel::Cancel))
1566 0 : .and_then(|x| x)?;
1567 :
1568 : // all good, disarm the guard and mark as success
1569 : ScopeGuard::into_inner(undo_deleted_at);
1570 : {
1571 : let mut locked = self.upload_queue.lock().unwrap();
1572 :
1573 : let stopped = locked
1574 : .stopped_mut()
1575 : .expect("there's no way out of Stopping, and we checked it's Stopping above");
1576 : stopped.deleted_at = SetDeletedFlagProgress::Successful(
1577 : index_part_with_deleted_at
1578 : .deleted_at
1579 : .expect("we set it above"),
1580 : );
1581 : }
1582 :
1583 : Ok(())
1584 : }
1585 :
1586 0 : pub(crate) fn is_deleting(&self) -> bool {
1587 0 : let mut locked = self.upload_queue.lock().unwrap();
1588 0 : locked.stopped_mut().is_ok()
1589 0 : }
1590 :
1591 0 : pub(crate) async fn preserve_initdb_archive(
1592 0 : self: &Arc<Self>,
1593 0 : tenant_id: &TenantId,
1594 0 : timeline_id: &TimelineId,
1595 0 : cancel: &CancellationToken,
1596 0 : ) -> anyhow::Result<()> {
1597 0 : backoff::retry(
1598 0 : || async {
1599 0 : upload::preserve_initdb_archive(&self.storage_impl, tenant_id, timeline_id, cancel)
1600 0 : .await
1601 0 : },
1602 0 : TimeoutOrCancel::caused_by_cancel,
1603 0 : FAILED_DOWNLOAD_WARN_THRESHOLD,
1604 0 : FAILED_REMOTE_OP_RETRIES,
1605 0 : "preserve_initdb_tar_zst",
1606 0 : &cancel.clone(),
1607 0 : )
1608 0 : .await
1609 0 : .ok_or_else(|| anyhow::Error::new(TimeoutOrCancel::Cancel))
1610 0 : .and_then(|x| x)
1611 0 : .context("backing up initdb archive")?;
1612 0 : Ok(())
1613 0 : }
1614 :
1615 : /// Uploads the given layer **without** adding it to be part of a future `index_part.json` upload.
1616 : ///
1617 : /// This is not normally needed.
1618 0 : pub(crate) async fn upload_layer_file(
1619 0 : self: &Arc<Self>,
1620 0 : uploaded: &ResidentLayer,
1621 0 : cancel: &CancellationToken,
1622 0 : ) -> anyhow::Result<()> {
1623 0 : let remote_path = remote_layer_path(
1624 0 : &self.tenant_shard_id.tenant_id,
1625 0 : &self.timeline_id,
1626 0 : uploaded.metadata().shard,
1627 0 : &uploaded.layer_desc().layer_name(),
1628 0 : uploaded.metadata().generation,
1629 0 : );
1630 0 :
1631 0 : backoff::retry(
1632 0 : || async {
1633 0 : upload::upload_timeline_layer(
1634 0 : &self.storage_impl,
1635 0 : uploaded.local_path(),
1636 0 : &remote_path,
1637 0 : uploaded.metadata().file_size,
1638 0 : cancel,
1639 0 : )
1640 0 : .await
1641 0 : },
1642 0 : TimeoutOrCancel::caused_by_cancel,
1643 0 : FAILED_UPLOAD_WARN_THRESHOLD,
1644 0 : FAILED_REMOTE_OP_RETRIES,
1645 0 : "upload a layer without adding it to latest files",
1646 0 : cancel,
1647 0 : )
1648 0 : .await
1649 0 : .ok_or_else(|| anyhow::Error::new(TimeoutOrCancel::Cancel))
1650 0 : .and_then(|x| x)
1651 0 : .context("upload a layer without adding it to latest files")
1652 0 : }
1653 :
1654 : /// Copies the `adopted` remote existing layer to the remote path of `adopted_as`. The layer is
1655 : /// not added to be part of a future `index_part.json` upload.
1656 0 : pub(crate) async fn copy_timeline_layer(
1657 0 : self: &Arc<Self>,
1658 0 : adopted: &Layer,
1659 0 : adopted_as: &Layer,
1660 0 : cancel: &CancellationToken,
1661 0 : ) -> anyhow::Result<()> {
1662 0 : let source_remote_path = remote_layer_path(
1663 0 : &self.tenant_shard_id.tenant_id,
1664 0 : &adopted
1665 0 : .get_timeline_id()
1666 0 : .expect("Source timeline should be alive"),
1667 0 : adopted.metadata().shard,
1668 0 : &adopted.layer_desc().layer_name(),
1669 0 : adopted.metadata().generation,
1670 0 : );
1671 0 :
1672 0 : let target_remote_path = remote_layer_path(
1673 0 : &self.tenant_shard_id.tenant_id,
1674 0 : &self.timeline_id,
1675 0 : adopted_as.metadata().shard,
1676 0 : &adopted_as.layer_desc().layer_name(),
1677 0 : adopted_as.metadata().generation,
1678 0 : );
1679 0 :
1680 0 : backoff::retry(
1681 0 : || async {
1682 0 : upload::copy_timeline_layer(
1683 0 : &self.storage_impl,
1684 0 : &source_remote_path,
1685 0 : &target_remote_path,
1686 0 : cancel,
1687 0 : )
1688 0 : .await
1689 0 : },
1690 0 : TimeoutOrCancel::caused_by_cancel,
1691 0 : FAILED_UPLOAD_WARN_THRESHOLD,
1692 0 : FAILED_REMOTE_OP_RETRIES,
1693 0 : "copy timeline layer",
1694 0 : cancel,
1695 0 : )
1696 0 : .await
1697 0 : .ok_or_else(|| anyhow::Error::new(TimeoutOrCancel::Cancel))
1698 0 : .and_then(|x| x)
1699 0 : .context("remote copy timeline layer")
1700 0 : }
1701 :
1702 0 : async fn flush_deletion_queue(&self) -> Result<(), DeletionQueueError> {
1703 0 : match tokio::time::timeout(
1704 0 : DELETION_QUEUE_FLUSH_TIMEOUT,
1705 0 : self.deletion_queue_client.flush_immediate(),
1706 0 : )
1707 0 : .await
1708 : {
1709 0 : Ok(result) => result,
1710 0 : Err(_timeout) => {
1711 0 : // Flushing remote deletions is not mandatory: we flush here to make the system easier to test, and
1712 0 : // to ensure that _usually_ objects are really gone after a DELETE is acked. However, in case of deletion
1713 0 : // queue issues (https://github.com/neondatabase/neon/issues/6440), we don't want to wait indefinitely here.
1714 0 : tracing::warn!(
1715 0 : "Timed out waiting for deletion queue flush, acking deletion anyway"
1716 : );
1717 0 : Ok(())
1718 : }
1719 : }
1720 0 : }
1721 :
1722 : /// Prerequisites: UploadQueue should be in stopped state and deleted_at should be successfuly set.
1723 : /// The function deletes layer files one by one, then lists the prefix to see if we leaked something
1724 : /// deletes leaked files if any and proceeds with deletion of index file at the end.
1725 0 : pub(crate) async fn delete_all(self: &Arc<Self>) -> Result<(), DeleteTimelineError> {
1726 0 : debug_assert_current_span_has_tenant_and_timeline_id();
1727 :
1728 0 : let layers: Vec<RemotePath> = {
1729 0 : let mut locked = self.upload_queue.lock().unwrap();
1730 0 : let stopped = locked.stopped_mut().map_err(DeleteTimelineError::Other)?;
1731 :
1732 0 : if !matches!(stopped.deleted_at, SetDeletedFlagProgress::Successful(_)) {
1733 0 : return Err(DeleteTimelineError::Other(anyhow::anyhow!(
1734 0 : "deleted_at is not set"
1735 0 : )));
1736 0 : }
1737 0 :
1738 0 : debug_assert!(stopped.upload_queue_for_deletion.no_pending_work());
1739 :
1740 0 : stopped
1741 0 : .upload_queue_for_deletion
1742 0 : .dirty
1743 0 : .layer_metadata
1744 0 : .drain()
1745 0 : .filter(|(_file_name, meta)| {
1746 0 : // Filter out layers that belonged to an ancestor shard. Since we are deleting the whole timeline from
1747 0 : // all shards anyway, we _could_ delete these, but
1748 0 : // - it creates a potential race if other shards are still
1749 0 : // using the layers while this shard deletes them.
1750 0 : // - it means that if we rolled back the shard split, the ancestor shards would be in a state where
1751 0 : // these timelines are present but corrupt (their index exists but some layers don't)
1752 0 : //
1753 0 : // These layers will eventually be cleaned up by the scrubber when it does physical GC.
1754 0 : meta.shard.shard_number == self.tenant_shard_id.shard_number
1755 0 : && meta.shard.shard_count == self.tenant_shard_id.shard_count
1756 0 : })
1757 0 : .map(|(file_name, meta)| {
1758 0 : remote_layer_path(
1759 0 : &self.tenant_shard_id.tenant_id,
1760 0 : &self.timeline_id,
1761 0 : meta.shard,
1762 0 : &file_name,
1763 0 : meta.generation,
1764 0 : )
1765 0 : })
1766 0 : .collect()
1767 0 : };
1768 0 :
1769 0 : let layer_deletion_count = layers.len();
1770 0 : self.deletion_queue_client
1771 0 : .push_immediate(layers)
1772 0 : .await
1773 0 : .map_err(|_| DeleteTimelineError::Cancelled)?;
1774 :
1775 : // Delete the initdb.tar.zst, which is not always present, but deletion attempts of
1776 : // inexistant objects are not considered errors.
1777 0 : let initdb_path =
1778 0 : remote_initdb_archive_path(&self.tenant_shard_id.tenant_id, &self.timeline_id);
1779 0 : self.deletion_queue_client
1780 0 : .push_immediate(vec![initdb_path])
1781 0 : .await
1782 0 : .map_err(|_| DeleteTimelineError::Cancelled)?;
1783 :
1784 : // Do not delete index part yet, it is needed for possible retry. If we remove it first
1785 : // and retry will arrive to different pageserver there wont be any traces of it on remote storage
1786 0 : let timeline_storage_path = remote_timeline_path(&self.tenant_shard_id, &self.timeline_id);
1787 0 :
1788 0 : // Execute all pending deletions, so that when we proceed to do a listing below, we aren't
1789 0 : // taking the burden of listing all the layers that we already know we should delete.
1790 0 : self.flush_deletion_queue()
1791 0 : .await
1792 0 : .map_err(|_| DeleteTimelineError::Cancelled)?;
1793 :
1794 0 : let cancel = shutdown_token();
1795 :
1796 0 : let remaining = download_retry(
1797 0 : || async {
1798 0 : self.storage_impl
1799 0 : .list(
1800 0 : Some(&timeline_storage_path),
1801 0 : ListingMode::NoDelimiter,
1802 0 : None,
1803 0 : &cancel,
1804 0 : )
1805 0 : .await
1806 0 : },
1807 0 : "list remaining files",
1808 0 : &cancel,
1809 0 : )
1810 0 : .await
1811 0 : .context("list files remaining files")?
1812 : .keys;
1813 :
1814 : // We will delete the current index_part object last, since it acts as a deletion
1815 : // marker via its deleted_at attribute
1816 0 : let latest_index = remaining
1817 0 : .iter()
1818 0 : .filter(|o| {
1819 0 : o.key
1820 0 : .object_name()
1821 0 : .map(|n| n.starts_with(IndexPart::FILE_NAME))
1822 0 : .unwrap_or(false)
1823 0 : })
1824 0 : .filter_map(|o| parse_remote_index_path(o.key.clone()).map(|gen| (o.key.clone(), gen)))
1825 0 : .max_by_key(|i| i.1)
1826 0 : .map(|i| i.0.clone())
1827 0 : .unwrap_or(
1828 0 : // No generation-suffixed indices, assume we are dealing with
1829 0 : // a legacy index.
1830 0 : remote_index_path(&self.tenant_shard_id, &self.timeline_id, Generation::none()),
1831 0 : );
1832 0 :
1833 0 : let remaining_layers: Vec<RemotePath> = remaining
1834 0 : .into_iter()
1835 0 : .filter_map(|o| {
1836 0 : if o.key == latest_index || o.key.object_name() == Some(INITDB_PRESERVED_PATH) {
1837 0 : None
1838 : } else {
1839 0 : Some(o.key)
1840 : }
1841 0 : })
1842 0 : .inspect(|path| {
1843 0 : if let Some(name) = path.object_name() {
1844 0 : info!(%name, "deleting a file not referenced from index_part.json");
1845 : } else {
1846 0 : warn!(%path, "deleting a nameless or non-utf8 object not referenced from index_part.json");
1847 : }
1848 0 : })
1849 0 : .collect();
1850 0 :
1851 0 : let not_referenced_count = remaining_layers.len();
1852 0 : if !remaining_layers.is_empty() {
1853 0 : self.deletion_queue_client
1854 0 : .push_immediate(remaining_layers)
1855 0 : .await
1856 0 : .map_err(|_| DeleteTimelineError::Cancelled)?;
1857 0 : }
1858 :
1859 0 : fail::fail_point!("timeline-delete-before-index-delete", |_| {
1860 0 : Err(DeleteTimelineError::Other(anyhow::anyhow!(
1861 0 : "failpoint: timeline-delete-before-index-delete"
1862 0 : )))?
1863 0 : });
1864 :
1865 0 : debug!("enqueuing index part deletion");
1866 0 : self.deletion_queue_client
1867 0 : .push_immediate([latest_index].to_vec())
1868 0 : .await
1869 0 : .map_err(|_| DeleteTimelineError::Cancelled)?;
1870 :
1871 : // Timeline deletion is rare and we have probably emitted a reasonably number of objects: wait
1872 : // for a flush to a persistent deletion list so that we may be sure deletion will occur.
1873 0 : self.flush_deletion_queue()
1874 0 : .await
1875 0 : .map_err(|_| DeleteTimelineError::Cancelled)?;
1876 :
1877 0 : fail::fail_point!("timeline-delete-after-index-delete", |_| {
1878 0 : Err(DeleteTimelineError::Other(anyhow::anyhow!(
1879 0 : "failpoint: timeline-delete-after-index-delete"
1880 0 : )))?
1881 0 : });
1882 :
1883 0 : info!(prefix=%timeline_storage_path, referenced=layer_deletion_count, not_referenced=%not_referenced_count, "done deleting in timeline prefix, including index_part.json");
1884 :
1885 0 : Ok(())
1886 0 : }
1887 :
1888 : /// Pick next tasks from the queue, and start as many of them as possible without violating
1889 : /// the ordering constraints.
1890 : ///
1891 : /// TODO: consider limiting the number of in-progress tasks, beyond what remote_storage does.
1892 : /// This can launch an unbounded number of queued tasks. `UploadQueue::next_ready()` also has
1893 : /// worst-case quadratic cost in the number of tasks, and may struggle beyond 10,000 tasks.
1894 16499 : fn launch_queued_tasks(self: &Arc<Self>, upload_queue: &mut UploadQueueInitialized) {
1895 26423 : while let Some((mut next_op, coalesced_ops)) = upload_queue.next_ready() {
1896 9924 : debug!("starting op: {next_op}");
1897 :
1898 : // Prepare upload.
1899 9924 : match &mut next_op {
1900 3456 : UploadOp::UploadLayer(layer, meta, mode) => {
1901 3456 : if upload_queue
1902 3456 : .recently_deleted
1903 3456 : .remove(&(layer.layer_desc().layer_name().clone(), meta.generation))
1904 0 : {
1905 0 : *mode = Some(OpType::FlushDeletion);
1906 0 : } else {
1907 3456 : *mode = Some(OpType::MayReorder)
1908 : }
1909 : }
1910 3005 : UploadOp::UploadMetadata { .. } => {}
1911 679 : UploadOp::Delete(Delete { layers }) => {
1912 1358 : for (name, meta) in layers {
1913 679 : upload_queue
1914 679 : .recently_deleted
1915 679 : .insert((name.clone(), meta.generation));
1916 679 : }
1917 : }
1918 2784 : UploadOp::Barrier(sender) => {
1919 2784 : sender.send_replace(());
1920 2784 : continue;
1921 : }
1922 0 : UploadOp::Shutdown => unreachable!("shutdown is intentionally never popped off"),
1923 : };
1924 :
1925 : // Assign unique ID to this task
1926 7140 : upload_queue.task_counter += 1;
1927 7140 : let upload_task_id = upload_queue.task_counter;
1928 7140 :
1929 7140 : // Add it to the in-progress map
1930 7140 : let task = Arc::new(UploadTask {
1931 7140 : task_id: upload_task_id,
1932 7140 : op: next_op,
1933 7140 : coalesced_ops,
1934 7140 : retries: AtomicU32::new(0),
1935 7140 : });
1936 7140 : upload_queue
1937 7140 : .inprogress_tasks
1938 7140 : .insert(task.task_id, Arc::clone(&task));
1939 7140 :
1940 7140 : // Spawn task to perform the task
1941 7140 : let self_rc = Arc::clone(self);
1942 7140 : let tenant_shard_id = self.tenant_shard_id;
1943 7140 : let timeline_id = self.timeline_id;
1944 7140 : task_mgr::spawn(
1945 7140 : &self.runtime,
1946 7140 : TaskKind::RemoteUploadTask,
1947 7140 : self.tenant_shard_id,
1948 7140 : Some(self.timeline_id),
1949 7140 : "remote upload",
1950 6987 : async move {
1951 6987 : self_rc.perform_upload_task(task).await;
1952 6696 : Ok(())
1953 6696 : }
1954 7140 : .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)),
1955 : );
1956 :
1957 : // Loop back to process next task
1958 : }
1959 16499 : }
1960 :
1961 : ///
1962 : /// Perform an upload task.
1963 : ///
1964 : /// The task is in the `inprogress_tasks` list. This function will try to
1965 : /// execute it, retrying forever. On successful completion, the task is
1966 : /// removed it from the `inprogress_tasks` list, and any next task(s) in the
1967 : /// queue that were waiting by the completion are launched.
1968 : ///
1969 : /// The task can be shut down, however. That leads to stopping the whole
1970 : /// queue.
1971 : ///
1972 6987 : async fn perform_upload_task(self: &Arc<Self>, task: Arc<UploadTask>) {
1973 6987 : let cancel = shutdown_token();
1974 : // Loop to retry until it completes.
1975 : loop {
1976 : // If we're requested to shut down, close up shop and exit.
1977 : //
1978 : // Note: We only check for the shutdown requests between retries, so
1979 : // if a shutdown request arrives while we're busy uploading, in the
1980 : // upload::upload:*() call below, we will wait not exit until it has
1981 : // finished. We probably could cancel the upload by simply dropping
1982 : // the Future, but we're not 100% sure if the remote storage library
1983 : // is cancellation safe, so we don't dare to do that. Hopefully, the
1984 : // upload finishes or times out soon enough.
1985 6987 : if cancel.is_cancelled() {
1986 0 : info!("upload task cancelled by shutdown request");
1987 0 : self.stop();
1988 0 : return;
1989 6987 : }
1990 6987 :
1991 6987 : // Assert that we don't modify a layer that's referenced by the current index.
1992 6987 : if cfg!(debug_assertions) {
1993 6987 : let modified = match &task.op {
1994 3326 : UploadOp::UploadLayer(layer, layer_metadata, _) => {
1995 3326 : vec![(layer.layer_desc().layer_name(), layer_metadata)]
1996 : }
1997 679 : UploadOp::Delete(delete) => {
1998 679 : delete.layers.iter().map(|(n, m)| (n.clone(), m)).collect()
1999 : }
2000 : // These don't modify layers.
2001 2982 : UploadOp::UploadMetadata { .. } => Vec::new(),
2002 0 : UploadOp::Barrier(_) => Vec::new(),
2003 0 : UploadOp::Shutdown => Vec::new(),
2004 : };
2005 6987 : if let Ok(queue) = self.upload_queue.lock().unwrap().initialized_mut() {
2006 10984 : for (ref name, metadata) in modified {
2007 4005 : debug_assert!(
2008 4005 : !queue.clean.0.references(name, metadata),
2009 8 : "layer {name} modified while referenced by index",
2010 : );
2011 : }
2012 0 : }
2013 0 : }
2014 :
2015 6979 : let upload_result: anyhow::Result<()> = match &task.op {
2016 3326 : UploadOp::UploadLayer(ref layer, ref layer_metadata, mode) => {
2017 : // TODO: check if this mechanism can be removed now that can_bypass() performs
2018 : // conflict checks during scheduling.
2019 3326 : if let Some(OpType::FlushDeletion) = mode {
2020 0 : if self.config.read().unwrap().block_deletions {
2021 : // Of course, this is not efficient... but usually the queue should be empty.
2022 0 : let mut queue_locked = self.upload_queue.lock().unwrap();
2023 0 : let mut detected = false;
2024 0 : if let Ok(queue) = queue_locked.initialized_mut() {
2025 0 : for list in queue.blocked_deletions.iter_mut() {
2026 0 : list.layers.retain(|(name, meta)| {
2027 0 : if name == &layer.layer_desc().layer_name()
2028 0 : && meta.generation == layer_metadata.generation
2029 : {
2030 0 : detected = true;
2031 0 : // remove the layer from deletion queue
2032 0 : false
2033 : } else {
2034 : // keep the layer
2035 0 : true
2036 : }
2037 0 : });
2038 0 : }
2039 0 : }
2040 0 : if detected {
2041 0 : info!(
2042 0 : "cancelled blocked deletion of layer {} at gen {:?}",
2043 0 : layer.layer_desc().layer_name(),
2044 : layer_metadata.generation
2045 : );
2046 0 : }
2047 : } else {
2048 : // TODO: we did not guarantee that upload task starts after deletion task, so there could be possibly race conditions
2049 : // that we still get the layer deleted. But this only happens if someone creates a layer immediately after it's deleted,
2050 : // which is not possible in the current system.
2051 0 : info!(
2052 0 : "waiting for deletion queue flush to complete before uploading layer {} at gen {:?}",
2053 0 : layer.layer_desc().layer_name(),
2054 : layer_metadata.generation
2055 : );
2056 : {
2057 : // We are going to flush, we can clean up the recently deleted list.
2058 0 : let mut queue_locked = self.upload_queue.lock().unwrap();
2059 0 : if let Ok(queue) = queue_locked.initialized_mut() {
2060 0 : queue.recently_deleted.clear();
2061 0 : }
2062 : }
2063 0 : if let Err(e) = self.deletion_queue_client.flush_execute().await {
2064 0 : warn!(
2065 0 : "failed to flush the deletion queue before uploading layer {} at gen {:?}, still proceeding to upload: {e:#} ",
2066 0 : layer.layer_desc().layer_name(),
2067 : layer_metadata.generation
2068 : );
2069 : } else {
2070 0 : info!(
2071 0 : "done flushing deletion queue before uploading layer {} at gen {:?}",
2072 0 : layer.layer_desc().layer_name(),
2073 : layer_metadata.generation
2074 : );
2075 : }
2076 : }
2077 3326 : }
2078 3326 : let local_path = layer.local_path();
2079 3326 :
2080 3326 : // We should only be uploading layers created by this `Tenant`'s lifetime, so
2081 3326 : // the metadata in the upload should always match our current generation.
2082 3326 : assert_eq!(layer_metadata.generation, self.generation);
2083 :
2084 3326 : let remote_path = remote_layer_path(
2085 3326 : &self.tenant_shard_id.tenant_id,
2086 3326 : &self.timeline_id,
2087 3326 : layer_metadata.shard,
2088 3326 : &layer.layer_desc().layer_name(),
2089 3326 : layer_metadata.generation,
2090 3326 : );
2091 3326 :
2092 3326 : upload::upload_timeline_layer(
2093 3326 : &self.storage_impl,
2094 3326 : local_path,
2095 3326 : &remote_path,
2096 3326 : layer_metadata.file_size,
2097 3326 : &self.cancel,
2098 3326 : )
2099 3326 : .measure_remote_op(
2100 3326 : RemoteOpFileKind::Layer,
2101 3326 : RemoteOpKind::Upload,
2102 3326 : Arc::clone(&self.metrics),
2103 3326 : )
2104 3326 : .await
2105 : }
2106 2982 : UploadOp::UploadMetadata { ref uploaded } => {
2107 2982 : let res = upload::upload_index_part(
2108 2982 : &self.storage_impl,
2109 2982 : &self.tenant_shard_id,
2110 2982 : &self.timeline_id,
2111 2982 : self.generation,
2112 2982 : uploaded,
2113 2982 : &self.cancel,
2114 2982 : )
2115 2982 : .measure_remote_op(
2116 2982 : RemoteOpFileKind::Index,
2117 2982 : RemoteOpKind::Upload,
2118 2982 : Arc::clone(&self.metrics),
2119 2982 : )
2120 2982 : .await;
2121 2962 : if res.is_ok() {
2122 2962 : self.update_remote_physical_size_gauge(Some(uploaded));
2123 2962 : let mention_having_future_layers = if cfg!(feature = "testing") {
2124 2962 : uploaded
2125 2962 : .layer_metadata
2126 2962 : .keys()
2127 35443 : .any(|x| x.is_in_future(uploaded.metadata.disk_consistent_lsn()))
2128 : } else {
2129 0 : false
2130 : };
2131 2962 : if mention_having_future_layers {
2132 : // find rationale near crate::tenant::timeline::init::cleanup_future_layer
2133 119 : tracing::info!(
2134 0 : disk_consistent_lsn = %uploaded.metadata.disk_consistent_lsn(),
2135 0 : "uploaded an index_part.json with future layers -- this is ok! if shutdown now, expect future layer cleanup"
2136 : );
2137 2843 : }
2138 0 : }
2139 2962 : res
2140 : }
2141 671 : UploadOp::Delete(delete) => {
2142 671 : if self.config.read().unwrap().block_deletions {
2143 0 : let mut queue_locked = self.upload_queue.lock().unwrap();
2144 0 : if let Ok(queue) = queue_locked.initialized_mut() {
2145 0 : queue.blocked_deletions.push(delete.clone());
2146 0 : }
2147 0 : Ok(())
2148 : } else {
2149 671 : pausable_failpoint!("before-delete-layer-pausable");
2150 671 : self.deletion_queue_client
2151 671 : .push_layers(
2152 671 : self.tenant_shard_id,
2153 671 : self.timeline_id,
2154 671 : self.generation,
2155 671 : delete.layers.clone(),
2156 671 : )
2157 671 : .await
2158 671 : .map_err(|e| anyhow::anyhow!(e))
2159 : }
2160 : }
2161 0 : unexpected @ UploadOp::Barrier(_) | unexpected @ UploadOp::Shutdown => {
2162 : // unreachable. Barrier operations are handled synchronously in
2163 : // launch_queued_tasks
2164 0 : warn!("unexpected {unexpected:?} operation in perform_upload_task");
2165 0 : break;
2166 : }
2167 : };
2168 :
2169 0 : match upload_result {
2170 : Ok(()) => {
2171 6696 : break;
2172 : }
2173 0 : Err(e) if TimeoutOrCancel::caused_by_cancel(&e) => {
2174 0 : // loop around to do the proper stopping
2175 0 : continue;
2176 : }
2177 0 : Err(e) => {
2178 0 : let retries = task.retries.fetch_add(1, Ordering::SeqCst);
2179 0 :
2180 0 : // Uploads can fail due to rate limits (IAM, S3), spurious network problems,
2181 0 : // or other external reasons. Such issues are relatively regular, so log them
2182 0 : // at info level at first, and only WARN if the operation fails repeatedly.
2183 0 : //
2184 0 : // (See similar logic for downloads in `download::download_retry`)
2185 0 : if retries < FAILED_UPLOAD_WARN_THRESHOLD {
2186 0 : info!(
2187 0 : "failed to perform remote task {}, will retry (attempt {}): {:#}",
2188 0 : task.op, retries, e
2189 : );
2190 : } else {
2191 0 : warn!(
2192 0 : "failed to perform remote task {}, will retry (attempt {}): {:?}",
2193 0 : task.op, retries, e
2194 : );
2195 : }
2196 :
2197 : // sleep until it's time to retry, or we're cancelled
2198 0 : exponential_backoff(
2199 0 : retries,
2200 0 : DEFAULT_BASE_BACKOFF_SECONDS,
2201 0 : DEFAULT_MAX_BACKOFF_SECONDS,
2202 0 : &cancel,
2203 0 : )
2204 0 : .await;
2205 : }
2206 : }
2207 : }
2208 :
2209 6696 : let retries = task.retries.load(Ordering::SeqCst);
2210 6696 : if retries > 0 {
2211 0 : info!(
2212 0 : "remote task {} completed successfully after {} retries",
2213 0 : task.op, retries
2214 : );
2215 : } else {
2216 6696 : debug!("remote task {} completed successfully", task.op);
2217 : }
2218 :
2219 : // The task has completed successfully. Remove it from the in-progress list.
2220 6696 : let lsn_update = {
2221 6696 : let mut upload_queue_guard = self.upload_queue.lock().unwrap();
2222 6696 : let upload_queue = match upload_queue_guard.deref_mut() {
2223 0 : UploadQueue::Uninitialized => panic!("callers are responsible for ensuring this is only called on an initialized queue"),
2224 0 : UploadQueue::Stopped(_stopped) => {
2225 0 : None
2226 : },
2227 6696 : UploadQueue::Initialized(qi) => { Some(qi) }
2228 : };
2229 :
2230 6696 : let upload_queue = match upload_queue {
2231 6696 : Some(upload_queue) => upload_queue,
2232 : None => {
2233 0 : info!("another concurrent task already stopped the queue");
2234 0 : return;
2235 : }
2236 : };
2237 :
2238 6696 : upload_queue.inprogress_tasks.remove(&task.task_id);
2239 :
2240 6696 : let lsn_update = match task.op {
2241 3063 : UploadOp::UploadLayer(_, _, _) => None,
2242 2962 : UploadOp::UploadMetadata { ref uploaded } => {
2243 2962 : // the task id is reused as a monotonicity check for storing the "clean"
2244 2962 : // IndexPart.
2245 2962 : let last_updater = upload_queue.clean.1;
2246 2962 : let is_later = last_updater.is_some_and(|task_id| task_id < task.task_id);
2247 2962 : let monotone = is_later || last_updater.is_none();
2248 :
2249 2962 : assert!(monotone, "no two index uploads should be completing at the same time, prev={last_updater:?}, task.task_id={}", task.task_id);
2250 :
2251 : // not taking ownership is wasteful
2252 2962 : upload_queue.clean.0.clone_from(uploaded);
2253 2962 : upload_queue.clean.1 = Some(task.task_id);
2254 2962 :
2255 2962 : let lsn = upload_queue.clean.0.metadata.disk_consistent_lsn();
2256 2962 : self.metrics
2257 2962 : .projected_remote_consistent_lsn_gauge
2258 2962 : .set(lsn.0);
2259 2962 :
2260 2962 : if self.generation.is_none() {
2261 : // Legacy mode: skip validating generation
2262 0 : upload_queue.visible_remote_consistent_lsn.store(lsn);
2263 0 : None
2264 2962 : } else if self
2265 2962 : .config
2266 2962 : .read()
2267 2962 : .unwrap()
2268 2962 : .process_remote_consistent_lsn_updates
2269 : {
2270 2962 : Some((lsn, upload_queue.visible_remote_consistent_lsn.clone()))
2271 : } else {
2272 : // Our config disables remote_consistent_lsn updates: drop it.
2273 0 : None
2274 : }
2275 : }
2276 671 : UploadOp::Delete(_) => None,
2277 0 : UploadOp::Barrier(..) | UploadOp::Shutdown => unreachable!(),
2278 : };
2279 :
2280 : // Launch any queued tasks that were unblocked by this one.
2281 6696 : self.launch_queued_tasks(upload_queue);
2282 6696 : lsn_update
2283 : };
2284 :
2285 6696 : if let Some((lsn, slot)) = lsn_update {
2286 : // Updates to the remote_consistent_lsn we advertise to pageservers
2287 : // are all routed through the DeletionQueue, to enforce important
2288 : // data safety guarantees (see docs/rfcs/025-generation-numbers.md)
2289 2962 : self.deletion_queue_client
2290 2962 : .update_remote_consistent_lsn(
2291 2962 : self.tenant_shard_id,
2292 2962 : self.timeline_id,
2293 2962 : self.generation,
2294 2962 : lsn,
2295 2962 : slot,
2296 2962 : )
2297 2962 : .await;
2298 3734 : }
2299 :
2300 6696 : self.metric_end(&task.op);
2301 6696 : for coalesced_op in &task.coalesced_ops {
2302 4 : self.metric_end(coalesced_op);
2303 4 : }
2304 6696 : }
2305 :
2306 14271 : fn metric_impl(
2307 14271 : &self,
2308 14271 : op: &UploadOp,
2309 14271 : ) -> Option<(
2310 14271 : RemoteOpFileKind,
2311 14271 : RemoteOpKind,
2312 14271 : RemoteTimelineClientMetricsCallTrackSize,
2313 14271 : )> {
2314 : use RemoteTimelineClientMetricsCallTrackSize::DontTrackSize;
2315 14271 : let res = match op {
2316 6519 : UploadOp::UploadLayer(_, m, _) => (
2317 6519 : RemoteOpFileKind::Layer,
2318 6519 : RemoteOpKind::Upload,
2319 6519 : RemoteTimelineClientMetricsCallTrackSize::Bytes(m.file_size),
2320 6519 : ),
2321 6046 : UploadOp::UploadMetadata { .. } => (
2322 6046 : RemoteOpFileKind::Index,
2323 6046 : RemoteOpKind::Upload,
2324 6046 : DontTrackSize {
2325 6046 : reason: "metadata uploads are tiny",
2326 6046 : },
2327 6046 : ),
2328 1690 : UploadOp::Delete(_delete) => (
2329 1690 : RemoteOpFileKind::Layer,
2330 1690 : RemoteOpKind::Delete,
2331 1690 : DontTrackSize {
2332 1690 : reason: "should we track deletes? positive or negative sign?",
2333 1690 : },
2334 1690 : ),
2335 : UploadOp::Barrier(..) | UploadOp::Shutdown => {
2336 : // we do not account these
2337 16 : return None;
2338 : }
2339 : };
2340 14255 : Some(res)
2341 14271 : }
2342 :
2343 7555 : fn metric_begin(&self, op: &UploadOp) {
2344 7555 : let (file_kind, op_kind, track_bytes) = match self.metric_impl(op) {
2345 7555 : Some(x) => x,
2346 0 : None => return,
2347 : };
2348 7555 : let guard = self.metrics.call_begin(&file_kind, &op_kind, track_bytes);
2349 7555 : guard.will_decrement_manually(); // in metric_end(), see right below
2350 7555 : }
2351 :
2352 6716 : fn metric_end(&self, op: &UploadOp) {
2353 6716 : let (file_kind, op_kind, track_bytes) = match self.metric_impl(op) {
2354 6700 : Some(x) => x,
2355 16 : None => return,
2356 : };
2357 6700 : self.metrics.call_end(&file_kind, &op_kind, track_bytes);
2358 6716 : }
2359 :
2360 : /// Close the upload queue for new operations and cancel queued operations.
2361 : ///
2362 : /// Use [`RemoteTimelineClient::shutdown`] for graceful stop.
2363 : ///
2364 : /// In-progress operations will still be running after this function returns.
2365 : /// Use `task_mgr::shutdown_tasks(Some(TaskKind::RemoteUploadTask), Some(self.tenant_shard_id), Some(timeline_id))`
2366 : /// to wait for them to complete, after calling this function.
2367 36 : pub(crate) fn stop(&self) {
2368 36 : // Whichever *task* for this RemoteTimelineClient grabs the mutex first will transition the queue
2369 36 : // into stopped state, thereby dropping all off the queued *ops* which haven't become *tasks* yet.
2370 36 : // The other *tasks* will come here and observe an already shut down queue and hence simply wrap up their business.
2371 36 : let mut guard = self.upload_queue.lock().unwrap();
2372 36 : self.stop_impl(&mut guard);
2373 36 : }
2374 :
2375 36 : fn stop_impl(&self, guard: &mut std::sync::MutexGuard<UploadQueue>) {
2376 36 : match &mut **guard {
2377 : UploadQueue::Uninitialized => {
2378 0 : info!("UploadQueue is in state Uninitialized, nothing to do");
2379 0 : **guard = UploadQueue::Stopped(UploadQueueStopped::Uninitialized);
2380 : }
2381 : UploadQueue::Stopped(_) => {
2382 : // nothing to do
2383 16 : info!("another concurrent task already shut down the queue");
2384 : }
2385 20 : UploadQueue::Initialized(initialized) => {
2386 20 : info!("shutting down upload queue");
2387 :
2388 : // Replace the queue with the Stopped state, taking ownership of the old
2389 : // Initialized queue. We will do some checks on it, and then drop it.
2390 20 : let qi = {
2391 : // Here we preserve working version of the upload queue for possible use during deletions.
2392 : // In-place replace of Initialized to Stopped can be done with the help of https://github.com/Sgeo/take_mut
2393 : // but for this use case it doesnt really makes sense to bring unsafe code only for this usage point.
2394 : // Deletion is not really perf sensitive so there shouldnt be any problems with cloning a fraction of it.
2395 20 : let upload_queue_for_deletion = UploadQueueInitialized {
2396 20 : inprogress_limit: initialized.inprogress_limit,
2397 20 : task_counter: 0,
2398 20 : dirty: initialized.dirty.clone(),
2399 20 : clean: initialized.clean.clone(),
2400 20 : latest_files_changes_since_metadata_upload_scheduled: 0,
2401 20 : visible_remote_consistent_lsn: initialized
2402 20 : .visible_remote_consistent_lsn
2403 20 : .clone(),
2404 20 : inprogress_tasks: HashMap::default(),
2405 20 : queued_operations: VecDeque::default(),
2406 20 : #[cfg(feature = "testing")]
2407 20 : dangling_files: HashMap::default(),
2408 20 : blocked_deletions: Vec::new(),
2409 20 : shutting_down: false,
2410 20 : shutdown_ready: Arc::new(tokio::sync::Semaphore::new(0)),
2411 20 : recently_deleted: HashSet::new(),
2412 20 : };
2413 20 :
2414 20 : let upload_queue = std::mem::replace(
2415 20 : &mut **guard,
2416 20 : UploadQueue::Stopped(UploadQueueStopped::Deletable(
2417 20 : UploadQueueStoppedDeletable {
2418 20 : upload_queue_for_deletion,
2419 20 : deleted_at: SetDeletedFlagProgress::NotRunning,
2420 20 : },
2421 20 : )),
2422 20 : );
2423 20 : if let UploadQueue::Initialized(qi) = upload_queue {
2424 20 : qi
2425 : } else {
2426 0 : unreachable!("we checked in the match above that it is Initialized");
2427 : }
2428 : };
2429 :
2430 : // We don't need to do anything here for in-progress tasks. They will finish
2431 : // on their own, decrement the unfinished-task counter themselves, and observe
2432 : // that the queue is Stopped.
2433 20 : drop(qi.inprogress_tasks);
2434 :
2435 : // Tear down queued ops
2436 20 : for op in qi.queued_operations.into_iter() {
2437 16 : self.metric_end(&op);
2438 16 : // Dropping UploadOp::Barrier() here will make wait_completion() return with an Err()
2439 16 : // which is exactly what we want to happen.
2440 16 : drop(op);
2441 16 : }
2442 : }
2443 : }
2444 36 : }
2445 :
2446 : /// Returns an accessor which will hold the UploadQueue mutex for accessing the upload queue
2447 : /// externally to RemoteTimelineClient.
2448 0 : pub(crate) fn initialized_upload_queue(
2449 0 : &self,
2450 0 : ) -> Result<UploadQueueAccessor<'_>, NotInitialized> {
2451 0 : let mut inner = self.upload_queue.lock().unwrap();
2452 0 : inner.initialized_mut()?;
2453 0 : Ok(UploadQueueAccessor { inner })
2454 0 : }
2455 :
2456 16 : pub(crate) fn no_pending_work(&self) -> bool {
2457 16 : let inner = self.upload_queue.lock().unwrap();
2458 16 : match &*inner {
2459 : UploadQueue::Uninitialized
2460 0 : | UploadQueue::Stopped(UploadQueueStopped::Uninitialized) => true,
2461 16 : UploadQueue::Stopped(UploadQueueStopped::Deletable(x)) => {
2462 16 : x.upload_queue_for_deletion.no_pending_work()
2463 : }
2464 0 : UploadQueue::Initialized(x) => x.no_pending_work(),
2465 : }
2466 16 : }
2467 :
2468 : /// 'foreign' in the sense that it does not belong to this tenant shard. This method
2469 : /// is used during GC for other shards to get the index of shard zero.
2470 0 : pub(crate) async fn download_foreign_index(
2471 0 : &self,
2472 0 : shard_number: ShardNumber,
2473 0 : cancel: &CancellationToken,
2474 0 : ) -> Result<(IndexPart, Generation, std::time::SystemTime), DownloadError> {
2475 0 : let foreign_shard_id = TenantShardId {
2476 0 : shard_number,
2477 0 : shard_count: self.tenant_shard_id.shard_count,
2478 0 : tenant_id: self.tenant_shard_id.tenant_id,
2479 0 : };
2480 0 : download_index_part(
2481 0 : &self.storage_impl,
2482 0 : &foreign_shard_id,
2483 0 : &self.timeline_id,
2484 0 : Generation::MAX,
2485 0 : cancel,
2486 0 : )
2487 0 : .await
2488 0 : }
2489 : }
2490 :
2491 : pub(crate) struct UploadQueueAccessor<'a> {
2492 : inner: std::sync::MutexGuard<'a, UploadQueue>,
2493 : }
2494 :
2495 : impl UploadQueueAccessor<'_> {
2496 0 : pub(crate) fn latest_uploaded_index_part(&self) -> &IndexPart {
2497 0 : match &*self.inner {
2498 0 : UploadQueue::Initialized(x) => &x.clean.0,
2499 : UploadQueue::Uninitialized | UploadQueue::Stopped(_) => {
2500 0 : unreachable!("checked before constructing")
2501 : }
2502 : }
2503 0 : }
2504 : }
2505 :
2506 0 : pub fn remote_tenant_path(tenant_shard_id: &TenantShardId) -> RemotePath {
2507 0 : let path = format!("tenants/{tenant_shard_id}");
2508 0 : RemotePath::from_string(&path).expect("Failed to construct path")
2509 0 : }
2510 :
2511 1324 : pub fn remote_tenant_manifest_path(
2512 1324 : tenant_shard_id: &TenantShardId,
2513 1324 : generation: Generation,
2514 1324 : ) -> RemotePath {
2515 1324 : let path = format!(
2516 1324 : "tenants/{tenant_shard_id}/tenant-manifest{}.json",
2517 1324 : generation.get_suffix()
2518 1324 : );
2519 1324 : RemotePath::from_string(&path).expect("Failed to construct path")
2520 1324 : }
2521 :
2522 : /// Prefix to all generations' manifest objects in a tenant shard
2523 440 : pub fn remote_tenant_manifest_prefix(tenant_shard_id: &TenantShardId) -> RemotePath {
2524 440 : let path = format!("tenants/{tenant_shard_id}/tenant-manifest",);
2525 440 : RemotePath::from_string(&path).expect("Failed to construct path")
2526 440 : }
2527 :
2528 500 : pub fn remote_timelines_path(tenant_shard_id: &TenantShardId) -> RemotePath {
2529 500 : let path = format!("tenants/{tenant_shard_id}/{TIMELINES_SEGMENT_NAME}");
2530 500 : RemotePath::from_string(&path).expect("Failed to construct path")
2531 500 : }
2532 :
2533 0 : fn remote_timelines_path_unsharded(tenant_id: &TenantId) -> RemotePath {
2534 0 : let path = format!("tenants/{tenant_id}/{TIMELINES_SEGMENT_NAME}");
2535 0 : RemotePath::from_string(&path).expect("Failed to construct path")
2536 0 : }
2537 :
2538 60 : pub fn remote_timeline_path(
2539 60 : tenant_shard_id: &TenantShardId,
2540 60 : timeline_id: &TimelineId,
2541 60 : ) -> RemotePath {
2542 60 : remote_timelines_path(tenant_shard_id).join(Utf8Path::new(&timeline_id.to_string()))
2543 60 : }
2544 :
2545 : /// Obtains the path of the given Layer in the remote
2546 : ///
2547 : /// Note that the shard component of a remote layer path is _not_ always the same
2548 : /// as in the TenantShardId of the caller: tenants may reference layers from a different
2549 : /// ShardIndex. Use the ShardIndex from the layer's metadata.
2550 3940 : pub fn remote_layer_path(
2551 3940 : tenant_id: &TenantId,
2552 3940 : timeline_id: &TimelineId,
2553 3940 : shard: ShardIndex,
2554 3940 : layer_file_name: &LayerName,
2555 3940 : generation: Generation,
2556 3940 : ) -> RemotePath {
2557 3940 : // Generation-aware key format
2558 3940 : let path = format!(
2559 3940 : "tenants/{tenant_id}{0}/{TIMELINES_SEGMENT_NAME}/{timeline_id}/{1}{2}",
2560 3940 : shard.get_suffix(),
2561 3940 : layer_file_name,
2562 3940 : generation.get_suffix()
2563 3940 : );
2564 3940 :
2565 3940 : RemotePath::from_string(&path).expect("Failed to construct path")
2566 3940 : }
2567 :
2568 : /// Returns true if a and b have the same layer path within a tenant/timeline. This is essentially
2569 : /// remote_layer_path(a) == remote_layer_path(b) without the string allocations.
2570 : ///
2571 : /// TODO: there should be a variant of LayerName for the physical path that contains information
2572 : /// about the shard and generation, such that this could be replaced by a simple comparison.
2573 2351567 : pub fn is_same_remote_layer_path(
2574 2351567 : aname: &LayerName,
2575 2351567 : ameta: &LayerFileMetadata,
2576 2351567 : bname: &LayerName,
2577 2351567 : bmeta: &LayerFileMetadata,
2578 2351567 : ) -> bool {
2579 2351567 : // NB: don't assert remote_layer_path(a) == remote_layer_path(b); too expensive even for debug.
2580 2351567 : aname == bname && ameta.shard == bmeta.shard && ameta.generation == bmeta.generation
2581 2351567 : }
2582 :
2583 8 : pub fn remote_initdb_archive_path(tenant_id: &TenantId, timeline_id: &TimelineId) -> RemotePath {
2584 8 : RemotePath::from_string(&format!(
2585 8 : "tenants/{tenant_id}/{TIMELINES_SEGMENT_NAME}/{timeline_id}/{INITDB_PATH}"
2586 8 : ))
2587 8 : .expect("Failed to construct path")
2588 8 : }
2589 :
2590 4 : pub fn remote_initdb_preserved_archive_path(
2591 4 : tenant_id: &TenantId,
2592 4 : timeline_id: &TimelineId,
2593 4 : ) -> RemotePath {
2594 4 : RemotePath::from_string(&format!(
2595 4 : "tenants/{tenant_id}/{TIMELINES_SEGMENT_NAME}/{timeline_id}/{INITDB_PRESERVED_PATH}"
2596 4 : ))
2597 4 : .expect("Failed to construct path")
2598 4 : }
2599 :
2600 3113 : pub fn remote_index_path(
2601 3113 : tenant_shard_id: &TenantShardId,
2602 3113 : timeline_id: &TimelineId,
2603 3113 : generation: Generation,
2604 3113 : ) -> RemotePath {
2605 3113 : RemotePath::from_string(&format!(
2606 3113 : "tenants/{tenant_shard_id}/{TIMELINES_SEGMENT_NAME}/{timeline_id}/{0}{1}",
2607 3113 : IndexPart::FILE_NAME,
2608 3113 : generation.get_suffix()
2609 3113 : ))
2610 3113 : .expect("Failed to construct path")
2611 3113 : }
2612 :
2613 0 : pub(crate) fn remote_heatmap_path(tenant_shard_id: &TenantShardId) -> RemotePath {
2614 0 : RemotePath::from_string(&format!(
2615 0 : "tenants/{tenant_shard_id}/{TENANT_HEATMAP_BASENAME}"
2616 0 : ))
2617 0 : .expect("Failed to construct path")
2618 0 : }
2619 :
2620 : /// Given the key of an index, parse out the generation part of the name
2621 36 : pub fn parse_remote_index_path(path: RemotePath) -> Option<Generation> {
2622 36 : let file_name = match path.get_path().file_name() {
2623 36 : Some(f) => f,
2624 : None => {
2625 : // Unexpected: we should be seeing index_part.json paths only
2626 0 : tracing::warn!("Malformed index key {}", path);
2627 0 : return None;
2628 : }
2629 : };
2630 :
2631 36 : match file_name.split_once('-') {
2632 24 : Some((_, gen_suffix)) => Generation::parse_suffix(gen_suffix),
2633 12 : None => None,
2634 : }
2635 36 : }
2636 :
2637 : /// Given the key of a tenant manifest, parse out the generation number
2638 0 : pub fn parse_remote_tenant_manifest_path(path: RemotePath) -> Option<Generation> {
2639 : static RE: OnceLock<Regex> = OnceLock::new();
2640 0 : let re = RE.get_or_init(|| Regex::new(r".*tenant-manifest-([0-9a-f]{8}).json").unwrap());
2641 0 : re.captures(path.get_path().as_str())
2642 0 : .and_then(|c| c.get(1))
2643 0 : .and_then(|m| Generation::parse_suffix(m.as_str()))
2644 0 : }
2645 :
2646 : #[cfg(test)]
2647 : mod tests {
2648 : use super::*;
2649 : use crate::{
2650 : context::RequestContext,
2651 : tenant::{
2652 : config::AttachmentMode,
2653 : harness::{TenantHarness, TIMELINE_ID},
2654 : storage_layer::layer::local_layer_path,
2655 : Tenant, Timeline,
2656 : },
2657 : DEFAULT_PG_VERSION,
2658 : };
2659 :
2660 : use std::collections::HashSet;
2661 :
2662 16 : pub(super) fn dummy_contents(name: &str) -> Vec<u8> {
2663 16 : format!("contents for {name}").into()
2664 16 : }
2665 :
2666 4 : pub(super) fn dummy_metadata(disk_consistent_lsn: Lsn) -> TimelineMetadata {
2667 4 : let metadata = TimelineMetadata::new(
2668 4 : disk_consistent_lsn,
2669 4 : None,
2670 4 : None,
2671 4 : Lsn(0),
2672 4 : Lsn(0),
2673 4 : Lsn(0),
2674 4 : // Any version will do
2675 4 : // but it should be consistent with the one in the tests
2676 4 : crate::DEFAULT_PG_VERSION,
2677 4 : );
2678 4 :
2679 4 : // go through serialize + deserialize to fix the header, including checksum
2680 4 : TimelineMetadata::from_bytes(&metadata.to_bytes().unwrap()).unwrap()
2681 4 : }
2682 :
2683 4 : fn assert_file_list(a: &HashSet<LayerName>, b: &[&str]) {
2684 12 : let mut avec: Vec<String> = a.iter().map(|x| x.to_string()).collect();
2685 4 : avec.sort();
2686 4 :
2687 4 : let mut bvec = b.to_vec();
2688 4 : bvec.sort_unstable();
2689 4 :
2690 4 : assert_eq!(avec, bvec);
2691 4 : }
2692 :
2693 8 : fn assert_remote_files(expected: &[&str], remote_path: &Utf8Path, generation: Generation) {
2694 8 : let mut expected: Vec<String> = expected
2695 8 : .iter()
2696 32 : .map(|x| format!("{}{}", x, generation.get_suffix()))
2697 8 : .collect();
2698 8 : expected.sort();
2699 8 :
2700 8 : let mut found: Vec<String> = Vec::new();
2701 32 : for entry in std::fs::read_dir(remote_path).unwrap().flatten() {
2702 32 : let entry_name = entry.file_name();
2703 32 : let fname = entry_name.to_str().unwrap();
2704 32 : found.push(String::from(fname));
2705 32 : }
2706 8 : found.sort();
2707 8 :
2708 8 : assert_eq!(found, expected);
2709 8 : }
2710 :
2711 : struct TestSetup {
2712 : harness: TenantHarness,
2713 : tenant: Arc<Tenant>,
2714 : timeline: Arc<Timeline>,
2715 : tenant_ctx: RequestContext,
2716 : }
2717 :
2718 : impl TestSetup {
2719 16 : async fn new(test_name: &str) -> anyhow::Result<Self> {
2720 16 : let test_name = Box::leak(Box::new(format!("remote_timeline_client__{test_name}")));
2721 16 : let harness = TenantHarness::create(test_name).await?;
2722 16 : let (tenant, ctx) = harness.load().await;
2723 :
2724 16 : let timeline = tenant
2725 16 : .create_test_timeline(TIMELINE_ID, Lsn(8), DEFAULT_PG_VERSION, &ctx)
2726 16 : .await?;
2727 :
2728 16 : Ok(Self {
2729 16 : harness,
2730 16 : tenant,
2731 16 : timeline,
2732 16 : tenant_ctx: ctx,
2733 16 : })
2734 16 : }
2735 :
2736 : /// Construct a RemoteTimelineClient in an arbitrary generation
2737 20 : fn build_client(&self, generation: Generation) -> Arc<RemoteTimelineClient> {
2738 20 : let location_conf = AttachedLocationConfig {
2739 20 : generation,
2740 20 : attach_mode: AttachmentMode::Single,
2741 20 : };
2742 20 : Arc::new(RemoteTimelineClient {
2743 20 : conf: self.harness.conf,
2744 20 : runtime: tokio::runtime::Handle::current(),
2745 20 : tenant_shard_id: self.harness.tenant_shard_id,
2746 20 : timeline_id: TIMELINE_ID,
2747 20 : generation,
2748 20 : storage_impl: self.harness.remote_storage.clone(),
2749 20 : deletion_queue_client: self.harness.deletion_queue.new_client(),
2750 20 : upload_queue: Mutex::new(UploadQueue::Uninitialized),
2751 20 : metrics: Arc::new(RemoteTimelineClientMetrics::new(
2752 20 : &self.harness.tenant_shard_id,
2753 20 : &TIMELINE_ID,
2754 20 : )),
2755 20 : config: std::sync::RwLock::new(RemoteTimelineClientConfig::from(&location_conf)),
2756 20 : cancel: CancellationToken::new(),
2757 20 : })
2758 20 : }
2759 :
2760 : /// A tracing::Span that satisfies remote_timeline_client methods that assert tenant_id
2761 : /// and timeline_id are present.
2762 12 : fn span(&self) -> tracing::Span {
2763 12 : tracing::info_span!(
2764 : "test",
2765 : tenant_id = %self.harness.tenant_shard_id.tenant_id,
2766 0 : shard_id = %self.harness.tenant_shard_id.shard_slug(),
2767 : timeline_id = %TIMELINE_ID
2768 : )
2769 12 : }
2770 : }
2771 :
2772 : // Test scheduling
2773 : #[tokio::test]
2774 4 : async fn upload_scheduling() {
2775 4 : // Test outline:
2776 4 : //
2777 4 : // Schedule upload of a bunch of layers. Check that they are started immediately, not queued
2778 4 : // Schedule upload of index. Check that it is queued
2779 4 : // let the layer file uploads finish. Check that the index-upload is now started
2780 4 : // let the index-upload finish.
2781 4 : //
2782 4 : // Download back the index.json. Check that the list of files is correct
2783 4 : //
2784 4 : // Schedule upload. Schedule deletion. Check that the deletion is queued
2785 4 : // let upload finish. Check that deletion is now started
2786 4 : // Schedule another deletion. Check that it's launched immediately.
2787 4 : // Schedule index upload. Check that it's queued
2788 4 :
2789 4 : let test_setup = TestSetup::new("upload_scheduling").await.unwrap();
2790 4 : let span = test_setup.span();
2791 4 : let _guard = span.enter();
2792 4 :
2793 4 : let TestSetup {
2794 4 : harness,
2795 4 : tenant: _tenant,
2796 4 : timeline,
2797 4 : tenant_ctx: _tenant_ctx,
2798 4 : } = test_setup;
2799 4 :
2800 4 : let client = &timeline.remote_client;
2801 4 :
2802 4 : // Download back the index.json, and check that the list of files is correct
2803 4 : let initial_index_part = match client
2804 4 : .download_index_file(&CancellationToken::new())
2805 4 : .await
2806 4 : .unwrap()
2807 4 : {
2808 4 : MaybeDeletedIndexPart::IndexPart(index_part) => index_part,
2809 4 : MaybeDeletedIndexPart::Deleted(_) => panic!("unexpectedly got deleted index part"),
2810 4 : };
2811 4 : let initial_layers = initial_index_part
2812 4 : .layer_metadata
2813 4 : .keys()
2814 4 : .map(|f| f.to_owned())
2815 4 : .collect::<HashSet<LayerName>>();
2816 4 : let initial_layer = {
2817 4 : assert!(initial_layers.len() == 1);
2818 4 : initial_layers.into_iter().next().unwrap()
2819 4 : };
2820 4 :
2821 4 : let timeline_path = harness.timeline_path(&TIMELINE_ID);
2822 4 :
2823 4 : println!("workdir: {}", harness.conf.workdir);
2824 4 :
2825 4 : let remote_timeline_dir = harness
2826 4 : .remote_fs_dir
2827 4 : .join(timeline_path.strip_prefix(&harness.conf.workdir).unwrap());
2828 4 : println!("remote_timeline_dir: {remote_timeline_dir}");
2829 4 :
2830 4 : let generation = harness.generation;
2831 4 : let shard = harness.shard;
2832 4 :
2833 4 : // Create a couple of dummy files, schedule upload for them
2834 4 :
2835 4 : let layers = [
2836 4 : ("000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__00000000016B59D8-00000000016B5A51".parse().unwrap(), dummy_contents("foo")),
2837 4 : ("000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__00000000016B59D9-00000000016B5A52".parse().unwrap(), dummy_contents("bar")),
2838 4 : ("000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__00000000016B59DA-00000000016B5A53".parse().unwrap(), dummy_contents("baz"))
2839 4 : ]
2840 4 : .into_iter()
2841 12 : .map(|(name, contents): (LayerName, Vec<u8>)| {
2842 12 :
2843 12 : let local_path = local_layer_path(
2844 12 : harness.conf,
2845 12 : &timeline.tenant_shard_id,
2846 12 : &timeline.timeline_id,
2847 12 : &name,
2848 12 : &generation,
2849 12 : );
2850 12 : std::fs::write(&local_path, &contents).unwrap();
2851 12 :
2852 12 : Layer::for_resident(
2853 12 : harness.conf,
2854 12 : &timeline,
2855 12 : local_path,
2856 12 : name,
2857 12 : LayerFileMetadata::new(contents.len() as u64, generation, shard),
2858 12 : )
2859 12 : }).collect::<Vec<_>>();
2860 4 :
2861 4 : client
2862 4 : .schedule_layer_file_upload(layers[0].clone())
2863 4 : .unwrap();
2864 4 : client
2865 4 : .schedule_layer_file_upload(layers[1].clone())
2866 4 : .unwrap();
2867 4 :
2868 4 : // Check that they are started immediately, not queued
2869 4 : //
2870 4 : // this works because we running within block_on, so any futures are now queued up until
2871 4 : // our next await point.
2872 4 : {
2873 4 : let mut guard = client.upload_queue.lock().unwrap();
2874 4 : let upload_queue = guard.initialized_mut().unwrap();
2875 4 : assert!(upload_queue.queued_operations.is_empty());
2876 4 : assert_eq!(upload_queue.inprogress_tasks.len(), 2);
2877 4 : assert_eq!(upload_queue.num_inprogress_layer_uploads(), 2);
2878 4 :
2879 4 : // also check that `latest_file_changes` was updated
2880 4 : assert!(upload_queue.latest_files_changes_since_metadata_upload_scheduled == 2);
2881 4 : }
2882 4 :
2883 4 : // Schedule upload of index. Check that it is queued
2884 4 : let metadata = dummy_metadata(Lsn(0x20));
2885 4 : client
2886 4 : .schedule_index_upload_for_full_metadata_update(&metadata)
2887 4 : .unwrap();
2888 4 : {
2889 4 : let mut guard = client.upload_queue.lock().unwrap();
2890 4 : let upload_queue = guard.initialized_mut().unwrap();
2891 4 : assert!(upload_queue.queued_operations.len() == 1);
2892 4 : assert!(upload_queue.latest_files_changes_since_metadata_upload_scheduled == 0);
2893 4 : }
2894 4 :
2895 4 : // Wait for the uploads to finish
2896 4 : client.wait_completion().await.unwrap();
2897 4 : {
2898 4 : let mut guard = client.upload_queue.lock().unwrap();
2899 4 : let upload_queue = guard.initialized_mut().unwrap();
2900 4 :
2901 4 : assert!(upload_queue.queued_operations.is_empty());
2902 4 : assert!(upload_queue.inprogress_tasks.is_empty());
2903 4 : }
2904 4 :
2905 4 : // Download back the index.json, and check that the list of files is correct
2906 4 : let index_part = match client
2907 4 : .download_index_file(&CancellationToken::new())
2908 4 : .await
2909 4 : .unwrap()
2910 4 : {
2911 4 : MaybeDeletedIndexPart::IndexPart(index_part) => index_part,
2912 4 : MaybeDeletedIndexPart::Deleted(_) => panic!("unexpectedly got deleted index part"),
2913 4 : };
2914 4 :
2915 4 : assert_file_list(
2916 4 : &index_part
2917 4 : .layer_metadata
2918 4 : .keys()
2919 12 : .map(|f| f.to_owned())
2920 4 : .collect(),
2921 4 : &[
2922 4 : &initial_layer.to_string(),
2923 4 : &layers[0].layer_desc().layer_name().to_string(),
2924 4 : &layers[1].layer_desc().layer_name().to_string(),
2925 4 : ],
2926 4 : );
2927 4 : assert_eq!(index_part.metadata, metadata);
2928 4 :
2929 4 : // Schedule upload and then a deletion. Check that the deletion is queued
2930 4 : client
2931 4 : .schedule_layer_file_upload(layers[2].clone())
2932 4 : .unwrap();
2933 4 :
2934 4 : // this is no longer consistent with how deletion works with Layer::drop, but in this test
2935 4 : // keep using schedule_layer_file_deletion because we don't have a way to wait for the
2936 4 : // spawn_blocking started by the drop.
2937 4 : client
2938 4 : .schedule_layer_file_deletion(&[layers[0].layer_desc().layer_name()])
2939 4 : .unwrap();
2940 4 : {
2941 4 : let mut guard = client.upload_queue.lock().unwrap();
2942 4 : let upload_queue = guard.initialized_mut().unwrap();
2943 4 :
2944 4 : // Deletion schedules upload of the index file, and the file deletion itself
2945 4 : assert_eq!(upload_queue.queued_operations.len(), 2);
2946 4 : assert_eq!(upload_queue.inprogress_tasks.len(), 1);
2947 4 : assert_eq!(upload_queue.num_inprogress_layer_uploads(), 1);
2948 4 : assert_eq!(upload_queue.num_inprogress_deletions(), 0);
2949 4 : assert_eq!(
2950 4 : upload_queue.latest_files_changes_since_metadata_upload_scheduled,
2951 4 : 0
2952 4 : );
2953 4 : }
2954 4 : assert_remote_files(
2955 4 : &[
2956 4 : &initial_layer.to_string(),
2957 4 : &layers[0].layer_desc().layer_name().to_string(),
2958 4 : &layers[1].layer_desc().layer_name().to_string(),
2959 4 : "index_part.json",
2960 4 : ],
2961 4 : &remote_timeline_dir,
2962 4 : generation,
2963 4 : );
2964 4 :
2965 4 : // Finish them
2966 4 : client.wait_completion().await.unwrap();
2967 4 : harness.deletion_queue.pump().await;
2968 4 :
2969 4 : assert_remote_files(
2970 4 : &[
2971 4 : &initial_layer.to_string(),
2972 4 : &layers[1].layer_desc().layer_name().to_string(),
2973 4 : &layers[2].layer_desc().layer_name().to_string(),
2974 4 : "index_part.json",
2975 4 : ],
2976 4 : &remote_timeline_dir,
2977 4 : generation,
2978 4 : );
2979 4 : }
2980 :
2981 : #[tokio::test]
2982 4 : async fn bytes_unfinished_gauge_for_layer_file_uploads() {
2983 4 : // Setup
2984 4 :
2985 4 : let TestSetup {
2986 4 : harness,
2987 4 : tenant: _tenant,
2988 4 : timeline,
2989 4 : ..
2990 4 : } = TestSetup::new("metrics").await.unwrap();
2991 4 : let client = &timeline.remote_client;
2992 4 :
2993 4 : let layer_file_name_1: LayerName = "000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__00000000016B59D8-00000000016B5A51".parse().unwrap();
2994 4 : let local_path = local_layer_path(
2995 4 : harness.conf,
2996 4 : &timeline.tenant_shard_id,
2997 4 : &timeline.timeline_id,
2998 4 : &layer_file_name_1,
2999 4 : &harness.generation,
3000 4 : );
3001 4 : let content_1 = dummy_contents("foo");
3002 4 : std::fs::write(&local_path, &content_1).unwrap();
3003 4 :
3004 4 : let layer_file_1 = Layer::for_resident(
3005 4 : harness.conf,
3006 4 : &timeline,
3007 4 : local_path,
3008 4 : layer_file_name_1.clone(),
3009 4 : LayerFileMetadata::new(content_1.len() as u64, harness.generation, harness.shard),
3010 4 : );
3011 4 :
3012 4 : #[derive(Debug, PartialEq, Clone, Copy)]
3013 4 : struct BytesStartedFinished {
3014 4 : started: Option<usize>,
3015 4 : finished: Option<usize>,
3016 4 : }
3017 4 : impl std::ops::Add for BytesStartedFinished {
3018 4 : type Output = Self;
3019 8 : fn add(self, rhs: Self) -> Self::Output {
3020 8 : Self {
3021 8 : started: self.started.map(|v| v + rhs.started.unwrap_or(0)),
3022 8 : finished: self.finished.map(|v| v + rhs.finished.unwrap_or(0)),
3023 8 : }
3024 8 : }
3025 4 : }
3026 12 : let get_bytes_started_stopped = || {
3027 12 : let started = client
3028 12 : .metrics
3029 12 : .get_bytes_started_counter_value(&RemoteOpFileKind::Layer, &RemoteOpKind::Upload)
3030 12 : .map(|v| v.try_into().unwrap());
3031 12 : let stopped = client
3032 12 : .metrics
3033 12 : .get_bytes_finished_counter_value(&RemoteOpFileKind::Layer, &RemoteOpKind::Upload)
3034 12 : .map(|v| v.try_into().unwrap());
3035 12 : BytesStartedFinished {
3036 12 : started,
3037 12 : finished: stopped,
3038 12 : }
3039 12 : };
3040 4 :
3041 4 : // Test
3042 4 : tracing::info!("now doing actual test");
3043 4 :
3044 4 : let actual_a = get_bytes_started_stopped();
3045 4 :
3046 4 : client
3047 4 : .schedule_layer_file_upload(layer_file_1.clone())
3048 4 : .unwrap();
3049 4 :
3050 4 : let actual_b = get_bytes_started_stopped();
3051 4 :
3052 4 : client.wait_completion().await.unwrap();
3053 4 :
3054 4 : let actual_c = get_bytes_started_stopped();
3055 4 :
3056 4 : // Validate
3057 4 :
3058 4 : let expected_b = actual_a
3059 4 : + BytesStartedFinished {
3060 4 : started: Some(content_1.len()),
3061 4 : // assert that the _finished metric is created eagerly so that subtractions work on first sample
3062 4 : finished: Some(0),
3063 4 : };
3064 4 : assert_eq!(actual_b, expected_b);
3065 4 :
3066 4 : let expected_c = actual_a
3067 4 : + BytesStartedFinished {
3068 4 : started: Some(content_1.len()),
3069 4 : finished: Some(content_1.len()),
3070 4 : };
3071 4 : assert_eq!(actual_c, expected_c);
3072 4 : }
3073 :
3074 24 : async fn inject_index_part(test_state: &TestSetup, generation: Generation) -> IndexPart {
3075 24 : // An empty IndexPart, just sufficient to ensure deserialization will succeed
3076 24 : let example_index_part = IndexPart::example();
3077 24 :
3078 24 : let index_part_bytes = serde_json::to_vec(&example_index_part).unwrap();
3079 24 :
3080 24 : let index_path = test_state.harness.remote_fs_dir.join(
3081 24 : remote_index_path(
3082 24 : &test_state.harness.tenant_shard_id,
3083 24 : &TIMELINE_ID,
3084 24 : generation,
3085 24 : )
3086 24 : .get_path(),
3087 24 : );
3088 24 :
3089 24 : std::fs::create_dir_all(index_path.parent().unwrap())
3090 24 : .expect("creating test dir should work");
3091 24 :
3092 24 : eprintln!("Writing {index_path}");
3093 24 : std::fs::write(&index_path, index_part_bytes).unwrap();
3094 24 : example_index_part
3095 24 : }
3096 :
3097 : /// Assert that when a RemoteTimelineclient in generation `get_generation` fetches its
3098 : /// index, the IndexPart returned is equal to `expected`
3099 20 : async fn assert_got_index_part(
3100 20 : test_state: &TestSetup,
3101 20 : get_generation: Generation,
3102 20 : expected: &IndexPart,
3103 20 : ) {
3104 20 : let client = test_state.build_client(get_generation);
3105 :
3106 20 : let download_r = client
3107 20 : .download_index_file(&CancellationToken::new())
3108 20 : .await
3109 20 : .expect("download should always succeed");
3110 20 : assert!(matches!(download_r, MaybeDeletedIndexPart::IndexPart(_)));
3111 20 : match download_r {
3112 20 : MaybeDeletedIndexPart::IndexPart(index_part) => {
3113 20 : assert_eq!(&index_part, expected);
3114 : }
3115 0 : MaybeDeletedIndexPart::Deleted(_index_part) => panic!("Test doesn't set deleted_at"),
3116 : }
3117 20 : }
3118 :
3119 : #[tokio::test]
3120 4 : async fn index_part_download_simple() -> anyhow::Result<()> {
3121 4 : let test_state = TestSetup::new("index_part_download_simple").await.unwrap();
3122 4 : let span = test_state.span();
3123 4 : let _guard = span.enter();
3124 4 :
3125 4 : // Simple case: we are in generation N, load the index from generation N - 1
3126 4 : let generation_n = 5;
3127 4 : let injected = inject_index_part(&test_state, Generation::new(generation_n - 1)).await;
3128 4 :
3129 4 : assert_got_index_part(&test_state, Generation::new(generation_n), &injected).await;
3130 4 :
3131 4 : Ok(())
3132 4 : }
3133 :
3134 : #[tokio::test]
3135 4 : async fn index_part_download_ordering() -> anyhow::Result<()> {
3136 4 : let test_state = TestSetup::new("index_part_download_ordering")
3137 4 : .await
3138 4 : .unwrap();
3139 4 :
3140 4 : let span = test_state.span();
3141 4 : let _guard = span.enter();
3142 4 :
3143 4 : // A generation-less IndexPart exists in the bucket, we should find it
3144 4 : let generation_n = 5;
3145 4 : let injected_none = inject_index_part(&test_state, Generation::none()).await;
3146 4 : assert_got_index_part(&test_state, Generation::new(generation_n), &injected_none).await;
3147 4 :
3148 4 : // If a more recent-than-none generation exists, we should prefer to load that
3149 4 : let injected_1 = inject_index_part(&test_state, Generation::new(1)).await;
3150 4 : assert_got_index_part(&test_state, Generation::new(generation_n), &injected_1).await;
3151 4 :
3152 4 : // If a more-recent-than-me generation exists, we should ignore it.
3153 4 : let _injected_10 = inject_index_part(&test_state, Generation::new(10)).await;
3154 4 : assert_got_index_part(&test_state, Generation::new(generation_n), &injected_1).await;
3155 4 :
3156 4 : // If a directly previous generation exists, _and_ an index exists in my own
3157 4 : // generation, I should prefer my own generation.
3158 4 : let _injected_prev =
3159 4 : inject_index_part(&test_state, Generation::new(generation_n - 1)).await;
3160 4 : let injected_current = inject_index_part(&test_state, Generation::new(generation_n)).await;
3161 4 : assert_got_index_part(
3162 4 : &test_state,
3163 4 : Generation::new(generation_n),
3164 4 : &injected_current,
3165 4 : )
3166 4 : .await;
3167 4 :
3168 4 : Ok(())
3169 4 : }
3170 : }
|
