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 std::collections::{HashMap, HashSet, VecDeque};
183 : use std::ops::DerefMut;
184 : use std::sync::atomic::{AtomicU32, Ordering};
185 : use std::sync::{Arc, Mutex, OnceLock};
186 : use std::time::Duration;
187 :
188 : use anyhow::Context;
189 : use camino::Utf8Path;
190 : use chrono::{NaiveDateTime, Utc};
191 : pub(crate) use download::{
192 : download_index_part, download_initdb_tar_zst, download_tenant_manifest, is_temp_download_file,
193 : list_remote_tenant_shards, list_remote_timelines,
194 : };
195 : use index::GcCompactionState;
196 : pub(crate) use index::LayerFileMetadata;
197 : use pageserver_api::models::{RelSizeMigration, TimelineArchivalState};
198 : use pageserver_api::shard::{ShardIndex, TenantShardId};
199 : use regex::Regex;
200 : use remote_storage::{
201 : DownloadError, GenericRemoteStorage, ListingMode, RemotePath, TimeoutOrCancel,
202 : };
203 : use scopeguard::ScopeGuard;
204 : use tokio_util::sync::CancellationToken;
205 : use tracing::{Instrument, debug, error, info, info_span, instrument, warn};
206 : pub(crate) use upload::upload_initdb_dir;
207 : use utils::backoff::{
208 : self, DEFAULT_BASE_BACKOFF_SECONDS, DEFAULT_MAX_BACKOFF_SECONDS, exponential_backoff,
209 : };
210 : use utils::id::{TenantId, TimelineId};
211 : use utils::lsn::Lsn;
212 : use utils::pausable_failpoint;
213 : use utils::shard::ShardNumber;
214 :
215 : use self::index::IndexPart;
216 : use super::config::AttachedLocationConfig;
217 : use super::metadata::MetadataUpdate;
218 : use super::storage_layer::{Layer, LayerName, ResidentLayer};
219 : use super::timeline::import_pgdata;
220 : use super::upload_queue::{NotInitialized, SetDeletedFlagProgress};
221 : use super::{DeleteTimelineError, Generation};
222 : use crate::config::PageServerConf;
223 : use crate::context::RequestContext;
224 : use crate::deletion_queue::{DeletionQueueClient, DeletionQueueError};
225 : use crate::metrics::{
226 : MeasureRemoteOp, REMOTE_ONDEMAND_DOWNLOADED_BYTES, REMOTE_ONDEMAND_DOWNLOADED_LAYERS,
227 : RemoteOpFileKind, RemoteOpKind, RemoteTimelineClientMetrics,
228 : RemoteTimelineClientMetricsCallTrackSize,
229 : };
230 : use crate::task_mgr::{BACKGROUND_RUNTIME, TaskKind, shutdown_token};
231 : use crate::tenant::metadata::TimelineMetadata;
232 : use crate::tenant::remote_timeline_client::download::download_retry;
233 : use crate::tenant::storage_layer::AsLayerDesc;
234 : use crate::tenant::upload_queue::{
235 : Delete, OpType, UploadOp, UploadQueue, UploadQueueInitialized, UploadQueueStopped,
236 : UploadQueueStoppedDeletable, UploadTask,
237 : };
238 : use crate::tenant::{TIMELINES_SEGMENT_NAME, debug_assert_current_span_has_tenant_and_timeline_id};
239 : use crate::{TENANT_HEATMAP_BASENAME, task_mgr};
240 :
241 : // Occasional network issues and such can cause remote operations to fail, and
242 : // that's expected. If a download fails, we log it at info-level, and retry.
243 : // But after FAILED_DOWNLOAD_WARN_THRESHOLD retries, we start to log it at WARN
244 : // level instead, as repeated failures can mean a more serious problem. If it
245 : // fails more than FAILED_DOWNLOAD_RETRIES times, we give up
246 : pub(crate) const FAILED_DOWNLOAD_WARN_THRESHOLD: u32 = 3;
247 : pub(crate) const FAILED_REMOTE_OP_RETRIES: u32 = 10;
248 :
249 : // Similarly log failed uploads and deletions at WARN level, after this many
250 : // retries. Uploads and deletions are retried forever, though.
251 : pub(crate) const FAILED_UPLOAD_WARN_THRESHOLD: u32 = 3;
252 :
253 : pub(crate) const INITDB_PATH: &str = "initdb.tar.zst";
254 :
255 : pub(crate) const INITDB_PRESERVED_PATH: &str = "initdb-preserved.tar.zst";
256 :
257 : /// Default buffer size when interfacing with [`tokio::fs::File`].
258 : pub(crate) const BUFFER_SIZE: usize = 32 * 1024;
259 :
260 : /// Doing non-essential flushes of deletion queue is subject to this timeout, after
261 : /// which we warn and skip.
262 : const DELETION_QUEUE_FLUSH_TIMEOUT: Duration = Duration::from_secs(10);
263 :
264 : pub enum MaybeDeletedIndexPart {
265 : IndexPart(IndexPart),
266 : Deleted(IndexPart),
267 : }
268 :
269 : #[derive(Debug, thiserror::Error)]
270 : pub enum PersistIndexPartWithDeletedFlagError {
271 : #[error("another task is already setting the deleted_flag, started at {0:?}")]
272 : AlreadyInProgress(NaiveDateTime),
273 : #[error("the deleted_flag was already set, value is {0:?}")]
274 : AlreadyDeleted(NaiveDateTime),
275 : #[error(transparent)]
276 : Other(#[from] anyhow::Error),
277 : }
278 :
279 : #[derive(Debug, thiserror::Error)]
280 : pub enum WaitCompletionError {
281 : #[error(transparent)]
282 : NotInitialized(NotInitialized),
283 : #[error("wait_completion aborted because upload queue was stopped")]
284 : UploadQueueShutDownOrStopped,
285 : }
286 :
287 : #[derive(Debug, thiserror::Error)]
288 : #[error("Upload queue either in unexpected state or hasn't downloaded manifest yet")]
289 : pub struct UploadQueueNotReadyError;
290 :
291 : #[derive(Debug, thiserror::Error)]
292 : pub enum ShutdownIfArchivedError {
293 : #[error(transparent)]
294 : NotInitialized(NotInitialized),
295 : #[error("timeline is not archived")]
296 : NotArchived,
297 : }
298 :
299 : /// Behavioral modes that enable seamless live migration.
300 : ///
301 : /// See docs/rfcs/028-pageserver-migration.md to understand how these fit in.
302 : struct RemoteTimelineClientConfig {
303 : /// If this is false, then update to remote_consistent_lsn are dropped rather
304 : /// than being submitted to DeletionQueue for validation. This behavior is
305 : /// used when a tenant attachment is known to have a stale generation number,
306 : /// such that validation attempts will always fail. This is not necessary
307 : /// for correctness, but avoids spamming error statistics with failed validations
308 : /// when doing migrations of tenants.
309 : process_remote_consistent_lsn_updates: bool,
310 :
311 : /// If this is true, then object deletions are held in a buffer in RemoteTimelineClient
312 : /// rather than being submitted to the DeletionQueue. This behavior is used when a tenant
313 : /// is known to be multi-attached, in order to avoid disrupting other attached tenants
314 : /// whose generations' metadata refers to the deleted objects.
315 : block_deletions: bool,
316 : }
317 :
318 : /// RemoteTimelineClientConfig's state is entirely driven by LocationConf, but we do
319 : /// not carry the entire LocationConf structure: it's much more than we need. The From
320 : /// impl extracts the subset of the LocationConf that is interesting to RemoteTimelineClient.
321 : impl From<&AttachedLocationConfig> for RemoteTimelineClientConfig {
322 924 : fn from(lc: &AttachedLocationConfig) -> Self {
323 924 : Self {
324 924 : block_deletions: !lc.may_delete_layers_hint(),
325 924 : process_remote_consistent_lsn_updates: lc.may_upload_layers_hint(),
326 924 : }
327 924 : }
328 : }
329 :
330 : /// A client for accessing a timeline's data in remote storage.
331 : ///
332 : /// This takes care of managing the number of connections, and balancing them
333 : /// across tenants. This also handles retries of failed uploads.
334 : ///
335 : /// Upload and delete requests are ordered so that before a deletion is
336 : /// performed, we wait for all preceding uploads to finish. This ensures sure
337 : /// that if you perform a compaction operation that reshuffles data in layer
338 : /// files, we don't have a transient state where the old files have already been
339 : /// deleted, but new files have not yet been uploaded.
340 : ///
341 : /// Similarly, this enforces an order between index-file uploads, and layer
342 : /// uploads. Before an index-file upload is performed, all preceding layer
343 : /// uploads must be finished.
344 : ///
345 : /// This also maintains a list of remote files, and automatically includes that
346 : /// in the index part file, whenever timeline metadata is uploaded.
347 : ///
348 : /// Downloads are not queued, they are performed immediately.
349 : pub(crate) struct RemoteTimelineClient {
350 : conf: &'static PageServerConf,
351 :
352 : runtime: tokio::runtime::Handle,
353 :
354 : tenant_shard_id: TenantShardId,
355 : timeline_id: TimelineId,
356 : generation: Generation,
357 :
358 : upload_queue: Mutex<UploadQueue>,
359 :
360 : pub(crate) metrics: Arc<RemoteTimelineClientMetrics>,
361 :
362 : storage_impl: GenericRemoteStorage,
363 :
364 : deletion_queue_client: DeletionQueueClient,
365 :
366 : /// Subset of tenant configuration used to control upload behaviors during migrations
367 : config: std::sync::RwLock<RemoteTimelineClientConfig>,
368 :
369 : cancel: CancellationToken,
370 : }
371 :
372 : impl Drop for RemoteTimelineClient {
373 40 : fn drop(&mut self) {
374 40 : debug!("dropping RemoteTimelineClient");
375 40 : }
376 : }
377 :
378 : impl RemoteTimelineClient {
379 : ///
380 : /// Create a remote storage client for given timeline
381 : ///
382 : /// Note: the caller must initialize the upload queue before any uploads can be scheduled,
383 : /// by calling init_upload_queue.
384 : ///
385 904 : pub(crate) fn new(
386 904 : remote_storage: GenericRemoteStorage,
387 904 : deletion_queue_client: DeletionQueueClient,
388 904 : conf: &'static PageServerConf,
389 904 : tenant_shard_id: TenantShardId,
390 904 : timeline_id: TimelineId,
391 904 : generation: Generation,
392 904 : location_conf: &AttachedLocationConfig,
393 904 : ) -> RemoteTimelineClient {
394 904 : RemoteTimelineClient {
395 904 : conf,
396 904 : runtime: if cfg!(test) {
397 : // remote_timeline_client.rs tests rely on current-thread runtime
398 904 : tokio::runtime::Handle::current()
399 : } else {
400 0 : BACKGROUND_RUNTIME.handle().clone()
401 : },
402 904 : tenant_shard_id,
403 904 : timeline_id,
404 904 : generation,
405 904 : storage_impl: remote_storage,
406 904 : deletion_queue_client,
407 904 : upload_queue: Mutex::new(UploadQueue::Uninitialized),
408 904 : metrics: Arc::new(RemoteTimelineClientMetrics::new(
409 904 : &tenant_shard_id,
410 904 : &timeline_id,
411 904 : )),
412 904 : config: std::sync::RwLock::new(RemoteTimelineClientConfig::from(location_conf)),
413 904 : cancel: CancellationToken::new(),
414 904 : }
415 904 : }
416 :
417 : /// Initialize the upload queue for a remote storage that already received
418 : /// an index file upload, i.e., it's not empty.
419 : /// The given `index_part` must be the one on the remote.
420 12 : pub fn init_upload_queue(&self, index_part: &IndexPart) -> anyhow::Result<()> {
421 12 : // Set the maximum number of inprogress tasks to the remote storage concurrency. There's
422 12 : // certainly no point in starting more upload tasks than this.
423 12 : let inprogress_limit = self
424 12 : .conf
425 12 : .remote_storage_config
426 12 : .as_ref()
427 12 : .map_or(0, |r| r.concurrency_limit());
428 12 : let mut upload_queue = self.upload_queue.lock().unwrap();
429 12 : upload_queue.initialize_with_current_remote_index_part(index_part, inprogress_limit)?;
430 12 : self.update_remote_physical_size_gauge(Some(index_part));
431 12 : info!(
432 0 : "initialized upload queue from remote index with {} layer files",
433 0 : index_part.layer_metadata.len()
434 : );
435 12 : Ok(())
436 12 : }
437 :
438 : /// Initialize the upload queue for the case where the remote storage is empty,
439 : /// i.e., it doesn't have an `IndexPart`.
440 : ///
441 : /// `rel_size_v2_status` needs to be carried over during branching, and that's why
442 : /// it's passed in here.
443 892 : pub fn init_upload_queue_for_empty_remote(
444 892 : &self,
445 892 : local_metadata: &TimelineMetadata,
446 892 : rel_size_v2_status: Option<RelSizeMigration>,
447 892 : ) -> anyhow::Result<()> {
448 892 : // Set the maximum number of inprogress tasks to the remote storage concurrency. There's
449 892 : // certainly no point in starting more upload tasks than this.
450 892 : let inprogress_limit = self
451 892 : .conf
452 892 : .remote_storage_config
453 892 : .as_ref()
454 892 : .map_or(0, |r| r.concurrency_limit());
455 892 : let mut upload_queue = self.upload_queue.lock().unwrap();
456 892 : let initialized_queue =
457 892 : upload_queue.initialize_empty_remote(local_metadata, inprogress_limit)?;
458 892 : initialized_queue.dirty.rel_size_migration = rel_size_v2_status;
459 892 : self.update_remote_physical_size_gauge(None);
460 892 : info!("initialized upload queue as empty");
461 892 : Ok(())
462 892 : }
463 :
464 : /// Initialize the queue in stopped state. Used in startup path
465 : /// to continue deletion operation interrupted by pageserver crash or restart.
466 0 : pub fn init_upload_queue_stopped_to_continue_deletion(
467 0 : &self,
468 0 : index_part: &IndexPart,
469 0 : ) -> anyhow::Result<()> {
470 : // FIXME: consider newtype for DeletedIndexPart.
471 0 : let deleted_at = index_part.deleted_at.ok_or(anyhow::anyhow!(
472 0 : "bug: it is responsibility of the caller to provide index part from MaybeDeletedIndexPart::Deleted"
473 0 : ))?;
474 0 : let inprogress_limit = self
475 0 : .conf
476 0 : .remote_storage_config
477 0 : .as_ref()
478 0 : .map_or(0, |r| r.concurrency_limit());
479 0 :
480 0 : let mut upload_queue = self.upload_queue.lock().unwrap();
481 0 : upload_queue.initialize_with_current_remote_index_part(index_part, inprogress_limit)?;
482 0 : self.update_remote_physical_size_gauge(Some(index_part));
483 0 : self.stop_impl(&mut upload_queue);
484 0 :
485 0 : upload_queue
486 0 : .stopped_mut()
487 0 : .expect("stopped above")
488 0 : .deleted_at = SetDeletedFlagProgress::Successful(deleted_at);
489 0 :
490 0 : Ok(())
491 0 : }
492 :
493 : /// Notify this client of a change to its parent tenant's config, as this may cause us to
494 : /// take action (unblocking deletions when transitioning from AttachedMulti to AttachedSingle)
495 0 : pub(super) fn update_config(&self, location_conf: &AttachedLocationConfig) {
496 0 : let new_conf = RemoteTimelineClientConfig::from(location_conf);
497 0 : let unblocked = !new_conf.block_deletions;
498 0 :
499 0 : // Update config before draining deletions, so that we don't race with more being
500 0 : // inserted. This can result in deletions happening our of order, but that does not
501 0 : // violate any invariants: deletions only need to be ordered relative to upload of the index
502 0 : // that dereferences the deleted objects, and we are not changing that order.
503 0 : *self.config.write().unwrap() = new_conf;
504 0 :
505 0 : if unblocked {
506 : // If we may now delete layers, drain any that were blocked in our old
507 : // configuration state
508 0 : let mut queue_locked = self.upload_queue.lock().unwrap();
509 :
510 0 : if let Ok(queue) = queue_locked.initialized_mut() {
511 0 : let blocked_deletions = std::mem::take(&mut queue.blocked_deletions);
512 0 : for d in blocked_deletions {
513 0 : if let Err(e) = self.deletion_queue_client.push_layers(
514 0 : self.tenant_shard_id,
515 0 : self.timeline_id,
516 0 : self.generation,
517 0 : d.layers,
518 0 : ) {
519 : // This could happen if the pageserver is shut down while a tenant
520 : // is transitioning from a deletion-blocked state: we will leak some
521 : // S3 objects in this case.
522 0 : warn!("Failed to drain blocked deletions: {}", e);
523 0 : break;
524 0 : }
525 : }
526 0 : }
527 0 : }
528 0 : }
529 :
530 : /// Returns `None` if nothing is yet uplodaded, `Some(disk_consistent_lsn)` otherwise.
531 0 : pub fn remote_consistent_lsn_projected(&self) -> Option<Lsn> {
532 0 : match &mut *self.upload_queue.lock().unwrap() {
533 0 : UploadQueue::Uninitialized => None,
534 0 : UploadQueue::Initialized(q) => q.get_last_remote_consistent_lsn_projected(),
535 0 : UploadQueue::Stopped(UploadQueueStopped::Uninitialized) => None,
536 0 : UploadQueue::Stopped(UploadQueueStopped::Deletable(q)) => q
537 0 : .upload_queue_for_deletion
538 0 : .get_last_remote_consistent_lsn_projected(),
539 : }
540 0 : }
541 :
542 0 : pub fn remote_consistent_lsn_visible(&self) -> Option<Lsn> {
543 0 : match &mut *self.upload_queue.lock().unwrap() {
544 0 : UploadQueue::Uninitialized => None,
545 0 : UploadQueue::Initialized(q) => Some(q.get_last_remote_consistent_lsn_visible()),
546 0 : UploadQueue::Stopped(UploadQueueStopped::Uninitialized) => None,
547 0 : UploadQueue::Stopped(UploadQueueStopped::Deletable(q)) => Some(
548 0 : q.upload_queue_for_deletion
549 0 : .get_last_remote_consistent_lsn_visible(),
550 0 : ),
551 : }
552 0 : }
553 :
554 : /// Returns true if this timeline was previously detached at this Lsn and the remote timeline
555 : /// client is currently initialized.
556 0 : pub(crate) fn is_previous_ancestor_lsn(&self, lsn: Lsn) -> bool {
557 0 : self.upload_queue
558 0 : .lock()
559 0 : .unwrap()
560 0 : .initialized_mut()
561 0 : .map(|uq| uq.clean.0.lineage.is_previous_ancestor_lsn(lsn))
562 0 : .unwrap_or(false)
563 0 : }
564 :
565 : /// Returns whether the timeline is archived.
566 : /// Return None if the remote index_part hasn't been downloaded yet.
567 0 : pub(crate) fn is_archived(&self) -> Option<bool> {
568 0 : self.upload_queue
569 0 : .lock()
570 0 : .unwrap()
571 0 : .initialized_mut()
572 0 : .map(|q| q.clean.0.archived_at.is_some())
573 0 : .ok()
574 0 : }
575 :
576 : /// Returns `Ok(Some(timestamp))` if the timeline has been archived, `Ok(None)` if the timeline hasn't been archived.
577 : ///
578 : /// Return Err(_) if the remote index_part hasn't been downloaded yet, or the timeline hasn't been stopped yet.
579 4 : pub(crate) fn archived_at_stopped_queue(
580 4 : &self,
581 4 : ) -> Result<Option<NaiveDateTime>, UploadQueueNotReadyError> {
582 4 : self.upload_queue
583 4 : .lock()
584 4 : .unwrap()
585 4 : .stopped_mut()
586 4 : .map(|q| q.upload_queue_for_deletion.clean.0.archived_at)
587 4 : .map_err(|_| UploadQueueNotReadyError)
588 4 : }
589 :
590 3889 : fn update_remote_physical_size_gauge(&self, current_remote_index_part: Option<&IndexPart>) {
591 3889 : let size: u64 = if let Some(current_remote_index_part) = current_remote_index_part {
592 2997 : current_remote_index_part
593 2997 : .layer_metadata
594 2997 : .values()
595 35875 : .map(|ilmd| ilmd.file_size)
596 2997 : .sum()
597 : } else {
598 892 : 0
599 : };
600 3889 : self.metrics.remote_physical_size_gauge.set(size);
601 3889 : }
602 :
603 4 : pub fn get_remote_physical_size(&self) -> u64 {
604 4 : self.metrics.remote_physical_size_gauge.get()
605 4 : }
606 :
607 : //
608 : // Download operations.
609 : //
610 : // These don't use the per-timeline queue. They do use the global semaphore in
611 : // S3Bucket, to limit the total number of concurrent operations, though.
612 : //
613 :
614 : /// Download index file
615 40 : pub async fn download_index_file(
616 40 : &self,
617 40 : cancel: &CancellationToken,
618 40 : ) -> Result<MaybeDeletedIndexPart, DownloadError> {
619 40 : let _unfinished_gauge_guard = self.metrics.call_begin(
620 40 : &RemoteOpFileKind::Index,
621 40 : &RemoteOpKind::Download,
622 40 : crate::metrics::RemoteTimelineClientMetricsCallTrackSize::DontTrackSize {
623 40 : reason: "no need for a downloads gauge",
624 40 : },
625 40 : );
626 :
627 40 : let (index_part, index_generation, index_last_modified) = download::download_index_part(
628 40 : &self.storage_impl,
629 40 : &self.tenant_shard_id,
630 40 : &self.timeline_id,
631 40 : self.generation,
632 40 : cancel,
633 40 : )
634 40 : .measure_remote_op(
635 40 : RemoteOpFileKind::Index,
636 40 : RemoteOpKind::Download,
637 40 : Arc::clone(&self.metrics),
638 40 : )
639 40 : .await?;
640 :
641 : // Defense in depth: monotonicity of generation numbers is an important correctness guarantee, so when we see a very
642 : // old index, we do extra checks in case this is the result of backward time-travel of the generation number (e.g.
643 : // in case of a bug in the service that issues generation numbers). Indices are allowed to be old, but we expect that
644 : // when we load an old index we are loading the _latest_ index: if we are asked to load an old index and there is
645 : // also a newer index available, that is surprising.
646 : const INDEX_AGE_CHECKS_THRESHOLD: Duration = Duration::from_secs(14 * 24 * 3600);
647 40 : let index_age = index_last_modified.elapsed().unwrap_or_else(|e| {
648 0 : if e.duration() > Duration::from_secs(5) {
649 : // We only warn if the S3 clock and our local clock are >5s out: because this is a low resolution
650 : // timestamp, it is common to be out by at least 1 second.
651 0 : tracing::warn!("Index has modification time in the future: {e}");
652 0 : }
653 0 : Duration::ZERO
654 40 : });
655 40 : if index_age > INDEX_AGE_CHECKS_THRESHOLD {
656 0 : tracing::info!(
657 : ?index_generation,
658 0 : age = index_age.as_secs_f64(),
659 0 : "Loaded an old index, checking for other indices..."
660 : );
661 :
662 : // Find the highest-generation index
663 0 : let (_latest_index_part, latest_index_generation, latest_index_mtime) =
664 0 : download::download_index_part(
665 0 : &self.storage_impl,
666 0 : &self.tenant_shard_id,
667 0 : &self.timeline_id,
668 0 : Generation::MAX,
669 0 : cancel,
670 0 : )
671 0 : .await?;
672 :
673 0 : if latest_index_generation > index_generation {
674 : // Unexpected! Why are we loading such an old index if a more recent one exists?
675 : // We will refuse to proceed, as there is no reasonable scenario where this should happen, but
676 : // there _is_ a clear bug/corruption scenario where it would happen (controller sets the generation
677 : // backwards).
678 0 : tracing::error!(
679 : ?index_generation,
680 : ?latest_index_generation,
681 : ?latest_index_mtime,
682 0 : "Found a newer index while loading an old one"
683 : );
684 0 : return Err(DownloadError::Fatal(
685 0 : "Index age exceeds threshold and a newer index exists".into(),
686 0 : ));
687 0 : }
688 40 : }
689 :
690 40 : if index_part.deleted_at.is_some() {
691 0 : Ok(MaybeDeletedIndexPart::Deleted(index_part))
692 : } else {
693 40 : Ok(MaybeDeletedIndexPart::IndexPart(index_part))
694 : }
695 40 : }
696 :
697 : /// Download a (layer) file from `path`, into local filesystem.
698 : ///
699 : /// 'layer_metadata' is the metadata from the remote index file.
700 : ///
701 : /// On success, returns the size of the downloaded file.
702 28 : pub async fn download_layer_file(
703 28 : &self,
704 28 : layer_file_name: &LayerName,
705 28 : layer_metadata: &LayerFileMetadata,
706 28 : local_path: &Utf8Path,
707 28 : gate: &utils::sync::gate::Gate,
708 28 : cancel: &CancellationToken,
709 28 : ctx: &RequestContext,
710 28 : ) -> Result<u64, DownloadError> {
711 28 : let downloaded_size = {
712 28 : let _unfinished_gauge_guard = self.metrics.call_begin(
713 28 : &RemoteOpFileKind::Layer,
714 28 : &RemoteOpKind::Download,
715 28 : crate::metrics::RemoteTimelineClientMetricsCallTrackSize::DontTrackSize {
716 28 : reason: "no need for a downloads gauge",
717 28 : },
718 28 : );
719 28 : download::download_layer_file(
720 28 : self.conf,
721 28 : &self.storage_impl,
722 28 : self.tenant_shard_id,
723 28 : self.timeline_id,
724 28 : layer_file_name,
725 28 : layer_metadata,
726 28 : local_path,
727 28 : gate,
728 28 : cancel,
729 28 : ctx,
730 28 : )
731 28 : .measure_remote_op(
732 28 : RemoteOpFileKind::Layer,
733 28 : RemoteOpKind::Download,
734 28 : Arc::clone(&self.metrics),
735 28 : )
736 28 : .await?
737 : };
738 :
739 28 : REMOTE_ONDEMAND_DOWNLOADED_LAYERS.inc();
740 28 : REMOTE_ONDEMAND_DOWNLOADED_BYTES.inc_by(downloaded_size);
741 28 :
742 28 : Ok(downloaded_size)
743 28 : }
744 :
745 : //
746 : // Upload operations.
747 : //
748 :
749 : /// Launch an index-file upload operation in the background, with
750 : /// fully updated metadata.
751 : ///
752 : /// This should only be used to upload initial metadata to remote storage.
753 : ///
754 : /// The upload will be added to the queue immediately, but it
755 : /// won't be performed until all previously scheduled layer file
756 : /// upload operations have completed successfully. This is to
757 : /// ensure that when the index file claims that layers X, Y and Z
758 : /// exist in remote storage, they really do. To wait for the upload
759 : /// to complete, use `wait_completion`.
760 : ///
761 : /// If there were any changes to the list of files, i.e. if any
762 : /// layer file uploads were scheduled, since the last index file
763 : /// upload, those will be included too.
764 460 : pub fn schedule_index_upload_for_full_metadata_update(
765 460 : self: &Arc<Self>,
766 460 : metadata: &TimelineMetadata,
767 460 : ) -> anyhow::Result<()> {
768 460 : let mut guard = self.upload_queue.lock().unwrap();
769 460 : let upload_queue = guard.initialized_mut()?;
770 :
771 : // As documented in the struct definition, it's ok for latest_metadata to be
772 : // ahead of what's _actually_ on the remote during index upload.
773 460 : upload_queue.dirty.metadata = metadata.clone();
774 460 :
775 460 : self.schedule_index_upload(upload_queue);
776 460 :
777 460 : Ok(())
778 460 : }
779 :
780 : /// Launch an index-file upload operation in the background, with only parts of the metadata
781 : /// updated.
782 : ///
783 : /// This is the regular way of updating metadata on layer flushes or Gc.
784 : ///
785 : /// Using this lighter update mechanism allows for reparenting and detaching without changes to
786 : /// `index_part.json`, while being more clear on what values update regularly.
787 2456 : pub(crate) fn schedule_index_upload_for_metadata_update(
788 2456 : self: &Arc<Self>,
789 2456 : update: &MetadataUpdate,
790 2456 : ) -> anyhow::Result<()> {
791 2456 : let mut guard = self.upload_queue.lock().unwrap();
792 2456 : let upload_queue = guard.initialized_mut()?;
793 :
794 2456 : upload_queue.dirty.metadata.apply(update);
795 2456 :
796 2456 : // Defense in depth: if we somehow generated invalid metadata, do not persist it.
797 2456 : upload_queue
798 2456 : .dirty
799 2456 : .validate()
800 2456 : .map_err(|e| anyhow::anyhow!(e))?;
801 :
802 2456 : self.schedule_index_upload(upload_queue);
803 2456 :
804 2456 : Ok(())
805 2456 : }
806 :
807 : /// Launch an index-file upload operation in the background, with only the `archived_at` field updated.
808 : ///
809 : /// Returns whether it is required to wait for the queue to be empty to ensure that the change is uploaded,
810 : /// so either if the change is already sitting in the queue, but not commited yet, or the change has not
811 : /// been in the queue yet.
812 4 : pub(crate) fn schedule_index_upload_for_timeline_archival_state(
813 4 : self: &Arc<Self>,
814 4 : state: TimelineArchivalState,
815 4 : ) -> anyhow::Result<bool> {
816 4 : let mut guard = self.upload_queue.lock().unwrap();
817 4 : let upload_queue = guard.initialized_mut()?;
818 :
819 : /// Returns Some(_) if a change is needed, and Some(true) if it's a
820 : /// change needed to set archived_at.
821 8 : fn need_change(
822 8 : archived_at: &Option<NaiveDateTime>,
823 8 : state: TimelineArchivalState,
824 8 : ) -> Option<bool> {
825 8 : match (archived_at, state) {
826 : (Some(_), TimelineArchivalState::Archived)
827 : | (None, TimelineArchivalState::Unarchived) => {
828 : // Nothing to do
829 0 : tracing::info!("intended state matches present state");
830 0 : None
831 : }
832 8 : (None, TimelineArchivalState::Archived) => Some(true),
833 0 : (Some(_), TimelineArchivalState::Unarchived) => Some(false),
834 : }
835 8 : }
836 4 : let need_upload_scheduled = need_change(&upload_queue.dirty.archived_at, state);
837 :
838 4 : if let Some(archived_at_set) = need_upload_scheduled {
839 4 : let intended_archived_at = archived_at_set.then(|| Utc::now().naive_utc());
840 4 : upload_queue.dirty.archived_at = intended_archived_at;
841 4 : self.schedule_index_upload(upload_queue);
842 4 : }
843 :
844 4 : let need_wait = need_change(&upload_queue.clean.0.archived_at, state).is_some();
845 4 : Ok(need_wait)
846 4 : }
847 :
848 : /// Shuts the timeline client down, but only if the timeline is archived.
849 : ///
850 : /// This function and [`Self::schedule_index_upload_for_timeline_archival_state`] use the
851 : /// same lock to prevent races between unarchival and offloading: unarchival requires the
852 : /// upload queue to be initialized, and leaves behind an upload queue where either dirty
853 : /// or clean has archived_at of `None`. offloading leaves behind an uninitialized upload
854 : /// queue.
855 4 : pub(crate) async fn shutdown_if_archived(
856 4 : self: &Arc<Self>,
857 4 : ) -> Result<(), ShutdownIfArchivedError> {
858 4 : {
859 4 : let mut guard = self.upload_queue.lock().unwrap();
860 4 : let upload_queue = guard
861 4 : .initialized_mut()
862 4 : .map_err(ShutdownIfArchivedError::NotInitialized)?;
863 :
864 4 : match (
865 4 : upload_queue.dirty.archived_at.is_none(),
866 4 : upload_queue.clean.0.archived_at.is_none(),
867 4 : ) {
868 : // The expected case: the timeline is archived and we don't want to unarchive
869 4 : (false, false) => {}
870 : (true, false) => {
871 0 : tracing::info!("can't shut down timeline: timeline slated for unarchival");
872 0 : return Err(ShutdownIfArchivedError::NotArchived);
873 : }
874 0 : (dirty_archived, true) => {
875 0 : tracing::info!(%dirty_archived, "can't shut down timeline: timeline not archived in remote storage");
876 0 : return Err(ShutdownIfArchivedError::NotArchived);
877 : }
878 : }
879 :
880 : // Set the shutting_down flag while the guard from the archival check is held.
881 : // This prevents a race with unarchival, as initialized_mut will not return
882 : // an upload queue from this point.
883 : // Also launch the queued tasks like shutdown() does.
884 4 : if !upload_queue.shutting_down {
885 4 : upload_queue.shutting_down = true;
886 4 : upload_queue.queued_operations.push_back(UploadOp::Shutdown);
887 4 : // this operation is not counted similar to Barrier
888 4 : self.launch_queued_tasks(upload_queue);
889 4 : }
890 : }
891 :
892 4 : self.shutdown().await;
893 :
894 4 : Ok(())
895 4 : }
896 :
897 : /// Launch an index-file upload operation in the background, setting `import_pgdata` field.
898 0 : pub(crate) fn schedule_index_upload_for_import_pgdata_state_update(
899 0 : self: &Arc<Self>,
900 0 : state: Option<import_pgdata::index_part_format::Root>,
901 0 : ) -> anyhow::Result<()> {
902 0 : let mut guard = self.upload_queue.lock().unwrap();
903 0 : let upload_queue = guard.initialized_mut()?;
904 0 : upload_queue.dirty.import_pgdata = state;
905 0 : self.schedule_index_upload(upload_queue);
906 0 : Ok(())
907 0 : }
908 :
909 : /// Launch an index-file upload operation in the background, setting `gc_compaction_state` field.
910 0 : pub(crate) fn schedule_index_upload_for_gc_compaction_state_update(
911 0 : self: &Arc<Self>,
912 0 : gc_compaction_state: GcCompactionState,
913 0 : ) -> anyhow::Result<()> {
914 0 : let mut guard = self.upload_queue.lock().unwrap();
915 0 : let upload_queue = guard.initialized_mut()?;
916 0 : upload_queue.dirty.gc_compaction = Some(gc_compaction_state);
917 0 : self.schedule_index_upload(upload_queue);
918 0 : Ok(())
919 0 : }
920 :
921 : /// Launch an index-file upload operation in the background, setting `rel_size_v2_status` field.
922 0 : pub(crate) fn schedule_index_upload_for_rel_size_v2_status_update(
923 0 : self: &Arc<Self>,
924 0 : rel_size_v2_status: RelSizeMigration,
925 0 : ) -> anyhow::Result<()> {
926 0 : let mut guard = self.upload_queue.lock().unwrap();
927 0 : let upload_queue = guard.initialized_mut()?;
928 0 : upload_queue.dirty.rel_size_migration = Some(rel_size_v2_status);
929 0 : // TODO: allow this operation to bypass the validation check because we might upload the index part
930 0 : // with no layers but the flag updated. For now, we just modify the index part in memory and the next
931 0 : // upload will include the flag.
932 0 : // self.schedule_index_upload(upload_queue);
933 0 : Ok(())
934 0 : }
935 :
936 : ///
937 : /// Launch an index-file upload operation in the background, if necessary.
938 : ///
939 : /// Use this function to schedule the update of the index file after
940 : /// scheduling file uploads or deletions. If no file uploads or deletions
941 : /// have been scheduled since the last index file upload, this does
942 : /// nothing.
943 : ///
944 : /// Like schedule_index_upload_for_metadata_update(), this merely adds
945 : /// the upload to the upload queue and returns quickly.
946 740 : pub fn schedule_index_upload_for_file_changes(self: &Arc<Self>) -> Result<(), NotInitialized> {
947 740 : let mut guard = self.upload_queue.lock().unwrap();
948 740 : let upload_queue = guard.initialized_mut()?;
949 :
950 740 : if upload_queue.latest_files_changes_since_metadata_upload_scheduled > 0 {
951 28 : self.schedule_index_upload(upload_queue);
952 712 : }
953 :
954 740 : Ok(())
955 740 : }
956 :
957 : /// Only used in the `patch_index_part` HTTP API to force trigger an index upload.
958 0 : pub fn force_schedule_index_upload(self: &Arc<Self>) -> Result<(), NotInitialized> {
959 0 : let mut guard = self.upload_queue.lock().unwrap();
960 0 : let upload_queue = guard.initialized_mut()?;
961 0 : self.schedule_index_upload(upload_queue);
962 0 : Ok(())
963 0 : }
964 :
965 : /// Launch an index-file upload operation in the background (internal function)
966 3092 : fn schedule_index_upload(self: &Arc<Self>, upload_queue: &mut UploadQueueInitialized) {
967 3092 : let disk_consistent_lsn = upload_queue.dirty.metadata.disk_consistent_lsn();
968 3092 : // fix up the duplicated field
969 3092 : upload_queue.dirty.disk_consistent_lsn = disk_consistent_lsn;
970 3092 :
971 3092 : // make sure it serializes before doing it in perform_upload_task so that it doesn't
972 3092 : // look like a retryable error
973 3092 : let void = std::io::sink();
974 3092 : serde_json::to_writer(void, &upload_queue.dirty).expect("serialize index_part.json");
975 3092 :
976 3092 : let index_part = &upload_queue.dirty;
977 3092 :
978 3092 : info!(
979 0 : "scheduling metadata upload up to consistent LSN {disk_consistent_lsn} with {} files ({} changed)",
980 0 : index_part.layer_metadata.len(),
981 : upload_queue.latest_files_changes_since_metadata_upload_scheduled,
982 : );
983 :
984 3092 : let op = UploadOp::UploadMetadata {
985 3092 : uploaded: Box::new(index_part.clone()),
986 3092 : };
987 3092 : self.metric_begin(&op);
988 3092 : upload_queue.queued_operations.push_back(op);
989 3092 : upload_queue.latest_files_changes_since_metadata_upload_scheduled = 0;
990 3092 :
991 3092 : // Launch the task immediately, if possible
992 3092 : self.launch_queued_tasks(upload_queue);
993 3092 : }
994 :
995 : /// Reparent this timeline to a new parent.
996 : ///
997 : /// A retryable step of timeline ancestor detach.
998 0 : pub(crate) async fn schedule_reparenting_and_wait(
999 0 : self: &Arc<Self>,
1000 0 : new_parent: &TimelineId,
1001 0 : ) -> anyhow::Result<()> {
1002 0 : let receiver = {
1003 0 : let mut guard = self.upload_queue.lock().unwrap();
1004 0 : let upload_queue = guard.initialized_mut()?;
1005 :
1006 0 : let Some(prev) = upload_queue.dirty.metadata.ancestor_timeline() else {
1007 0 : return Err(anyhow::anyhow!(
1008 0 : "cannot reparent without a current ancestor"
1009 0 : ));
1010 : };
1011 :
1012 0 : let uploaded = &upload_queue.clean.0.metadata;
1013 0 :
1014 0 : if uploaded.ancestor_timeline().is_none() && !uploaded.ancestor_lsn().is_valid() {
1015 : // nothing to do
1016 0 : None
1017 : } else {
1018 0 : upload_queue.dirty.metadata.reparent(new_parent);
1019 0 : upload_queue.dirty.lineage.record_previous_ancestor(&prev);
1020 0 :
1021 0 : self.schedule_index_upload(upload_queue);
1022 0 :
1023 0 : Some(self.schedule_barrier0(upload_queue))
1024 : }
1025 : };
1026 :
1027 0 : if let Some(receiver) = receiver {
1028 0 : Self::wait_completion0(receiver).await?;
1029 0 : }
1030 0 : Ok(())
1031 0 : }
1032 :
1033 : /// Schedules uploading a new version of `index_part.json` with the given layers added,
1034 : /// detaching from ancestor and waits for it to complete.
1035 : ///
1036 : /// This is used with `Timeline::detach_ancestor` functionality.
1037 0 : pub(crate) async fn schedule_adding_existing_layers_to_index_detach_and_wait(
1038 0 : self: &Arc<Self>,
1039 0 : layers: &[Layer],
1040 0 : adopted: (TimelineId, Lsn),
1041 0 : ) -> anyhow::Result<()> {
1042 0 : let barrier = {
1043 0 : let mut guard = self.upload_queue.lock().unwrap();
1044 0 : let upload_queue = guard.initialized_mut()?;
1045 :
1046 0 : if upload_queue.clean.0.lineage.detached_previous_ancestor() == Some(adopted) {
1047 0 : None
1048 : } else {
1049 0 : upload_queue.dirty.metadata.detach_from_ancestor(&adopted);
1050 0 : upload_queue.dirty.lineage.record_detaching(&adopted);
1051 :
1052 0 : for layer in layers {
1053 0 : let prev = upload_queue
1054 0 : .dirty
1055 0 : .layer_metadata
1056 0 : .insert(layer.layer_desc().layer_name(), layer.metadata());
1057 0 : assert!(prev.is_none(), "copied layer existed already {layer}");
1058 : }
1059 :
1060 0 : self.schedule_index_upload(upload_queue);
1061 0 :
1062 0 : Some(self.schedule_barrier0(upload_queue))
1063 : }
1064 : };
1065 :
1066 0 : if let Some(barrier) = barrier {
1067 0 : Self::wait_completion0(barrier).await?;
1068 0 : }
1069 0 : Ok(())
1070 0 : }
1071 :
1072 : /// Adds a gc blocking reason for this timeline if one does not exist already.
1073 : ///
1074 : /// A retryable step of timeline detach ancestor.
1075 : ///
1076 : /// Returns a future which waits until the completion of the upload.
1077 0 : pub(crate) fn schedule_insert_gc_block_reason(
1078 0 : self: &Arc<Self>,
1079 0 : reason: index::GcBlockingReason,
1080 0 : ) -> Result<impl std::future::Future<Output = Result<(), WaitCompletionError>>, NotInitialized>
1081 0 : {
1082 0 : let maybe_barrier = {
1083 0 : let mut guard = self.upload_queue.lock().unwrap();
1084 0 : let upload_queue = guard.initialized_mut()?;
1085 :
1086 0 : if let index::GcBlockingReason::DetachAncestor = reason {
1087 0 : if upload_queue.dirty.metadata.ancestor_timeline().is_none() {
1088 0 : drop(guard);
1089 0 : panic!("cannot start detach ancestor if there is nothing to detach from");
1090 0 : }
1091 0 : }
1092 :
1093 0 : let wanted = |x: Option<&index::GcBlocking>| x.is_some_and(|x| x.blocked_by(reason));
1094 :
1095 0 : let current = upload_queue.dirty.gc_blocking.as_ref();
1096 0 : let uploaded = upload_queue.clean.0.gc_blocking.as_ref();
1097 0 :
1098 0 : match (current, uploaded) {
1099 0 : (x, y) if wanted(x) && wanted(y) => None,
1100 0 : (x, y) if wanted(x) && !wanted(y) => Some(self.schedule_barrier0(upload_queue)),
1101 : // Usual case: !wanted(x) && !wanted(y)
1102 : //
1103 : // Unusual: !wanted(x) && wanted(y) which means we have two processes waiting to
1104 : // turn on and off some reason.
1105 0 : (x, y) => {
1106 0 : if !wanted(x) && wanted(y) {
1107 : // this could be avoided by having external in-memory synchronization, like
1108 : // timeline detach ancestor
1109 0 : warn!(
1110 : ?reason,
1111 : op = "insert",
1112 0 : "unexpected: two racing processes to enable and disable a gc blocking reason"
1113 : );
1114 0 : }
1115 :
1116 : // at this point, the metadata must always show that there is a parent
1117 0 : upload_queue.dirty.gc_blocking = current
1118 0 : .map(|x| x.with_reason(reason))
1119 0 : .or_else(|| Some(index::GcBlocking::started_now_for(reason)));
1120 0 : self.schedule_index_upload(upload_queue);
1121 0 : Some(self.schedule_barrier0(upload_queue))
1122 : }
1123 : }
1124 : };
1125 :
1126 0 : Ok(async move {
1127 0 : if let Some(barrier) = maybe_barrier {
1128 0 : Self::wait_completion0(barrier).await?;
1129 0 : }
1130 0 : Ok(())
1131 0 : })
1132 0 : }
1133 :
1134 : /// Removes a gc blocking reason for this timeline if one exists.
1135 : ///
1136 : /// A retryable step of timeline detach ancestor.
1137 : ///
1138 : /// Returns a future which waits until the completion of the upload.
1139 0 : pub(crate) fn schedule_remove_gc_block_reason(
1140 0 : self: &Arc<Self>,
1141 0 : reason: index::GcBlockingReason,
1142 0 : ) -> Result<impl std::future::Future<Output = Result<(), WaitCompletionError>>, NotInitialized>
1143 0 : {
1144 0 : let maybe_barrier = {
1145 0 : let mut guard = self.upload_queue.lock().unwrap();
1146 0 : let upload_queue = guard.initialized_mut()?;
1147 :
1148 0 : if let index::GcBlockingReason::DetachAncestor = reason {
1149 0 : if !upload_queue.clean.0.lineage.is_detached_from_ancestor() {
1150 0 : drop(guard);
1151 0 : panic!("cannot complete timeline_ancestor_detach while not detached");
1152 0 : }
1153 0 : }
1154 :
1155 0 : let wanted = |x: Option<&index::GcBlocking>| {
1156 0 : x.is_none() || x.is_some_and(|b| !b.blocked_by(reason))
1157 0 : };
1158 :
1159 0 : let current = upload_queue.dirty.gc_blocking.as_ref();
1160 0 : let uploaded = upload_queue.clean.0.gc_blocking.as_ref();
1161 0 :
1162 0 : match (current, uploaded) {
1163 0 : (x, y) if wanted(x) && wanted(y) => None,
1164 0 : (x, y) if wanted(x) && !wanted(y) => Some(self.schedule_barrier0(upload_queue)),
1165 0 : (x, y) => {
1166 0 : if !wanted(x) && wanted(y) {
1167 0 : warn!(
1168 : ?reason,
1169 : op = "remove",
1170 0 : "unexpected: two racing processes to enable and disable a gc blocking reason (remove)"
1171 : );
1172 0 : }
1173 :
1174 0 : upload_queue.dirty.gc_blocking =
1175 0 : current.as_ref().and_then(|x| x.without_reason(reason));
1176 0 : assert!(wanted(upload_queue.dirty.gc_blocking.as_ref()));
1177 0 : self.schedule_index_upload(upload_queue);
1178 0 : Some(self.schedule_barrier0(upload_queue))
1179 : }
1180 : }
1181 : };
1182 :
1183 0 : Ok(async move {
1184 0 : if let Some(barrier) = maybe_barrier {
1185 0 : Self::wait_completion0(barrier).await?;
1186 0 : }
1187 0 : Ok(())
1188 0 : })
1189 0 : }
1190 :
1191 : /// Launch an upload operation in the background; the file is added to be included in next
1192 : /// `index_part.json` upload.
1193 2760 : pub(crate) fn schedule_layer_file_upload(
1194 2760 : self: &Arc<Self>,
1195 2760 : layer: ResidentLayer,
1196 2760 : ) -> Result<(), NotInitialized> {
1197 2760 : let mut guard = self.upload_queue.lock().unwrap();
1198 2760 : let upload_queue = guard.initialized_mut()?;
1199 :
1200 2760 : self.schedule_layer_file_upload0(upload_queue, layer);
1201 2760 : self.launch_queued_tasks(upload_queue);
1202 2760 : Ok(())
1203 2760 : }
1204 :
1205 3484 : fn schedule_layer_file_upload0(
1206 3484 : self: &Arc<Self>,
1207 3484 : upload_queue: &mut UploadQueueInitialized,
1208 3484 : layer: ResidentLayer,
1209 3484 : ) {
1210 3484 : let metadata = layer.metadata();
1211 3484 :
1212 3484 : upload_queue
1213 3484 : .dirty
1214 3484 : .layer_metadata
1215 3484 : .insert(layer.layer_desc().layer_name(), metadata.clone());
1216 3484 : upload_queue.latest_files_changes_since_metadata_upload_scheduled += 1;
1217 3484 :
1218 3484 : info!(
1219 : gen=?metadata.generation,
1220 : shard=?metadata.shard,
1221 0 : "scheduled layer file upload {layer}",
1222 : );
1223 :
1224 3484 : let op = UploadOp::UploadLayer(layer, metadata, None);
1225 3484 : self.metric_begin(&op);
1226 3484 : upload_queue.queued_operations.push_back(op);
1227 3484 : }
1228 :
1229 : /// Launch a delete operation in the background.
1230 : ///
1231 : /// The operation does not modify local filesystem state.
1232 : ///
1233 : /// Note: This schedules an index file upload before the deletions. The
1234 : /// deletion won't actually be performed, until all previously scheduled
1235 : /// upload operations, and the index file upload, have completed
1236 : /// successfully.
1237 16 : pub fn schedule_layer_file_deletion(
1238 16 : self: &Arc<Self>,
1239 16 : names: &[LayerName],
1240 16 : ) -> anyhow::Result<()> {
1241 16 : let mut guard = self.upload_queue.lock().unwrap();
1242 16 : let upload_queue = guard.initialized_mut()?;
1243 :
1244 16 : let with_metadata =
1245 16 : self.schedule_unlinking_of_layers_from_index_part0(upload_queue, names.iter().cloned());
1246 16 :
1247 16 : self.schedule_deletion_of_unlinked0(upload_queue, with_metadata);
1248 16 :
1249 16 : // Launch the tasks immediately, if possible
1250 16 : self.launch_queued_tasks(upload_queue);
1251 16 : Ok(())
1252 16 : }
1253 :
1254 : /// Unlinks the layer files from `index_part.json` but does not yet schedule deletion for the
1255 : /// layer files, leaving them dangling.
1256 : ///
1257 : /// The files will be leaked in remote storage unless [`Self::schedule_deletion_of_unlinked`]
1258 : /// is invoked on them.
1259 8 : pub(crate) fn schedule_gc_update(
1260 8 : self: &Arc<Self>,
1261 8 : gc_layers: &[Layer],
1262 8 : ) -> Result<(), NotInitialized> {
1263 8 : let mut guard = self.upload_queue.lock().unwrap();
1264 8 : let upload_queue = guard.initialized_mut()?;
1265 :
1266 : // just forget the return value; after uploading the next index_part.json, we can consider
1267 : // the layer files as "dangling". this is fine, at worst case we create work for the
1268 : // scrubber.
1269 :
1270 8 : let names = gc_layers.iter().map(|x| x.layer_desc().layer_name());
1271 8 :
1272 8 : self.schedule_unlinking_of_layers_from_index_part0(upload_queue, names);
1273 8 :
1274 8 : self.launch_queued_tasks(upload_queue);
1275 8 :
1276 8 : Ok(())
1277 8 : }
1278 :
1279 : /// Update the remote index file, removing the to-be-deleted files from the index,
1280 : /// allowing scheduling of actual deletions later.
1281 176 : fn schedule_unlinking_of_layers_from_index_part0<I>(
1282 176 : self: &Arc<Self>,
1283 176 : upload_queue: &mut UploadQueueInitialized,
1284 176 : names: I,
1285 176 : ) -> Vec<(LayerName, LayerFileMetadata)>
1286 176 : where
1287 176 : I: IntoIterator<Item = LayerName>,
1288 176 : {
1289 176 : // Decorate our list of names with each name's metadata, dropping
1290 176 : // names that are unexpectedly missing from our metadata. This metadata
1291 176 : // is later used when physically deleting layers, to construct key paths.
1292 176 : let with_metadata: Vec<_> = names
1293 176 : .into_iter()
1294 1024 : .filter_map(|name| {
1295 1024 : let meta = upload_queue.dirty.layer_metadata.remove(&name);
1296 :
1297 1024 : if let Some(meta) = meta {
1298 1024 : upload_queue.latest_files_changes_since_metadata_upload_scheduled += 1;
1299 1024 : Some((name, meta))
1300 : } else {
1301 : // This can only happen if we forgot to to schedule the file upload
1302 : // before scheduling the delete. Log it because it is a rare/strange
1303 : // situation, and in case something is misbehaving, we'd like to know which
1304 : // layers experienced this.
1305 0 : info!("Deleting layer {name} not found in latest_files list, never uploaded?");
1306 0 : None
1307 : }
1308 1024 : })
1309 176 : .collect();
1310 :
1311 : #[cfg(feature = "testing")]
1312 1200 : for (name, metadata) in &with_metadata {
1313 1024 : let gen_ = metadata.generation;
1314 1024 : if let Some(unexpected) = upload_queue.dangling_files.insert(name.to_owned(), gen_) {
1315 0 : if unexpected == gen_ {
1316 0 : tracing::error!("{name} was unlinked twice with same generation");
1317 : } else {
1318 0 : tracing::error!(
1319 0 : "{name} was unlinked twice with different generations {gen_:?} and {unexpected:?}"
1320 : );
1321 : }
1322 1024 : }
1323 : }
1324 :
1325 : // after unlinking files from the upload_queue.latest_files we must always schedule an
1326 : // index_part update, because that needs to be uploaded before we can actually delete the
1327 : // files.
1328 176 : if upload_queue.latest_files_changes_since_metadata_upload_scheduled > 0 {
1329 144 : self.schedule_index_upload(upload_queue);
1330 144 : }
1331 :
1332 176 : with_metadata
1333 176 : }
1334 :
1335 : /// Schedules deletion for layer files which have previously been unlinked from the
1336 : /// `index_part.json` with [`Self::schedule_gc_update`] or [`Self::schedule_compaction_update`].
1337 1026 : pub(crate) fn schedule_deletion_of_unlinked(
1338 1026 : self: &Arc<Self>,
1339 1026 : layers: Vec<(LayerName, LayerFileMetadata)>,
1340 1026 : ) -> anyhow::Result<()> {
1341 1026 : let mut guard = self.upload_queue.lock().unwrap();
1342 1026 : let upload_queue = guard.initialized_mut()?;
1343 :
1344 1026 : self.schedule_deletion_of_unlinked0(upload_queue, layers);
1345 1026 : self.launch_queued_tasks(upload_queue);
1346 1026 : Ok(())
1347 1026 : }
1348 :
1349 1038 : fn schedule_deletion_of_unlinked0(
1350 1038 : self: &Arc<Self>,
1351 1038 : upload_queue: &mut UploadQueueInitialized,
1352 1038 : mut with_metadata: Vec<(LayerName, LayerFileMetadata)>,
1353 1038 : ) {
1354 1038 : // Filter out any layers which were not created by this tenant shard. These are
1355 1038 : // layers that originate from some ancestor shard after a split, and may still
1356 1038 : // be referenced by other shards. We are free to delete them locally and remove
1357 1038 : // them from our index (and would have already done so when we reach this point
1358 1038 : // in the code), but we may not delete them remotely.
1359 1038 : with_metadata.retain(|(name, meta)| {
1360 1026 : let retain = meta.shard.shard_number == self.tenant_shard_id.shard_number
1361 1026 : && meta.shard.shard_count == self.tenant_shard_id.shard_count;
1362 1026 : if !retain {
1363 0 : tracing::debug!(
1364 0 : "Skipping deletion of ancestor-shard layer {name}, from shard {}",
1365 : meta.shard
1366 : );
1367 1026 : }
1368 1026 : retain
1369 1038 : });
1370 :
1371 2064 : for (name, meta) in &with_metadata {
1372 1026 : info!(
1373 0 : "scheduling deletion of layer {}{} (shard {})",
1374 0 : name,
1375 0 : meta.generation.get_suffix(),
1376 : meta.shard
1377 : );
1378 : }
1379 :
1380 : #[cfg(feature = "testing")]
1381 2064 : for (name, meta) in &with_metadata {
1382 1026 : let gen_ = meta.generation;
1383 1026 : match upload_queue.dangling_files.remove(name) {
1384 1018 : Some(same) if same == gen_ => { /* expected */ }
1385 0 : Some(other) => {
1386 0 : tracing::error!("{name} was unlinked with {other:?} but deleted with {gen_:?}");
1387 : }
1388 : None => {
1389 8 : tracing::error!("{name} was unlinked but was not dangling");
1390 : }
1391 : }
1392 : }
1393 :
1394 : // schedule the actual deletions
1395 1038 : if with_metadata.is_empty() {
1396 : // avoid scheduling the op & bumping the metric
1397 12 : return;
1398 1026 : }
1399 1026 : let op = UploadOp::Delete(Delete {
1400 1026 : layers: with_metadata,
1401 1026 : });
1402 1026 : self.metric_begin(&op);
1403 1026 : upload_queue.queued_operations.push_back(op);
1404 1038 : }
1405 :
1406 : /// Schedules a compaction update to the remote `index_part.json`.
1407 : ///
1408 : /// `compacted_from` represent the L0 names which have been `compacted_to` L1 layers.
1409 152 : pub(crate) fn schedule_compaction_update(
1410 152 : self: &Arc<Self>,
1411 152 : compacted_from: &[Layer],
1412 152 : compacted_to: &[ResidentLayer],
1413 152 : ) -> Result<(), NotInitialized> {
1414 152 : let mut guard = self.upload_queue.lock().unwrap();
1415 152 : let upload_queue = guard.initialized_mut()?;
1416 :
1417 876 : for layer in compacted_to {
1418 724 : self.schedule_layer_file_upload0(upload_queue, layer.clone());
1419 724 : }
1420 :
1421 1012 : let names = compacted_from.iter().map(|x| x.layer_desc().layer_name());
1422 152 :
1423 152 : self.schedule_unlinking_of_layers_from_index_part0(upload_queue, names);
1424 152 : self.launch_queued_tasks(upload_queue);
1425 152 :
1426 152 : Ok(())
1427 152 : }
1428 :
1429 : /// Wait for all previously scheduled uploads/deletions to complete
1430 2796 : pub(crate) async fn wait_completion(self: &Arc<Self>) -> Result<(), WaitCompletionError> {
1431 2796 : let receiver = {
1432 2796 : let mut guard = self.upload_queue.lock().unwrap();
1433 2796 : let upload_queue = guard
1434 2796 : .initialized_mut()
1435 2796 : .map_err(WaitCompletionError::NotInitialized)?;
1436 2796 : self.schedule_barrier0(upload_queue)
1437 2796 : };
1438 2796 :
1439 2796 : Self::wait_completion0(receiver).await
1440 2796 : }
1441 :
1442 2796 : async fn wait_completion0(
1443 2796 : mut receiver: tokio::sync::watch::Receiver<()>,
1444 2796 : ) -> Result<(), WaitCompletionError> {
1445 2796 : if receiver.changed().await.is_err() {
1446 0 : return Err(WaitCompletionError::UploadQueueShutDownOrStopped);
1447 2796 : }
1448 2796 :
1449 2796 : Ok(())
1450 2796 : }
1451 :
1452 12 : pub(crate) fn schedule_barrier(self: &Arc<Self>) -> anyhow::Result<()> {
1453 12 : let mut guard = self.upload_queue.lock().unwrap();
1454 12 : let upload_queue = guard.initialized_mut()?;
1455 12 : self.schedule_barrier0(upload_queue);
1456 12 : Ok(())
1457 12 : }
1458 :
1459 2808 : fn schedule_barrier0(
1460 2808 : self: &Arc<Self>,
1461 2808 : upload_queue: &mut UploadQueueInitialized,
1462 2808 : ) -> tokio::sync::watch::Receiver<()> {
1463 2808 : let (sender, receiver) = tokio::sync::watch::channel(());
1464 2808 : let barrier_op = UploadOp::Barrier(sender);
1465 2808 :
1466 2808 : upload_queue.queued_operations.push_back(barrier_op);
1467 2808 : // Don't count this kind of operation!
1468 2808 :
1469 2808 : // Launch the task immediately, if possible
1470 2808 : self.launch_queued_tasks(upload_queue);
1471 2808 :
1472 2808 : receiver
1473 2808 : }
1474 :
1475 : /// Wait for all previously scheduled operations to complete, and then stop.
1476 : ///
1477 : /// Not cancellation safe
1478 20 : pub(crate) async fn shutdown(self: &Arc<Self>) {
1479 20 : // On cancellation the queue is left in ackward state of refusing new operations but
1480 20 : // proper stop is yet to be called. On cancel the original or some later task must call
1481 20 : // `stop` or `shutdown`.
1482 20 : let sg = scopeguard::guard((), |_| {
1483 0 : tracing::error!(
1484 0 : "RemoteTimelineClient::shutdown was cancelled; this should not happen, do not make this into an allowed_error"
1485 : )
1486 20 : });
1487 :
1488 16 : let fut = {
1489 20 : let mut guard = self.upload_queue.lock().unwrap();
1490 20 : let upload_queue = match &mut *guard {
1491 : UploadQueue::Stopped(_) => {
1492 4 : scopeguard::ScopeGuard::into_inner(sg);
1493 4 : return;
1494 : }
1495 : UploadQueue::Uninitialized => {
1496 : // transition into Stopped state
1497 0 : self.stop_impl(&mut guard);
1498 0 : scopeguard::ScopeGuard::into_inner(sg);
1499 0 : return;
1500 : }
1501 16 : UploadQueue::Initialized(init) => init,
1502 16 : };
1503 16 :
1504 16 : // if the queue is already stuck due to a shutdown operation which was cancelled, then
1505 16 : // just don't add more of these as they would never complete.
1506 16 : //
1507 16 : // TODO: if launch_queued_tasks were to be refactored to accept a &mut UploadQueue
1508 16 : // in every place we would not have to jump through this hoop, and this method could be
1509 16 : // made cancellable.
1510 16 : if !upload_queue.shutting_down {
1511 12 : upload_queue.shutting_down = true;
1512 12 : upload_queue.queued_operations.push_back(UploadOp::Shutdown);
1513 12 : // this operation is not counted similar to Barrier
1514 12 :
1515 12 : self.launch_queued_tasks(upload_queue);
1516 12 : }
1517 :
1518 16 : upload_queue.shutdown_ready.clone().acquire_owned()
1519 : };
1520 :
1521 16 : let res = fut.await;
1522 :
1523 16 : scopeguard::ScopeGuard::into_inner(sg);
1524 16 :
1525 16 : match res {
1526 0 : Ok(_permit) => unreachable!("shutdown_ready should not have been added permits"),
1527 16 : Err(_closed) => {
1528 16 : // expected
1529 16 : }
1530 16 : }
1531 16 :
1532 16 : self.stop();
1533 20 : }
1534 :
1535 : /// Set the deleted_at field in the remote index file.
1536 : ///
1537 : /// This fails if the upload queue has not been `stop()`ed.
1538 : ///
1539 : /// The caller is responsible for calling `stop()` AND for waiting
1540 : /// for any ongoing upload tasks to finish after `stop()` has succeeded.
1541 : /// Check method [`RemoteTimelineClient::stop`] for details.
1542 : #[instrument(skip_all)]
1543 : pub(crate) async fn persist_index_part_with_deleted_flag(
1544 : self: &Arc<Self>,
1545 : ) -> Result<(), PersistIndexPartWithDeletedFlagError> {
1546 : let index_part_with_deleted_at = {
1547 : let mut locked = self.upload_queue.lock().unwrap();
1548 :
1549 : // We must be in stopped state because otherwise
1550 : // we can have inprogress index part upload that can overwrite the file
1551 : // with missing is_deleted flag that we going to set below
1552 : let stopped = locked.stopped_mut()?;
1553 :
1554 : match stopped.deleted_at {
1555 : SetDeletedFlagProgress::NotRunning => (), // proceed
1556 : SetDeletedFlagProgress::InProgress(at) => {
1557 : return Err(PersistIndexPartWithDeletedFlagError::AlreadyInProgress(at));
1558 : }
1559 : SetDeletedFlagProgress::Successful(at) => {
1560 : return Err(PersistIndexPartWithDeletedFlagError::AlreadyDeleted(at));
1561 : }
1562 : };
1563 : let deleted_at = Utc::now().naive_utc();
1564 : stopped.deleted_at = SetDeletedFlagProgress::InProgress(deleted_at);
1565 :
1566 : let mut index_part = stopped.upload_queue_for_deletion.dirty.clone();
1567 : index_part.deleted_at = Some(deleted_at);
1568 : index_part
1569 : };
1570 :
1571 0 : let undo_deleted_at = scopeguard::guard(Arc::clone(self), |self_clone| {
1572 0 : let mut locked = self_clone.upload_queue.lock().unwrap();
1573 0 : let stopped = locked
1574 0 : .stopped_mut()
1575 0 : .expect("there's no way out of Stopping, and we checked it's Stopping above");
1576 0 : stopped.deleted_at = SetDeletedFlagProgress::NotRunning;
1577 0 : });
1578 :
1579 : pausable_failpoint!("persist_deleted_index_part");
1580 :
1581 : backoff::retry(
1582 0 : || {
1583 0 : upload::upload_index_part(
1584 0 : &self.storage_impl,
1585 0 : &self.tenant_shard_id,
1586 0 : &self.timeline_id,
1587 0 : self.generation,
1588 0 : &index_part_with_deleted_at,
1589 0 : &self.cancel,
1590 0 : )
1591 0 : },
1592 0 : |_e| false,
1593 : 1,
1594 : // have just a couple of attempts
1595 : // when executed as part of timeline deletion this happens in context of api call
1596 : // when executed as part of tenant deletion this happens in the background
1597 : 2,
1598 : "persist_index_part_with_deleted_flag",
1599 : &self.cancel,
1600 : )
1601 : .await
1602 0 : .ok_or_else(|| anyhow::Error::new(TimeoutOrCancel::Cancel))
1603 0 : .and_then(|x| x)?;
1604 :
1605 : // all good, disarm the guard and mark as success
1606 : ScopeGuard::into_inner(undo_deleted_at);
1607 : {
1608 : let mut locked = self.upload_queue.lock().unwrap();
1609 :
1610 : let stopped = locked
1611 : .stopped_mut()
1612 : .expect("there's no way out of Stopping, and we checked it's Stopping above");
1613 : stopped.deleted_at = SetDeletedFlagProgress::Successful(
1614 : index_part_with_deleted_at
1615 : .deleted_at
1616 : .expect("we set it above"),
1617 : );
1618 : }
1619 :
1620 : Ok(())
1621 : }
1622 :
1623 0 : pub(crate) fn is_deleting(&self) -> bool {
1624 0 : let mut locked = self.upload_queue.lock().unwrap();
1625 0 : locked.stopped_mut().is_ok()
1626 0 : }
1627 :
1628 0 : pub(crate) async fn preserve_initdb_archive(
1629 0 : self: &Arc<Self>,
1630 0 : tenant_id: &TenantId,
1631 0 : timeline_id: &TimelineId,
1632 0 : cancel: &CancellationToken,
1633 0 : ) -> anyhow::Result<()> {
1634 0 : backoff::retry(
1635 0 : || async {
1636 0 : upload::preserve_initdb_archive(&self.storage_impl, tenant_id, timeline_id, cancel)
1637 0 : .await
1638 0 : },
1639 0 : TimeoutOrCancel::caused_by_cancel,
1640 0 : FAILED_DOWNLOAD_WARN_THRESHOLD,
1641 0 : FAILED_REMOTE_OP_RETRIES,
1642 0 : "preserve_initdb_tar_zst",
1643 0 : &cancel.clone(),
1644 0 : )
1645 0 : .await
1646 0 : .ok_or_else(|| anyhow::Error::new(TimeoutOrCancel::Cancel))
1647 0 : .and_then(|x| x)
1648 0 : .context("backing up initdb archive")?;
1649 0 : Ok(())
1650 0 : }
1651 :
1652 : /// Uploads the given layer **without** adding it to be part of a future `index_part.json` upload.
1653 : ///
1654 : /// This is not normally needed.
1655 0 : pub(crate) async fn upload_layer_file(
1656 0 : self: &Arc<Self>,
1657 0 : uploaded: &ResidentLayer,
1658 0 : cancel: &CancellationToken,
1659 0 : ) -> anyhow::Result<()> {
1660 0 : let remote_path = remote_layer_path(
1661 0 : &self.tenant_shard_id.tenant_id,
1662 0 : &self.timeline_id,
1663 0 : uploaded.metadata().shard,
1664 0 : &uploaded.layer_desc().layer_name(),
1665 0 : uploaded.metadata().generation,
1666 0 : );
1667 0 :
1668 0 : backoff::retry(
1669 0 : || async {
1670 0 : upload::upload_timeline_layer(
1671 0 : &self.storage_impl,
1672 0 : uploaded.local_path(),
1673 0 : &remote_path,
1674 0 : uploaded.metadata().file_size,
1675 0 : cancel,
1676 0 : )
1677 0 : .await
1678 0 : },
1679 0 : TimeoutOrCancel::caused_by_cancel,
1680 0 : FAILED_UPLOAD_WARN_THRESHOLD,
1681 0 : FAILED_REMOTE_OP_RETRIES,
1682 0 : "upload a layer without adding it to latest files",
1683 0 : cancel,
1684 0 : )
1685 0 : .await
1686 0 : .ok_or_else(|| anyhow::Error::new(TimeoutOrCancel::Cancel))
1687 0 : .and_then(|x| x)
1688 0 : .context("upload a layer without adding it to latest files")
1689 0 : }
1690 :
1691 : /// Copies the `adopted` remote existing layer to the remote path of `adopted_as`. The layer is
1692 : /// not added to be part of a future `index_part.json` upload.
1693 0 : pub(crate) async fn copy_timeline_layer(
1694 0 : self: &Arc<Self>,
1695 0 : adopted: &Layer,
1696 0 : adopted_as: &Layer,
1697 0 : cancel: &CancellationToken,
1698 0 : ) -> anyhow::Result<()> {
1699 0 : let source_remote_path = remote_layer_path(
1700 0 : &self.tenant_shard_id.tenant_id,
1701 0 : &adopted
1702 0 : .get_timeline_id()
1703 0 : .expect("Source timeline should be alive"),
1704 0 : adopted.metadata().shard,
1705 0 : &adopted.layer_desc().layer_name(),
1706 0 : adopted.metadata().generation,
1707 0 : );
1708 0 :
1709 0 : let target_remote_path = remote_layer_path(
1710 0 : &self.tenant_shard_id.tenant_id,
1711 0 : &self.timeline_id,
1712 0 : adopted_as.metadata().shard,
1713 0 : &adopted_as.layer_desc().layer_name(),
1714 0 : adopted_as.metadata().generation,
1715 0 : );
1716 0 :
1717 0 : backoff::retry(
1718 0 : || async {
1719 0 : upload::copy_timeline_layer(
1720 0 : &self.storage_impl,
1721 0 : &source_remote_path,
1722 0 : &target_remote_path,
1723 0 : cancel,
1724 0 : )
1725 0 : .await
1726 0 : },
1727 0 : TimeoutOrCancel::caused_by_cancel,
1728 0 : FAILED_UPLOAD_WARN_THRESHOLD,
1729 0 : FAILED_REMOTE_OP_RETRIES,
1730 0 : "copy timeline layer",
1731 0 : cancel,
1732 0 : )
1733 0 : .await
1734 0 : .ok_or_else(|| anyhow::Error::new(TimeoutOrCancel::Cancel))
1735 0 : .and_then(|x| x)
1736 0 : .context("remote copy timeline layer")
1737 0 : }
1738 :
1739 0 : async fn flush_deletion_queue(&self) -> Result<(), DeletionQueueError> {
1740 0 : match tokio::time::timeout(
1741 0 : DELETION_QUEUE_FLUSH_TIMEOUT,
1742 0 : self.deletion_queue_client.flush_immediate(),
1743 0 : )
1744 0 : .await
1745 : {
1746 0 : Ok(result) => result,
1747 0 : Err(_timeout) => {
1748 0 : // Flushing remote deletions is not mandatory: we flush here to make the system easier to test, and
1749 0 : // to ensure that _usually_ objects are really gone after a DELETE is acked. However, in case of deletion
1750 0 : // queue issues (https://github.com/neondatabase/neon/issues/6440), we don't want to wait indefinitely here.
1751 0 : tracing::warn!(
1752 0 : "Timed out waiting for deletion queue flush, acking deletion anyway"
1753 : );
1754 0 : Ok(())
1755 : }
1756 : }
1757 0 : }
1758 :
1759 : /// Prerequisites: UploadQueue should be in stopped state and deleted_at should be successfuly set.
1760 : /// The function deletes layer files one by one, then lists the prefix to see if we leaked something
1761 : /// deletes leaked files if any and proceeds with deletion of index file at the end.
1762 0 : pub(crate) async fn delete_all(self: &Arc<Self>) -> Result<(), DeleteTimelineError> {
1763 0 : debug_assert_current_span_has_tenant_and_timeline_id();
1764 :
1765 0 : let layers: Vec<RemotePath> = {
1766 0 : let mut locked = self.upload_queue.lock().unwrap();
1767 0 : let stopped = locked.stopped_mut().map_err(DeleteTimelineError::Other)?;
1768 :
1769 0 : if !matches!(stopped.deleted_at, SetDeletedFlagProgress::Successful(_)) {
1770 0 : return Err(DeleteTimelineError::Other(anyhow::anyhow!(
1771 0 : "deleted_at is not set"
1772 0 : )));
1773 0 : }
1774 0 :
1775 0 : debug_assert!(stopped.upload_queue_for_deletion.no_pending_work());
1776 :
1777 0 : stopped
1778 0 : .upload_queue_for_deletion
1779 0 : .dirty
1780 0 : .layer_metadata
1781 0 : .drain()
1782 0 : .filter(|(_file_name, meta)| {
1783 0 : // Filter out layers that belonged to an ancestor shard. Since we are deleting the whole timeline from
1784 0 : // all shards anyway, we _could_ delete these, but
1785 0 : // - it creates a potential race if other shards are still
1786 0 : // using the layers while this shard deletes them.
1787 0 : // - it means that if we rolled back the shard split, the ancestor shards would be in a state where
1788 0 : // these timelines are present but corrupt (their index exists but some layers don't)
1789 0 : //
1790 0 : // These layers will eventually be cleaned up by the scrubber when it does physical GC.
1791 0 : meta.shard.shard_number == self.tenant_shard_id.shard_number
1792 0 : && meta.shard.shard_count == self.tenant_shard_id.shard_count
1793 0 : })
1794 0 : .map(|(file_name, meta)| {
1795 0 : remote_layer_path(
1796 0 : &self.tenant_shard_id.tenant_id,
1797 0 : &self.timeline_id,
1798 0 : meta.shard,
1799 0 : &file_name,
1800 0 : meta.generation,
1801 0 : )
1802 0 : })
1803 0 : .collect()
1804 0 : };
1805 0 :
1806 0 : let layer_deletion_count = layers.len();
1807 0 : self.deletion_queue_client
1808 0 : .push_immediate(layers)
1809 0 : .await
1810 0 : .map_err(|_| DeleteTimelineError::Cancelled)?;
1811 :
1812 : // Delete the initdb.tar.zst, which is not always present, but deletion attempts of
1813 : // inexistant objects are not considered errors.
1814 0 : let initdb_path =
1815 0 : remote_initdb_archive_path(&self.tenant_shard_id.tenant_id, &self.timeline_id);
1816 0 : self.deletion_queue_client
1817 0 : .push_immediate(vec![initdb_path])
1818 0 : .await
1819 0 : .map_err(|_| DeleteTimelineError::Cancelled)?;
1820 :
1821 : // Do not delete index part yet, it is needed for possible retry. If we remove it first
1822 : // and retry will arrive to different pageserver there wont be any traces of it on remote storage
1823 0 : let timeline_storage_path = remote_timeline_path(&self.tenant_shard_id, &self.timeline_id);
1824 0 :
1825 0 : // Execute all pending deletions, so that when we proceed to do a listing below, we aren't
1826 0 : // taking the burden of listing all the layers that we already know we should delete.
1827 0 : self.flush_deletion_queue()
1828 0 : .await
1829 0 : .map_err(|_| DeleteTimelineError::Cancelled)?;
1830 :
1831 0 : let cancel = shutdown_token();
1832 :
1833 0 : let remaining = download_retry(
1834 0 : || async {
1835 0 : self.storage_impl
1836 0 : .list(
1837 0 : Some(&timeline_storage_path),
1838 0 : ListingMode::NoDelimiter,
1839 0 : None,
1840 0 : &cancel,
1841 0 : )
1842 0 : .await
1843 0 : },
1844 0 : "list remaining files",
1845 0 : &cancel,
1846 0 : )
1847 0 : .await
1848 0 : .context("list files remaining files")?
1849 : .keys;
1850 :
1851 : // We will delete the current index_part object last, since it acts as a deletion
1852 : // marker via its deleted_at attribute
1853 0 : let latest_index = remaining
1854 0 : .iter()
1855 0 : .filter(|o| {
1856 0 : o.key
1857 0 : .object_name()
1858 0 : .map(|n| n.starts_with(IndexPart::FILE_NAME))
1859 0 : .unwrap_or(false)
1860 0 : })
1861 0 : .filter_map(|o| {
1862 0 : parse_remote_index_path(o.key.clone()).map(|gen_| (o.key.clone(), gen_))
1863 0 : })
1864 0 : .max_by_key(|i| i.1)
1865 0 : .map(|i| i.0.clone())
1866 0 : .unwrap_or(
1867 0 : // No generation-suffixed indices, assume we are dealing with
1868 0 : // a legacy index.
1869 0 : remote_index_path(&self.tenant_shard_id, &self.timeline_id, Generation::none()),
1870 0 : );
1871 0 :
1872 0 : let remaining_layers: Vec<RemotePath> = remaining
1873 0 : .into_iter()
1874 0 : .filter_map(|o| {
1875 0 : if o.key == latest_index || o.key.object_name() == Some(INITDB_PRESERVED_PATH) {
1876 0 : None
1877 : } else {
1878 0 : Some(o.key)
1879 : }
1880 0 : })
1881 0 : .inspect(|path| {
1882 0 : if let Some(name) = path.object_name() {
1883 0 : info!(%name, "deleting a file not referenced from index_part.json");
1884 : } else {
1885 0 : warn!(%path, "deleting a nameless or non-utf8 object not referenced from index_part.json");
1886 : }
1887 0 : })
1888 0 : .collect();
1889 0 :
1890 0 : let not_referenced_count = remaining_layers.len();
1891 0 : if !remaining_layers.is_empty() {
1892 0 : self.deletion_queue_client
1893 0 : .push_immediate(remaining_layers)
1894 0 : .await
1895 0 : .map_err(|_| DeleteTimelineError::Cancelled)?;
1896 0 : }
1897 :
1898 0 : fail::fail_point!("timeline-delete-before-index-delete", |_| {
1899 0 : Err(DeleteTimelineError::Other(anyhow::anyhow!(
1900 0 : "failpoint: timeline-delete-before-index-delete"
1901 0 : )))?
1902 0 : });
1903 :
1904 0 : debug!("enqueuing index part deletion");
1905 0 : self.deletion_queue_client
1906 0 : .push_immediate([latest_index].to_vec())
1907 0 : .await
1908 0 : .map_err(|_| DeleteTimelineError::Cancelled)?;
1909 :
1910 : // Timeline deletion is rare and we have probably emitted a reasonably number of objects: wait
1911 : // for a flush to a persistent deletion list so that we may be sure deletion will occur.
1912 0 : self.flush_deletion_queue()
1913 0 : .await
1914 0 : .map_err(|_| DeleteTimelineError::Cancelled)?;
1915 :
1916 0 : fail::fail_point!("timeline-delete-after-index-delete", |_| {
1917 0 : Err(DeleteTimelineError::Other(anyhow::anyhow!(
1918 0 : "failpoint: timeline-delete-after-index-delete"
1919 0 : )))?
1920 0 : });
1921 :
1922 0 : info!(prefix=%timeline_storage_path, referenced=layer_deletion_count, not_referenced=%not_referenced_count, "done deleting in timeline prefix, including index_part.json");
1923 :
1924 0 : Ok(())
1925 0 : }
1926 :
1927 : /// Pick next tasks from the queue, and start as many of them as possible without violating
1928 : /// the ordering constraints.
1929 : ///
1930 : /// TODO: consider limiting the number of in-progress tasks, beyond what remote_storage does.
1931 : /// This can launch an unbounded number of queued tasks. `UploadQueue::next_ready()` also has
1932 : /// worst-case quadratic cost in the number of tasks, and may struggle beyond 10,000 tasks.
1933 16653 : fn launch_queued_tasks(self: &Arc<Self>, upload_queue: &mut UploadQueueInitialized) {
1934 26665 : while let Some((mut next_op, coalesced_ops)) = upload_queue.next_ready() {
1935 10012 : debug!("starting op: {next_op}");
1936 :
1937 : // Prepare upload.
1938 10012 : match &mut next_op {
1939 3484 : UploadOp::UploadLayer(layer, meta, mode) => {
1940 3484 : if upload_queue
1941 3484 : .recently_deleted
1942 3484 : .remove(&(layer.layer_desc().layer_name().clone(), meta.generation))
1943 0 : {
1944 0 : *mode = Some(OpType::FlushDeletion);
1945 0 : } else {
1946 3484 : *mode = Some(OpType::MayReorder)
1947 : }
1948 : }
1949 3025 : UploadOp::UploadMetadata { .. } => {}
1950 695 : UploadOp::Delete(Delete { layers }) => {
1951 1390 : for (name, meta) in layers {
1952 695 : upload_queue
1953 695 : .recently_deleted
1954 695 : .insert((name.clone(), meta.generation));
1955 695 : }
1956 : }
1957 2808 : UploadOp::Barrier(sender) => {
1958 2808 : sender.send_replace(());
1959 2808 : continue;
1960 : }
1961 0 : UploadOp::Shutdown => unreachable!("shutdown is intentionally never popped off"),
1962 : };
1963 :
1964 : // Assign unique ID to this task
1965 7204 : upload_queue.task_counter += 1;
1966 7204 : let upload_task_id = upload_queue.task_counter;
1967 7204 :
1968 7204 : // Add it to the in-progress map
1969 7204 : let task = Arc::new(UploadTask {
1970 7204 : task_id: upload_task_id,
1971 7204 : op: next_op,
1972 7204 : coalesced_ops,
1973 7204 : retries: AtomicU32::new(0),
1974 7204 : });
1975 7204 : upload_queue
1976 7204 : .inprogress_tasks
1977 7204 : .insert(task.task_id, Arc::clone(&task));
1978 7204 :
1979 7204 : // Spawn task to perform the task
1980 7204 : let self_rc = Arc::clone(self);
1981 7204 : let tenant_shard_id = self.tenant_shard_id;
1982 7204 : let timeline_id = self.timeline_id;
1983 7204 : task_mgr::spawn(
1984 7204 : &self.runtime,
1985 7204 : TaskKind::RemoteUploadTask,
1986 7204 : self.tenant_shard_id,
1987 7204 : Some(self.timeline_id),
1988 7204 : "remote upload",
1989 7062 : async move {
1990 7062 : self_rc.perform_upload_task(task).await;
1991 6779 : Ok(())
1992 6779 : }
1993 7204 : .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)),
1994 : );
1995 :
1996 : // Loop back to process next task
1997 : }
1998 16653 : }
1999 :
2000 : ///
2001 : /// Perform an upload task.
2002 : ///
2003 : /// The task is in the `inprogress_tasks` list. This function will try to
2004 : /// execute it, retrying forever. On successful completion, the task is
2005 : /// removed it from the `inprogress_tasks` list, and any next task(s) in the
2006 : /// queue that were waiting by the completion are launched.
2007 : ///
2008 : /// The task can be shut down, however. That leads to stopping the whole
2009 : /// queue.
2010 : ///
2011 7062 : async fn perform_upload_task(self: &Arc<Self>, task: Arc<UploadTask>) {
2012 7062 : let cancel = shutdown_token();
2013 : // Loop to retry until it completes.
2014 : loop {
2015 : // If we're requested to shut down, close up shop and exit.
2016 : //
2017 : // Note: We only check for the shutdown requests between retries, so
2018 : // if a shutdown request arrives while we're busy uploading, in the
2019 : // upload::upload:*() call below, we will wait not exit until it has
2020 : // finished. We probably could cancel the upload by simply dropping
2021 : // the Future, but we're not 100% sure if the remote storage library
2022 : // is cancellation safe, so we don't dare to do that. Hopefully, the
2023 : // upload finishes or times out soon enough.
2024 7062 : if cancel.is_cancelled() {
2025 0 : info!("upload task cancelled by shutdown request");
2026 0 : self.stop();
2027 0 : return;
2028 7062 : }
2029 7062 :
2030 7062 : // Assert that we don't modify a layer that's referenced by the current index.
2031 7062 : if cfg!(debug_assertions) {
2032 7062 : let modified = match &task.op {
2033 3366 : UploadOp::UploadLayer(layer, layer_metadata, _) => {
2034 3366 : vec![(layer.layer_desc().layer_name(), layer_metadata)]
2035 : }
2036 695 : UploadOp::Delete(delete) => {
2037 695 : delete.layers.iter().map(|(n, m)| (n.clone(), m)).collect()
2038 : }
2039 : // These don't modify layers.
2040 3001 : UploadOp::UploadMetadata { .. } => Vec::new(),
2041 0 : UploadOp::Barrier(_) => Vec::new(),
2042 0 : UploadOp::Shutdown => Vec::new(),
2043 : };
2044 7062 : if let Ok(queue) = self.upload_queue.lock().unwrap().initialized_mut() {
2045 11115 : for (ref name, metadata) in modified {
2046 4061 : debug_assert!(
2047 4061 : !queue.clean.0.references(name, metadata),
2048 8 : "layer {name} modified while referenced by index",
2049 : );
2050 : }
2051 0 : }
2052 0 : }
2053 :
2054 7054 : let upload_result: anyhow::Result<()> = match &task.op {
2055 3366 : UploadOp::UploadLayer(layer, layer_metadata, mode) => {
2056 : // TODO: check if this mechanism can be removed now that can_bypass() performs
2057 : // conflict checks during scheduling.
2058 3366 : if let Some(OpType::FlushDeletion) = mode {
2059 0 : if self.config.read().unwrap().block_deletions {
2060 : // Of course, this is not efficient... but usually the queue should be empty.
2061 0 : let mut queue_locked = self.upload_queue.lock().unwrap();
2062 0 : let mut detected = false;
2063 0 : if let Ok(queue) = queue_locked.initialized_mut() {
2064 0 : for list in queue.blocked_deletions.iter_mut() {
2065 0 : list.layers.retain(|(name, meta)| {
2066 0 : if name == &layer.layer_desc().layer_name()
2067 0 : && meta.generation == layer_metadata.generation
2068 : {
2069 0 : detected = true;
2070 0 : // remove the layer from deletion queue
2071 0 : false
2072 : } else {
2073 : // keep the layer
2074 0 : true
2075 : }
2076 0 : });
2077 0 : }
2078 0 : }
2079 0 : if detected {
2080 0 : info!(
2081 0 : "cancelled blocked deletion of layer {} at gen {:?}",
2082 0 : layer.layer_desc().layer_name(),
2083 : layer_metadata.generation
2084 : );
2085 0 : }
2086 : } else {
2087 : // TODO: we did not guarantee that upload task starts after deletion task, so there could be possibly race conditions
2088 : // that we still get the layer deleted. But this only happens if someone creates a layer immediately after it's deleted,
2089 : // which is not possible in the current system.
2090 0 : info!(
2091 0 : "waiting for deletion queue flush to complete before uploading layer {} at gen {:?}",
2092 0 : layer.layer_desc().layer_name(),
2093 : layer_metadata.generation
2094 : );
2095 : {
2096 : // We are going to flush, we can clean up the recently deleted list.
2097 0 : let mut queue_locked = self.upload_queue.lock().unwrap();
2098 0 : if let Ok(queue) = queue_locked.initialized_mut() {
2099 0 : queue.recently_deleted.clear();
2100 0 : }
2101 : }
2102 0 : if let Err(e) = self.deletion_queue_client.flush_execute().await {
2103 0 : warn!(
2104 0 : "failed to flush the deletion queue before uploading layer {} at gen {:?}, still proceeding to upload: {e:#} ",
2105 0 : layer.layer_desc().layer_name(),
2106 : layer_metadata.generation
2107 : );
2108 : } else {
2109 0 : info!(
2110 0 : "done flushing deletion queue before uploading layer {} at gen {:?}",
2111 0 : layer.layer_desc().layer_name(),
2112 : layer_metadata.generation
2113 : );
2114 : }
2115 : }
2116 3366 : }
2117 3366 : let local_path = layer.local_path();
2118 3366 :
2119 3366 : // We should only be uploading layers created by this `Tenant`'s lifetime, so
2120 3366 : // the metadata in the upload should always match our current generation.
2121 3366 : assert_eq!(layer_metadata.generation, self.generation);
2122 :
2123 3366 : let remote_path = remote_layer_path(
2124 3366 : &self.tenant_shard_id.tenant_id,
2125 3366 : &self.timeline_id,
2126 3366 : layer_metadata.shard,
2127 3366 : &layer.layer_desc().layer_name(),
2128 3366 : layer_metadata.generation,
2129 3366 : );
2130 3366 :
2131 3366 : upload::upload_timeline_layer(
2132 3366 : &self.storage_impl,
2133 3366 : local_path,
2134 3366 : &remote_path,
2135 3366 : layer_metadata.file_size,
2136 3366 : &self.cancel,
2137 3366 : )
2138 3366 : .measure_remote_op(
2139 3366 : RemoteOpFileKind::Layer,
2140 3366 : RemoteOpKind::Upload,
2141 3366 : Arc::clone(&self.metrics),
2142 3366 : )
2143 3366 : .await
2144 : }
2145 3001 : UploadOp::UploadMetadata { uploaded } => {
2146 3001 : let res = upload::upload_index_part(
2147 3001 : &self.storage_impl,
2148 3001 : &self.tenant_shard_id,
2149 3001 : &self.timeline_id,
2150 3001 : self.generation,
2151 3001 : uploaded,
2152 3001 : &self.cancel,
2153 3001 : )
2154 3001 : .measure_remote_op(
2155 3001 : RemoteOpFileKind::Index,
2156 3001 : RemoteOpKind::Upload,
2157 3001 : Arc::clone(&self.metrics),
2158 3001 : )
2159 3001 : .await;
2160 2985 : if res.is_ok() {
2161 2985 : self.update_remote_physical_size_gauge(Some(uploaded));
2162 2985 : let mention_having_future_layers = if cfg!(feature = "testing") {
2163 2985 : uploaded
2164 2985 : .layer_metadata
2165 2985 : .keys()
2166 35473 : .any(|x| x.is_in_future(uploaded.metadata.disk_consistent_lsn()))
2167 : } else {
2168 0 : false
2169 : };
2170 2985 : if mention_having_future_layers {
2171 : // find rationale near crate::tenant::timeline::init::cleanup_future_layer
2172 134 : tracing::info!(
2173 0 : disk_consistent_lsn = %uploaded.metadata.disk_consistent_lsn(),
2174 0 : "uploaded an index_part.json with future layers -- this is ok! if shutdown now, expect future layer cleanup"
2175 : );
2176 2851 : }
2177 0 : }
2178 2985 : res
2179 : }
2180 687 : UploadOp::Delete(delete) => {
2181 687 : if self.config.read().unwrap().block_deletions {
2182 0 : let mut queue_locked = self.upload_queue.lock().unwrap();
2183 0 : if let Ok(queue) = queue_locked.initialized_mut() {
2184 0 : queue.blocked_deletions.push(delete.clone());
2185 0 : }
2186 0 : Ok(())
2187 : } else {
2188 687 : pausable_failpoint!("before-delete-layer-pausable");
2189 687 : self.deletion_queue_client
2190 687 : .push_layers(
2191 687 : self.tenant_shard_id,
2192 687 : self.timeline_id,
2193 687 : self.generation,
2194 687 : delete.layers.clone(),
2195 687 : )
2196 687 : .map_err(|e| anyhow::anyhow!(e))
2197 : }
2198 : }
2199 0 : unexpected @ UploadOp::Barrier(_) | unexpected @ UploadOp::Shutdown => {
2200 : // unreachable. Barrier operations are handled synchronously in
2201 : // launch_queued_tasks
2202 0 : warn!("unexpected {unexpected:?} operation in perform_upload_task");
2203 0 : break;
2204 : }
2205 : };
2206 :
2207 0 : match upload_result {
2208 : Ok(()) => {
2209 6779 : break;
2210 : }
2211 0 : Err(e) if TimeoutOrCancel::caused_by_cancel(&e) => {
2212 0 : // loop around to do the proper stopping
2213 0 : continue;
2214 : }
2215 0 : Err(e) => {
2216 0 : let retries = task.retries.fetch_add(1, Ordering::SeqCst);
2217 0 :
2218 0 : // Uploads can fail due to rate limits (IAM, S3), spurious network problems,
2219 0 : // or other external reasons. Such issues are relatively regular, so log them
2220 0 : // at info level at first, and only WARN if the operation fails repeatedly.
2221 0 : //
2222 0 : // (See similar logic for downloads in `download::download_retry`)
2223 0 : if retries < FAILED_UPLOAD_WARN_THRESHOLD {
2224 0 : info!(
2225 0 : "failed to perform remote task {}, will retry (attempt {}): {:#}",
2226 0 : task.op, retries, e
2227 : );
2228 : } else {
2229 0 : warn!(
2230 0 : "failed to perform remote task {}, will retry (attempt {}): {:?}",
2231 0 : task.op, retries, e
2232 : );
2233 : }
2234 :
2235 : // sleep until it's time to retry, or we're cancelled
2236 0 : exponential_backoff(
2237 0 : retries,
2238 0 : DEFAULT_BASE_BACKOFF_SECONDS,
2239 0 : DEFAULT_MAX_BACKOFF_SECONDS,
2240 0 : &cancel,
2241 0 : )
2242 0 : .await;
2243 : }
2244 : }
2245 : }
2246 :
2247 6779 : let retries = task.retries.load(Ordering::SeqCst);
2248 6779 : if retries > 0 {
2249 0 : info!(
2250 0 : "remote task {} completed successfully after {} retries",
2251 0 : task.op, retries
2252 : );
2253 : } else {
2254 6779 : debug!("remote task {} completed successfully", task.op);
2255 : }
2256 :
2257 : // The task has completed successfully. Remove it from the in-progress list.
2258 6779 : let lsn_update = {
2259 6779 : let mut upload_queue_guard = self.upload_queue.lock().unwrap();
2260 6779 : let upload_queue = match upload_queue_guard.deref_mut() {
2261 0 : UploadQueue::Uninitialized => panic!(
2262 0 : "callers are responsible for ensuring this is only called on an initialized queue"
2263 0 : ),
2264 0 : UploadQueue::Stopped(_stopped) => None,
2265 6779 : UploadQueue::Initialized(qi) => Some(qi),
2266 : };
2267 :
2268 6779 : let upload_queue = match upload_queue {
2269 6779 : Some(upload_queue) => upload_queue,
2270 : None => {
2271 0 : info!("another concurrent task already stopped the queue");
2272 0 : return;
2273 : }
2274 : };
2275 :
2276 6779 : upload_queue.inprogress_tasks.remove(&task.task_id);
2277 :
2278 6779 : let lsn_update = match task.op {
2279 3107 : UploadOp::UploadLayer(_, _, _) => None,
2280 2985 : UploadOp::UploadMetadata { ref uploaded } => {
2281 2985 : // the task id is reused as a monotonicity check for storing the "clean"
2282 2985 : // IndexPart.
2283 2985 : let last_updater = upload_queue.clean.1;
2284 2985 : let is_later = last_updater.is_some_and(|task_id| task_id < task.task_id);
2285 2985 : let monotone = is_later || last_updater.is_none();
2286 :
2287 2985 : assert!(
2288 2985 : monotone,
2289 0 : "no two index uploads should be completing at the same time, prev={last_updater:?}, task.task_id={}",
2290 0 : task.task_id
2291 : );
2292 :
2293 : // not taking ownership is wasteful
2294 2985 : upload_queue.clean.0.clone_from(uploaded);
2295 2985 : upload_queue.clean.1 = Some(task.task_id);
2296 2985 :
2297 2985 : let lsn = upload_queue.clean.0.metadata.disk_consistent_lsn();
2298 2985 : self.metrics
2299 2985 : .projected_remote_consistent_lsn_gauge
2300 2985 : .set(lsn.0);
2301 2985 :
2302 2985 : if self.generation.is_none() {
2303 : // Legacy mode: skip validating generation
2304 0 : upload_queue.visible_remote_consistent_lsn.store(lsn);
2305 0 : None
2306 2985 : } else if self
2307 2985 : .config
2308 2985 : .read()
2309 2985 : .unwrap()
2310 2985 : .process_remote_consistent_lsn_updates
2311 : {
2312 2985 : Some((lsn, upload_queue.visible_remote_consistent_lsn.clone()))
2313 : } else {
2314 : // Our config disables remote_consistent_lsn updates: drop it.
2315 0 : None
2316 : }
2317 : }
2318 687 : UploadOp::Delete(_) => None,
2319 0 : UploadOp::Barrier(..) | UploadOp::Shutdown => unreachable!(),
2320 : };
2321 :
2322 : // Launch any queued tasks that were unblocked by this one.
2323 6779 : self.launch_queued_tasks(upload_queue);
2324 6779 : lsn_update
2325 : };
2326 :
2327 6779 : if let Some((lsn, slot)) = lsn_update {
2328 : // Updates to the remote_consistent_lsn we advertise to pageservers
2329 : // are all routed through the DeletionQueue, to enforce important
2330 : // data safety guarantees (see docs/rfcs/025-generation-numbers.md)
2331 2985 : self.deletion_queue_client
2332 2985 : .update_remote_consistent_lsn(
2333 2985 : self.tenant_shard_id,
2334 2985 : self.timeline_id,
2335 2985 : self.generation,
2336 2985 : lsn,
2337 2985 : slot,
2338 2985 : )
2339 2985 : .await;
2340 3794 : }
2341 :
2342 6779 : self.metric_end(&task.op);
2343 6779 : for coalesced_op in &task.coalesced_ops {
2344 8 : self.metric_end(coalesced_op);
2345 8 : }
2346 6779 : }
2347 :
2348 14405 : fn metric_impl(
2349 14405 : &self,
2350 14405 : op: &UploadOp,
2351 14405 : ) -> Option<(
2352 14405 : RemoteOpFileKind,
2353 14405 : RemoteOpKind,
2354 14405 : RemoteTimelineClientMetricsCallTrackSize,
2355 14405 : )> {
2356 : use RemoteTimelineClientMetricsCallTrackSize::DontTrackSize;
2357 14405 : let res = match op {
2358 6591 : UploadOp::UploadLayer(_, m, _) => (
2359 6591 : RemoteOpFileKind::Layer,
2360 6591 : RemoteOpKind::Upload,
2361 6591 : RemoteTimelineClientMetricsCallTrackSize::Bytes(m.file_size),
2362 6591 : ),
2363 6085 : UploadOp::UploadMetadata { .. } => (
2364 6085 : RemoteOpFileKind::Index,
2365 6085 : RemoteOpKind::Upload,
2366 6085 : DontTrackSize {
2367 6085 : reason: "metadata uploads are tiny",
2368 6085 : },
2369 6085 : ),
2370 1713 : UploadOp::Delete(_delete) => (
2371 1713 : RemoteOpFileKind::Layer,
2372 1713 : RemoteOpKind::Delete,
2373 1713 : DontTrackSize {
2374 1713 : reason: "should we track deletes? positive or negative sign?",
2375 1713 : },
2376 1713 : ),
2377 : UploadOp::Barrier(..) | UploadOp::Shutdown => {
2378 : // we do not account these
2379 16 : return None;
2380 : }
2381 : };
2382 14389 : Some(res)
2383 14405 : }
2384 :
2385 7602 : fn metric_begin(&self, op: &UploadOp) {
2386 7602 : let (file_kind, op_kind, track_bytes) = match self.metric_impl(op) {
2387 7602 : Some(x) => x,
2388 0 : None => return,
2389 : };
2390 7602 : let guard = self.metrics.call_begin(&file_kind, &op_kind, track_bytes);
2391 7602 : guard.will_decrement_manually(); // in metric_end(), see right below
2392 7602 : }
2393 :
2394 6803 : fn metric_end(&self, op: &UploadOp) {
2395 6803 : let (file_kind, op_kind, track_bytes) = match self.metric_impl(op) {
2396 6787 : Some(x) => x,
2397 16 : None => return,
2398 : };
2399 6787 : self.metrics.call_end(&file_kind, &op_kind, track_bytes);
2400 6803 : }
2401 :
2402 : /// Close the upload queue for new operations and cancel queued operations.
2403 : ///
2404 : /// Use [`RemoteTimelineClient::shutdown`] for graceful stop.
2405 : ///
2406 : /// In-progress operations will still be running after this function returns.
2407 : /// Use `task_mgr::shutdown_tasks(Some(TaskKind::RemoteUploadTask), Some(self.tenant_shard_id), Some(timeline_id))`
2408 : /// to wait for them to complete, after calling this function.
2409 36 : pub(crate) fn stop(&self) {
2410 36 : // Whichever *task* for this RemoteTimelineClient grabs the mutex first will transition the queue
2411 36 : // into stopped state, thereby dropping all off the queued *ops* which haven't become *tasks* yet.
2412 36 : // The other *tasks* will come here and observe an already shut down queue and hence simply wrap up their business.
2413 36 : let mut guard = self.upload_queue.lock().unwrap();
2414 36 : self.stop_impl(&mut guard);
2415 36 : }
2416 :
2417 36 : fn stop_impl(&self, guard: &mut std::sync::MutexGuard<UploadQueue>) {
2418 36 : match &mut **guard {
2419 : UploadQueue::Uninitialized => {
2420 0 : info!("UploadQueue is in state Uninitialized, nothing to do");
2421 0 : **guard = UploadQueue::Stopped(UploadQueueStopped::Uninitialized);
2422 : }
2423 : UploadQueue::Stopped(_) => {
2424 : // nothing to do
2425 16 : info!("another concurrent task already shut down the queue");
2426 : }
2427 20 : UploadQueue::Initialized(initialized) => {
2428 20 : info!("shutting down upload queue");
2429 :
2430 : // Replace the queue with the Stopped state, taking ownership of the old
2431 : // Initialized queue. We will do some checks on it, and then drop it.
2432 20 : let qi = {
2433 : // Here we preserve working version of the upload queue for possible use during deletions.
2434 : // In-place replace of Initialized to Stopped can be done with the help of https://github.com/Sgeo/take_mut
2435 : // but for this use case it doesnt really makes sense to bring unsafe code only for this usage point.
2436 : // Deletion is not really perf sensitive so there shouldnt be any problems with cloning a fraction of it.
2437 20 : let upload_queue_for_deletion = UploadQueueInitialized {
2438 20 : inprogress_limit: initialized.inprogress_limit,
2439 20 : task_counter: 0,
2440 20 : dirty: initialized.dirty.clone(),
2441 20 : clean: initialized.clean.clone(),
2442 20 : latest_files_changes_since_metadata_upload_scheduled: 0,
2443 20 : visible_remote_consistent_lsn: initialized
2444 20 : .visible_remote_consistent_lsn
2445 20 : .clone(),
2446 20 : inprogress_tasks: HashMap::default(),
2447 20 : queued_operations: VecDeque::default(),
2448 20 : #[cfg(feature = "testing")]
2449 20 : dangling_files: HashMap::default(),
2450 20 : blocked_deletions: Vec::new(),
2451 20 : shutting_down: false,
2452 20 : shutdown_ready: Arc::new(tokio::sync::Semaphore::new(0)),
2453 20 : recently_deleted: HashSet::new(),
2454 20 : };
2455 20 :
2456 20 : let upload_queue = std::mem::replace(
2457 20 : &mut **guard,
2458 20 : UploadQueue::Stopped(UploadQueueStopped::Deletable(
2459 20 : UploadQueueStoppedDeletable {
2460 20 : upload_queue_for_deletion,
2461 20 : deleted_at: SetDeletedFlagProgress::NotRunning,
2462 20 : },
2463 20 : )),
2464 20 : );
2465 20 : if let UploadQueue::Initialized(qi) = upload_queue {
2466 20 : qi
2467 : } else {
2468 0 : unreachable!("we checked in the match above that it is Initialized");
2469 : }
2470 : };
2471 :
2472 : // We don't need to do anything here for in-progress tasks. They will finish
2473 : // on their own, decrement the unfinished-task counter themselves, and observe
2474 : // that the queue is Stopped.
2475 20 : drop(qi.inprogress_tasks);
2476 :
2477 : // Tear down queued ops
2478 20 : for op in qi.queued_operations.into_iter() {
2479 16 : self.metric_end(&op);
2480 16 : // Dropping UploadOp::Barrier() here will make wait_completion() return with an Err()
2481 16 : // which is exactly what we want to happen.
2482 16 : drop(op);
2483 16 : }
2484 : }
2485 : }
2486 36 : }
2487 :
2488 : /// Returns an accessor which will hold the UploadQueue mutex for accessing the upload queue
2489 : /// externally to RemoteTimelineClient.
2490 0 : pub(crate) fn initialized_upload_queue(
2491 0 : &self,
2492 0 : ) -> Result<UploadQueueAccessor<'_>, NotInitialized> {
2493 0 : let mut inner = self.upload_queue.lock().unwrap();
2494 0 : inner.initialized_mut()?;
2495 0 : Ok(UploadQueueAccessor { inner })
2496 0 : }
2497 :
2498 16 : pub(crate) fn no_pending_work(&self) -> bool {
2499 16 : let inner = self.upload_queue.lock().unwrap();
2500 16 : match &*inner {
2501 : UploadQueue::Uninitialized
2502 0 : | UploadQueue::Stopped(UploadQueueStopped::Uninitialized) => true,
2503 16 : UploadQueue::Stopped(UploadQueueStopped::Deletable(x)) => {
2504 16 : x.upload_queue_for_deletion.no_pending_work()
2505 : }
2506 0 : UploadQueue::Initialized(x) => x.no_pending_work(),
2507 : }
2508 16 : }
2509 :
2510 : /// 'foreign' in the sense that it does not belong to this tenant shard. This method
2511 : /// is used during GC for other shards to get the index of shard zero.
2512 0 : pub(crate) async fn download_foreign_index(
2513 0 : &self,
2514 0 : shard_number: ShardNumber,
2515 0 : cancel: &CancellationToken,
2516 0 : ) -> Result<(IndexPart, Generation, std::time::SystemTime), DownloadError> {
2517 0 : let foreign_shard_id = TenantShardId {
2518 0 : shard_number,
2519 0 : shard_count: self.tenant_shard_id.shard_count,
2520 0 : tenant_id: self.tenant_shard_id.tenant_id,
2521 0 : };
2522 0 : download_index_part(
2523 0 : &self.storage_impl,
2524 0 : &foreign_shard_id,
2525 0 : &self.timeline_id,
2526 0 : Generation::MAX,
2527 0 : cancel,
2528 0 : )
2529 0 : .await
2530 0 : }
2531 : }
2532 :
2533 : pub(crate) struct UploadQueueAccessor<'a> {
2534 : inner: std::sync::MutexGuard<'a, UploadQueue>,
2535 : }
2536 :
2537 : impl UploadQueueAccessor<'_> {
2538 0 : pub(crate) fn latest_uploaded_index_part(&self) -> &IndexPart {
2539 0 : match &*self.inner {
2540 0 : UploadQueue::Initialized(x) => &x.clean.0,
2541 : UploadQueue::Uninitialized | UploadQueue::Stopped(_) => {
2542 0 : unreachable!("checked before constructing")
2543 : }
2544 : }
2545 0 : }
2546 : }
2547 :
2548 0 : pub fn remote_tenant_path(tenant_shard_id: &TenantShardId) -> RemotePath {
2549 0 : let path = format!("tenants/{tenant_shard_id}");
2550 0 : RemotePath::from_string(&path).expect("Failed to construct path")
2551 0 : }
2552 :
2553 1360 : pub fn remote_tenant_manifest_path(
2554 1360 : tenant_shard_id: &TenantShardId,
2555 1360 : generation: Generation,
2556 1360 : ) -> RemotePath {
2557 1360 : let path = format!(
2558 1360 : "tenants/{tenant_shard_id}/tenant-manifest{}.json",
2559 1360 : generation.get_suffix()
2560 1360 : );
2561 1360 : RemotePath::from_string(&path).expect("Failed to construct path")
2562 1360 : }
2563 :
2564 : /// Prefix to all generations' manifest objects in a tenant shard
2565 452 : pub fn remote_tenant_manifest_prefix(tenant_shard_id: &TenantShardId) -> RemotePath {
2566 452 : let path = format!("tenants/{tenant_shard_id}/tenant-manifest",);
2567 452 : RemotePath::from_string(&path).expect("Failed to construct path")
2568 452 : }
2569 :
2570 512 : pub fn remote_timelines_path(tenant_shard_id: &TenantShardId) -> RemotePath {
2571 512 : let path = format!("tenants/{tenant_shard_id}/{TIMELINES_SEGMENT_NAME}");
2572 512 : RemotePath::from_string(&path).expect("Failed to construct path")
2573 512 : }
2574 :
2575 0 : fn remote_timelines_path_unsharded(tenant_id: &TenantId) -> RemotePath {
2576 0 : let path = format!("tenants/{tenant_id}/{TIMELINES_SEGMENT_NAME}");
2577 0 : RemotePath::from_string(&path).expect("Failed to construct path")
2578 0 : }
2579 :
2580 60 : pub fn remote_timeline_path(
2581 60 : tenant_shard_id: &TenantShardId,
2582 60 : timeline_id: &TimelineId,
2583 60 : ) -> RemotePath {
2584 60 : remote_timelines_path(tenant_shard_id).join(Utf8Path::new(&timeline_id.to_string()))
2585 60 : }
2586 :
2587 : /// Obtains the path of the given Layer in the remote
2588 : ///
2589 : /// Note that the shard component of a remote layer path is _not_ always the same
2590 : /// as in the TenantShardId of the caller: tenants may reference layers from a different
2591 : /// ShardIndex. Use the ShardIndex from the layer's metadata.
2592 4051 : pub fn remote_layer_path(
2593 4051 : tenant_id: &TenantId,
2594 4051 : timeline_id: &TimelineId,
2595 4051 : shard: ShardIndex,
2596 4051 : layer_file_name: &LayerName,
2597 4051 : generation: Generation,
2598 4051 : ) -> RemotePath {
2599 4051 : // Generation-aware key format
2600 4051 : let path = format!(
2601 4051 : "tenants/{tenant_id}{0}/{TIMELINES_SEGMENT_NAME}/{timeline_id}/{1}{2}",
2602 4051 : shard.get_suffix(),
2603 4051 : layer_file_name,
2604 4051 : generation.get_suffix()
2605 4051 : );
2606 4051 :
2607 4051 : RemotePath::from_string(&path).expect("Failed to construct path")
2608 4051 : }
2609 :
2610 : /// Returns true if a and b have the same layer path within a tenant/timeline. This is essentially
2611 : /// remote_layer_path(a) == remote_layer_path(b) without the string allocations.
2612 : ///
2613 : /// TODO: there should be a variant of LayerName for the physical path that contains information
2614 : /// about the shard and generation, such that this could be replaced by a simple comparison.
2615 2351785 : pub fn is_same_remote_layer_path(
2616 2351785 : aname: &LayerName,
2617 2351785 : ameta: &LayerFileMetadata,
2618 2351785 : bname: &LayerName,
2619 2351785 : bmeta: &LayerFileMetadata,
2620 2351785 : ) -> bool {
2621 2351785 : // NB: don't assert remote_layer_path(a) == remote_layer_path(b); too expensive even for debug.
2622 2351785 : aname == bname && ameta.shard == bmeta.shard && ameta.generation == bmeta.generation
2623 2351785 : }
2624 :
2625 8 : pub fn remote_initdb_archive_path(tenant_id: &TenantId, timeline_id: &TimelineId) -> RemotePath {
2626 8 : RemotePath::from_string(&format!(
2627 8 : "tenants/{tenant_id}/{TIMELINES_SEGMENT_NAME}/{timeline_id}/{INITDB_PATH}"
2628 8 : ))
2629 8 : .expect("Failed to construct path")
2630 8 : }
2631 :
2632 4 : pub fn remote_initdb_preserved_archive_path(
2633 4 : tenant_id: &TenantId,
2634 4 : timeline_id: &TimelineId,
2635 4 : ) -> RemotePath {
2636 4 : RemotePath::from_string(&format!(
2637 4 : "tenants/{tenant_id}/{TIMELINES_SEGMENT_NAME}/{timeline_id}/{INITDB_PRESERVED_PATH}"
2638 4 : ))
2639 4 : .expect("Failed to construct path")
2640 4 : }
2641 :
2642 3133 : pub fn remote_index_path(
2643 3133 : tenant_shard_id: &TenantShardId,
2644 3133 : timeline_id: &TimelineId,
2645 3133 : generation: Generation,
2646 3133 : ) -> RemotePath {
2647 3133 : RemotePath::from_string(&format!(
2648 3133 : "tenants/{tenant_shard_id}/{TIMELINES_SEGMENT_NAME}/{timeline_id}/{0}{1}",
2649 3133 : IndexPart::FILE_NAME,
2650 3133 : generation.get_suffix()
2651 3133 : ))
2652 3133 : .expect("Failed to construct path")
2653 3133 : }
2654 :
2655 0 : pub(crate) fn remote_heatmap_path(tenant_shard_id: &TenantShardId) -> RemotePath {
2656 0 : RemotePath::from_string(&format!(
2657 0 : "tenants/{tenant_shard_id}/{TENANT_HEATMAP_BASENAME}"
2658 0 : ))
2659 0 : .expect("Failed to construct path")
2660 0 : }
2661 :
2662 : /// Given the key of an index, parse out the generation part of the name
2663 36 : pub fn parse_remote_index_path(path: RemotePath) -> Option<Generation> {
2664 36 : let file_name = match path.get_path().file_name() {
2665 36 : Some(f) => f,
2666 : None => {
2667 : // Unexpected: we should be seeing index_part.json paths only
2668 0 : tracing::warn!("Malformed index key {}", path);
2669 0 : return None;
2670 : }
2671 : };
2672 :
2673 36 : match file_name.split_once('-') {
2674 24 : Some((_, gen_suffix)) => Generation::parse_suffix(gen_suffix),
2675 12 : None => None,
2676 : }
2677 36 : }
2678 :
2679 : /// Given the key of a tenant manifest, parse out the generation number
2680 0 : pub fn parse_remote_tenant_manifest_path(path: RemotePath) -> Option<Generation> {
2681 : static RE: OnceLock<Regex> = OnceLock::new();
2682 0 : let re = RE.get_or_init(|| Regex::new(r".*tenant-manifest-([0-9a-f]{8}).json").unwrap());
2683 0 : re.captures(path.get_path().as_str())
2684 0 : .and_then(|c| c.get(1))
2685 0 : .and_then(|m| Generation::parse_suffix(m.as_str()))
2686 0 : }
2687 :
2688 : #[cfg(test)]
2689 : mod tests {
2690 : use std::collections::HashSet;
2691 :
2692 : use super::*;
2693 : use crate::DEFAULT_PG_VERSION;
2694 : use crate::context::RequestContext;
2695 : use crate::tenant::config::AttachmentMode;
2696 : use crate::tenant::harness::{TIMELINE_ID, TenantHarness};
2697 : use crate::tenant::storage_layer::layer::local_layer_path;
2698 : use crate::tenant::{Tenant, Timeline};
2699 :
2700 16 : pub(super) fn dummy_contents(name: &str) -> Vec<u8> {
2701 16 : format!("contents for {name}").into()
2702 16 : }
2703 :
2704 4 : pub(super) fn dummy_metadata(disk_consistent_lsn: Lsn) -> TimelineMetadata {
2705 4 : let metadata = TimelineMetadata::new(
2706 4 : disk_consistent_lsn,
2707 4 : None,
2708 4 : None,
2709 4 : Lsn(0),
2710 4 : Lsn(0),
2711 4 : Lsn(0),
2712 4 : // Any version will do
2713 4 : // but it should be consistent with the one in the tests
2714 4 : crate::DEFAULT_PG_VERSION,
2715 4 : );
2716 4 :
2717 4 : // go through serialize + deserialize to fix the header, including checksum
2718 4 : TimelineMetadata::from_bytes(&metadata.to_bytes().unwrap()).unwrap()
2719 4 : }
2720 :
2721 4 : fn assert_file_list(a: &HashSet<LayerName>, b: &[&str]) {
2722 12 : let mut avec: Vec<String> = a.iter().map(|x| x.to_string()).collect();
2723 4 : avec.sort();
2724 4 :
2725 4 : let mut bvec = b.to_vec();
2726 4 : bvec.sort_unstable();
2727 4 :
2728 4 : assert_eq!(avec, bvec);
2729 4 : }
2730 :
2731 8 : fn assert_remote_files(expected: &[&str], remote_path: &Utf8Path, generation: Generation) {
2732 8 : let mut expected: Vec<String> = expected
2733 8 : .iter()
2734 32 : .map(|x| format!("{}{}", x, generation.get_suffix()))
2735 8 : .collect();
2736 8 : expected.sort();
2737 8 :
2738 8 : let mut found: Vec<String> = Vec::new();
2739 32 : for entry in std::fs::read_dir(remote_path).unwrap().flatten() {
2740 32 : let entry_name = entry.file_name();
2741 32 : let fname = entry_name.to_str().unwrap();
2742 32 : found.push(String::from(fname));
2743 32 : }
2744 8 : found.sort();
2745 8 :
2746 8 : assert_eq!(found, expected);
2747 8 : }
2748 :
2749 : struct TestSetup {
2750 : harness: TenantHarness,
2751 : tenant: Arc<Tenant>,
2752 : timeline: Arc<Timeline>,
2753 : tenant_ctx: RequestContext,
2754 : }
2755 :
2756 : impl TestSetup {
2757 16 : async fn new(test_name: &str) -> anyhow::Result<Self> {
2758 16 : let test_name = Box::leak(Box::new(format!("remote_timeline_client__{test_name}")));
2759 16 : let harness = TenantHarness::create(test_name).await?;
2760 16 : let (tenant, ctx) = harness.load().await;
2761 :
2762 16 : let timeline = tenant
2763 16 : .create_test_timeline(TIMELINE_ID, Lsn(8), DEFAULT_PG_VERSION, &ctx)
2764 16 : .await?;
2765 :
2766 16 : Ok(Self {
2767 16 : harness,
2768 16 : tenant,
2769 16 : timeline,
2770 16 : tenant_ctx: ctx,
2771 16 : })
2772 16 : }
2773 :
2774 : /// Construct a RemoteTimelineClient in an arbitrary generation
2775 20 : fn build_client(&self, generation: Generation) -> Arc<RemoteTimelineClient> {
2776 20 : let location_conf = AttachedLocationConfig {
2777 20 : generation,
2778 20 : attach_mode: AttachmentMode::Single,
2779 20 : };
2780 20 : Arc::new(RemoteTimelineClient {
2781 20 : conf: self.harness.conf,
2782 20 : runtime: tokio::runtime::Handle::current(),
2783 20 : tenant_shard_id: self.harness.tenant_shard_id,
2784 20 : timeline_id: TIMELINE_ID,
2785 20 : generation,
2786 20 : storage_impl: self.harness.remote_storage.clone(),
2787 20 : deletion_queue_client: self.harness.deletion_queue.new_client(),
2788 20 : upload_queue: Mutex::new(UploadQueue::Uninitialized),
2789 20 : metrics: Arc::new(RemoteTimelineClientMetrics::new(
2790 20 : &self.harness.tenant_shard_id,
2791 20 : &TIMELINE_ID,
2792 20 : )),
2793 20 : config: std::sync::RwLock::new(RemoteTimelineClientConfig::from(&location_conf)),
2794 20 : cancel: CancellationToken::new(),
2795 20 : })
2796 20 : }
2797 :
2798 : /// A tracing::Span that satisfies remote_timeline_client methods that assert tenant_id
2799 : /// and timeline_id are present.
2800 12 : fn span(&self) -> tracing::Span {
2801 12 : tracing::info_span!(
2802 : "test",
2803 : tenant_id = %self.harness.tenant_shard_id.tenant_id,
2804 0 : shard_id = %self.harness.tenant_shard_id.shard_slug(),
2805 : timeline_id = %TIMELINE_ID
2806 : )
2807 12 : }
2808 : }
2809 :
2810 : // Test scheduling
2811 : #[tokio::test]
2812 4 : async fn upload_scheduling() {
2813 4 : // Test outline:
2814 4 : //
2815 4 : // Schedule upload of a bunch of layers. Check that they are started immediately, not queued
2816 4 : // Schedule upload of index. Check that it is queued
2817 4 : // let the layer file uploads finish. Check that the index-upload is now started
2818 4 : // let the index-upload finish.
2819 4 : //
2820 4 : // Download back the index.json. Check that the list of files is correct
2821 4 : //
2822 4 : // Schedule upload. Schedule deletion. Check that the deletion is queued
2823 4 : // let upload finish. Check that deletion is now started
2824 4 : // Schedule another deletion. Check that it's launched immediately.
2825 4 : // Schedule index upload. Check that it's queued
2826 4 :
2827 4 : let test_setup = TestSetup::new("upload_scheduling").await.unwrap();
2828 4 : let span = test_setup.span();
2829 4 : let _guard = span.enter();
2830 4 :
2831 4 : let TestSetup {
2832 4 : harness,
2833 4 : tenant: _tenant,
2834 4 : timeline,
2835 4 : tenant_ctx: _tenant_ctx,
2836 4 : } = test_setup;
2837 4 :
2838 4 : let client = &timeline.remote_client;
2839 4 :
2840 4 : // Download back the index.json, and check that the list of files is correct
2841 4 : let initial_index_part = match client
2842 4 : .download_index_file(&CancellationToken::new())
2843 4 : .await
2844 4 : .unwrap()
2845 4 : {
2846 4 : MaybeDeletedIndexPart::IndexPart(index_part) => index_part,
2847 4 : MaybeDeletedIndexPart::Deleted(_) => panic!("unexpectedly got deleted index part"),
2848 4 : };
2849 4 : let initial_layers = initial_index_part
2850 4 : .layer_metadata
2851 4 : .keys()
2852 4 : .map(|f| f.to_owned())
2853 4 : .collect::<HashSet<LayerName>>();
2854 4 : let initial_layer = {
2855 4 : assert!(initial_layers.len() == 1);
2856 4 : initial_layers.into_iter().next().unwrap()
2857 4 : };
2858 4 :
2859 4 : let timeline_path = harness.timeline_path(&TIMELINE_ID);
2860 4 :
2861 4 : println!("workdir: {}", harness.conf.workdir);
2862 4 :
2863 4 : let remote_timeline_dir = harness
2864 4 : .remote_fs_dir
2865 4 : .join(timeline_path.strip_prefix(&harness.conf.workdir).unwrap());
2866 4 : println!("remote_timeline_dir: {remote_timeline_dir}");
2867 4 :
2868 4 : let generation = harness.generation;
2869 4 : let shard = harness.shard;
2870 4 :
2871 4 : // Create a couple of dummy files, schedule upload for them
2872 4 :
2873 4 : let layers = [
2874 4 : ("000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__00000000016B59D8-00000000016B5A51".parse().unwrap(), dummy_contents("foo")),
2875 4 : ("000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__00000000016B59D9-00000000016B5A52".parse().unwrap(), dummy_contents("bar")),
2876 4 : ("000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__00000000016B59DA-00000000016B5A53".parse().unwrap(), dummy_contents("baz"))
2877 4 : ]
2878 4 : .into_iter()
2879 12 : .map(|(name, contents): (LayerName, Vec<u8>)| {
2880 12 :
2881 12 : let local_path = local_layer_path(
2882 12 : harness.conf,
2883 12 : &timeline.tenant_shard_id,
2884 12 : &timeline.timeline_id,
2885 12 : &name,
2886 12 : &generation,
2887 12 : );
2888 12 : std::fs::write(&local_path, &contents).unwrap();
2889 12 :
2890 12 : Layer::for_resident(
2891 12 : harness.conf,
2892 12 : &timeline,
2893 12 : local_path,
2894 12 : name,
2895 12 : LayerFileMetadata::new(contents.len() as u64, generation, shard),
2896 12 : )
2897 12 : }).collect::<Vec<_>>();
2898 4 :
2899 4 : client
2900 4 : .schedule_layer_file_upload(layers[0].clone())
2901 4 : .unwrap();
2902 4 : client
2903 4 : .schedule_layer_file_upload(layers[1].clone())
2904 4 : .unwrap();
2905 4 :
2906 4 : // Check that they are started immediately, not queued
2907 4 : //
2908 4 : // this works because we running within block_on, so any futures are now queued up until
2909 4 : // our next await point.
2910 4 : {
2911 4 : let mut guard = client.upload_queue.lock().unwrap();
2912 4 : let upload_queue = guard.initialized_mut().unwrap();
2913 4 : assert!(upload_queue.queued_operations.is_empty());
2914 4 : assert_eq!(upload_queue.inprogress_tasks.len(), 2);
2915 4 : assert_eq!(upload_queue.num_inprogress_layer_uploads(), 2);
2916 4 :
2917 4 : // also check that `latest_file_changes` was updated
2918 4 : assert!(upload_queue.latest_files_changes_since_metadata_upload_scheduled == 2);
2919 4 : }
2920 4 :
2921 4 : // Schedule upload of index. Check that it is queued
2922 4 : let metadata = dummy_metadata(Lsn(0x20));
2923 4 : client
2924 4 : .schedule_index_upload_for_full_metadata_update(&metadata)
2925 4 : .unwrap();
2926 4 : {
2927 4 : let mut guard = client.upload_queue.lock().unwrap();
2928 4 : let upload_queue = guard.initialized_mut().unwrap();
2929 4 : assert!(upload_queue.queued_operations.len() == 1);
2930 4 : assert!(upload_queue.latest_files_changes_since_metadata_upload_scheduled == 0);
2931 4 : }
2932 4 :
2933 4 : // Wait for the uploads to finish
2934 4 : client.wait_completion().await.unwrap();
2935 4 : {
2936 4 : let mut guard = client.upload_queue.lock().unwrap();
2937 4 : let upload_queue = guard.initialized_mut().unwrap();
2938 4 :
2939 4 : assert!(upload_queue.queued_operations.is_empty());
2940 4 : assert!(upload_queue.inprogress_tasks.is_empty());
2941 4 : }
2942 4 :
2943 4 : // Download back the index.json, and check that the list of files is correct
2944 4 : let index_part = match client
2945 4 : .download_index_file(&CancellationToken::new())
2946 4 : .await
2947 4 : .unwrap()
2948 4 : {
2949 4 : MaybeDeletedIndexPart::IndexPart(index_part) => index_part,
2950 4 : MaybeDeletedIndexPart::Deleted(_) => panic!("unexpectedly got deleted index part"),
2951 4 : };
2952 4 :
2953 4 : assert_file_list(
2954 4 : &index_part
2955 4 : .layer_metadata
2956 4 : .keys()
2957 12 : .map(|f| f.to_owned())
2958 4 : .collect(),
2959 4 : &[
2960 4 : &initial_layer.to_string(),
2961 4 : &layers[0].layer_desc().layer_name().to_string(),
2962 4 : &layers[1].layer_desc().layer_name().to_string(),
2963 4 : ],
2964 4 : );
2965 4 : assert_eq!(index_part.metadata, metadata);
2966 4 :
2967 4 : // Schedule upload and then a deletion. Check that the deletion is queued
2968 4 : client
2969 4 : .schedule_layer_file_upload(layers[2].clone())
2970 4 : .unwrap();
2971 4 :
2972 4 : // this is no longer consistent with how deletion works with Layer::drop, but in this test
2973 4 : // keep using schedule_layer_file_deletion because we don't have a way to wait for the
2974 4 : // spawn_blocking started by the drop.
2975 4 : client
2976 4 : .schedule_layer_file_deletion(&[layers[0].layer_desc().layer_name()])
2977 4 : .unwrap();
2978 4 : {
2979 4 : let mut guard = client.upload_queue.lock().unwrap();
2980 4 : let upload_queue = guard.initialized_mut().unwrap();
2981 4 :
2982 4 : // Deletion schedules upload of the index file, and the file deletion itself
2983 4 : assert_eq!(upload_queue.queued_operations.len(), 2);
2984 4 : assert_eq!(upload_queue.inprogress_tasks.len(), 1);
2985 4 : assert_eq!(upload_queue.num_inprogress_layer_uploads(), 1);
2986 4 : assert_eq!(upload_queue.num_inprogress_deletions(), 0);
2987 4 : assert_eq!(
2988 4 : upload_queue.latest_files_changes_since_metadata_upload_scheduled,
2989 4 : 0
2990 4 : );
2991 4 : }
2992 4 : assert_remote_files(
2993 4 : &[
2994 4 : &initial_layer.to_string(),
2995 4 : &layers[0].layer_desc().layer_name().to_string(),
2996 4 : &layers[1].layer_desc().layer_name().to_string(),
2997 4 : "index_part.json",
2998 4 : ],
2999 4 : &remote_timeline_dir,
3000 4 : generation,
3001 4 : );
3002 4 :
3003 4 : // Finish them
3004 4 : client.wait_completion().await.unwrap();
3005 4 : harness.deletion_queue.pump().await;
3006 4 :
3007 4 : assert_remote_files(
3008 4 : &[
3009 4 : &initial_layer.to_string(),
3010 4 : &layers[1].layer_desc().layer_name().to_string(),
3011 4 : &layers[2].layer_desc().layer_name().to_string(),
3012 4 : "index_part.json",
3013 4 : ],
3014 4 : &remote_timeline_dir,
3015 4 : generation,
3016 4 : );
3017 4 : }
3018 :
3019 : #[tokio::test]
3020 4 : async fn bytes_unfinished_gauge_for_layer_file_uploads() {
3021 4 : // Setup
3022 4 :
3023 4 : let TestSetup {
3024 4 : harness,
3025 4 : tenant: _tenant,
3026 4 : timeline,
3027 4 : ..
3028 4 : } = TestSetup::new("metrics").await.unwrap();
3029 4 : let client = &timeline.remote_client;
3030 4 :
3031 4 : let layer_file_name_1: LayerName = "000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__00000000016B59D8-00000000016B5A51".parse().unwrap();
3032 4 : let local_path = local_layer_path(
3033 4 : harness.conf,
3034 4 : &timeline.tenant_shard_id,
3035 4 : &timeline.timeline_id,
3036 4 : &layer_file_name_1,
3037 4 : &harness.generation,
3038 4 : );
3039 4 : let content_1 = dummy_contents("foo");
3040 4 : std::fs::write(&local_path, &content_1).unwrap();
3041 4 :
3042 4 : let layer_file_1 = Layer::for_resident(
3043 4 : harness.conf,
3044 4 : &timeline,
3045 4 : local_path,
3046 4 : layer_file_name_1.clone(),
3047 4 : LayerFileMetadata::new(content_1.len() as u64, harness.generation, harness.shard),
3048 4 : );
3049 4 :
3050 4 : #[derive(Debug, PartialEq, Clone, Copy)]
3051 4 : struct BytesStartedFinished {
3052 4 : started: Option<usize>,
3053 4 : finished: Option<usize>,
3054 4 : }
3055 4 : impl std::ops::Add for BytesStartedFinished {
3056 4 : type Output = Self;
3057 8 : fn add(self, rhs: Self) -> Self::Output {
3058 8 : Self {
3059 8 : started: self.started.map(|v| v + rhs.started.unwrap_or(0)),
3060 8 : finished: self.finished.map(|v| v + rhs.finished.unwrap_or(0)),
3061 8 : }
3062 8 : }
3063 4 : }
3064 12 : let get_bytes_started_stopped = || {
3065 12 : let started = client
3066 12 : .metrics
3067 12 : .get_bytes_started_counter_value(&RemoteOpFileKind::Layer, &RemoteOpKind::Upload)
3068 12 : .map(|v| v.try_into().unwrap());
3069 12 : let stopped = client
3070 12 : .metrics
3071 12 : .get_bytes_finished_counter_value(&RemoteOpFileKind::Layer, &RemoteOpKind::Upload)
3072 12 : .map(|v| v.try_into().unwrap());
3073 12 : BytesStartedFinished {
3074 12 : started,
3075 12 : finished: stopped,
3076 12 : }
3077 12 : };
3078 4 :
3079 4 : // Test
3080 4 : tracing::info!("now doing actual test");
3081 4 :
3082 4 : let actual_a = get_bytes_started_stopped();
3083 4 :
3084 4 : client
3085 4 : .schedule_layer_file_upload(layer_file_1.clone())
3086 4 : .unwrap();
3087 4 :
3088 4 : let actual_b = get_bytes_started_stopped();
3089 4 :
3090 4 : client.wait_completion().await.unwrap();
3091 4 :
3092 4 : let actual_c = get_bytes_started_stopped();
3093 4 :
3094 4 : // Validate
3095 4 :
3096 4 : let expected_b = actual_a
3097 4 : + BytesStartedFinished {
3098 4 : started: Some(content_1.len()),
3099 4 : // assert that the _finished metric is created eagerly so that subtractions work on first sample
3100 4 : finished: Some(0),
3101 4 : };
3102 4 : assert_eq!(actual_b, expected_b);
3103 4 :
3104 4 : let expected_c = actual_a
3105 4 : + BytesStartedFinished {
3106 4 : started: Some(content_1.len()),
3107 4 : finished: Some(content_1.len()),
3108 4 : };
3109 4 : assert_eq!(actual_c, expected_c);
3110 4 : }
3111 :
3112 24 : async fn inject_index_part(test_state: &TestSetup, generation: Generation) -> IndexPart {
3113 24 : // An empty IndexPart, just sufficient to ensure deserialization will succeed
3114 24 : let example_index_part = IndexPart::example();
3115 24 :
3116 24 : let index_part_bytes = serde_json::to_vec(&example_index_part).unwrap();
3117 24 :
3118 24 : let index_path = test_state.harness.remote_fs_dir.join(
3119 24 : remote_index_path(
3120 24 : &test_state.harness.tenant_shard_id,
3121 24 : &TIMELINE_ID,
3122 24 : generation,
3123 24 : )
3124 24 : .get_path(),
3125 24 : );
3126 24 :
3127 24 : std::fs::create_dir_all(index_path.parent().unwrap())
3128 24 : .expect("creating test dir should work");
3129 24 :
3130 24 : eprintln!("Writing {index_path}");
3131 24 : std::fs::write(&index_path, index_part_bytes).unwrap();
3132 24 : example_index_part
3133 24 : }
3134 :
3135 : /// Assert that when a RemoteTimelineclient in generation `get_generation` fetches its
3136 : /// index, the IndexPart returned is equal to `expected`
3137 20 : async fn assert_got_index_part(
3138 20 : test_state: &TestSetup,
3139 20 : get_generation: Generation,
3140 20 : expected: &IndexPart,
3141 20 : ) {
3142 20 : let client = test_state.build_client(get_generation);
3143 :
3144 20 : let download_r = client
3145 20 : .download_index_file(&CancellationToken::new())
3146 20 : .await
3147 20 : .expect("download should always succeed");
3148 20 : assert!(matches!(download_r, MaybeDeletedIndexPart::IndexPart(_)));
3149 20 : match download_r {
3150 20 : MaybeDeletedIndexPart::IndexPart(index_part) => {
3151 20 : assert_eq!(&index_part, expected);
3152 : }
3153 0 : MaybeDeletedIndexPart::Deleted(_index_part) => panic!("Test doesn't set deleted_at"),
3154 : }
3155 20 : }
3156 :
3157 : #[tokio::test]
3158 4 : async fn index_part_download_simple() -> anyhow::Result<()> {
3159 4 : let test_state = TestSetup::new("index_part_download_simple").await.unwrap();
3160 4 : let span = test_state.span();
3161 4 : let _guard = span.enter();
3162 4 :
3163 4 : // Simple case: we are in generation N, load the index from generation N - 1
3164 4 : let generation_n = 5;
3165 4 : let injected = inject_index_part(&test_state, Generation::new(generation_n - 1)).await;
3166 4 :
3167 4 : assert_got_index_part(&test_state, Generation::new(generation_n), &injected).await;
3168 4 :
3169 4 : Ok(())
3170 4 : }
3171 :
3172 : #[tokio::test]
3173 4 : async fn index_part_download_ordering() -> anyhow::Result<()> {
3174 4 : let test_state = TestSetup::new("index_part_download_ordering")
3175 4 : .await
3176 4 : .unwrap();
3177 4 :
3178 4 : let span = test_state.span();
3179 4 : let _guard = span.enter();
3180 4 :
3181 4 : // A generation-less IndexPart exists in the bucket, we should find it
3182 4 : let generation_n = 5;
3183 4 : let injected_none = inject_index_part(&test_state, Generation::none()).await;
3184 4 : assert_got_index_part(&test_state, Generation::new(generation_n), &injected_none).await;
3185 4 :
3186 4 : // If a more recent-than-none generation exists, we should prefer to load that
3187 4 : let injected_1 = inject_index_part(&test_state, Generation::new(1)).await;
3188 4 : assert_got_index_part(&test_state, Generation::new(generation_n), &injected_1).await;
3189 4 :
3190 4 : // If a more-recent-than-me generation exists, we should ignore it.
3191 4 : let _injected_10 = inject_index_part(&test_state, Generation::new(10)).await;
3192 4 : assert_got_index_part(&test_state, Generation::new(generation_n), &injected_1).await;
3193 4 :
3194 4 : // If a directly previous generation exists, _and_ an index exists in my own
3195 4 : // generation, I should prefer my own generation.
3196 4 : let _injected_prev =
3197 4 : inject_index_part(&test_state, Generation::new(generation_n - 1)).await;
3198 4 : let injected_current = inject_index_part(&test_state, Generation::new(generation_n)).await;
3199 4 : assert_got_index_part(
3200 4 : &test_state,
3201 4 : Generation::new(generation_n),
3202 4 : &injected_current,
3203 4 : )
3204 4 : .await;
3205 4 :
3206 4 : Ok(())
3207 4 : }
3208 : }
|
