Line data Source code
1 : //!
2 : //! Timeline repository implementation that keeps old data in files on disk, and
3 : //! the recent changes in memory. See tenant/*_layer.rs files.
4 : //! The functions here are responsible for locating the correct layer for the
5 : //! get/put call, walking back the timeline branching history as needed.
6 : //!
7 : //! The files are stored in the .neon/tenants/<tenant_id>/timelines/<timeline_id>
8 : //! directory. See docs/pageserver-storage.md for how the files are managed.
9 : //! In addition to the layer files, there is a metadata file in the same
10 : //! directory that contains information about the timeline, in particular its
11 : //! parent timeline, and the last LSN that has been written to disk.
12 : //!
13 :
14 : use anyhow::{bail, Context};
15 : use arc_swap::ArcSwap;
16 : use camino::Utf8Path;
17 : use camino::Utf8PathBuf;
18 : use enumset::EnumSet;
19 : use futures::stream::FuturesUnordered;
20 : use futures::FutureExt;
21 : use futures::StreamExt;
22 : use pageserver_api::models;
23 : use pageserver_api::models::AuxFilePolicy;
24 : use pageserver_api::models::TimelineArchivalState;
25 : use pageserver_api::models::TimelineState;
26 : use pageserver_api::models::TopTenantShardItem;
27 : use pageserver_api::models::WalRedoManagerStatus;
28 : use pageserver_api::shard::ShardIdentity;
29 : use pageserver_api::shard::ShardStripeSize;
30 : use pageserver_api::shard::TenantShardId;
31 : use remote_storage::DownloadError;
32 : use remote_storage::GenericRemoteStorage;
33 : use remote_storage::TimeoutOrCancel;
34 : use std::fmt;
35 : use std::time::SystemTime;
36 : use storage_broker::BrokerClientChannel;
37 : use tokio::io::BufReader;
38 : use tokio::sync::watch;
39 : use tokio::task::JoinSet;
40 : use tokio_util::sync::CancellationToken;
41 : use tracing::*;
42 : use utils::backoff;
43 : use utils::circuit_breaker::CircuitBreaker;
44 : use utils::completion;
45 : use utils::crashsafe::path_with_suffix_extension;
46 : use utils::failpoint_support;
47 : use utils::fs_ext;
48 : use utils::pausable_failpoint;
49 : use utils::sync::gate::Gate;
50 : use utils::sync::gate::GateGuard;
51 : use utils::timeout::timeout_cancellable;
52 : use utils::timeout::TimeoutCancellableError;
53 : use utils::zstd::create_zst_tarball;
54 : use utils::zstd::extract_zst_tarball;
55 :
56 : use self::config::AttachedLocationConfig;
57 : use self::config::AttachmentMode;
58 : use self::config::LocationConf;
59 : use self::config::TenantConf;
60 : use self::metadata::TimelineMetadata;
61 : use self::mgr::GetActiveTenantError;
62 : use self::mgr::GetTenantError;
63 : use self::remote_timeline_client::upload::upload_index_part;
64 : use self::remote_timeline_client::RemoteTimelineClient;
65 : use self::timeline::uninit::TimelineCreateGuard;
66 : use self::timeline::uninit::TimelineExclusionError;
67 : use self::timeline::uninit::UninitializedTimeline;
68 : use self::timeline::EvictionTaskTenantState;
69 : use self::timeline::GcCutoffs;
70 : use self::timeline::TimelineResources;
71 : use self::timeline::WaitLsnError;
72 : use crate::config::PageServerConf;
73 : use crate::context::{DownloadBehavior, RequestContext};
74 : use crate::deletion_queue::DeletionQueueClient;
75 : use crate::deletion_queue::DeletionQueueError;
76 : use crate::import_datadir;
77 : use crate::is_uninit_mark;
78 : use crate::l0_flush::L0FlushGlobalState;
79 : use crate::metrics::TENANT;
80 : use crate::metrics::{
81 : remove_tenant_metrics, BROKEN_TENANTS_SET, CIRCUIT_BREAKERS_BROKEN, CIRCUIT_BREAKERS_UNBROKEN,
82 : TENANT_STATE_METRIC, TENANT_SYNTHETIC_SIZE_METRIC,
83 : };
84 : use crate::repository::GcResult;
85 : use crate::task_mgr;
86 : use crate::task_mgr::TaskKind;
87 : use crate::tenant::config::LocationMode;
88 : use crate::tenant::config::TenantConfOpt;
89 : pub use crate::tenant::remote_timeline_client::index::IndexPart;
90 : use crate::tenant::remote_timeline_client::remote_initdb_archive_path;
91 : use crate::tenant::remote_timeline_client::MaybeDeletedIndexPart;
92 : use crate::tenant::remote_timeline_client::INITDB_PATH;
93 : use crate::tenant::storage_layer::DeltaLayer;
94 : use crate::tenant::storage_layer::ImageLayer;
95 : use crate::walredo;
96 : use crate::InitializationOrder;
97 : use std::collections::hash_map::Entry;
98 : use std::collections::BTreeSet;
99 : use std::collections::HashMap;
100 : use std::collections::HashSet;
101 : use std::fmt::Debug;
102 : use std::fmt::Display;
103 : use std::fs;
104 : use std::fs::File;
105 : use std::ops::Bound::Included;
106 : use std::sync::atomic::AtomicU64;
107 : use std::sync::atomic::Ordering;
108 : use std::sync::Arc;
109 : use std::sync::Mutex;
110 : use std::time::{Duration, Instant};
111 :
112 : use crate::span;
113 : use crate::tenant::timeline::delete::DeleteTimelineFlow;
114 : use crate::tenant::timeline::uninit::cleanup_timeline_directory;
115 : use crate::virtual_file::VirtualFile;
116 : use crate::walredo::PostgresRedoManager;
117 : use crate::TEMP_FILE_SUFFIX;
118 : use once_cell::sync::Lazy;
119 : pub use pageserver_api::models::TenantState;
120 : use tokio::sync::Semaphore;
121 :
122 0 : static INIT_DB_SEMAPHORE: Lazy<Semaphore> = Lazy::new(|| Semaphore::new(8));
123 : use utils::{
124 : crashsafe,
125 : generation::Generation,
126 : id::TimelineId,
127 : lsn::{Lsn, RecordLsn},
128 : };
129 :
130 : pub mod blob_io;
131 : pub mod block_io;
132 : pub mod vectored_blob_io;
133 :
134 : pub mod disk_btree;
135 : pub(crate) mod ephemeral_file;
136 : pub mod layer_map;
137 :
138 : pub mod metadata;
139 : pub mod remote_timeline_client;
140 : pub mod storage_layer;
141 :
142 : pub mod config;
143 : pub mod mgr;
144 : pub mod secondary;
145 : pub mod tasks;
146 : pub mod upload_queue;
147 :
148 : pub(crate) mod timeline;
149 :
150 : pub mod size;
151 :
152 : pub(crate) mod throttle;
153 :
154 : pub(crate) use crate::span::debug_assert_current_span_has_tenant_and_timeline_id;
155 : pub(crate) use timeline::{LogicalSizeCalculationCause, PageReconstructError, Timeline};
156 :
157 : // re-export for use in walreceiver
158 : pub use crate::tenant::timeline::WalReceiverInfo;
159 :
160 : /// The "tenants" part of `tenants/<tenant>/timelines...`
161 : pub const TENANTS_SEGMENT_NAME: &str = "tenants";
162 :
163 : /// Parts of the `.neon/tenants/<tenant_id>/timelines/<timeline_id>` directory prefix.
164 : pub const TIMELINES_SEGMENT_NAME: &str = "timelines";
165 :
166 : /// References to shared objects that are passed into each tenant, such
167 : /// as the shared remote storage client and process initialization state.
168 : #[derive(Clone)]
169 : pub struct TenantSharedResources {
170 : pub broker_client: storage_broker::BrokerClientChannel,
171 : pub remote_storage: GenericRemoteStorage,
172 : pub deletion_queue_client: DeletionQueueClient,
173 : pub l0_flush_global_state: L0FlushGlobalState,
174 : }
175 :
176 : /// A [`Tenant`] is really an _attached_ tenant. The configuration
177 : /// for an attached tenant is a subset of the [`LocationConf`], represented
178 : /// in this struct.
179 : pub(super) struct AttachedTenantConf {
180 : tenant_conf: TenantConfOpt,
181 : location: AttachedLocationConfig,
182 : }
183 :
184 : impl AttachedTenantConf {
185 0 : fn new(tenant_conf: TenantConfOpt, location: AttachedLocationConfig) -> Self {
186 0 : Self {
187 0 : tenant_conf,
188 0 : location,
189 0 : }
190 0 : }
191 :
192 176 : fn try_from(location_conf: LocationConf) -> anyhow::Result<Self> {
193 176 : match &location_conf.mode {
194 176 : LocationMode::Attached(attach_conf) => Ok(Self {
195 176 : tenant_conf: location_conf.tenant_conf,
196 176 : location: *attach_conf,
197 176 : }),
198 : LocationMode::Secondary(_) => {
199 0 : anyhow::bail!("Attempted to construct AttachedTenantConf from a LocationConf in secondary mode")
200 : }
201 : }
202 176 : }
203 : }
204 : struct TimelinePreload {
205 : timeline_id: TimelineId,
206 : client: RemoteTimelineClient,
207 : index_part: Result<MaybeDeletedIndexPart, DownloadError>,
208 : }
209 :
210 : pub(crate) struct TenantPreload {
211 : timelines: HashMap<TimelineId, TimelinePreload>,
212 : }
213 :
214 : /// When we spawn a tenant, there is a special mode for tenant creation that
215 : /// avoids trying to read anything from remote storage.
216 : pub(crate) enum SpawnMode {
217 : /// Activate as soon as possible
218 : Eager,
219 : /// Lazy activation in the background, with the option to skip the queue if the need comes up
220 : Lazy,
221 : }
222 :
223 : ///
224 : /// Tenant consists of multiple timelines. Keep them in a hash table.
225 : ///
226 : pub struct Tenant {
227 : // Global pageserver config parameters
228 : pub conf: &'static PageServerConf,
229 :
230 : /// The value creation timestamp, used to measure activation delay, see:
231 : /// <https://github.com/neondatabase/neon/issues/4025>
232 : constructed_at: Instant,
233 :
234 : state: watch::Sender<TenantState>,
235 :
236 : // Overridden tenant-specific config parameters.
237 : // We keep TenantConfOpt sturct here to preserve the information
238 : // about parameters that are not set.
239 : // This is necessary to allow global config updates.
240 : tenant_conf: Arc<ArcSwap<AttachedTenantConf>>,
241 :
242 : tenant_shard_id: TenantShardId,
243 :
244 : // The detailed sharding information, beyond the number/count in tenant_shard_id
245 : shard_identity: ShardIdentity,
246 :
247 : /// The remote storage generation, used to protect S3 objects from split-brain.
248 : /// Does not change over the lifetime of the [`Tenant`] object.
249 : ///
250 : /// This duplicates the generation stored in LocationConf, but that structure is mutable:
251 : /// this copy enforces the invariant that generatio doesn't change during a Tenant's lifetime.
252 : generation: Generation,
253 :
254 : timelines: Mutex<HashMap<TimelineId, Arc<Timeline>>>,
255 :
256 : /// During timeline creation, we first insert the TimelineId to the
257 : /// creating map, then `timelines`, then remove it from the creating map.
258 : /// **Lock order**: if acquring both, acquire`timelines` before `timelines_creating`
259 : timelines_creating: std::sync::Mutex<HashSet<TimelineId>>,
260 :
261 : // This mutex prevents creation of new timelines during GC.
262 : // Adding yet another mutex (in addition to `timelines`) is needed because holding
263 : // `timelines` mutex during all GC iteration
264 : // may block for a long time `get_timeline`, `get_timelines_state`,... and other operations
265 : // with timelines, which in turn may cause dropping replication connection, expiration of wait_for_lsn
266 : // timeout...
267 : gc_cs: tokio::sync::Mutex<()>,
268 : walredo_mgr: Option<Arc<WalRedoManager>>,
269 :
270 : // provides access to timeline data sitting in the remote storage
271 : pub(crate) remote_storage: GenericRemoteStorage,
272 :
273 : // Access to global deletion queue for when this tenant wants to schedule a deletion
274 : deletion_queue_client: DeletionQueueClient,
275 :
276 : /// Cached logical sizes updated updated on each [`Tenant::gather_size_inputs`].
277 : cached_logical_sizes: tokio::sync::Mutex<HashMap<(TimelineId, Lsn), u64>>,
278 : cached_synthetic_tenant_size: Arc<AtomicU64>,
279 :
280 : eviction_task_tenant_state: tokio::sync::Mutex<EvictionTaskTenantState>,
281 :
282 : /// Track repeated failures to compact, so that we can back off.
283 : /// Overhead of mutex is acceptable because compaction is done with a multi-second period.
284 : compaction_circuit_breaker: std::sync::Mutex<CircuitBreaker>,
285 :
286 : /// If the tenant is in Activating state, notify this to encourage it
287 : /// to proceed to Active as soon as possible, rather than waiting for lazy
288 : /// background warmup.
289 : pub(crate) activate_now_sem: tokio::sync::Semaphore,
290 :
291 : // Cancellation token fires when we have entered shutdown(). This is a parent of
292 : // Timelines' cancellation token.
293 : pub(crate) cancel: CancellationToken,
294 :
295 : // Users of the Tenant such as the page service must take this Gate to avoid
296 : // trying to use a Tenant which is shutting down.
297 : pub(crate) gate: Gate,
298 :
299 : /// Throttle applied at the top of [`Timeline::get`].
300 : /// All [`Tenant::timelines`] of a given [`Tenant`] instance share the same [`throttle::Throttle`] instance.
301 : pub(crate) timeline_get_throttle:
302 : Arc<throttle::Throttle<&'static crate::metrics::tenant_throttling::TimelineGet>>,
303 :
304 : /// An ongoing timeline detach must be checked during attempts to GC or compact a timeline.
305 : ongoing_timeline_detach: std::sync::Mutex<Option<(TimelineId, utils::completion::Barrier)>>,
306 :
307 : l0_flush_global_state: L0FlushGlobalState,
308 : }
309 :
310 : impl std::fmt::Debug for Tenant {
311 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
312 0 : write!(f, "{} ({})", self.tenant_shard_id, self.current_state())
313 0 : }
314 : }
315 :
316 : pub(crate) enum WalRedoManager {
317 : Prod(PostgresRedoManager),
318 : #[cfg(test)]
319 : Test(harness::TestRedoManager),
320 : }
321 :
322 : impl From<PostgresRedoManager> for WalRedoManager {
323 0 : fn from(mgr: PostgresRedoManager) -> Self {
324 0 : Self::Prod(mgr)
325 0 : }
326 : }
327 :
328 : #[cfg(test)]
329 : impl From<harness::TestRedoManager> for WalRedoManager {
330 168 : fn from(mgr: harness::TestRedoManager) -> Self {
331 168 : Self::Test(mgr)
332 168 : }
333 : }
334 :
335 : impl WalRedoManager {
336 6 : pub(crate) async fn shutdown(&self) {
337 6 : match self {
338 0 : Self::Prod(mgr) => mgr.shutdown().await,
339 : #[cfg(test)]
340 6 : Self::Test(_) => {
341 6 : // Not applicable to test redo manager
342 6 : }
343 : }
344 6 : }
345 :
346 0 : pub(crate) fn maybe_quiesce(&self, idle_timeout: Duration) {
347 0 : match self {
348 0 : Self::Prod(mgr) => mgr.maybe_quiesce(idle_timeout),
349 0 : #[cfg(test)]
350 0 : Self::Test(_) => {
351 0 : // Not applicable to test redo manager
352 0 : }
353 0 : }
354 0 : }
355 :
356 : /// # Cancel-Safety
357 : ///
358 : /// This method is cancellation-safe.
359 58 : pub async fn request_redo(
360 58 : &self,
361 58 : key: crate::repository::Key,
362 58 : lsn: Lsn,
363 58 : base_img: Option<(Lsn, bytes::Bytes)>,
364 58 : records: Vec<(Lsn, crate::walrecord::NeonWalRecord)>,
365 58 : pg_version: u32,
366 58 : ) -> Result<bytes::Bytes, walredo::Error> {
367 58 : match self {
368 0 : Self::Prod(mgr) => {
369 0 : mgr.request_redo(key, lsn, base_img, records, pg_version)
370 0 : .await
371 : }
372 : #[cfg(test)]
373 58 : Self::Test(mgr) => {
374 58 : mgr.request_redo(key, lsn, base_img, records, pg_version)
375 0 : .await
376 : }
377 : }
378 58 : }
379 :
380 0 : pub(crate) fn status(&self) -> Option<WalRedoManagerStatus> {
381 0 : match self {
382 0 : WalRedoManager::Prod(m) => Some(m.status()),
383 0 : #[cfg(test)]
384 0 : WalRedoManager::Test(_) => None,
385 0 : }
386 0 : }
387 : }
388 :
389 0 : #[derive(Debug, thiserror::Error, PartialEq, Eq)]
390 : pub enum GetTimelineError {
391 : #[error("Timeline {tenant_id}/{timeline_id} is not active, state: {state:?}")]
392 : NotActive {
393 : tenant_id: TenantShardId,
394 : timeline_id: TimelineId,
395 : state: TimelineState,
396 : },
397 : #[error("Timeline {tenant_id}/{timeline_id} was not found")]
398 : NotFound {
399 : tenant_id: TenantShardId,
400 : timeline_id: TimelineId,
401 : },
402 : }
403 :
404 0 : #[derive(Debug, thiserror::Error)]
405 : pub enum LoadLocalTimelineError {
406 : #[error("FailedToLoad")]
407 : Load(#[source] anyhow::Error),
408 : #[error("FailedToResumeDeletion")]
409 : ResumeDeletion(#[source] anyhow::Error),
410 : }
411 :
412 0 : #[derive(thiserror::Error)]
413 : pub enum DeleteTimelineError {
414 : #[error("NotFound")]
415 : NotFound,
416 :
417 : #[error("HasChildren")]
418 : HasChildren(Vec<TimelineId>),
419 :
420 : #[error("Timeline deletion is already in progress")]
421 : AlreadyInProgress(Arc<tokio::sync::Mutex<DeleteTimelineFlow>>),
422 :
423 : #[error(transparent)]
424 : Other(#[from] anyhow::Error),
425 : }
426 :
427 : impl Debug for DeleteTimelineError {
428 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
429 0 : match self {
430 0 : Self::NotFound => write!(f, "NotFound"),
431 0 : Self::HasChildren(c) => f.debug_tuple("HasChildren").field(c).finish(),
432 0 : Self::AlreadyInProgress(_) => f.debug_tuple("AlreadyInProgress").finish(),
433 0 : Self::Other(e) => f.debug_tuple("Other").field(e).finish(),
434 : }
435 0 : }
436 : }
437 :
438 : pub enum SetStoppingError {
439 : AlreadyStopping(completion::Barrier),
440 : Broken,
441 : }
442 :
443 : impl Debug for SetStoppingError {
444 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
445 0 : match self {
446 0 : Self::AlreadyStopping(_) => f.debug_tuple("AlreadyStopping").finish(),
447 0 : Self::Broken => write!(f, "Broken"),
448 : }
449 0 : }
450 : }
451 :
452 0 : #[derive(thiserror::Error, Debug)]
453 : pub enum CreateTimelineError {
454 : #[error("creation of timeline with the given ID is in progress")]
455 : AlreadyCreating,
456 : #[error("timeline already exists with different parameters")]
457 : Conflict,
458 : #[error(transparent)]
459 : AncestorLsn(anyhow::Error),
460 : #[error("ancestor timeline is not active")]
461 : AncestorNotActive,
462 : #[error("tenant shutting down")]
463 : ShuttingDown,
464 : #[error(transparent)]
465 : Other(#[from] anyhow::Error),
466 : }
467 :
468 : #[derive(thiserror::Error, Debug)]
469 : enum InitdbError {
470 : Other(anyhow::Error),
471 : Cancelled,
472 : Spawn(std::io::Result<()>),
473 : Failed(std::process::ExitStatus, Vec<u8>),
474 : }
475 :
476 : impl fmt::Display for InitdbError {
477 0 : fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
478 0 : match self {
479 0 : InitdbError::Cancelled => write!(f, "Operation was cancelled"),
480 0 : InitdbError::Spawn(e) => write!(f, "Spawn error: {:?}", e),
481 0 : InitdbError::Failed(status, stderr) => write!(
482 0 : f,
483 0 : "Command failed with status {:?}: {}",
484 0 : status,
485 0 : String::from_utf8_lossy(stderr)
486 0 : ),
487 0 : InitdbError::Other(e) => write!(f, "Error: {:?}", e),
488 : }
489 0 : }
490 : }
491 :
492 : impl From<std::io::Error> for InitdbError {
493 0 : fn from(error: std::io::Error) -> Self {
494 0 : InitdbError::Spawn(Err(error))
495 0 : }
496 : }
497 :
498 : enum CreateTimelineCause {
499 : Load,
500 : Delete,
501 : }
502 :
503 0 : #[derive(thiserror::Error, Debug)]
504 : pub(crate) enum GcError {
505 : // The tenant is shutting down
506 : #[error("tenant shutting down")]
507 : TenantCancelled,
508 :
509 : // The tenant is shutting down
510 : #[error("timeline shutting down")]
511 : TimelineCancelled,
512 :
513 : // The tenant is in a state inelegible to run GC
514 : #[error("not active")]
515 : NotActive,
516 :
517 : // A requested GC cutoff LSN was invalid, for example it tried to move backwards
518 : #[error("not active")]
519 : BadLsn { why: String },
520 :
521 : // A remote storage error while scheduling updates after compaction
522 : #[error(transparent)]
523 : Remote(anyhow::Error),
524 :
525 : // An error reading while calculating GC cutoffs
526 : #[error(transparent)]
527 : GcCutoffs(PageReconstructError),
528 :
529 : // If GC was invoked for a particular timeline, this error means it didn't exist
530 : #[error("timeline not found")]
531 : TimelineNotFound,
532 : }
533 :
534 : impl From<PageReconstructError> for GcError {
535 0 : fn from(value: PageReconstructError) -> Self {
536 0 : match value {
537 0 : PageReconstructError::Cancelled => Self::TimelineCancelled,
538 0 : other => Self::GcCutoffs(other),
539 : }
540 0 : }
541 : }
542 :
543 0 : #[derive(thiserror::Error, Debug)]
544 : pub(crate) enum LoadConfigError {
545 : #[error("TOML deserialization error: '{0}'")]
546 : DeserializeToml(#[from] toml_edit::de::Error),
547 :
548 : #[error("Config not found at {0}")]
549 : NotFound(Utf8PathBuf),
550 : }
551 :
552 : impl Tenant {
553 : /// Yet another helper for timeline initialization.
554 : ///
555 : /// - Initializes the Timeline struct and inserts it into the tenant's hash map
556 : /// - Scans the local timeline directory for layer files and builds the layer map
557 : /// - Downloads remote index file and adds remote files to the layer map
558 : /// - Schedules remote upload tasks for any files that are present locally but missing from remote storage.
559 : ///
560 : /// If the operation fails, the timeline is left in the tenant's hash map in Broken state. On success,
561 : /// it is marked as Active.
562 : #[allow(clippy::too_many_arguments)]
563 6 : async fn timeline_init_and_sync(
564 6 : &self,
565 6 : timeline_id: TimelineId,
566 6 : resources: TimelineResources,
567 6 : index_part: Option<IndexPart>,
568 6 : metadata: TimelineMetadata,
569 6 : ancestor: Option<Arc<Timeline>>,
570 6 : last_aux_file_policy: Option<AuxFilePolicy>,
571 6 : _ctx: &RequestContext,
572 6 : ) -> anyhow::Result<()> {
573 6 : let tenant_id = self.tenant_shard_id;
574 :
575 6 : let timeline = self.create_timeline_struct(
576 6 : timeline_id,
577 6 : &metadata,
578 6 : ancestor.clone(),
579 6 : resources,
580 6 : CreateTimelineCause::Load,
581 6 : // This could be derived from ancestor branch + index part. Though the only caller of `timeline_init_and_sync` is `load_remote_timeline`,
582 6 : // there will potentially be other caller of this function in the future, and we don't know whether `index_part` or `ancestor` takes precedence.
583 6 : // Therefore, we pass this field explicitly for now, and remove it once we fully migrate to aux file v2.
584 6 : last_aux_file_policy,
585 6 : )?;
586 6 : let disk_consistent_lsn = timeline.get_disk_consistent_lsn();
587 6 : anyhow::ensure!(
588 6 : disk_consistent_lsn.is_valid(),
589 0 : "Timeline {tenant_id}/{timeline_id} has invalid disk_consistent_lsn"
590 : );
591 6 : assert_eq!(
592 6 : disk_consistent_lsn,
593 6 : metadata.disk_consistent_lsn(),
594 0 : "these are used interchangeably"
595 : );
596 :
597 6 : if let Some(index_part) = index_part.as_ref() {
598 6 : timeline.remote_client.init_upload_queue(index_part)?;
599 :
600 6 : timeline
601 6 : .last_aux_file_policy
602 6 : .store(index_part.last_aux_file_policy());
603 : } else {
604 : // No data on the remote storage, but we have local metadata file. We can end up
605 : // here with timeline_create being interrupted before finishing index part upload.
606 : // By doing what we do here, the index part upload is retried.
607 : // If control plane retries timeline creation in the meantime, the mgmt API handler
608 : // for timeline creation will coalesce on the upload we queue here.
609 :
610 : // FIXME: this branch should be dead code as we no longer write local metadata.
611 :
612 0 : timeline
613 0 : .remote_client
614 0 : .init_upload_queue_for_empty_remote(&metadata)?;
615 0 : timeline
616 0 : .remote_client
617 0 : .schedule_index_upload_for_full_metadata_update(&metadata)?;
618 : }
619 :
620 6 : timeline
621 6 : .load_layer_map(disk_consistent_lsn, index_part)
622 6 : .await
623 6 : .with_context(|| {
624 0 : format!("Failed to load layermap for timeline {tenant_id}/{timeline_id}")
625 6 : })?;
626 :
627 : {
628 : // avoiding holding it across awaits
629 6 : let mut timelines_accessor = self.timelines.lock().unwrap();
630 6 : match timelines_accessor.entry(timeline_id) {
631 : // We should never try and load the same timeline twice during startup
632 : Entry::Occupied(_) => {
633 0 : unreachable!(
634 0 : "Timeline {tenant_id}/{timeline_id} already exists in the tenant map"
635 0 : );
636 : }
637 6 : Entry::Vacant(v) => {
638 6 : v.insert(Arc::clone(&timeline));
639 6 : timeline.maybe_spawn_flush_loop();
640 6 : }
641 6 : }
642 6 : };
643 6 :
644 6 : // Sanity check: a timeline should have some content.
645 6 : anyhow::ensure!(
646 6 : ancestor.is_some()
647 4 : || timeline
648 4 : .layers
649 4 : .read()
650 0 : .await
651 4 : .layer_map()
652 4 : .iter_historic_layers()
653 4 : .next()
654 4 : .is_some(),
655 0 : "Timeline has no ancestor and no layer files"
656 : );
657 :
658 6 : Ok(())
659 6 : }
660 :
661 : /// Attach a tenant that's available in cloud storage.
662 : ///
663 : /// This returns quickly, after just creating the in-memory object
664 : /// Tenant struct and launching a background task to download
665 : /// the remote index files. On return, the tenant is most likely still in
666 : /// Attaching state, and it will become Active once the background task
667 : /// finishes. You can use wait_until_active() to wait for the task to
668 : /// complete.
669 : ///
670 : #[allow(clippy::too_many_arguments)]
671 0 : pub(crate) fn spawn(
672 0 : conf: &'static PageServerConf,
673 0 : tenant_shard_id: TenantShardId,
674 0 : resources: TenantSharedResources,
675 0 : attached_conf: AttachedTenantConf,
676 0 : shard_identity: ShardIdentity,
677 0 : init_order: Option<InitializationOrder>,
678 0 : mode: SpawnMode,
679 0 : ctx: &RequestContext,
680 0 : ) -> Arc<Tenant> {
681 0 : let wal_redo_manager = Arc::new(WalRedoManager::from(PostgresRedoManager::new(
682 0 : conf,
683 0 : tenant_shard_id,
684 0 : )));
685 0 :
686 0 : let TenantSharedResources {
687 0 : broker_client,
688 0 : remote_storage,
689 0 : deletion_queue_client,
690 0 : l0_flush_global_state,
691 0 : } = resources;
692 0 :
693 0 : let attach_mode = attached_conf.location.attach_mode;
694 0 : let generation = attached_conf.location.generation;
695 0 :
696 0 : let tenant = Arc::new(Tenant::new(
697 0 : TenantState::Attaching,
698 0 : conf,
699 0 : attached_conf,
700 0 : shard_identity,
701 0 : Some(wal_redo_manager),
702 0 : tenant_shard_id,
703 0 : remote_storage.clone(),
704 0 : deletion_queue_client,
705 0 : l0_flush_global_state,
706 0 : ));
707 0 :
708 0 : // The attach task will carry a GateGuard, so that shutdown() reliably waits for it to drop out if
709 0 : // we shut down while attaching.
710 0 : let attach_gate_guard = tenant
711 0 : .gate
712 0 : .enter()
713 0 : .expect("We just created the Tenant: nothing else can have shut it down yet");
714 0 :
715 0 : // Do all the hard work in the background
716 0 : let tenant_clone = Arc::clone(&tenant);
717 0 : let ctx = ctx.detached_child(TaskKind::Attach, DownloadBehavior::Warn);
718 0 : task_mgr::spawn(
719 0 : &tokio::runtime::Handle::current(),
720 0 : TaskKind::Attach,
721 0 : Some(tenant_shard_id),
722 0 : None,
723 0 : "attach tenant",
724 : false,
725 0 : async move {
726 0 :
727 0 : info!(
728 : ?attach_mode,
729 0 : "Attaching tenant"
730 : );
731 :
732 0 : let _gate_guard = attach_gate_guard;
733 0 :
734 0 : // Is this tenant being spawned as part of process startup?
735 0 : let starting_up = init_order.is_some();
736 : scopeguard::defer! {
737 : if starting_up {
738 : TENANT.startup_complete.inc();
739 : }
740 : }
741 :
742 : // Ideally we should use Tenant::set_broken_no_wait, but it is not supposed to be used when tenant is in loading state.
743 : enum BrokenVerbosity {
744 : Error,
745 : Info
746 : }
747 0 : let make_broken =
748 0 : |t: &Tenant, err: anyhow::Error, verbosity: BrokenVerbosity| {
749 0 : match verbosity {
750 : BrokenVerbosity::Info => {
751 0 : info!("attach cancelled, setting tenant state to Broken: {err}");
752 : },
753 : BrokenVerbosity::Error => {
754 0 : error!("attach failed, setting tenant state to Broken: {err:?}");
755 : }
756 : }
757 0 : t.state.send_modify(|state| {
758 0 : // The Stopping case is for when we have passed control on to DeleteTenantFlow:
759 0 : // if it errors, we will call make_broken when tenant is already in Stopping.
760 0 : assert!(
761 0 : matches!(*state, TenantState::Attaching | TenantState::Stopping { .. }),
762 0 : "the attach task owns the tenant state until activation is complete"
763 : );
764 :
765 0 : *state = TenantState::broken_from_reason(err.to_string());
766 0 : });
767 0 : };
768 :
769 0 : let mut init_order = init_order;
770 0 : // take the completion because initial tenant loading will complete when all of
771 0 : // these tasks complete.
772 0 : let _completion = init_order
773 0 : .as_mut()
774 0 : .and_then(|x| x.initial_tenant_load.take());
775 0 : let remote_load_completion = init_order
776 0 : .as_mut()
777 0 : .and_then(|x| x.initial_tenant_load_remote.take());
778 :
779 : enum AttachType<'a> {
780 : /// We are attaching this tenant lazily in the background.
781 : Warmup {
782 : _permit: tokio::sync::SemaphorePermit<'a>,
783 : during_startup: bool
784 : },
785 : /// We are attaching this tenant as soon as we can, because for example an
786 : /// endpoint tried to access it.
787 : OnDemand,
788 : /// During normal operations after startup, we are attaching a tenant, and
789 : /// eager attach was requested.
790 : Normal,
791 : }
792 :
793 0 : let attach_type = if matches!(mode, SpawnMode::Lazy) {
794 : // Before doing any I/O, wait for at least one of:
795 : // - A client attempting to access to this tenant (on-demand loading)
796 : // - A permit becoming available in the warmup semaphore (background warmup)
797 :
798 : tokio::select!(
799 : permit = tenant_clone.activate_now_sem.acquire() => {
800 : let _ = permit.expect("activate_now_sem is never closed");
801 : tracing::info!("Activating tenant (on-demand)");
802 : AttachType::OnDemand
803 : },
804 : permit = conf.concurrent_tenant_warmup.inner().acquire() => {
805 : let _permit = permit.expect("concurrent_tenant_warmup semaphore is never closed");
806 : tracing::info!("Activating tenant (warmup)");
807 : AttachType::Warmup {
808 : _permit,
809 : during_startup: init_order.is_some()
810 : }
811 : }
812 : _ = tenant_clone.cancel.cancelled() => {
813 : // This is safe, but should be pretty rare: it is interesting if a tenant
814 : // stayed in Activating for such a long time that shutdown found it in
815 : // that state.
816 : tracing::info!(state=%tenant_clone.current_state(), "Tenant shut down before activation");
817 : // Make the tenant broken so that set_stopping will not hang waiting for it to leave
818 : // the Attaching state. This is an over-reaction (nothing really broke, the tenant is
819 : // just shutting down), but ensures progress.
820 : make_broken(&tenant_clone, anyhow::anyhow!("Shut down while Attaching"), BrokenVerbosity::Info);
821 : return Ok(());
822 : },
823 : )
824 : } else {
825 : // SpawnMode::{Create,Eager} always cause jumping ahead of the
826 : // concurrent_tenant_warmup queue
827 0 : AttachType::Normal
828 : };
829 :
830 0 : let preload = match &mode {
831 : SpawnMode::Eager | SpawnMode::Lazy => {
832 0 : let _preload_timer = TENANT.preload.start_timer();
833 0 : let res = tenant_clone
834 0 : .preload(&remote_storage, task_mgr::shutdown_token())
835 0 : .await;
836 0 : match res {
837 0 : Ok(p) => Some(p),
838 0 : Err(e) => {
839 0 : make_broken(&tenant_clone, anyhow::anyhow!(e), BrokenVerbosity::Error);
840 0 : return Ok(());
841 : }
842 : }
843 : }
844 :
845 : };
846 :
847 : // Remote preload is complete.
848 0 : drop(remote_load_completion);
849 :
850 : // We will time the duration of the attach phase unless this is a creation (attach will do no work)
851 0 : let attached = {
852 0 : let _attach_timer = Some(TENANT.attach.start_timer());
853 0 : tenant_clone.attach(preload, &ctx).await
854 : };
855 :
856 0 : match attached {
857 : Ok(()) => {
858 0 : info!("attach finished, activating");
859 0 : tenant_clone.activate(broker_client, None, &ctx);
860 : }
861 0 : Err(e) => {
862 0 : make_broken(&tenant_clone, anyhow::anyhow!(e), BrokenVerbosity::Error);
863 0 : }
864 : }
865 :
866 : // If we are doing an opportunistic warmup attachment at startup, initialize
867 : // logical size at the same time. This is better than starting a bunch of idle tenants
868 : // with cold caches and then coming back later to initialize their logical sizes.
869 : //
870 : // It also prevents the warmup proccess competing with the concurrency limit on
871 : // logical size calculations: if logical size calculation semaphore is saturated,
872 : // then warmup will wait for that before proceeding to the next tenant.
873 0 : if matches!(attach_type, AttachType::Warmup { during_startup: true, .. }) {
874 0 : let mut futs: FuturesUnordered<_> = tenant_clone.timelines.lock().unwrap().values().cloned().map(|t| t.await_initial_logical_size()).collect();
875 0 : tracing::info!("Waiting for initial logical sizes while warming up...");
876 0 : while futs.next().await.is_some() {}
877 0 : tracing::info!("Warm-up complete");
878 0 : }
879 :
880 0 : Ok(())
881 0 : }
882 0 : .instrument(tracing::info_span!(parent: None, "attach", tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), gen=?generation)),
883 : );
884 0 : tenant
885 0 : }
886 :
887 336 : #[instrument(skip_all)]
888 : pub(crate) async fn preload(
889 : self: &Arc<Self>,
890 : remote_storage: &GenericRemoteStorage,
891 : cancel: CancellationToken,
892 : ) -> anyhow::Result<TenantPreload> {
893 : span::debug_assert_current_span_has_tenant_id();
894 : // Get list of remote timelines
895 : // download index files for every tenant timeline
896 : info!("listing remote timelines");
897 : let (remote_timeline_ids, other_keys) = remote_timeline_client::list_remote_timelines(
898 : remote_storage,
899 : self.tenant_shard_id,
900 : cancel.clone(),
901 : )
902 : .await?;
903 :
904 : info!("found {} timelines", remote_timeline_ids.len(),);
905 :
906 : for k in other_keys {
907 : warn!("Unexpected non timeline key {k}");
908 : }
909 :
910 : Ok(TenantPreload {
911 : timelines: Self::load_timeline_metadata(
912 : self,
913 : remote_timeline_ids,
914 : remote_storage,
915 : cancel,
916 : )
917 : .await?,
918 : })
919 : }
920 :
921 : ///
922 : /// Background task that downloads all data for a tenant and brings it to Active state.
923 : ///
924 : /// No background tasks are started as part of this routine.
925 : ///
926 168 : async fn attach(
927 168 : self: &Arc<Tenant>,
928 168 : preload: Option<TenantPreload>,
929 168 : ctx: &RequestContext,
930 168 : ) -> anyhow::Result<()> {
931 168 : span::debug_assert_current_span_has_tenant_id();
932 168 :
933 168 : failpoint_support::sleep_millis_async!("before-attaching-tenant");
934 :
935 168 : let Some(preload) = preload else {
936 0 : anyhow::bail!("local-only deployment is no longer supported, https://github.com/neondatabase/neon/issues/5624");
937 : };
938 :
939 168 : let mut timelines_to_resume_deletions = vec![];
940 168 :
941 168 : let mut remote_index_and_client = HashMap::new();
942 168 : let mut timeline_ancestors = HashMap::new();
943 168 : let mut existent_timelines = HashSet::new();
944 174 : for (timeline_id, preload) in preload.timelines {
945 6 : let index_part = match preload.index_part {
946 6 : Ok(i) => {
947 6 : debug!("remote index part exists for timeline {timeline_id}");
948 : // We found index_part on the remote, this is the standard case.
949 6 : existent_timelines.insert(timeline_id);
950 6 : i
951 : }
952 : Err(DownloadError::NotFound) => {
953 : // There is no index_part on the remote. We only get here
954 : // if there is some prefix for the timeline in the remote storage.
955 : // This can e.g. be the initdb.tar.zst archive, maybe a
956 : // remnant from a prior incomplete creation or deletion attempt.
957 : // Delete the local directory as the deciding criterion for a
958 : // timeline's existence is presence of index_part.
959 0 : info!(%timeline_id, "index_part not found on remote");
960 0 : continue;
961 : }
962 0 : Err(e) => {
963 0 : // Some (possibly ephemeral) error happened during index_part download.
964 0 : // Pretend the timeline exists to not delete the timeline directory,
965 0 : // as it might be a temporary issue and we don't want to re-download
966 0 : // everything after it resolves.
967 0 : warn!(%timeline_id, "Failed to load index_part from remote storage, failed creation? ({e})");
968 :
969 0 : existent_timelines.insert(timeline_id);
970 0 : continue;
971 : }
972 : };
973 6 : match index_part {
974 6 : MaybeDeletedIndexPart::IndexPart(index_part) => {
975 6 : timeline_ancestors.insert(timeline_id, index_part.metadata.clone());
976 6 : remote_index_and_client.insert(timeline_id, (index_part, preload.client));
977 6 : }
978 0 : MaybeDeletedIndexPart::Deleted(index_part) => {
979 0 : info!(
980 0 : "timeline {} is deleted, picking to resume deletion",
981 : timeline_id
982 : );
983 0 : timelines_to_resume_deletions.push((timeline_id, index_part, preload.client));
984 : }
985 : }
986 : }
987 :
988 : // For every timeline, download the metadata file, scan the local directory,
989 : // and build a layer map that contains an entry for each remote and local
990 : // layer file.
991 168 : let sorted_timelines = tree_sort_timelines(timeline_ancestors, |m| m.ancestor_timeline())?;
992 174 : for (timeline_id, remote_metadata) in sorted_timelines {
993 6 : let (index_part, remote_client) = remote_index_and_client
994 6 : .remove(&timeline_id)
995 6 : .expect("just put it in above");
996 6 :
997 6 : // TODO again handle early failure
998 6 : self.load_remote_timeline(
999 6 : timeline_id,
1000 6 : index_part,
1001 6 : remote_metadata,
1002 6 : TimelineResources {
1003 6 : remote_client,
1004 6 : timeline_get_throttle: self.timeline_get_throttle.clone(),
1005 6 : l0_flush_global_state: self.l0_flush_global_state.clone(),
1006 6 : },
1007 6 : ctx,
1008 6 : )
1009 12 : .await
1010 6 : .with_context(|| {
1011 0 : format!(
1012 0 : "failed to load remote timeline {} for tenant {}",
1013 0 : timeline_id, self.tenant_shard_id
1014 0 : )
1015 6 : })?;
1016 : }
1017 :
1018 : // Walk through deleted timelines, resume deletion
1019 168 : for (timeline_id, index_part, remote_timeline_client) in timelines_to_resume_deletions {
1020 0 : remote_timeline_client
1021 0 : .init_upload_queue_stopped_to_continue_deletion(&index_part)
1022 0 : .context("init queue stopped")
1023 0 : .map_err(LoadLocalTimelineError::ResumeDeletion)?;
1024 :
1025 0 : DeleteTimelineFlow::resume_deletion(
1026 0 : Arc::clone(self),
1027 0 : timeline_id,
1028 0 : &index_part.metadata,
1029 0 : remote_timeline_client,
1030 0 : )
1031 0 : .instrument(tracing::info_span!("timeline_delete", %timeline_id))
1032 0 : .await
1033 0 : .context("resume_deletion")
1034 0 : .map_err(LoadLocalTimelineError::ResumeDeletion)?;
1035 : }
1036 :
1037 : // The local filesystem contents are a cache of what's in the remote IndexPart;
1038 : // IndexPart is the source of truth.
1039 168 : self.clean_up_timelines(&existent_timelines)?;
1040 :
1041 168 : fail::fail_point!("attach-before-activate", |_| {
1042 0 : anyhow::bail!("attach-before-activate");
1043 168 : });
1044 168 : failpoint_support::sleep_millis_async!("attach-before-activate-sleep", &self.cancel);
1045 :
1046 168 : info!("Done");
1047 :
1048 168 : Ok(())
1049 168 : }
1050 :
1051 : /// Check for any local timeline directories that are temporary, or do not correspond to a
1052 : /// timeline that still exists: this can happen if we crashed during a deletion/creation, or
1053 : /// if a timeline was deleted while the tenant was attached to a different pageserver.
1054 168 : fn clean_up_timelines(&self, existent_timelines: &HashSet<TimelineId>) -> anyhow::Result<()> {
1055 168 : let timelines_dir = self.conf.timelines_path(&self.tenant_shard_id);
1056 :
1057 168 : let entries = match timelines_dir.read_dir_utf8() {
1058 168 : Ok(d) => d,
1059 0 : Err(e) => {
1060 0 : if e.kind() == std::io::ErrorKind::NotFound {
1061 0 : return Ok(());
1062 : } else {
1063 0 : return Err(e).context("list timelines directory for tenant");
1064 : }
1065 : }
1066 : };
1067 :
1068 176 : for entry in entries {
1069 8 : let entry = entry.context("read timeline dir entry")?;
1070 8 : let entry_path = entry.path();
1071 :
1072 8 : let purge = if crate::is_temporary(entry_path)
1073 : // TODO: remove uninit mark code (https://github.com/neondatabase/neon/issues/5718)
1074 8 : || is_uninit_mark(entry_path)
1075 8 : || crate::is_delete_mark(entry_path)
1076 : {
1077 0 : true
1078 : } else {
1079 8 : match TimelineId::try_from(entry_path.file_name()) {
1080 8 : Ok(i) => {
1081 8 : // Purge if the timeline ID does not exist in remote storage: remote storage is the authority.
1082 8 : !existent_timelines.contains(&i)
1083 : }
1084 0 : Err(e) => {
1085 0 : tracing::warn!(
1086 0 : "Unparseable directory in timelines directory: {entry_path}, ignoring ({e})"
1087 : );
1088 : // Do not purge junk: if we don't recognize it, be cautious and leave it for a human.
1089 0 : false
1090 : }
1091 : }
1092 : };
1093 :
1094 8 : if purge {
1095 2 : tracing::info!("Purging stale timeline dentry {entry_path}");
1096 2 : if let Err(e) = match entry.file_type() {
1097 2 : Ok(t) => if t.is_dir() {
1098 2 : std::fs::remove_dir_all(entry_path)
1099 : } else {
1100 0 : std::fs::remove_file(entry_path)
1101 : }
1102 2 : .or_else(fs_ext::ignore_not_found),
1103 0 : Err(e) => Err(e),
1104 : } {
1105 0 : tracing::warn!("Failed to purge stale timeline dentry {entry_path}: {e}");
1106 2 : }
1107 6 : }
1108 : }
1109 :
1110 168 : Ok(())
1111 168 : }
1112 :
1113 : /// Get sum of all remote timelines sizes
1114 : ///
1115 : /// This function relies on the index_part instead of listing the remote storage
1116 0 : pub fn remote_size(&self) -> u64 {
1117 0 : let mut size = 0;
1118 :
1119 0 : for timeline in self.list_timelines() {
1120 0 : size += timeline.remote_client.get_remote_physical_size();
1121 0 : }
1122 :
1123 0 : size
1124 0 : }
1125 :
1126 12 : #[instrument(skip_all, fields(timeline_id=%timeline_id))]
1127 : async fn load_remote_timeline(
1128 : &self,
1129 : timeline_id: TimelineId,
1130 : index_part: IndexPart,
1131 : remote_metadata: TimelineMetadata,
1132 : resources: TimelineResources,
1133 : ctx: &RequestContext,
1134 : ) -> anyhow::Result<()> {
1135 : span::debug_assert_current_span_has_tenant_id();
1136 :
1137 : info!("downloading index file for timeline {}", timeline_id);
1138 : tokio::fs::create_dir_all(self.conf.timeline_path(&self.tenant_shard_id, &timeline_id))
1139 : .await
1140 : .context("Failed to create new timeline directory")?;
1141 :
1142 : let ancestor = if let Some(ancestor_id) = remote_metadata.ancestor_timeline() {
1143 : let timelines = self.timelines.lock().unwrap();
1144 : Some(Arc::clone(timelines.get(&ancestor_id).ok_or_else(
1145 0 : || {
1146 0 : anyhow::anyhow!(
1147 0 : "cannot find ancestor timeline {ancestor_id} for timeline {timeline_id}"
1148 0 : )
1149 0 : },
1150 : )?))
1151 : } else {
1152 : None
1153 : };
1154 :
1155 : let last_aux_file_policy = index_part.last_aux_file_policy();
1156 :
1157 : self.timeline_init_and_sync(
1158 : timeline_id,
1159 : resources,
1160 : Some(index_part),
1161 : remote_metadata,
1162 : ancestor,
1163 : last_aux_file_policy,
1164 : ctx,
1165 : )
1166 : .await
1167 : }
1168 :
1169 168 : async fn load_timeline_metadata(
1170 168 : self: &Arc<Tenant>,
1171 168 : timeline_ids: HashSet<TimelineId>,
1172 168 : remote_storage: &GenericRemoteStorage,
1173 168 : cancel: CancellationToken,
1174 168 : ) -> anyhow::Result<HashMap<TimelineId, TimelinePreload>> {
1175 168 : let mut part_downloads = JoinSet::new();
1176 174 : for timeline_id in timeline_ids {
1177 6 : let client = RemoteTimelineClient::new(
1178 6 : remote_storage.clone(),
1179 6 : self.deletion_queue_client.clone(),
1180 6 : self.conf,
1181 6 : self.tenant_shard_id,
1182 6 : timeline_id,
1183 6 : self.generation,
1184 6 : );
1185 6 : let cancel_clone = cancel.clone();
1186 6 : part_downloads.spawn(
1187 6 : async move {
1188 6 : debug!("starting index part download");
1189 :
1190 23 : let index_part = client.download_index_file(&cancel_clone).await;
1191 :
1192 6 : debug!("finished index part download");
1193 :
1194 6 : Result::<_, anyhow::Error>::Ok(TimelinePreload {
1195 6 : client,
1196 6 : timeline_id,
1197 6 : index_part,
1198 6 : })
1199 6 : }
1200 6 : .map(move |res| {
1201 6 : res.with_context(|| format!("download index part for timeline {timeline_id}"))
1202 6 : })
1203 6 : .instrument(info_span!("download_index_part", %timeline_id)),
1204 : );
1205 : }
1206 :
1207 168 : let mut timeline_preloads: HashMap<TimelineId, TimelinePreload> = HashMap::new();
1208 :
1209 174 : loop {
1210 174 : tokio::select!(
1211 : next = part_downloads.join_next() => {
1212 : match next {
1213 : Some(result) => {
1214 : let preload_result = result.context("join preload task")?;
1215 : let preload = preload_result?;
1216 : timeline_preloads.insert(preload.timeline_id, preload);
1217 : },
1218 : None => {
1219 : break;
1220 : }
1221 : }
1222 : },
1223 : _ = cancel.cancelled() => {
1224 : anyhow::bail!("Cancelled while waiting for remote index download")
1225 : }
1226 174 : )
1227 174 : }
1228 :
1229 168 : Ok(timeline_preloads)
1230 168 : }
1231 :
1232 0 : pub async fn apply_timeline_archival_config(
1233 0 : &self,
1234 0 : _timeline_id: TimelineId,
1235 0 : _config: TimelineArchivalState,
1236 0 : ) -> anyhow::Result<()> {
1237 0 : Ok(())
1238 0 : }
1239 :
1240 4 : pub(crate) fn tenant_shard_id(&self) -> TenantShardId {
1241 4 : self.tenant_shard_id
1242 4 : }
1243 :
1244 : /// Get Timeline handle for given Neon timeline ID.
1245 : /// This function is idempotent. It doesn't change internal state in any way.
1246 222 : pub fn get_timeline(
1247 222 : &self,
1248 222 : timeline_id: TimelineId,
1249 222 : active_only: bool,
1250 222 : ) -> Result<Arc<Timeline>, GetTimelineError> {
1251 222 : let timelines_accessor = self.timelines.lock().unwrap();
1252 222 : let timeline = timelines_accessor
1253 222 : .get(&timeline_id)
1254 222 : .ok_or(GetTimelineError::NotFound {
1255 222 : tenant_id: self.tenant_shard_id,
1256 222 : timeline_id,
1257 222 : })?;
1258 :
1259 220 : if active_only && !timeline.is_active() {
1260 0 : Err(GetTimelineError::NotActive {
1261 0 : tenant_id: self.tenant_shard_id,
1262 0 : timeline_id,
1263 0 : state: timeline.current_state(),
1264 0 : })
1265 : } else {
1266 220 : Ok(Arc::clone(timeline))
1267 : }
1268 222 : }
1269 :
1270 : /// Lists timelines the tenant contains.
1271 : /// Up to tenant's implementation to omit certain timelines that ar not considered ready for use.
1272 8 : pub fn list_timelines(&self) -> Vec<Arc<Timeline>> {
1273 8 : self.timelines
1274 8 : .lock()
1275 8 : .unwrap()
1276 8 : .values()
1277 8 : .map(Arc::clone)
1278 8 : .collect()
1279 8 : }
1280 :
1281 0 : pub fn list_timeline_ids(&self) -> Vec<TimelineId> {
1282 0 : self.timelines.lock().unwrap().keys().cloned().collect()
1283 0 : }
1284 :
1285 : /// This is used to create the initial 'main' timeline during bootstrapping,
1286 : /// or when importing a new base backup. The caller is expected to load an
1287 : /// initial image of the datadir to the new timeline after this.
1288 : ///
1289 : /// Until that happens, the on-disk state is invalid (disk_consistent_lsn=Lsn(0))
1290 : /// and the timeline will fail to load at a restart.
1291 : ///
1292 : /// For tests, use `DatadirModification::init_empty_test_timeline` + `commit` to setup the
1293 : /// minimum amount of keys required to get a writable timeline.
1294 : /// (Without it, `put` might fail due to `repartition` failing.)
1295 160 : pub(crate) async fn create_empty_timeline(
1296 160 : &self,
1297 160 : new_timeline_id: TimelineId,
1298 160 : initdb_lsn: Lsn,
1299 160 : pg_version: u32,
1300 160 : _ctx: &RequestContext,
1301 160 : ) -> anyhow::Result<UninitializedTimeline> {
1302 160 : anyhow::ensure!(
1303 160 : self.is_active(),
1304 0 : "Cannot create empty timelines on inactive tenant"
1305 : );
1306 :
1307 : // Protect against concurrent attempts to use this TimelineId
1308 160 : let create_guard = self.create_timeline_create_guard(new_timeline_id)?;
1309 :
1310 158 : let new_metadata = TimelineMetadata::new(
1311 158 : // Initialize disk_consistent LSN to 0, The caller must import some data to
1312 158 : // make it valid, before calling finish_creation()
1313 158 : Lsn(0),
1314 158 : None,
1315 158 : None,
1316 158 : Lsn(0),
1317 158 : initdb_lsn,
1318 158 : initdb_lsn,
1319 158 : pg_version,
1320 158 : );
1321 158 : self.prepare_new_timeline(
1322 158 : new_timeline_id,
1323 158 : &new_metadata,
1324 158 : create_guard,
1325 158 : initdb_lsn,
1326 158 : None,
1327 158 : None,
1328 158 : )
1329 0 : .await
1330 160 : }
1331 :
1332 : /// Helper for unit tests to create an empty timeline.
1333 : ///
1334 : /// The timeline is has state value `Active` but its background loops are not running.
1335 : // This makes the various functions which anyhow::ensure! for Active state work in tests.
1336 : // Our current tests don't need the background loops.
1337 : #[cfg(test)]
1338 150 : pub async fn create_test_timeline(
1339 150 : &self,
1340 150 : new_timeline_id: TimelineId,
1341 150 : initdb_lsn: Lsn,
1342 150 : pg_version: u32,
1343 150 : ctx: &RequestContext,
1344 150 : ) -> anyhow::Result<Arc<Timeline>> {
1345 150 : let uninit_tl = self
1346 150 : .create_empty_timeline(new_timeline_id, initdb_lsn, pg_version, ctx)
1347 0 : .await?;
1348 150 : let tline = uninit_tl.raw_timeline().expect("we just created it");
1349 150 : assert_eq!(tline.get_last_record_lsn(), Lsn(0));
1350 :
1351 : // Setup minimum keys required for the timeline to be usable.
1352 150 : let mut modification = tline.begin_modification(initdb_lsn);
1353 150 : modification
1354 150 : .init_empty_test_timeline()
1355 150 : .context("init_empty_test_timeline")?;
1356 150 : modification
1357 150 : .commit(ctx)
1358 144 : .await
1359 150 : .context("commit init_empty_test_timeline modification")?;
1360 :
1361 : // Flush to disk so that uninit_tl's check for valid disk_consistent_lsn passes.
1362 150 : tline.maybe_spawn_flush_loop();
1363 150 : tline.freeze_and_flush().await.context("freeze_and_flush")?;
1364 :
1365 : // Make sure the freeze_and_flush reaches remote storage.
1366 150 : tline.remote_client.wait_completion().await.unwrap();
1367 :
1368 150 : let tl = uninit_tl.finish_creation()?;
1369 : // The non-test code would call tl.activate() here.
1370 150 : tl.set_state(TimelineState::Active);
1371 150 : Ok(tl)
1372 150 : }
1373 :
1374 : /// Helper for unit tests to create a timeline with some pre-loaded states.
1375 : #[cfg(test)]
1376 : #[allow(clippy::too_many_arguments)]
1377 18 : pub async fn create_test_timeline_with_layers(
1378 18 : &self,
1379 18 : new_timeline_id: TimelineId,
1380 18 : initdb_lsn: Lsn,
1381 18 : pg_version: u32,
1382 18 : ctx: &RequestContext,
1383 18 : delta_layer_desc: Vec<timeline::DeltaLayerTestDesc>,
1384 18 : image_layer_desc: Vec<(Lsn, Vec<(pageserver_api::key::Key, bytes::Bytes)>)>,
1385 18 : end_lsn: Lsn,
1386 18 : ) -> anyhow::Result<Arc<Timeline>> {
1387 18 : let tline = self
1388 18 : .create_test_timeline(new_timeline_id, initdb_lsn, pg_version, ctx)
1389 54 : .await?;
1390 18 : tline.force_advance_lsn(end_lsn);
1391 52 : for deltas in delta_layer_desc {
1392 34 : tline
1393 34 : .force_create_delta_layer(deltas, Some(initdb_lsn), ctx)
1394 102 : .await?;
1395 : }
1396 52 : for (lsn, images) in image_layer_desc {
1397 34 : tline
1398 34 : .force_create_image_layer(lsn, images, Some(initdb_lsn), ctx)
1399 157 : .await?;
1400 : }
1401 18 : Ok(tline)
1402 18 : }
1403 :
1404 : /// Create a new timeline.
1405 : ///
1406 : /// Returns the new timeline ID and reference to its Timeline object.
1407 : ///
1408 : /// If the caller specified the timeline ID to use (`new_timeline_id`), and timeline with
1409 : /// the same timeline ID already exists, returns CreateTimelineError::AlreadyExists.
1410 : #[allow(clippy::too_many_arguments)]
1411 0 : pub(crate) async fn create_timeline(
1412 0 : self: &Arc<Tenant>,
1413 0 : new_timeline_id: TimelineId,
1414 0 : ancestor_timeline_id: Option<TimelineId>,
1415 0 : mut ancestor_start_lsn: Option<Lsn>,
1416 0 : pg_version: u32,
1417 0 : load_existing_initdb: Option<TimelineId>,
1418 0 : broker_client: storage_broker::BrokerClientChannel,
1419 0 : ctx: &RequestContext,
1420 0 : ) -> Result<Arc<Timeline>, CreateTimelineError> {
1421 0 : if !self.is_active() {
1422 0 : if matches!(self.current_state(), TenantState::Stopping { .. }) {
1423 0 : return Err(CreateTimelineError::ShuttingDown);
1424 : } else {
1425 0 : return Err(CreateTimelineError::Other(anyhow::anyhow!(
1426 0 : "Cannot create timelines on inactive tenant"
1427 0 : )));
1428 : }
1429 0 : }
1430 :
1431 0 : let _gate = self
1432 0 : .gate
1433 0 : .enter()
1434 0 : .map_err(|_| CreateTimelineError::ShuttingDown)?;
1435 :
1436 : // Get exclusive access to the timeline ID: this ensures that it does not already exist,
1437 : // and that no other creation attempts will be allowed in while we are working.
1438 0 : let create_guard = match self.create_timeline_create_guard(new_timeline_id) {
1439 0 : Ok(m) => m,
1440 : Err(TimelineExclusionError::AlreadyCreating) => {
1441 : // Creation is in progress, we cannot create it again, and we cannot
1442 : // check if this request matches the existing one, so caller must try
1443 : // again later.
1444 0 : return Err(CreateTimelineError::AlreadyCreating);
1445 : }
1446 0 : Err(TimelineExclusionError::Other(e)) => {
1447 0 : return Err(CreateTimelineError::Other(e));
1448 : }
1449 0 : Err(TimelineExclusionError::AlreadyExists(existing)) => {
1450 0 : debug!("timeline {new_timeline_id} already exists");
1451 :
1452 : // Idempotency: creating the same timeline twice is not an error, unless
1453 : // the second creation has different parameters.
1454 0 : if existing.get_ancestor_timeline_id() != ancestor_timeline_id
1455 0 : || existing.pg_version != pg_version
1456 0 : || (ancestor_start_lsn.is_some()
1457 0 : && ancestor_start_lsn != Some(existing.get_ancestor_lsn()))
1458 : {
1459 0 : return Err(CreateTimelineError::Conflict);
1460 0 : }
1461 0 :
1462 0 : // Wait for uploads to complete, so that when we return Ok, the timeline
1463 0 : // is known to be durable on remote storage. Just like we do at the end of
1464 0 : // this function, after we have created the timeline ourselves.
1465 0 : //
1466 0 : // We only really care that the initial version of `index_part.json` has
1467 0 : // been uploaded. That's enough to remember that the timeline
1468 0 : // exists. However, there is no function to wait specifically for that so
1469 0 : // we just wait for all in-progress uploads to finish.
1470 0 : existing
1471 0 : .remote_client
1472 0 : .wait_completion()
1473 0 : .await
1474 0 : .context("wait for timeline uploads to complete")?;
1475 :
1476 0 : return Ok(existing);
1477 : }
1478 : };
1479 :
1480 : pausable_failpoint!("timeline-creation-after-uninit");
1481 :
1482 0 : let loaded_timeline = match ancestor_timeline_id {
1483 0 : Some(ancestor_timeline_id) => {
1484 0 : let ancestor_timeline = self
1485 0 : .get_timeline(ancestor_timeline_id, false)
1486 0 : .context("Cannot branch off the timeline that's not present in pageserver")?;
1487 :
1488 : // instead of waiting around, just deny the request because ancestor is not yet
1489 : // ready for other purposes either.
1490 0 : if !ancestor_timeline.is_active() {
1491 0 : return Err(CreateTimelineError::AncestorNotActive);
1492 0 : }
1493 :
1494 0 : if let Some(lsn) = ancestor_start_lsn.as_mut() {
1495 0 : *lsn = lsn.align();
1496 0 :
1497 0 : let ancestor_ancestor_lsn = ancestor_timeline.get_ancestor_lsn();
1498 0 : if ancestor_ancestor_lsn > *lsn {
1499 : // can we safely just branch from the ancestor instead?
1500 0 : return Err(CreateTimelineError::AncestorLsn(anyhow::anyhow!(
1501 0 : "invalid start lsn {} for ancestor timeline {}: less than timeline ancestor lsn {}",
1502 0 : lsn,
1503 0 : ancestor_timeline_id,
1504 0 : ancestor_ancestor_lsn,
1505 0 : )));
1506 0 : }
1507 0 :
1508 0 : // Wait for the WAL to arrive and be processed on the parent branch up
1509 0 : // to the requested branch point. The repository code itself doesn't
1510 0 : // require it, but if we start to receive WAL on the new timeline,
1511 0 : // decoding the new WAL might need to look up previous pages, relation
1512 0 : // sizes etc. and that would get confused if the previous page versions
1513 0 : // are not in the repository yet.
1514 0 : ancestor_timeline
1515 0 : .wait_lsn(*lsn, timeline::WaitLsnWaiter::Tenant, ctx)
1516 0 : .await
1517 0 : .map_err(|e| match e {
1518 0 : e @ (WaitLsnError::Timeout(_) | WaitLsnError::BadState { .. }) => {
1519 0 : CreateTimelineError::AncestorLsn(anyhow::anyhow!(e))
1520 : }
1521 0 : WaitLsnError::Shutdown => CreateTimelineError::ShuttingDown,
1522 0 : })?;
1523 0 : }
1524 :
1525 0 : self.branch_timeline(
1526 0 : &ancestor_timeline,
1527 0 : new_timeline_id,
1528 0 : ancestor_start_lsn,
1529 0 : create_guard,
1530 0 : ctx,
1531 0 : )
1532 0 : .await?
1533 : }
1534 : None => {
1535 0 : self.bootstrap_timeline(
1536 0 : new_timeline_id,
1537 0 : pg_version,
1538 0 : load_existing_initdb,
1539 0 : create_guard,
1540 0 : ctx,
1541 0 : )
1542 0 : .await?
1543 : }
1544 : };
1545 :
1546 : // At this point we have dropped our guard on [`Self::timelines_creating`], and
1547 : // the timeline is visible in [`Self::timelines`], but it is _not_ durable yet. We must
1548 : // not send a success to the caller until it is. The same applies to handling retries,
1549 : // see the handling of [`TimelineExclusionError::AlreadyExists`] above.
1550 0 : let kind = ancestor_timeline_id
1551 0 : .map(|_| "branched")
1552 0 : .unwrap_or("bootstrapped");
1553 0 : loaded_timeline
1554 0 : .remote_client
1555 0 : .wait_completion()
1556 0 : .await
1557 0 : .with_context(|| format!("wait for {} timeline initial uploads to complete", kind))?;
1558 :
1559 0 : loaded_timeline.activate(self.clone(), broker_client, None, ctx);
1560 0 :
1561 0 : Ok(loaded_timeline)
1562 0 : }
1563 :
1564 0 : pub(crate) async fn delete_timeline(
1565 0 : self: Arc<Self>,
1566 0 : timeline_id: TimelineId,
1567 0 : ) -> Result<(), DeleteTimelineError> {
1568 0 : DeleteTimelineFlow::run(&self, timeline_id, false).await?;
1569 :
1570 0 : Ok(())
1571 0 : }
1572 :
1573 : /// perform one garbage collection iteration, removing old data files from disk.
1574 : /// this function is periodically called by gc task.
1575 : /// also it can be explicitly requested through page server api 'do_gc' command.
1576 : ///
1577 : /// `target_timeline_id` specifies the timeline to GC, or None for all.
1578 : ///
1579 : /// The `horizon` an `pitr` parameters determine how much WAL history needs to be retained.
1580 : /// Also known as the retention period, or the GC cutoff point. `horizon` specifies
1581 : /// the amount of history, as LSN difference from current latest LSN on each timeline.
1582 : /// `pitr` specifies the same as a time difference from the current time. The effective
1583 : /// GC cutoff point is determined conservatively by either `horizon` and `pitr`, whichever
1584 : /// requires more history to be retained.
1585 : //
1586 754 : pub(crate) async fn gc_iteration(
1587 754 : &self,
1588 754 : target_timeline_id: Option<TimelineId>,
1589 754 : horizon: u64,
1590 754 : pitr: Duration,
1591 754 : cancel: &CancellationToken,
1592 754 : ctx: &RequestContext,
1593 754 : ) -> Result<GcResult, GcError> {
1594 754 : // Don't start doing work during shutdown
1595 754 : if let TenantState::Stopping { .. } = self.current_state() {
1596 0 : return Ok(GcResult::default());
1597 754 : }
1598 754 :
1599 754 : // there is a global allowed_error for this
1600 754 : if !self.is_active() {
1601 0 : return Err(GcError::NotActive);
1602 754 : }
1603 754 :
1604 754 : {
1605 754 : let conf = self.tenant_conf.load();
1606 754 :
1607 754 : if !conf.location.may_delete_layers_hint() {
1608 0 : info!("Skipping GC in location state {:?}", conf.location);
1609 0 : return Ok(GcResult::default());
1610 754 : }
1611 754 : }
1612 754 :
1613 754 : self.gc_iteration_internal(target_timeline_id, horizon, pitr, cancel, ctx)
1614 743 : .await
1615 754 : }
1616 :
1617 : /// Perform one compaction iteration.
1618 : /// This function is periodically called by compactor task.
1619 : /// Also it can be explicitly requested per timeline through page server
1620 : /// api's 'compact' command.
1621 0 : async fn compaction_iteration(
1622 0 : &self,
1623 0 : cancel: &CancellationToken,
1624 0 : ctx: &RequestContext,
1625 0 : ) -> anyhow::Result<(), timeline::CompactionError> {
1626 0 : // Don't start doing work during shutdown, or when broken, we do not need those in the logs
1627 0 : if !self.is_active() {
1628 0 : return Ok(());
1629 0 : }
1630 0 :
1631 0 : {
1632 0 : let conf = self.tenant_conf.load();
1633 0 : if !conf.location.may_delete_layers_hint() || !conf.location.may_upload_layers_hint() {
1634 0 : info!("Skipping compaction in location state {:?}", conf.location);
1635 0 : return Ok(());
1636 0 : }
1637 0 : }
1638 0 :
1639 0 : // Scan through the hashmap and collect a list of all the timelines,
1640 0 : // while holding the lock. Then drop the lock and actually perform the
1641 0 : // compactions. We don't want to block everything else while the
1642 0 : // compaction runs.
1643 0 : let timelines_to_compact = {
1644 0 : let timelines = self.timelines.lock().unwrap();
1645 0 : let timelines_to_compact = timelines
1646 0 : .iter()
1647 0 : .filter_map(|(timeline_id, timeline)| {
1648 0 : if timeline.is_active() {
1649 0 : Some((*timeline_id, timeline.clone()))
1650 : } else {
1651 0 : None
1652 : }
1653 0 : })
1654 0 : .collect::<Vec<_>>();
1655 0 : drop(timelines);
1656 0 : timelines_to_compact
1657 0 : };
1658 0 :
1659 0 : // Before doing any I/O work, check our circuit breaker
1660 0 : if self.compaction_circuit_breaker.lock().unwrap().is_broken() {
1661 0 : info!("Skipping compaction due to previous failures");
1662 0 : return Ok(());
1663 0 : }
1664 :
1665 0 : for (timeline_id, timeline) in &timelines_to_compact {
1666 0 : timeline
1667 0 : .compact(cancel, EnumSet::empty(), ctx)
1668 0 : .instrument(info_span!("compact_timeline", %timeline_id))
1669 0 : .await
1670 0 : .map_err(|e| {
1671 0 : self.compaction_circuit_breaker
1672 0 : .lock()
1673 0 : .unwrap()
1674 0 : .fail(&CIRCUIT_BREAKERS_BROKEN, &e);
1675 0 : e
1676 0 : })?;
1677 : }
1678 :
1679 0 : self.compaction_circuit_breaker
1680 0 : .lock()
1681 0 : .unwrap()
1682 0 : .success(&CIRCUIT_BREAKERS_UNBROKEN);
1683 0 :
1684 0 : Ok(())
1685 0 : }
1686 :
1687 : // Call through to all timelines to freeze ephemeral layers if needed. Usually
1688 : // this happens during ingest: this background housekeeping is for freezing layers
1689 : // that are open but haven't been written to for some time.
1690 0 : async fn ingest_housekeeping(&self) {
1691 0 : // Scan through the hashmap and collect a list of all the timelines,
1692 0 : // while holding the lock. Then drop the lock and actually perform the
1693 0 : // compactions. We don't want to block everything else while the
1694 0 : // compaction runs.
1695 0 : let timelines = {
1696 0 : self.timelines
1697 0 : .lock()
1698 0 : .unwrap()
1699 0 : .values()
1700 0 : .filter_map(|timeline| {
1701 0 : if timeline.is_active() {
1702 0 : Some(timeline.clone())
1703 : } else {
1704 0 : None
1705 : }
1706 0 : })
1707 0 : .collect::<Vec<_>>()
1708 : };
1709 :
1710 0 : for timeline in &timelines {
1711 0 : timeline.maybe_freeze_ephemeral_layer().await;
1712 : }
1713 0 : }
1714 :
1715 2428 : pub fn current_state(&self) -> TenantState {
1716 2428 : self.state.borrow().clone()
1717 2428 : }
1718 :
1719 1668 : pub fn is_active(&self) -> bool {
1720 1668 : self.current_state() == TenantState::Active
1721 1668 : }
1722 :
1723 0 : pub fn generation(&self) -> Generation {
1724 0 : self.generation
1725 0 : }
1726 :
1727 0 : pub(crate) fn wal_redo_manager_status(&self) -> Option<WalRedoManagerStatus> {
1728 0 : self.walredo_mgr.as_ref().and_then(|mgr| mgr.status())
1729 0 : }
1730 :
1731 : /// Changes tenant status to active, unless shutdown was already requested.
1732 : ///
1733 : /// `background_jobs_can_start` is an optional barrier set to a value during pageserver startup
1734 : /// to delay background jobs. Background jobs can be started right away when None is given.
1735 0 : fn activate(
1736 0 : self: &Arc<Self>,
1737 0 : broker_client: BrokerClientChannel,
1738 0 : background_jobs_can_start: Option<&completion::Barrier>,
1739 0 : ctx: &RequestContext,
1740 0 : ) {
1741 0 : span::debug_assert_current_span_has_tenant_id();
1742 0 :
1743 0 : let mut activating = false;
1744 0 : self.state.send_modify(|current_state| {
1745 0 : use pageserver_api::models::ActivatingFrom;
1746 0 : match &*current_state {
1747 : TenantState::Activating(_) | TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => {
1748 0 : panic!("caller is responsible for calling activate() only on Loading / Attaching tenants, got {state:?}", state = current_state);
1749 : }
1750 0 : TenantState::Loading => {
1751 0 : *current_state = TenantState::Activating(ActivatingFrom::Loading);
1752 0 : }
1753 0 : TenantState::Attaching => {
1754 0 : *current_state = TenantState::Activating(ActivatingFrom::Attaching);
1755 0 : }
1756 : }
1757 0 : debug!(tenant_id = %self.tenant_shard_id.tenant_id, shard_id = %self.tenant_shard_id.shard_slug(), "Activating tenant");
1758 0 : activating = true;
1759 0 : // Continue outside the closure. We need to grab timelines.lock()
1760 0 : // and we plan to turn it into a tokio::sync::Mutex in a future patch.
1761 0 : });
1762 0 :
1763 0 : if activating {
1764 0 : let timelines_accessor = self.timelines.lock().unwrap();
1765 0 : let timelines_to_activate = timelines_accessor
1766 0 : .values()
1767 0 : .filter(|timeline| !(timeline.is_broken() || timeline.is_stopping()));
1768 0 :
1769 0 : // Spawn gc and compaction loops. The loops will shut themselves
1770 0 : // down when they notice that the tenant is inactive.
1771 0 : tasks::start_background_loops(self, background_jobs_can_start);
1772 0 :
1773 0 : let mut activated_timelines = 0;
1774 :
1775 0 : for timeline in timelines_to_activate {
1776 0 : timeline.activate(
1777 0 : self.clone(),
1778 0 : broker_client.clone(),
1779 0 : background_jobs_can_start,
1780 0 : ctx,
1781 0 : );
1782 0 : activated_timelines += 1;
1783 0 : }
1784 :
1785 0 : self.state.send_modify(move |current_state| {
1786 0 : assert!(
1787 0 : matches!(current_state, TenantState::Activating(_)),
1788 0 : "set_stopping and set_broken wait for us to leave Activating state",
1789 : );
1790 0 : *current_state = TenantState::Active;
1791 0 :
1792 0 : let elapsed = self.constructed_at.elapsed();
1793 0 : let total_timelines = timelines_accessor.len();
1794 0 :
1795 0 : // log a lot of stuff, because some tenants sometimes suffer from user-visible
1796 0 : // times to activate. see https://github.com/neondatabase/neon/issues/4025
1797 0 : info!(
1798 0 : since_creation_millis = elapsed.as_millis(),
1799 0 : tenant_id = %self.tenant_shard_id.tenant_id,
1800 0 : shard_id = %self.tenant_shard_id.shard_slug(),
1801 0 : activated_timelines,
1802 0 : total_timelines,
1803 0 : post_state = <&'static str>::from(&*current_state),
1804 0 : "activation attempt finished"
1805 : );
1806 :
1807 0 : TENANT.activation.observe(elapsed.as_secs_f64());
1808 0 : });
1809 0 : }
1810 0 : }
1811 :
1812 : /// Shutdown the tenant and join all of the spawned tasks.
1813 : ///
1814 : /// The method caters for all use-cases:
1815 : /// - pageserver shutdown (freeze_and_flush == true)
1816 : /// - detach + ignore (freeze_and_flush == false)
1817 : ///
1818 : /// This will attempt to shutdown even if tenant is broken.
1819 : ///
1820 : /// `shutdown_progress` is a [`completion::Barrier`] for the shutdown initiated by this call.
1821 : /// If the tenant is already shutting down, we return a clone of the first shutdown call's
1822 : /// `Barrier` as an `Err`. This not-first caller can use the returned barrier to join with
1823 : /// the ongoing shutdown.
1824 6 : async fn shutdown(
1825 6 : &self,
1826 6 : shutdown_progress: completion::Barrier,
1827 6 : shutdown_mode: timeline::ShutdownMode,
1828 6 : ) -> Result<(), completion::Barrier> {
1829 6 : span::debug_assert_current_span_has_tenant_id();
1830 :
1831 : // Set tenant (and its timlines) to Stoppping state.
1832 : //
1833 : // Since we can only transition into Stopping state after activation is complete,
1834 : // run it in a JoinSet so all tenants have a chance to stop before we get SIGKILLed.
1835 : //
1836 : // Transitioning tenants to Stopping state has a couple of non-obvious side effects:
1837 : // 1. Lock out any new requests to the tenants.
1838 : // 2. Signal cancellation to WAL receivers (we wait on it below).
1839 : // 3. Signal cancellation for other tenant background loops.
1840 : // 4. ???
1841 : //
1842 : // The waiting for the cancellation is not done uniformly.
1843 : // We certainly wait for WAL receivers to shut down.
1844 : // That is necessary so that no new data comes in before the freeze_and_flush.
1845 : // But the tenant background loops are joined-on in our caller.
1846 : // It's mesed up.
1847 : // we just ignore the failure to stop
1848 :
1849 : // If we're still attaching, fire the cancellation token early to drop out: this
1850 : // will prevent us flushing, but ensures timely shutdown if some I/O during attach
1851 : // is very slow.
1852 6 : let shutdown_mode = if matches!(self.current_state(), TenantState::Attaching) {
1853 0 : self.cancel.cancel();
1854 0 :
1855 0 : // Having fired our cancellation token, do not try and flush timelines: their cancellation tokens
1856 0 : // are children of ours, so their flush loops will have shut down already
1857 0 : timeline::ShutdownMode::Hard
1858 : } else {
1859 6 : shutdown_mode
1860 : };
1861 :
1862 6 : match self.set_stopping(shutdown_progress, false, false).await {
1863 6 : Ok(()) => {}
1864 0 : Err(SetStoppingError::Broken) => {
1865 0 : // assume that this is acceptable
1866 0 : }
1867 0 : Err(SetStoppingError::AlreadyStopping(other)) => {
1868 0 : // give caller the option to wait for this this shutdown
1869 0 : info!("Tenant::shutdown: AlreadyStopping");
1870 0 : return Err(other);
1871 : }
1872 : };
1873 :
1874 6 : let mut js = tokio::task::JoinSet::new();
1875 6 : {
1876 6 : let timelines = self.timelines.lock().unwrap();
1877 6 : timelines.values().for_each(|timeline| {
1878 6 : let timeline = Arc::clone(timeline);
1879 6 : let timeline_id = timeline.timeline_id;
1880 6 : let span = tracing::info_span!("timeline_shutdown", %timeline_id, ?shutdown_mode);
1881 16 : js.spawn(async move { timeline.shutdown(shutdown_mode).instrument(span).await });
1882 6 : })
1883 6 : };
1884 6 : // test_long_timeline_create_then_tenant_delete is leaning on this message
1885 6 : tracing::info!("Waiting for timelines...");
1886 12 : while let Some(res) = js.join_next().await {
1887 0 : match res {
1888 6 : Ok(()) => {}
1889 0 : Err(je) if je.is_cancelled() => unreachable!("no cancelling used"),
1890 0 : Err(je) if je.is_panic() => { /* logged already */ }
1891 0 : Err(je) => warn!("unexpected JoinError: {je:?}"),
1892 : }
1893 : }
1894 :
1895 : // We cancel the Tenant's cancellation token _after_ the timelines have all shut down. This permits
1896 : // them to continue to do work during their shutdown methods, e.g. flushing data.
1897 6 : tracing::debug!("Cancelling CancellationToken");
1898 6 : self.cancel.cancel();
1899 6 :
1900 6 : // shutdown all tenant and timeline tasks: gc, compaction, page service
1901 6 : // No new tasks will be started for this tenant because it's in `Stopping` state.
1902 6 : //
1903 6 : // this will additionally shutdown and await all timeline tasks.
1904 6 : tracing::debug!("Waiting for tasks...");
1905 6 : task_mgr::shutdown_tasks(None, Some(self.tenant_shard_id), None).await;
1906 :
1907 6 : if let Some(walredo_mgr) = self.walredo_mgr.as_ref() {
1908 6 : walredo_mgr.shutdown().await;
1909 0 : }
1910 :
1911 : // Wait for any in-flight operations to complete
1912 6 : self.gate.close().await;
1913 :
1914 6 : remove_tenant_metrics(&self.tenant_shard_id);
1915 6 :
1916 6 : Ok(())
1917 6 : }
1918 :
1919 : /// Change tenant status to Stopping, to mark that it is being shut down.
1920 : ///
1921 : /// This function waits for the tenant to become active if it isn't already, before transitioning it into Stopping state.
1922 : ///
1923 : /// This function is not cancel-safe!
1924 : ///
1925 : /// `allow_transition_from_loading` is needed for the special case of loading task deleting the tenant.
1926 : /// `allow_transition_from_attaching` is needed for the special case of attaching deleted tenant.
1927 6 : async fn set_stopping(
1928 6 : &self,
1929 6 : progress: completion::Barrier,
1930 6 : allow_transition_from_loading: bool,
1931 6 : allow_transition_from_attaching: bool,
1932 6 : ) -> Result<(), SetStoppingError> {
1933 6 : let mut rx = self.state.subscribe();
1934 6 :
1935 6 : // cannot stop before we're done activating, so wait out until we're done activating
1936 6 : rx.wait_for(|state| match state {
1937 0 : TenantState::Attaching if allow_transition_from_attaching => true,
1938 : TenantState::Activating(_) | TenantState::Attaching => {
1939 0 : info!(
1940 0 : "waiting for {} to turn Active|Broken|Stopping",
1941 0 : <&'static str>::from(state)
1942 : );
1943 0 : false
1944 : }
1945 0 : TenantState::Loading => allow_transition_from_loading,
1946 6 : TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => true,
1947 6 : })
1948 0 : .await
1949 6 : .expect("cannot drop self.state while on a &self method");
1950 6 :
1951 6 : // we now know we're done activating, let's see whether this task is the winner to transition into Stopping
1952 6 : let mut err = None;
1953 6 : let stopping = self.state.send_if_modified(|current_state| match current_state {
1954 : TenantState::Activating(_) => {
1955 0 : unreachable!("1we ensured above that we're done with activation, and, there is no re-activation")
1956 : }
1957 : TenantState::Attaching => {
1958 0 : if !allow_transition_from_attaching {
1959 0 : unreachable!("2we ensured above that we're done with activation, and, there is no re-activation")
1960 0 : };
1961 0 : *current_state = TenantState::Stopping { progress };
1962 0 : true
1963 : }
1964 : TenantState::Loading => {
1965 0 : if !allow_transition_from_loading {
1966 0 : unreachable!("3we ensured above that we're done with activation, and, there is no re-activation")
1967 0 : };
1968 0 : *current_state = TenantState::Stopping { progress };
1969 0 : true
1970 : }
1971 : TenantState::Active => {
1972 : // FIXME: due to time-of-check vs time-of-use issues, it can happen that new timelines
1973 : // are created after the transition to Stopping. That's harmless, as the Timelines
1974 : // won't be accessible to anyone afterwards, because the Tenant is in Stopping state.
1975 6 : *current_state = TenantState::Stopping { progress };
1976 6 : // Continue stopping outside the closure. We need to grab timelines.lock()
1977 6 : // and we plan to turn it into a tokio::sync::Mutex in a future patch.
1978 6 : true
1979 : }
1980 0 : TenantState::Broken { reason, .. } => {
1981 0 : info!(
1982 0 : "Cannot set tenant to Stopping state, it is in Broken state due to: {reason}"
1983 : );
1984 0 : err = Some(SetStoppingError::Broken);
1985 0 : false
1986 : }
1987 0 : TenantState::Stopping { progress } => {
1988 0 : info!("Tenant is already in Stopping state");
1989 0 : err = Some(SetStoppingError::AlreadyStopping(progress.clone()));
1990 0 : false
1991 : }
1992 6 : });
1993 6 : match (stopping, err) {
1994 6 : (true, None) => {} // continue
1995 0 : (false, Some(err)) => return Err(err),
1996 0 : (true, Some(_)) => unreachable!(
1997 0 : "send_if_modified closure must error out if not transitioning to Stopping"
1998 0 : ),
1999 0 : (false, None) => unreachable!(
2000 0 : "send_if_modified closure must return true if transitioning to Stopping"
2001 0 : ),
2002 : }
2003 :
2004 6 : let timelines_accessor = self.timelines.lock().unwrap();
2005 6 : let not_broken_timelines = timelines_accessor
2006 6 : .values()
2007 6 : .filter(|timeline| !timeline.is_broken());
2008 12 : for timeline in not_broken_timelines {
2009 6 : timeline.set_state(TimelineState::Stopping);
2010 6 : }
2011 6 : Ok(())
2012 6 : }
2013 :
2014 : /// Method for tenant::mgr to transition us into Broken state in case of a late failure in
2015 : /// `remove_tenant_from_memory`
2016 : ///
2017 : /// This function waits for the tenant to become active if it isn't already, before transitioning it into Stopping state.
2018 : ///
2019 : /// In tests, we also use this to set tenants to Broken state on purpose.
2020 0 : pub(crate) async fn set_broken(&self, reason: String) {
2021 0 : let mut rx = self.state.subscribe();
2022 0 :
2023 0 : // The load & attach routines own the tenant state until it has reached `Active`.
2024 0 : // So, wait until it's done.
2025 0 : rx.wait_for(|state| match state {
2026 : TenantState::Activating(_) | TenantState::Loading | TenantState::Attaching => {
2027 0 : info!(
2028 0 : "waiting for {} to turn Active|Broken|Stopping",
2029 0 : <&'static str>::from(state)
2030 : );
2031 0 : false
2032 : }
2033 0 : TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => true,
2034 0 : })
2035 0 : .await
2036 0 : .expect("cannot drop self.state while on a &self method");
2037 0 :
2038 0 : // we now know we're done activating, let's see whether this task is the winner to transition into Broken
2039 0 : self.set_broken_no_wait(reason)
2040 0 : }
2041 :
2042 0 : pub(crate) fn set_broken_no_wait(&self, reason: impl Display) {
2043 0 : let reason = reason.to_string();
2044 0 : self.state.send_modify(|current_state| {
2045 0 : match *current_state {
2046 : TenantState::Activating(_) | TenantState::Loading | TenantState::Attaching => {
2047 0 : unreachable!("we ensured above that we're done with activation, and, there is no re-activation")
2048 : }
2049 : TenantState::Active => {
2050 0 : if cfg!(feature = "testing") {
2051 0 : warn!("Changing Active tenant to Broken state, reason: {}", reason);
2052 0 : *current_state = TenantState::broken_from_reason(reason);
2053 : } else {
2054 0 : unreachable!("not allowed to call set_broken on Active tenants in non-testing builds")
2055 : }
2056 : }
2057 : TenantState::Broken { .. } => {
2058 0 : warn!("Tenant is already in Broken state");
2059 : }
2060 : // This is the only "expected" path, any other path is a bug.
2061 : TenantState::Stopping { .. } => {
2062 0 : warn!(
2063 0 : "Marking Stopping tenant as Broken state, reason: {}",
2064 : reason
2065 : );
2066 0 : *current_state = TenantState::broken_from_reason(reason);
2067 : }
2068 : }
2069 0 : });
2070 0 : }
2071 :
2072 0 : pub fn subscribe_for_state_updates(&self) -> watch::Receiver<TenantState> {
2073 0 : self.state.subscribe()
2074 0 : }
2075 :
2076 : /// The activate_now semaphore is initialized with zero units. As soon as
2077 : /// we add a unit, waiters will be able to acquire a unit and proceed.
2078 0 : pub(crate) fn activate_now(&self) {
2079 0 : self.activate_now_sem.add_permits(1);
2080 0 : }
2081 :
2082 0 : pub(crate) async fn wait_to_become_active(
2083 0 : &self,
2084 0 : timeout: Duration,
2085 0 : ) -> Result<(), GetActiveTenantError> {
2086 0 : let mut receiver = self.state.subscribe();
2087 0 : loop {
2088 0 : let current_state = receiver.borrow_and_update().clone();
2089 0 : match current_state {
2090 : TenantState::Loading | TenantState::Attaching | TenantState::Activating(_) => {
2091 : // in these states, there's a chance that we can reach ::Active
2092 0 : self.activate_now();
2093 0 : match timeout_cancellable(timeout, &self.cancel, receiver.changed()).await {
2094 0 : Ok(r) => {
2095 0 : r.map_err(
2096 0 : |_e: tokio::sync::watch::error::RecvError|
2097 : // Tenant existed but was dropped: report it as non-existent
2098 0 : GetActiveTenantError::NotFound(GetTenantError::NotFound(self.tenant_shard_id.tenant_id))
2099 0 : )?
2100 : }
2101 : Err(TimeoutCancellableError::Cancelled) => {
2102 0 : return Err(GetActiveTenantError::Cancelled);
2103 : }
2104 : Err(TimeoutCancellableError::Timeout) => {
2105 0 : return Err(GetActiveTenantError::WaitForActiveTimeout {
2106 0 : latest_state: Some(self.current_state()),
2107 0 : wait_time: timeout,
2108 0 : });
2109 : }
2110 : }
2111 : }
2112 : TenantState::Active { .. } => {
2113 0 : return Ok(());
2114 : }
2115 0 : TenantState::Broken { reason, .. } => {
2116 0 : // This is fatal, and reported distinctly from the general case of "will never be active" because
2117 0 : // it's logically a 500 to external API users (broken is always a bug).
2118 0 : return Err(GetActiveTenantError::Broken(reason));
2119 : }
2120 : TenantState::Stopping { .. } => {
2121 : // There's no chance the tenant can transition back into ::Active
2122 0 : return Err(GetActiveTenantError::WillNotBecomeActive(current_state));
2123 : }
2124 : }
2125 : }
2126 0 : }
2127 :
2128 0 : pub(crate) fn get_attach_mode(&self) -> AttachmentMode {
2129 0 : self.tenant_conf.load().location.attach_mode
2130 0 : }
2131 :
2132 : /// For API access: generate a LocationConfig equivalent to the one that would be used to
2133 : /// create a Tenant in the same state. Do not use this in hot paths: it's for relatively
2134 : /// rare external API calls, like a reconciliation at startup.
2135 0 : pub(crate) fn get_location_conf(&self) -> models::LocationConfig {
2136 0 : let conf = self.tenant_conf.load();
2137 :
2138 0 : let location_config_mode = match conf.location.attach_mode {
2139 0 : AttachmentMode::Single => models::LocationConfigMode::AttachedSingle,
2140 0 : AttachmentMode::Multi => models::LocationConfigMode::AttachedMulti,
2141 0 : AttachmentMode::Stale => models::LocationConfigMode::AttachedStale,
2142 : };
2143 :
2144 : // We have a pageserver TenantConf, we need the API-facing TenantConfig.
2145 0 : let tenant_config: models::TenantConfig = conf.tenant_conf.clone().into();
2146 0 :
2147 0 : models::LocationConfig {
2148 0 : mode: location_config_mode,
2149 0 : generation: self.generation.into(),
2150 0 : secondary_conf: None,
2151 0 : shard_number: self.shard_identity.number.0,
2152 0 : shard_count: self.shard_identity.count.literal(),
2153 0 : shard_stripe_size: self.shard_identity.stripe_size.0,
2154 0 : tenant_conf: tenant_config,
2155 0 : }
2156 0 : }
2157 :
2158 0 : pub(crate) fn get_tenant_shard_id(&self) -> &TenantShardId {
2159 0 : &self.tenant_shard_id
2160 0 : }
2161 :
2162 0 : pub(crate) fn get_shard_stripe_size(&self) -> ShardStripeSize {
2163 0 : self.shard_identity.stripe_size
2164 0 : }
2165 :
2166 0 : pub(crate) fn get_generation(&self) -> Generation {
2167 0 : self.generation
2168 0 : }
2169 :
2170 : /// This function partially shuts down the tenant (it shuts down the Timelines) and is fallible,
2171 : /// and can leave the tenant in a bad state if it fails. The caller is responsible for
2172 : /// resetting this tenant to a valid state if we fail.
2173 0 : pub(crate) async fn split_prepare(
2174 0 : &self,
2175 0 : child_shards: &Vec<TenantShardId>,
2176 0 : ) -> anyhow::Result<()> {
2177 0 : let timelines = self.timelines.lock().unwrap().clone();
2178 0 : for timeline in timelines.values() {
2179 : // We do not block timeline creation/deletion during splits inside the pageserver: it is up to higher levels
2180 : // to ensure that they do not start a split if currently in the process of doing these.
2181 :
2182 : // Upload an index from the parent: this is partly to provide freshness for the
2183 : // child tenants that will copy it, and partly for general ease-of-debugging: there will
2184 : // always be a parent shard index in the same generation as we wrote the child shard index.
2185 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Uploading index");
2186 0 : timeline
2187 0 : .remote_client
2188 0 : .schedule_index_upload_for_file_changes()?;
2189 0 : timeline.remote_client.wait_completion().await?;
2190 :
2191 : // Shut down the timeline's remote client: this means that the indices we write
2192 : // for child shards will not be invalidated by the parent shard deleting layers.
2193 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Shutting down remote storage client");
2194 0 : timeline.remote_client.shutdown().await;
2195 :
2196 : // Download methods can still be used after shutdown, as they don't flow through the remote client's
2197 : // queue. In principal the RemoteTimelineClient could provide this without downloading it, but this
2198 : // operation is rare, so it's simpler to just download it (and robustly guarantees that the index
2199 : // we use here really is the remotely persistent one).
2200 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Downloading index_part from parent");
2201 0 : let result = timeline.remote_client
2202 0 : .download_index_file(&self.cancel)
2203 0 : .instrument(info_span!("download_index_file", tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug(), timeline_id=%timeline.timeline_id))
2204 0 : .await?;
2205 0 : let index_part = match result {
2206 : MaybeDeletedIndexPart::Deleted(_) => {
2207 0 : anyhow::bail!("Timeline deletion happened concurrently with split")
2208 : }
2209 0 : MaybeDeletedIndexPart::IndexPart(p) => p,
2210 : };
2211 :
2212 0 : for child_shard in child_shards {
2213 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Uploading index_part for child {}", child_shard.to_index());
2214 0 : upload_index_part(
2215 0 : &self.remote_storage,
2216 0 : child_shard,
2217 0 : &timeline.timeline_id,
2218 0 : self.generation,
2219 0 : &index_part,
2220 0 : &self.cancel,
2221 0 : )
2222 0 : .await?;
2223 : }
2224 : }
2225 :
2226 0 : Ok(())
2227 0 : }
2228 :
2229 0 : pub(crate) fn get_sizes(&self) -> TopTenantShardItem {
2230 0 : let mut result = TopTenantShardItem {
2231 0 : id: self.tenant_shard_id,
2232 0 : resident_size: 0,
2233 0 : physical_size: 0,
2234 0 : max_logical_size: 0,
2235 0 : };
2236 :
2237 0 : for timeline in self.timelines.lock().unwrap().values() {
2238 0 : result.resident_size += timeline.metrics.resident_physical_size_gauge.get();
2239 0 :
2240 0 : result.physical_size += timeline
2241 0 : .remote_client
2242 0 : .metrics
2243 0 : .remote_physical_size_gauge
2244 0 : .get();
2245 0 : result.max_logical_size = std::cmp::max(
2246 0 : result.max_logical_size,
2247 0 : timeline.metrics.current_logical_size_gauge.get(),
2248 0 : );
2249 0 : }
2250 :
2251 0 : result
2252 0 : }
2253 : }
2254 :
2255 : /// Given a Vec of timelines and their ancestors (timeline_id, ancestor_id),
2256 : /// perform a topological sort, so that the parent of each timeline comes
2257 : /// before the children.
2258 : /// E extracts the ancestor from T
2259 : /// This allows for T to be different. It can be TimelineMetadata, can be Timeline itself, etc.
2260 168 : fn tree_sort_timelines<T, E>(
2261 168 : timelines: HashMap<TimelineId, T>,
2262 168 : extractor: E,
2263 168 : ) -> anyhow::Result<Vec<(TimelineId, T)>>
2264 168 : where
2265 168 : E: Fn(&T) -> Option<TimelineId>,
2266 168 : {
2267 168 : let mut result = Vec::with_capacity(timelines.len());
2268 168 :
2269 168 : let mut now = Vec::with_capacity(timelines.len());
2270 168 : // (ancestor, children)
2271 168 : let mut later: HashMap<TimelineId, Vec<(TimelineId, T)>> =
2272 168 : HashMap::with_capacity(timelines.len());
2273 :
2274 174 : for (timeline_id, value) in timelines {
2275 6 : if let Some(ancestor_id) = extractor(&value) {
2276 2 : let children = later.entry(ancestor_id).or_default();
2277 2 : children.push((timeline_id, value));
2278 4 : } else {
2279 4 : now.push((timeline_id, value));
2280 4 : }
2281 : }
2282 :
2283 174 : while let Some((timeline_id, metadata)) = now.pop() {
2284 6 : result.push((timeline_id, metadata));
2285 : // All children of this can be loaded now
2286 6 : if let Some(mut children) = later.remove(&timeline_id) {
2287 2 : now.append(&mut children);
2288 4 : }
2289 : }
2290 :
2291 : // All timelines should be visited now. Unless there were timelines with missing ancestors.
2292 168 : if !later.is_empty() {
2293 0 : for (missing_id, orphan_ids) in later {
2294 0 : for (orphan_id, _) in orphan_ids {
2295 0 : error!("could not load timeline {orphan_id} because its ancestor timeline {missing_id} could not be loaded");
2296 : }
2297 : }
2298 0 : bail!("could not load tenant because some timelines are missing ancestors");
2299 168 : }
2300 168 :
2301 168 : Ok(result)
2302 168 : }
2303 :
2304 : impl Tenant {
2305 0 : pub fn tenant_specific_overrides(&self) -> TenantConfOpt {
2306 0 : self.tenant_conf.load().tenant_conf.clone()
2307 0 : }
2308 :
2309 0 : pub fn effective_config(&self) -> TenantConf {
2310 0 : self.tenant_specific_overrides()
2311 0 : .merge(self.conf.default_tenant_conf.clone())
2312 0 : }
2313 :
2314 0 : pub fn get_checkpoint_distance(&self) -> u64 {
2315 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2316 0 : tenant_conf
2317 0 : .checkpoint_distance
2318 0 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_distance)
2319 0 : }
2320 :
2321 0 : pub fn get_checkpoint_timeout(&self) -> Duration {
2322 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2323 0 : tenant_conf
2324 0 : .checkpoint_timeout
2325 0 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_timeout)
2326 0 : }
2327 :
2328 0 : pub fn get_compaction_target_size(&self) -> u64 {
2329 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2330 0 : tenant_conf
2331 0 : .compaction_target_size
2332 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_target_size)
2333 0 : }
2334 :
2335 0 : pub fn get_compaction_period(&self) -> Duration {
2336 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2337 0 : tenant_conf
2338 0 : .compaction_period
2339 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_period)
2340 0 : }
2341 :
2342 0 : pub fn get_compaction_threshold(&self) -> usize {
2343 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2344 0 : tenant_conf
2345 0 : .compaction_threshold
2346 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_threshold)
2347 0 : }
2348 :
2349 0 : pub fn get_gc_horizon(&self) -> u64 {
2350 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2351 0 : tenant_conf
2352 0 : .gc_horizon
2353 0 : .unwrap_or(self.conf.default_tenant_conf.gc_horizon)
2354 0 : }
2355 :
2356 0 : pub fn get_gc_period(&self) -> Duration {
2357 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2358 0 : tenant_conf
2359 0 : .gc_period
2360 0 : .unwrap_or(self.conf.default_tenant_conf.gc_period)
2361 0 : }
2362 :
2363 0 : pub fn get_image_creation_threshold(&self) -> usize {
2364 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2365 0 : tenant_conf
2366 0 : .image_creation_threshold
2367 0 : .unwrap_or(self.conf.default_tenant_conf.image_creation_threshold)
2368 0 : }
2369 :
2370 0 : pub fn get_pitr_interval(&self) -> Duration {
2371 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2372 0 : tenant_conf
2373 0 : .pitr_interval
2374 0 : .unwrap_or(self.conf.default_tenant_conf.pitr_interval)
2375 0 : }
2376 :
2377 0 : pub fn get_min_resident_size_override(&self) -> Option<u64> {
2378 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2379 0 : tenant_conf
2380 0 : .min_resident_size_override
2381 0 : .or(self.conf.default_tenant_conf.min_resident_size_override)
2382 0 : }
2383 :
2384 0 : pub fn get_heatmap_period(&self) -> Option<Duration> {
2385 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2386 0 : let heatmap_period = tenant_conf
2387 0 : .heatmap_period
2388 0 : .unwrap_or(self.conf.default_tenant_conf.heatmap_period);
2389 0 : if heatmap_period.is_zero() {
2390 0 : None
2391 : } else {
2392 0 : Some(heatmap_period)
2393 : }
2394 0 : }
2395 :
2396 0 : pub fn get_lsn_lease_length(&self) -> Duration {
2397 0 : let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
2398 0 : tenant_conf
2399 0 : .lsn_lease_length
2400 0 : .unwrap_or(self.conf.default_tenant_conf.lsn_lease_length)
2401 0 : }
2402 :
2403 0 : pub fn set_new_tenant_config(&self, new_tenant_conf: TenantConfOpt) {
2404 0 : // Use read-copy-update in order to avoid overwriting the location config
2405 0 : // state if this races with [`Tenant::set_new_location_config`]. Note that
2406 0 : // this race is not possible if both request types come from the storage
2407 0 : // controller (as they should!) because an exclusive op lock is required
2408 0 : // on the storage controller side.
2409 0 : self.tenant_conf.rcu(|inner| {
2410 0 : Arc::new(AttachedTenantConf {
2411 0 : tenant_conf: new_tenant_conf.clone(),
2412 0 : location: inner.location,
2413 0 : })
2414 0 : });
2415 0 :
2416 0 : self.tenant_conf_updated(&new_tenant_conf);
2417 0 : // Don't hold self.timelines.lock() during the notifies.
2418 0 : // There's no risk of deadlock right now, but there could be if we consolidate
2419 0 : // mutexes in struct Timeline in the future.
2420 0 : let timelines = self.list_timelines();
2421 0 : for timeline in timelines {
2422 0 : timeline.tenant_conf_updated(&new_tenant_conf);
2423 0 : }
2424 0 : }
2425 :
2426 8 : pub(crate) fn set_new_location_config(&self, new_conf: AttachedTenantConf) {
2427 8 : let new_tenant_conf = new_conf.tenant_conf.clone();
2428 8 :
2429 8 : self.tenant_conf.store(Arc::new(new_conf));
2430 8 :
2431 8 : self.tenant_conf_updated(&new_tenant_conf);
2432 8 : // Don't hold self.timelines.lock() during the notifies.
2433 8 : // There's no risk of deadlock right now, but there could be if we consolidate
2434 8 : // mutexes in struct Timeline in the future.
2435 8 : let timelines = self.list_timelines();
2436 16 : for timeline in timelines {
2437 8 : timeline.tenant_conf_updated(&new_tenant_conf);
2438 8 : }
2439 8 : }
2440 :
2441 176 : fn get_timeline_get_throttle_config(
2442 176 : psconf: &'static PageServerConf,
2443 176 : overrides: &TenantConfOpt,
2444 176 : ) -> throttle::Config {
2445 176 : overrides
2446 176 : .timeline_get_throttle
2447 176 : .clone()
2448 176 : .unwrap_or(psconf.default_tenant_conf.timeline_get_throttle.clone())
2449 176 : }
2450 :
2451 8 : pub(crate) fn tenant_conf_updated(&self, new_conf: &TenantConfOpt) {
2452 8 : let conf = Self::get_timeline_get_throttle_config(self.conf, new_conf);
2453 8 : self.timeline_get_throttle.reconfigure(conf)
2454 8 : }
2455 :
2456 : /// Helper function to create a new Timeline struct.
2457 : ///
2458 : /// The returned Timeline is in Loading state. The caller is responsible for
2459 : /// initializing any on-disk state, and for inserting the Timeline to the 'timelines'
2460 : /// map.
2461 : ///
2462 : /// `validate_ancestor == false` is used when a timeline is created for deletion
2463 : /// and we might not have the ancestor present anymore which is fine for to be
2464 : /// deleted timelines.
2465 390 : fn create_timeline_struct(
2466 390 : &self,
2467 390 : new_timeline_id: TimelineId,
2468 390 : new_metadata: &TimelineMetadata,
2469 390 : ancestor: Option<Arc<Timeline>>,
2470 390 : resources: TimelineResources,
2471 390 : cause: CreateTimelineCause,
2472 390 : last_aux_file_policy: Option<AuxFilePolicy>,
2473 390 : ) -> anyhow::Result<Arc<Timeline>> {
2474 390 : let state = match cause {
2475 : CreateTimelineCause::Load => {
2476 390 : let ancestor_id = new_metadata.ancestor_timeline();
2477 390 : anyhow::ensure!(
2478 390 : ancestor_id == ancestor.as_ref().map(|t| t.timeline_id),
2479 0 : "Timeline's {new_timeline_id} ancestor {ancestor_id:?} was not found"
2480 : );
2481 390 : TimelineState::Loading
2482 : }
2483 0 : CreateTimelineCause::Delete => TimelineState::Stopping,
2484 : };
2485 :
2486 390 : let pg_version = new_metadata.pg_version();
2487 390 :
2488 390 : let timeline = Timeline::new(
2489 390 : self.conf,
2490 390 : Arc::clone(&self.tenant_conf),
2491 390 : new_metadata,
2492 390 : ancestor,
2493 390 : new_timeline_id,
2494 390 : self.tenant_shard_id,
2495 390 : self.generation,
2496 390 : self.shard_identity,
2497 390 : self.walredo_mgr.clone(),
2498 390 : resources,
2499 390 : pg_version,
2500 390 : state,
2501 390 : last_aux_file_policy,
2502 390 : self.cancel.child_token(),
2503 390 : );
2504 390 :
2505 390 : Ok(timeline)
2506 390 : }
2507 :
2508 : // Allow too_many_arguments because a constructor's argument list naturally grows with the
2509 : // number of attributes in the struct: breaking these out into a builder wouldn't be helpful.
2510 : #[allow(clippy::too_many_arguments)]
2511 168 : fn new(
2512 168 : state: TenantState,
2513 168 : conf: &'static PageServerConf,
2514 168 : attached_conf: AttachedTenantConf,
2515 168 : shard_identity: ShardIdentity,
2516 168 : walredo_mgr: Option<Arc<WalRedoManager>>,
2517 168 : tenant_shard_id: TenantShardId,
2518 168 : remote_storage: GenericRemoteStorage,
2519 168 : deletion_queue_client: DeletionQueueClient,
2520 168 : l0_flush_global_state: L0FlushGlobalState,
2521 168 : ) -> Tenant {
2522 168 : debug_assert!(
2523 168 : !attached_conf.location.generation.is_none() || conf.control_plane_api.is_none()
2524 : );
2525 :
2526 168 : let (state, mut rx) = watch::channel(state);
2527 168 :
2528 168 : tokio::spawn(async move {
2529 168 : // reflect tenant state in metrics:
2530 168 : // - global per tenant state: TENANT_STATE_METRIC
2531 168 : // - "set" of broken tenants: BROKEN_TENANTS_SET
2532 168 : //
2533 168 : // set of broken tenants should not have zero counts so that it remains accessible for
2534 168 : // alerting.
2535 168 :
2536 168 : let tid = tenant_shard_id.to_string();
2537 168 : let shard_id = tenant_shard_id.shard_slug().to_string();
2538 168 : let set_key = &[tid.as_str(), shard_id.as_str()][..];
2539 168 :
2540 334 : fn inspect_state(state: &TenantState) -> ([&'static str; 1], bool) {
2541 334 : ([state.into()], matches!(state, TenantState::Broken { .. }))
2542 334 : }
2543 168 :
2544 168 : let mut tuple = inspect_state(&rx.borrow_and_update());
2545 168 :
2546 168 : let is_broken = tuple.1;
2547 168 : let mut counted_broken = if is_broken {
2548 : // add the id to the set right away, there should not be any updates on the channel
2549 : // after before tenant is removed, if ever
2550 0 : BROKEN_TENANTS_SET.with_label_values(set_key).set(1);
2551 0 : true
2552 : } else {
2553 168 : false
2554 : };
2555 :
2556 334 : loop {
2557 334 : let labels = &tuple.0;
2558 334 : let current = TENANT_STATE_METRIC.with_label_values(labels);
2559 334 : current.inc();
2560 334 :
2561 334 : if rx.changed().await.is_err() {
2562 : // tenant has been dropped
2563 16 : current.dec();
2564 16 : drop(BROKEN_TENANTS_SET.remove_label_values(set_key));
2565 16 : break;
2566 166 : }
2567 166 :
2568 166 : current.dec();
2569 166 : tuple = inspect_state(&rx.borrow_and_update());
2570 166 :
2571 166 : let is_broken = tuple.1;
2572 166 : if is_broken && !counted_broken {
2573 0 : counted_broken = true;
2574 0 : // insert the tenant_id (back) into the set while avoiding needless counter
2575 0 : // access
2576 0 : BROKEN_TENANTS_SET.with_label_values(set_key).set(1);
2577 166 : }
2578 : }
2579 168 : });
2580 168 :
2581 168 : Tenant {
2582 168 : tenant_shard_id,
2583 168 : shard_identity,
2584 168 : generation: attached_conf.location.generation,
2585 168 : conf,
2586 168 : // using now here is good enough approximation to catch tenants with really long
2587 168 : // activation times.
2588 168 : constructed_at: Instant::now(),
2589 168 : timelines: Mutex::new(HashMap::new()),
2590 168 : timelines_creating: Mutex::new(HashSet::new()),
2591 168 : gc_cs: tokio::sync::Mutex::new(()),
2592 168 : walredo_mgr,
2593 168 : remote_storage,
2594 168 : deletion_queue_client,
2595 168 : state,
2596 168 : cached_logical_sizes: tokio::sync::Mutex::new(HashMap::new()),
2597 168 : cached_synthetic_tenant_size: Arc::new(AtomicU64::new(0)),
2598 168 : eviction_task_tenant_state: tokio::sync::Mutex::new(EvictionTaskTenantState::default()),
2599 168 : compaction_circuit_breaker: std::sync::Mutex::new(CircuitBreaker::new(
2600 168 : format!("compaction-{tenant_shard_id}"),
2601 168 : 5,
2602 168 : // Compaction can be a very expensive operation, and might leak disk space. It also ought
2603 168 : // to be infallible, as long as remote storage is available. So if it repeatedly fails,
2604 168 : // use an extremely long backoff.
2605 168 : Some(Duration::from_secs(3600 * 24)),
2606 168 : )),
2607 168 : activate_now_sem: tokio::sync::Semaphore::new(0),
2608 168 : cancel: CancellationToken::default(),
2609 168 : gate: Gate::default(),
2610 168 : timeline_get_throttle: Arc::new(throttle::Throttle::new(
2611 168 : Tenant::get_timeline_get_throttle_config(conf, &attached_conf.tenant_conf),
2612 168 : &crate::metrics::tenant_throttling::TIMELINE_GET,
2613 168 : )),
2614 168 : tenant_conf: Arc::new(ArcSwap::from_pointee(attached_conf)),
2615 168 : ongoing_timeline_detach: std::sync::Mutex::default(),
2616 168 : l0_flush_global_state,
2617 168 : }
2618 168 : }
2619 :
2620 : /// Locate and load config
2621 0 : pub(super) fn load_tenant_config(
2622 0 : conf: &'static PageServerConf,
2623 0 : tenant_shard_id: &TenantShardId,
2624 0 : ) -> Result<LocationConf, LoadConfigError> {
2625 0 : let config_path = conf.tenant_location_config_path(tenant_shard_id);
2626 0 :
2627 0 : info!("loading tenant configuration from {config_path}");
2628 :
2629 : // load and parse file
2630 0 : let config = fs::read_to_string(&config_path).map_err(|e| {
2631 0 : match e.kind() {
2632 : std::io::ErrorKind::NotFound => {
2633 : // The config should almost always exist for a tenant directory:
2634 : // - When attaching a tenant, the config is the first thing we write
2635 : // - When detaching a tenant, we atomically move the directory to a tmp location
2636 : // before deleting contents.
2637 : //
2638 : // The very rare edge case that can result in a missing config is if we crash during attach
2639 : // between creating directory and writing config. Callers should handle that as if the
2640 : // directory didn't exist.
2641 :
2642 0 : LoadConfigError::NotFound(config_path)
2643 : }
2644 : _ => {
2645 : // No IO errors except NotFound are acceptable here: other kinds of error indicate local storage or permissions issues
2646 : // that we cannot cleanly recover
2647 0 : crate::virtual_file::on_fatal_io_error(&e, "Reading tenant config file")
2648 : }
2649 : }
2650 0 : })?;
2651 :
2652 0 : Ok(toml_edit::de::from_str::<LocationConf>(&config)?)
2653 0 : }
2654 :
2655 0 : #[tracing::instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug()))]
2656 : pub(super) async fn persist_tenant_config(
2657 : conf: &'static PageServerConf,
2658 : tenant_shard_id: &TenantShardId,
2659 : location_conf: &LocationConf,
2660 : ) -> std::io::Result<()> {
2661 : let config_path = conf.tenant_location_config_path(tenant_shard_id);
2662 :
2663 : Self::persist_tenant_config_at(tenant_shard_id, &config_path, location_conf).await
2664 : }
2665 :
2666 0 : #[tracing::instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug()))]
2667 : pub(super) async fn persist_tenant_config_at(
2668 : tenant_shard_id: &TenantShardId,
2669 : config_path: &Utf8Path,
2670 : location_conf: &LocationConf,
2671 : ) -> std::io::Result<()> {
2672 : debug!("persisting tenantconf to {config_path}");
2673 :
2674 : let mut conf_content = r#"# This file contains a specific per-tenant's config.
2675 : # It is read in case of pageserver restart.
2676 : "#
2677 : .to_string();
2678 :
2679 0 : fail::fail_point!("tenant-config-before-write", |_| {
2680 0 : Err(std::io::Error::new(
2681 0 : std::io::ErrorKind::Other,
2682 0 : "tenant-config-before-write",
2683 0 : ))
2684 0 : });
2685 :
2686 : // Convert the config to a toml file.
2687 : conf_content +=
2688 : &toml_edit::ser::to_string_pretty(&location_conf).expect("Config serialization failed");
2689 :
2690 : let temp_path = path_with_suffix_extension(config_path, TEMP_FILE_SUFFIX);
2691 :
2692 : let conf_content = conf_content.into_bytes();
2693 : VirtualFile::crashsafe_overwrite(config_path.to_owned(), temp_path, conf_content).await
2694 : }
2695 :
2696 : //
2697 : // How garbage collection works:
2698 : //
2699 : // +--bar------------->
2700 : // /
2701 : // +----+-----foo---------------->
2702 : // /
2703 : // ----main--+-------------------------->
2704 : // \
2705 : // +-----baz-------->
2706 : //
2707 : //
2708 : // 1. Grab 'gc_cs' mutex to prevent new timelines from being created while Timeline's
2709 : // `gc_infos` are being refreshed
2710 : // 2. Scan collected timelines, and on each timeline, make note of the
2711 : // all the points where other timelines have been branched off.
2712 : // We will refrain from removing page versions at those LSNs.
2713 : // 3. For each timeline, scan all layer files on the timeline.
2714 : // Remove all files for which a newer file exists and which
2715 : // don't cover any branch point LSNs.
2716 : //
2717 : // TODO:
2718 : // - if a relation has a non-incremental persistent layer on a child branch, then we
2719 : // don't need to keep that in the parent anymore. But currently
2720 : // we do.
2721 754 : async fn gc_iteration_internal(
2722 754 : &self,
2723 754 : target_timeline_id: Option<TimelineId>,
2724 754 : horizon: u64,
2725 754 : pitr: Duration,
2726 754 : cancel: &CancellationToken,
2727 754 : ctx: &RequestContext,
2728 754 : ) -> Result<GcResult, GcError> {
2729 754 : let mut totals: GcResult = Default::default();
2730 754 : let now = Instant::now();
2731 :
2732 754 : let gc_timelines = self
2733 754 : .refresh_gc_info_internal(target_timeline_id, horizon, pitr, cancel, ctx)
2734 743 : .await?;
2735 :
2736 754 : failpoint_support::sleep_millis_async!("gc_iteration_internal_after_getting_gc_timelines");
2737 :
2738 : // If there is nothing to GC, we don't want any messages in the INFO log.
2739 754 : if !gc_timelines.is_empty() {
2740 754 : info!("{} timelines need GC", gc_timelines.len());
2741 : } else {
2742 0 : debug!("{} timelines need GC", gc_timelines.len());
2743 : }
2744 :
2745 : // Perform GC for each timeline.
2746 : //
2747 : // Note that we don't hold the `Tenant::gc_cs` lock here because we don't want to delay the
2748 : // branch creation task, which requires the GC lock. A GC iteration can run concurrently
2749 : // with branch creation.
2750 : //
2751 : // See comments in [`Tenant::branch_timeline`] for more information about why branch
2752 : // creation task can run concurrently with timeline's GC iteration.
2753 1508 : for timeline in gc_timelines {
2754 754 : if cancel.is_cancelled() {
2755 : // We were requested to shut down. Stop and return with the progress we
2756 : // made.
2757 0 : break;
2758 754 : }
2759 754 : let result = match timeline.gc().await {
2760 : Err(GcError::TimelineCancelled) => {
2761 0 : if target_timeline_id.is_some() {
2762 : // If we were targetting this specific timeline, surface cancellation to caller
2763 0 : return Err(GcError::TimelineCancelled);
2764 : } else {
2765 : // A timeline may be shutting down independently of the tenant's lifecycle: we should
2766 : // skip past this and proceed to try GC on other timelines.
2767 0 : continue;
2768 : }
2769 : }
2770 754 : r => r?,
2771 : };
2772 754 : totals += result;
2773 : }
2774 :
2775 754 : totals.elapsed = now.elapsed();
2776 754 : Ok(totals)
2777 754 : }
2778 :
2779 : /// Refreshes the Timeline::gc_info for all timelines, returning the
2780 : /// vector of timelines which have [`Timeline::get_last_record_lsn`] past
2781 : /// [`Tenant::get_gc_horizon`].
2782 : ///
2783 : /// This is usually executed as part of periodic gc, but can now be triggered more often.
2784 0 : pub(crate) async fn refresh_gc_info(
2785 0 : &self,
2786 0 : cancel: &CancellationToken,
2787 0 : ctx: &RequestContext,
2788 0 : ) -> Result<Vec<Arc<Timeline>>, GcError> {
2789 0 : // since this method can now be called at different rates than the configured gc loop, it
2790 0 : // might be that these configuration values get applied faster than what it was previously,
2791 0 : // since these were only read from the gc task.
2792 0 : let horizon = self.get_gc_horizon();
2793 0 : let pitr = self.get_pitr_interval();
2794 0 :
2795 0 : // refresh all timelines
2796 0 : let target_timeline_id = None;
2797 0 :
2798 0 : self.refresh_gc_info_internal(target_timeline_id, horizon, pitr, cancel, ctx)
2799 0 : .await
2800 0 : }
2801 :
2802 754 : async fn refresh_gc_info_internal(
2803 754 : &self,
2804 754 : target_timeline_id: Option<TimelineId>,
2805 754 : horizon: u64,
2806 754 : pitr: Duration,
2807 754 : cancel: &CancellationToken,
2808 754 : ctx: &RequestContext,
2809 754 : ) -> Result<Vec<Arc<Timeline>>, GcError> {
2810 754 : // before taking the gc_cs lock, do the heavier weight finding of gc_cutoff points for
2811 754 : // currently visible timelines.
2812 754 : let timelines = self
2813 754 : .timelines
2814 754 : .lock()
2815 754 : .unwrap()
2816 754 : .values()
2817 3310 : .filter(|tl| match target_timeline_id.as_ref() {
2818 3310 : Some(target) => &tl.timeline_id == target,
2819 0 : None => true,
2820 3310 : })
2821 754 : .cloned()
2822 754 : .collect::<Vec<_>>();
2823 754 :
2824 754 : let mut gc_cutoffs: HashMap<TimelineId, GcCutoffs> =
2825 754 : HashMap::with_capacity(timelines.len());
2826 :
2827 754 : for timeline in timelines.iter() {
2828 754 : let cutoff = timeline
2829 754 : .get_last_record_lsn()
2830 754 : .checked_sub(horizon)
2831 754 : .unwrap_or(Lsn(0));
2832 :
2833 754 : let cutoffs = timeline.find_gc_cutoffs(cutoff, pitr, cancel, ctx).await?;
2834 754 : let old = gc_cutoffs.insert(timeline.timeline_id, cutoffs);
2835 754 : assert!(old.is_none());
2836 : }
2837 :
2838 754 : if !self.is_active() || self.cancel.is_cancelled() {
2839 0 : return Err(GcError::TenantCancelled);
2840 754 : }
2841 :
2842 : // grab mutex to prevent new timelines from being created here; avoid doing long operations
2843 : // because that will stall branch creation.
2844 754 : let gc_cs = self.gc_cs.lock().await;
2845 :
2846 : // Scan all timelines. For each timeline, remember the timeline ID and
2847 : // the branch point where it was created.
2848 754 : let (all_branchpoints, timelines): (BTreeSet<(TimelineId, Lsn)>, _) = {
2849 754 : let timelines = self.timelines.lock().unwrap();
2850 754 : let mut all_branchpoints = BTreeSet::new();
2851 754 : let timelines = {
2852 754 : if let Some(target_timeline_id) = target_timeline_id.as_ref() {
2853 754 : if timelines.get(target_timeline_id).is_none() {
2854 0 : return Err(GcError::TimelineNotFound);
2855 754 : }
2856 0 : };
2857 :
2858 754 : timelines
2859 754 : .iter()
2860 3310 : .map(|(_timeline_id, timeline_entry)| {
2861 2556 : if let Some(ancestor_timeline_id) =
2862 3310 : &timeline_entry.get_ancestor_timeline_id()
2863 : {
2864 : // If target_timeline is specified, we only need to know branchpoints of its children
2865 2556 : if let Some(timeline_id) = target_timeline_id {
2866 2556 : if ancestor_timeline_id == &timeline_id {
2867 6 : all_branchpoints.insert((
2868 6 : *ancestor_timeline_id,
2869 6 : timeline_entry.get_ancestor_lsn(),
2870 6 : ));
2871 2550 : }
2872 : }
2873 : // Collect branchpoints for all timelines
2874 0 : else {
2875 0 : all_branchpoints.insert((
2876 0 : *ancestor_timeline_id,
2877 0 : timeline_entry.get_ancestor_lsn(),
2878 0 : ));
2879 0 : }
2880 754 : }
2881 :
2882 3310 : timeline_entry.clone()
2883 3310 : })
2884 754 : .collect::<Vec<_>>()
2885 754 : };
2886 754 : (all_branchpoints, timelines)
2887 754 : };
2888 754 :
2889 754 : // Ok, we now know all the branch points.
2890 754 : // Update the GC information for each timeline.
2891 754 : let mut gc_timelines = Vec::with_capacity(timelines.len());
2892 4064 : for timeline in timelines {
2893 : // If target_timeline is specified, ignore all other timelines
2894 3310 : if let Some(target_timeline_id) = target_timeline_id {
2895 3310 : if timeline.timeline_id != target_timeline_id {
2896 2556 : continue;
2897 754 : }
2898 0 : }
2899 :
2900 754 : let branchpoints: Vec<Lsn> = all_branchpoints
2901 754 : .range((
2902 754 : Included((timeline.timeline_id, Lsn(0))),
2903 754 : Included((timeline.timeline_id, Lsn(u64::MAX))),
2904 754 : ))
2905 754 : .map(|&x| x.1)
2906 754 : .collect();
2907 754 :
2908 754 : {
2909 754 : let mut target = timeline.gc_info.write().unwrap();
2910 754 :
2911 754 : // Cull any expired leases
2912 754 : let now = SystemTime::now();
2913 754 : target.leases.retain(|_, lease| !lease.is_expired(&now));
2914 754 :
2915 754 : timeline
2916 754 : .metrics
2917 754 : .valid_lsn_lease_count_gauge
2918 754 : .set(target.leases.len() as u64);
2919 :
2920 : // Look up parent's PITR cutoff to update the child's knowledge of whether it is within parent's PITR
2921 754 : if let Some(ancestor_id) = timeline.get_ancestor_timeline_id() {
2922 100 : if let Some(ancestor_gc_cutoffs) = gc_cutoffs.get(&ancestor_id) {
2923 0 : target.within_ancestor_pitr =
2924 0 : timeline.get_ancestor_lsn() >= ancestor_gc_cutoffs.time;
2925 100 : }
2926 654 : }
2927 :
2928 : // Update metrics that depend on GC state
2929 754 : timeline
2930 754 : .metrics
2931 754 : .archival_size
2932 754 : .set(if target.within_ancestor_pitr {
2933 0 : timeline.metrics.current_logical_size_gauge.get()
2934 : } else {
2935 754 : 0
2936 : });
2937 754 : timeline.metrics.pitr_history_size.set(
2938 754 : timeline
2939 754 : .get_last_record_lsn()
2940 754 : .checked_sub(target.cutoffs.time)
2941 754 : .unwrap_or(Lsn(0))
2942 754 : .0,
2943 754 : );
2944 754 :
2945 754 : match gc_cutoffs.remove(&timeline.timeline_id) {
2946 754 : Some(cutoffs) => {
2947 754 : target.retain_lsns = branchpoints;
2948 754 : target.cutoffs = cutoffs;
2949 754 : }
2950 0 : None => {
2951 0 : // reasons for this being unavailable:
2952 0 : // - this timeline was created while we were finding cutoffs
2953 0 : // - lsn for timestamp search fails for this timeline repeatedly
2954 0 : //
2955 0 : // in both cases, refreshing the branchpoints is correct.
2956 0 : target.retain_lsns = branchpoints;
2957 0 : }
2958 : };
2959 : }
2960 :
2961 754 : gc_timelines.push(timeline);
2962 : }
2963 754 : drop(gc_cs);
2964 754 : Ok(gc_timelines)
2965 754 : }
2966 :
2967 : /// A substitute for `branch_timeline` for use in unit tests.
2968 : /// The returned timeline will have state value `Active` to make various `anyhow::ensure!()`
2969 : /// calls pass, but, we do not actually call `.activate()` under the hood. So, none of the
2970 : /// timeline background tasks are launched, except the flush loop.
2971 : #[cfg(test)]
2972 228 : async fn branch_timeline_test(
2973 228 : &self,
2974 228 : src_timeline: &Arc<Timeline>,
2975 228 : dst_id: TimelineId,
2976 228 : ancestor_lsn: Option<Lsn>,
2977 228 : ctx: &RequestContext,
2978 228 : ) -> Result<Arc<Timeline>, CreateTimelineError> {
2979 228 : let create_guard = self.create_timeline_create_guard(dst_id).unwrap();
2980 228 : let tl = self
2981 228 : .branch_timeline_impl(src_timeline, dst_id, ancestor_lsn, create_guard, ctx)
2982 4 : .await?;
2983 224 : tl.set_state(TimelineState::Active);
2984 224 : Ok(tl)
2985 228 : }
2986 :
2987 : /// Helper for unit tests to branch a timeline with some pre-loaded states.
2988 : #[cfg(test)]
2989 : #[allow(clippy::too_many_arguments)]
2990 4 : pub async fn branch_timeline_test_with_layers(
2991 4 : &self,
2992 4 : src_timeline: &Arc<Timeline>,
2993 4 : dst_id: TimelineId,
2994 4 : ancestor_lsn: Option<Lsn>,
2995 4 : ctx: &RequestContext,
2996 4 : delta_layer_desc: Vec<timeline::DeltaLayerTestDesc>,
2997 4 : image_layer_desc: Vec<(Lsn, Vec<(pageserver_api::key::Key, bytes::Bytes)>)>,
2998 4 : end_lsn: Lsn,
2999 4 : ) -> anyhow::Result<Arc<Timeline>> {
3000 4 : let tline = self
3001 4 : .branch_timeline_test(src_timeline, dst_id, ancestor_lsn, ctx)
3002 0 : .await?;
3003 4 : let ancestor_lsn = if let Some(ancestor_lsn) = ancestor_lsn {
3004 4 : ancestor_lsn
3005 : } else {
3006 0 : tline.get_last_record_lsn()
3007 : };
3008 4 : assert!(end_lsn >= ancestor_lsn);
3009 4 : tline.force_advance_lsn(end_lsn);
3010 4 : for deltas in delta_layer_desc {
3011 0 : tline
3012 0 : .force_create_delta_layer(deltas, Some(ancestor_lsn), ctx)
3013 0 : .await?;
3014 : }
3015 8 : for (lsn, images) in image_layer_desc {
3016 4 : tline
3017 4 : .force_create_image_layer(lsn, images, Some(ancestor_lsn), ctx)
3018 14 : .await?;
3019 : }
3020 4 : Ok(tline)
3021 4 : }
3022 :
3023 : /// Branch an existing timeline.
3024 : ///
3025 : /// The caller is responsible for activating the returned timeline.
3026 0 : async fn branch_timeline(
3027 0 : &self,
3028 0 : src_timeline: &Arc<Timeline>,
3029 0 : dst_id: TimelineId,
3030 0 : start_lsn: Option<Lsn>,
3031 0 : timeline_create_guard: TimelineCreateGuard<'_>,
3032 0 : ctx: &RequestContext,
3033 0 : ) -> Result<Arc<Timeline>, CreateTimelineError> {
3034 0 : self.branch_timeline_impl(src_timeline, dst_id, start_lsn, timeline_create_guard, ctx)
3035 0 : .await
3036 0 : }
3037 :
3038 228 : async fn branch_timeline_impl(
3039 228 : &self,
3040 228 : src_timeline: &Arc<Timeline>,
3041 228 : dst_id: TimelineId,
3042 228 : start_lsn: Option<Lsn>,
3043 228 : timeline_create_guard: TimelineCreateGuard<'_>,
3044 228 : _ctx: &RequestContext,
3045 228 : ) -> Result<Arc<Timeline>, CreateTimelineError> {
3046 228 : let src_id = src_timeline.timeline_id;
3047 :
3048 : // We will validate our ancestor LSN in this function. Acquire the GC lock so that
3049 : // this check cannot race with GC, and the ancestor LSN is guaranteed to remain
3050 : // valid while we are creating the branch.
3051 228 : let _gc_cs = self.gc_cs.lock().await;
3052 :
3053 : // If no start LSN is specified, we branch the new timeline from the source timeline's last record LSN
3054 228 : let start_lsn = start_lsn.unwrap_or_else(|| {
3055 2 : let lsn = src_timeline.get_last_record_lsn();
3056 2 : info!("branching timeline {dst_id} from timeline {src_id} at last record LSN: {lsn}");
3057 2 : lsn
3058 228 : });
3059 228 :
3060 228 : // Ensure that `start_lsn` is valid, i.e. the LSN is within the PITR
3061 228 : // horizon on the source timeline
3062 228 : //
3063 228 : // We check it against both the planned GC cutoff stored in 'gc_info',
3064 228 : // and the 'latest_gc_cutoff' of the last GC that was performed. The
3065 228 : // planned GC cutoff in 'gc_info' is normally larger than
3066 228 : // 'latest_gc_cutoff_lsn', but beware of corner cases like if you just
3067 228 : // changed the GC settings for the tenant to make the PITR window
3068 228 : // larger, but some of the data was already removed by an earlier GC
3069 228 : // iteration.
3070 228 :
3071 228 : // check against last actual 'latest_gc_cutoff' first
3072 228 : let latest_gc_cutoff_lsn = src_timeline.get_latest_gc_cutoff_lsn();
3073 228 : src_timeline
3074 228 : .check_lsn_is_in_scope(start_lsn, &latest_gc_cutoff_lsn)
3075 228 : .context(format!(
3076 228 : "invalid branch start lsn: less than latest GC cutoff {}",
3077 228 : *latest_gc_cutoff_lsn,
3078 228 : ))
3079 228 : .map_err(CreateTimelineError::AncestorLsn)?;
3080 :
3081 : // and then the planned GC cutoff
3082 : {
3083 224 : let gc_info = src_timeline.gc_info.read().unwrap();
3084 224 : let cutoff = gc_info.min_cutoff();
3085 224 : if start_lsn < cutoff {
3086 0 : return Err(CreateTimelineError::AncestorLsn(anyhow::anyhow!(
3087 0 : "invalid branch start lsn: less than planned GC cutoff {cutoff}"
3088 0 : )));
3089 224 : }
3090 224 : }
3091 224 :
3092 224 : //
3093 224 : // The branch point is valid, and we are still holding the 'gc_cs' lock
3094 224 : // so that GC cannot advance the GC cutoff until we are finished.
3095 224 : // Proceed with the branch creation.
3096 224 : //
3097 224 :
3098 224 : // Determine prev-LSN for the new timeline. We can only determine it if
3099 224 : // the timeline was branched at the current end of the source timeline.
3100 224 : let RecordLsn {
3101 224 : last: src_last,
3102 224 : prev: src_prev,
3103 224 : } = src_timeline.get_last_record_rlsn();
3104 224 : let dst_prev = if src_last == start_lsn {
3105 214 : Some(src_prev)
3106 : } else {
3107 10 : None
3108 : };
3109 :
3110 : // Create the metadata file, noting the ancestor of the new timeline.
3111 : // There is initially no data in it, but all the read-calls know to look
3112 : // into the ancestor.
3113 224 : let metadata = TimelineMetadata::new(
3114 224 : start_lsn,
3115 224 : dst_prev,
3116 224 : Some(src_id),
3117 224 : start_lsn,
3118 224 : *src_timeline.latest_gc_cutoff_lsn.read(), // FIXME: should we hold onto this guard longer?
3119 224 : src_timeline.initdb_lsn,
3120 224 : src_timeline.pg_version,
3121 224 : );
3122 :
3123 224 : let uninitialized_timeline = self
3124 224 : .prepare_new_timeline(
3125 224 : dst_id,
3126 224 : &metadata,
3127 224 : timeline_create_guard,
3128 224 : start_lsn + 1,
3129 224 : Some(Arc::clone(src_timeline)),
3130 224 : src_timeline.last_aux_file_policy.load(),
3131 224 : )
3132 0 : .await?;
3133 :
3134 224 : let new_timeline = uninitialized_timeline.finish_creation()?;
3135 :
3136 : // Root timeline gets its layers during creation and uploads them along with the metadata.
3137 : // A branch timeline though, when created, can get no writes for some time, hence won't get any layers created.
3138 : // We still need to upload its metadata eagerly: if other nodes `attach` the tenant and miss this timeline, their GC
3139 : // could get incorrect information and remove more layers, than needed.
3140 : // See also https://github.com/neondatabase/neon/issues/3865
3141 224 : new_timeline
3142 224 : .remote_client
3143 224 : .schedule_index_upload_for_full_metadata_update(&metadata)
3144 224 : .context("branch initial metadata upload")?;
3145 :
3146 224 : Ok(new_timeline)
3147 228 : }
3148 :
3149 : /// For unit tests, make this visible so that other modules can directly create timelines
3150 : #[cfg(test)]
3151 4 : #[tracing::instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug(), %timeline_id))]
3152 : pub(crate) async fn bootstrap_timeline_test(
3153 : &self,
3154 : timeline_id: TimelineId,
3155 : pg_version: u32,
3156 : load_existing_initdb: Option<TimelineId>,
3157 : ctx: &RequestContext,
3158 : ) -> anyhow::Result<Arc<Timeline>> {
3159 : let create_guard = self.create_timeline_create_guard(timeline_id).unwrap();
3160 : self.bootstrap_timeline(
3161 : timeline_id,
3162 : pg_version,
3163 : load_existing_initdb,
3164 : create_guard,
3165 : ctx,
3166 : )
3167 : .await
3168 : }
3169 :
3170 0 : async fn upload_initdb(
3171 0 : &self,
3172 0 : timelines_path: &Utf8PathBuf,
3173 0 : pgdata_path: &Utf8PathBuf,
3174 0 : timeline_id: &TimelineId,
3175 0 : ) -> anyhow::Result<()> {
3176 0 : let temp_path = timelines_path.join(format!(
3177 0 : "{INITDB_PATH}.upload-{timeline_id}.{TEMP_FILE_SUFFIX}"
3178 0 : ));
3179 :
3180 : scopeguard::defer! {
3181 : if let Err(e) = fs::remove_file(&temp_path) {
3182 : error!("Failed to remove temporary initdb archive '{temp_path}': {e}");
3183 : }
3184 : }
3185 :
3186 0 : let (pgdata_zstd, tar_zst_size) = create_zst_tarball(pgdata_path, &temp_path).await?;
3187 : const INITDB_TAR_ZST_WARN_LIMIT: u64 = 2 * 1024 * 1024;
3188 0 : if tar_zst_size > INITDB_TAR_ZST_WARN_LIMIT {
3189 0 : warn!(
3190 0 : "compressed {temp_path} size of {tar_zst_size} is above limit {INITDB_TAR_ZST_WARN_LIMIT}."
3191 : );
3192 0 : }
3193 :
3194 : pausable_failpoint!("before-initdb-upload");
3195 :
3196 0 : backoff::retry(
3197 0 : || async {
3198 0 : self::remote_timeline_client::upload_initdb_dir(
3199 0 : &self.remote_storage,
3200 0 : &self.tenant_shard_id.tenant_id,
3201 0 : timeline_id,
3202 0 : pgdata_zstd.try_clone().await?,
3203 0 : tar_zst_size,
3204 0 : &self.cancel,
3205 0 : )
3206 0 : .await
3207 0 : },
3208 0 : |_| false,
3209 0 : 3,
3210 0 : u32::MAX,
3211 0 : "persist_initdb_tar_zst",
3212 0 : &self.cancel,
3213 0 : )
3214 0 : .await
3215 0 : .ok_or_else(|| anyhow::Error::new(TimeoutOrCancel::Cancel))
3216 0 : .and_then(|x| x)
3217 0 : }
3218 :
3219 : /// - run initdb to init temporary instance and get bootstrap data
3220 : /// - after initialization completes, tar up the temp dir and upload it to S3.
3221 : ///
3222 : /// The caller is responsible for activating the returned timeline.
3223 2 : async fn bootstrap_timeline(
3224 2 : &self,
3225 2 : timeline_id: TimelineId,
3226 2 : pg_version: u32,
3227 2 : load_existing_initdb: Option<TimelineId>,
3228 2 : timeline_create_guard: TimelineCreateGuard<'_>,
3229 2 : ctx: &RequestContext,
3230 2 : ) -> anyhow::Result<Arc<Timeline>> {
3231 2 : // create a `tenant/{tenant_id}/timelines/basebackup-{timeline_id}.{TEMP_FILE_SUFFIX}/`
3232 2 : // temporary directory for basebackup files for the given timeline.
3233 2 :
3234 2 : let timelines_path = self.conf.timelines_path(&self.tenant_shard_id);
3235 2 : let pgdata_path = path_with_suffix_extension(
3236 2 : timelines_path.join(format!("basebackup-{timeline_id}")),
3237 2 : TEMP_FILE_SUFFIX,
3238 2 : );
3239 2 :
3240 2 : // Remove whatever was left from the previous runs: safe because TimelineCreateGuard guarantees
3241 2 : // we won't race with other creations or existent timelines with the same path.
3242 2 : if pgdata_path.exists() {
3243 0 : fs::remove_dir_all(&pgdata_path).with_context(|| {
3244 0 : format!("Failed to remove already existing initdb directory: {pgdata_path}")
3245 0 : })?;
3246 2 : }
3247 :
3248 : // this new directory is very temporary, set to remove it immediately after bootstrap, we don't need it
3249 : scopeguard::defer! {
3250 : if let Err(e) = fs::remove_dir_all(&pgdata_path) {
3251 : // this is unlikely, but we will remove the directory on pageserver restart or another bootstrap call
3252 : error!("Failed to remove temporary initdb directory '{pgdata_path}': {e}");
3253 : }
3254 : }
3255 2 : if let Some(existing_initdb_timeline_id) = load_existing_initdb {
3256 2 : if existing_initdb_timeline_id != timeline_id {
3257 0 : let source_path = &remote_initdb_archive_path(
3258 0 : &self.tenant_shard_id.tenant_id,
3259 0 : &existing_initdb_timeline_id,
3260 0 : );
3261 0 : let dest_path =
3262 0 : &remote_initdb_archive_path(&self.tenant_shard_id.tenant_id, &timeline_id);
3263 0 :
3264 0 : // if this fails, it will get retried by retried control plane requests
3265 0 : self.remote_storage
3266 0 : .copy_object(source_path, dest_path, &self.cancel)
3267 0 : .await
3268 0 : .context("copy initdb tar")?;
3269 2 : }
3270 2 : let (initdb_tar_zst_path, initdb_tar_zst) =
3271 2 : self::remote_timeline_client::download_initdb_tar_zst(
3272 2 : self.conf,
3273 2 : &self.remote_storage,
3274 2 : &self.tenant_shard_id,
3275 2 : &existing_initdb_timeline_id,
3276 2 : &self.cancel,
3277 2 : )
3278 738 : .await
3279 2 : .context("download initdb tar")?;
3280 :
3281 : scopeguard::defer! {
3282 : if let Err(e) = fs::remove_file(&initdb_tar_zst_path) {
3283 : error!("Failed to remove temporary initdb archive '{initdb_tar_zst_path}': {e}");
3284 : }
3285 : }
3286 :
3287 2 : let buf_read =
3288 2 : BufReader::with_capacity(remote_timeline_client::BUFFER_SIZE, initdb_tar_zst);
3289 2 : extract_zst_tarball(&pgdata_path, buf_read)
3290 10537 : .await
3291 2 : .context("extract initdb tar")?;
3292 : } else {
3293 : // Init temporarily repo to get bootstrap data, this creates a directory in the `pgdata_path` path
3294 0 : run_initdb(self.conf, &pgdata_path, pg_version, &self.cancel).await?;
3295 :
3296 : // Upload the created data dir to S3
3297 0 : if self.tenant_shard_id().is_shard_zero() {
3298 0 : self.upload_initdb(&timelines_path, &pgdata_path, &timeline_id)
3299 0 : .await?;
3300 0 : }
3301 : }
3302 2 : let pgdata_lsn = import_datadir::get_lsn_from_controlfile(&pgdata_path)?.align();
3303 2 :
3304 2 : // Import the contents of the data directory at the initial checkpoint
3305 2 : // LSN, and any WAL after that.
3306 2 : // Initdb lsn will be equal to last_record_lsn which will be set after import.
3307 2 : // Because we know it upfront avoid having an option or dummy zero value by passing it to the metadata.
3308 2 : let new_metadata = TimelineMetadata::new(
3309 2 : Lsn(0),
3310 2 : None,
3311 2 : None,
3312 2 : Lsn(0),
3313 2 : pgdata_lsn,
3314 2 : pgdata_lsn,
3315 2 : pg_version,
3316 2 : );
3317 2 : let raw_timeline = self
3318 2 : .prepare_new_timeline(
3319 2 : timeline_id,
3320 2 : &new_metadata,
3321 2 : timeline_create_guard,
3322 2 : pgdata_lsn,
3323 2 : None,
3324 2 : None,
3325 2 : )
3326 0 : .await?;
3327 :
3328 2 : let tenant_shard_id = raw_timeline.owning_tenant.tenant_shard_id;
3329 2 : let unfinished_timeline = raw_timeline.raw_timeline()?;
3330 :
3331 : // Flush the new layer files to disk, before we make the timeline as available to
3332 : // the outside world.
3333 : //
3334 : // Flush loop needs to be spawned in order to be able to flush.
3335 2 : unfinished_timeline.maybe_spawn_flush_loop();
3336 2 :
3337 2 : import_datadir::import_timeline_from_postgres_datadir(
3338 2 : unfinished_timeline,
3339 2 : &pgdata_path,
3340 2 : pgdata_lsn,
3341 2 : ctx,
3342 2 : )
3343 9214 : .await
3344 2 : .with_context(|| {
3345 0 : format!("Failed to import pgdatadir for timeline {tenant_shard_id}/{timeline_id}")
3346 2 : })?;
3347 :
3348 2 : fail::fail_point!("before-checkpoint-new-timeline", |_| {
3349 0 : anyhow::bail!("failpoint before-checkpoint-new-timeline");
3350 2 : });
3351 :
3352 2 : unfinished_timeline
3353 2 : .freeze_and_flush()
3354 2 : .await
3355 2 : .with_context(|| {
3356 0 : format!(
3357 0 : "Failed to flush after pgdatadir import for timeline {tenant_shard_id}/{timeline_id}"
3358 0 : )
3359 2 : })?;
3360 :
3361 : // All done!
3362 2 : let timeline = raw_timeline.finish_creation()?;
3363 :
3364 2 : Ok(timeline)
3365 2 : }
3366 :
3367 : /// Call this before constructing a timeline, to build its required structures
3368 384 : fn build_timeline_resources(&self, timeline_id: TimelineId) -> TimelineResources {
3369 384 : let remote_client = RemoteTimelineClient::new(
3370 384 : self.remote_storage.clone(),
3371 384 : self.deletion_queue_client.clone(),
3372 384 : self.conf,
3373 384 : self.tenant_shard_id,
3374 384 : timeline_id,
3375 384 : self.generation,
3376 384 : );
3377 384 : TimelineResources {
3378 384 : remote_client,
3379 384 : timeline_get_throttle: self.timeline_get_throttle.clone(),
3380 384 : l0_flush_global_state: self.l0_flush_global_state.clone(),
3381 384 : }
3382 384 : }
3383 :
3384 : /// Creates intermediate timeline structure and its files.
3385 : ///
3386 : /// An empty layer map is initialized, and new data and WAL can be imported starting
3387 : /// at 'disk_consistent_lsn'. After any initial data has been imported, call
3388 : /// `finish_creation` to insert the Timeline into the timelines map.
3389 384 : async fn prepare_new_timeline<'a>(
3390 384 : &'a self,
3391 384 : new_timeline_id: TimelineId,
3392 384 : new_metadata: &TimelineMetadata,
3393 384 : create_guard: TimelineCreateGuard<'a>,
3394 384 : start_lsn: Lsn,
3395 384 : ancestor: Option<Arc<Timeline>>,
3396 384 : last_aux_file_policy: Option<AuxFilePolicy>,
3397 384 : ) -> anyhow::Result<UninitializedTimeline> {
3398 384 : let tenant_shard_id = self.tenant_shard_id;
3399 384 :
3400 384 : let resources = self.build_timeline_resources(new_timeline_id);
3401 384 : resources
3402 384 : .remote_client
3403 384 : .init_upload_queue_for_empty_remote(new_metadata)?;
3404 :
3405 384 : let timeline_struct = self
3406 384 : .create_timeline_struct(
3407 384 : new_timeline_id,
3408 384 : new_metadata,
3409 384 : ancestor,
3410 384 : resources,
3411 384 : CreateTimelineCause::Load,
3412 384 : last_aux_file_policy,
3413 384 : )
3414 384 : .context("Failed to create timeline data structure")?;
3415 :
3416 384 : timeline_struct.init_empty_layer_map(start_lsn);
3417 :
3418 384 : if let Err(e) = self
3419 384 : .create_timeline_files(&create_guard.timeline_path)
3420 0 : .await
3421 : {
3422 0 : error!("Failed to create initial files for timeline {tenant_shard_id}/{new_timeline_id}, cleaning up: {e:?}");
3423 0 : cleanup_timeline_directory(create_guard);
3424 0 : return Err(e);
3425 384 : }
3426 384 :
3427 384 : debug!(
3428 0 : "Successfully created initial files for timeline {tenant_shard_id}/{new_timeline_id}"
3429 : );
3430 :
3431 384 : Ok(UninitializedTimeline::new(
3432 384 : self,
3433 384 : new_timeline_id,
3434 384 : Some((timeline_struct, create_guard)),
3435 384 : ))
3436 384 : }
3437 :
3438 384 : async fn create_timeline_files(&self, timeline_path: &Utf8Path) -> anyhow::Result<()> {
3439 384 : crashsafe::create_dir(timeline_path).context("Failed to create timeline directory")?;
3440 :
3441 384 : fail::fail_point!("after-timeline-dir-creation", |_| {
3442 0 : anyhow::bail!("failpoint after-timeline-dir-creation");
3443 384 : });
3444 :
3445 384 : Ok(())
3446 384 : }
3447 :
3448 : /// Get a guard that provides exclusive access to the timeline directory, preventing
3449 : /// concurrent attempts to create the same timeline.
3450 390 : fn create_timeline_create_guard(
3451 390 : &self,
3452 390 : timeline_id: TimelineId,
3453 390 : ) -> Result<TimelineCreateGuard, TimelineExclusionError> {
3454 390 : let tenant_shard_id = self.tenant_shard_id;
3455 390 :
3456 390 : let timeline_path = self.conf.timeline_path(&tenant_shard_id, &timeline_id);
3457 :
3458 390 : let create_guard = TimelineCreateGuard::new(self, timeline_id, timeline_path.clone())?;
3459 :
3460 : // At this stage, we have got exclusive access to in-memory state for this timeline ID
3461 : // for creation.
3462 : // A timeline directory should never exist on disk already:
3463 : // - a previous failed creation would have cleaned up after itself
3464 : // - a pageserver restart would clean up timeline directories that don't have valid remote state
3465 : //
3466 : // Therefore it is an unexpected internal error to encounter a timeline directory already existing here,
3467 : // this error may indicate a bug in cleanup on failed creations.
3468 388 : if timeline_path.exists() {
3469 0 : return Err(TimelineExclusionError::Other(anyhow::anyhow!(
3470 0 : "Timeline directory already exists! This is a bug."
3471 0 : )));
3472 388 : }
3473 388 :
3474 388 : Ok(create_guard)
3475 390 : }
3476 :
3477 : /// Gathers inputs from all of the timelines to produce a sizing model input.
3478 : ///
3479 : /// Future is cancellation safe. Only one calculation can be running at once per tenant.
3480 0 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
3481 : pub async fn gather_size_inputs(
3482 : &self,
3483 : // `max_retention_period` overrides the cutoff that is used to calculate the size
3484 : // (only if it is shorter than the real cutoff).
3485 : max_retention_period: Option<u64>,
3486 : cause: LogicalSizeCalculationCause,
3487 : cancel: &CancellationToken,
3488 : ctx: &RequestContext,
3489 : ) -> Result<size::ModelInputs, size::CalculateSyntheticSizeError> {
3490 : let logical_sizes_at_once = self
3491 : .conf
3492 : .concurrent_tenant_size_logical_size_queries
3493 : .inner();
3494 :
3495 : // TODO: Having a single mutex block concurrent reads is not great for performance.
3496 : //
3497 : // But the only case where we need to run multiple of these at once is when we
3498 : // request a size for a tenant manually via API, while another background calculation
3499 : // is in progress (which is not a common case).
3500 : //
3501 : // See more for on the issue #2748 condenced out of the initial PR review.
3502 : let mut shared_cache = tokio::select! {
3503 : locked = self.cached_logical_sizes.lock() => locked,
3504 : _ = cancel.cancelled() => return Err(size::CalculateSyntheticSizeError::Cancelled),
3505 : _ = self.cancel.cancelled() => return Err(size::CalculateSyntheticSizeError::Cancelled),
3506 : };
3507 :
3508 : size::gather_inputs(
3509 : self,
3510 : logical_sizes_at_once,
3511 : max_retention_period,
3512 : &mut shared_cache,
3513 : cause,
3514 : cancel,
3515 : ctx,
3516 : )
3517 : .await
3518 : }
3519 :
3520 : /// Calculate synthetic tenant size and cache the result.
3521 : /// This is periodically called by background worker.
3522 : /// result is cached in tenant struct
3523 0 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
3524 : pub async fn calculate_synthetic_size(
3525 : &self,
3526 : cause: LogicalSizeCalculationCause,
3527 : cancel: &CancellationToken,
3528 : ctx: &RequestContext,
3529 : ) -> Result<u64, size::CalculateSyntheticSizeError> {
3530 : let inputs = self.gather_size_inputs(None, cause, cancel, ctx).await?;
3531 :
3532 : let size = inputs.calculate();
3533 :
3534 : self.set_cached_synthetic_size(size);
3535 :
3536 : Ok(size)
3537 : }
3538 :
3539 : /// Cache given synthetic size and update the metric value
3540 0 : pub fn set_cached_synthetic_size(&self, size: u64) {
3541 0 : self.cached_synthetic_tenant_size
3542 0 : .store(size, Ordering::Relaxed);
3543 0 :
3544 0 : // Only shard zero should be calculating synthetic sizes
3545 0 : debug_assert!(self.shard_identity.is_shard_zero());
3546 :
3547 0 : TENANT_SYNTHETIC_SIZE_METRIC
3548 0 : .get_metric_with_label_values(&[&self.tenant_shard_id.tenant_id.to_string()])
3549 0 : .unwrap()
3550 0 : .set(size);
3551 0 : }
3552 :
3553 0 : pub fn cached_synthetic_size(&self) -> u64 {
3554 0 : self.cached_synthetic_tenant_size.load(Ordering::Relaxed)
3555 0 : }
3556 :
3557 : /// Flush any in-progress layers, schedule uploads, and wait for uploads to complete.
3558 : ///
3559 : /// This function can take a long time: callers should wrap it in a timeout if calling
3560 : /// from an external API handler.
3561 : ///
3562 : /// Cancel-safety: cancelling this function may leave I/O running, but such I/O is
3563 : /// still bounded by tenant/timeline shutdown.
3564 0 : #[tracing::instrument(skip_all)]
3565 : pub(crate) async fn flush_remote(&self) -> anyhow::Result<()> {
3566 : let timelines = self.timelines.lock().unwrap().clone();
3567 :
3568 0 : async fn flush_timeline(_gate: GateGuard, timeline: Arc<Timeline>) -> anyhow::Result<()> {
3569 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Flushing...");
3570 0 : timeline.freeze_and_flush().await?;
3571 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Waiting for uploads...");
3572 0 : timeline.remote_client.wait_completion().await?;
3573 :
3574 0 : Ok(())
3575 0 : }
3576 :
3577 : // We do not use a JoinSet for these tasks, because we don't want them to be
3578 : // aborted when this function's future is cancelled: they should stay alive
3579 : // holding their GateGuard until they complete, to ensure their I/Os complete
3580 : // before Timeline shutdown completes.
3581 : let mut results = FuturesUnordered::new();
3582 :
3583 : for (_timeline_id, timeline) in timelines {
3584 : // Run each timeline's flush in a task holding the timeline's gate: this
3585 : // means that if this function's future is cancelled, the Timeline shutdown
3586 : // will still wait for any I/O in here to complete.
3587 : let Ok(gate) = timeline.gate.enter() else {
3588 : continue;
3589 : };
3590 0 : let jh = tokio::task::spawn(async move { flush_timeline(gate, timeline).await });
3591 : results.push(jh);
3592 : }
3593 :
3594 : while let Some(r) = results.next().await {
3595 : if let Err(e) = r {
3596 : if !e.is_cancelled() && !e.is_panic() {
3597 : tracing::error!("unexpected join error: {e:?}");
3598 : }
3599 : }
3600 : }
3601 :
3602 : // The flushes we did above were just writes, but the Tenant might have had
3603 : // pending deletions as well from recent compaction/gc: we want to flush those
3604 : // as well. This requires flushing the global delete queue. This is cheap
3605 : // because it's typically a no-op.
3606 : match self.deletion_queue_client.flush_execute().await {
3607 : Ok(_) => {}
3608 : Err(DeletionQueueError::ShuttingDown) => {}
3609 : }
3610 :
3611 : Ok(())
3612 : }
3613 :
3614 0 : pub(crate) fn get_tenant_conf(&self) -> TenantConfOpt {
3615 0 : self.tenant_conf.load().tenant_conf.clone()
3616 0 : }
3617 : }
3618 :
3619 : /// Create the cluster temporarily in 'initdbpath' directory inside the repository
3620 : /// to get bootstrap data for timeline initialization.
3621 0 : async fn run_initdb(
3622 0 : conf: &'static PageServerConf,
3623 0 : initdb_target_dir: &Utf8Path,
3624 0 : pg_version: u32,
3625 0 : cancel: &CancellationToken,
3626 0 : ) -> Result<(), InitdbError> {
3627 0 : let initdb_bin_path = conf
3628 0 : .pg_bin_dir(pg_version)
3629 0 : .map_err(InitdbError::Other)?
3630 0 : .join("initdb");
3631 0 : let initdb_lib_dir = conf.pg_lib_dir(pg_version).map_err(InitdbError::Other)?;
3632 0 : info!(
3633 0 : "running {} in {}, libdir: {}",
3634 : initdb_bin_path, initdb_target_dir, initdb_lib_dir,
3635 : );
3636 :
3637 0 : let _permit = INIT_DB_SEMAPHORE.acquire().await;
3638 :
3639 0 : let initdb_command = tokio::process::Command::new(&initdb_bin_path)
3640 0 : .args(["-D", initdb_target_dir.as_ref()])
3641 0 : .args(["-U", &conf.superuser])
3642 0 : .args(["-E", "utf8"])
3643 0 : .arg("--no-instructions")
3644 0 : .arg("--no-sync")
3645 0 : .env_clear()
3646 0 : .env("LD_LIBRARY_PATH", &initdb_lib_dir)
3647 0 : .env("DYLD_LIBRARY_PATH", &initdb_lib_dir)
3648 0 : .stdin(std::process::Stdio::null())
3649 0 : // stdout invocation produces the same output every time, we don't need it
3650 0 : .stdout(std::process::Stdio::null())
3651 0 : // we would be interested in the stderr output, if there was any
3652 0 : .stderr(std::process::Stdio::piped())
3653 0 : .spawn()?;
3654 :
3655 : // Ideally we'd select here with the cancellation token, but the problem is that
3656 : // we can't safely terminate initdb: it launches processes of its own, and killing
3657 : // initdb doesn't kill them. After we return from this function, we want the target
3658 : // directory to be able to be cleaned up.
3659 : // See https://github.com/neondatabase/neon/issues/6385
3660 0 : let initdb_output = initdb_command.wait_with_output().await?;
3661 0 : if !initdb_output.status.success() {
3662 0 : return Err(InitdbError::Failed(
3663 0 : initdb_output.status,
3664 0 : initdb_output.stderr,
3665 0 : ));
3666 0 : }
3667 0 :
3668 0 : // This isn't true cancellation support, see above. Still return an error to
3669 0 : // excercise the cancellation code path.
3670 0 : if cancel.is_cancelled() {
3671 0 : return Err(InitdbError::Cancelled);
3672 0 : }
3673 0 :
3674 0 : Ok(())
3675 0 : }
3676 :
3677 : /// Dump contents of a layer file to stdout.
3678 0 : pub async fn dump_layerfile_from_path(
3679 0 : path: &Utf8Path,
3680 0 : verbose: bool,
3681 0 : ctx: &RequestContext,
3682 0 : ) -> anyhow::Result<()> {
3683 : use std::os::unix::fs::FileExt;
3684 :
3685 : // All layer files start with a two-byte "magic" value, to identify the kind of
3686 : // file.
3687 0 : let file = File::open(path)?;
3688 0 : let mut header_buf = [0u8; 2];
3689 0 : file.read_exact_at(&mut header_buf, 0)?;
3690 :
3691 0 : match u16::from_be_bytes(header_buf) {
3692 : crate::IMAGE_FILE_MAGIC => {
3693 0 : ImageLayer::new_for_path(path, file)?
3694 0 : .dump(verbose, ctx)
3695 0 : .await?
3696 : }
3697 : crate::DELTA_FILE_MAGIC => {
3698 0 : DeltaLayer::new_for_path(path, file)?
3699 0 : .dump(verbose, ctx)
3700 0 : .await?
3701 : }
3702 0 : magic => bail!("unrecognized magic identifier: {:?}", magic),
3703 : }
3704 :
3705 0 : Ok(())
3706 0 : }
3707 :
3708 : #[cfg(test)]
3709 : pub(crate) mod harness {
3710 : use bytes::{Bytes, BytesMut};
3711 : use once_cell::sync::OnceCell;
3712 : use pageserver_api::models::ShardParameters;
3713 : use pageserver_api::shard::ShardIndex;
3714 : use utils::logging;
3715 :
3716 : use crate::deletion_queue::mock::MockDeletionQueue;
3717 : use crate::l0_flush::L0FlushConfig;
3718 : use crate::walredo::apply_neon;
3719 : use crate::{repository::Key, walrecord::NeonWalRecord};
3720 :
3721 : use super::*;
3722 : use hex_literal::hex;
3723 : use utils::id::TenantId;
3724 :
3725 : pub const TIMELINE_ID: TimelineId =
3726 : TimelineId::from_array(hex!("11223344556677881122334455667788"));
3727 : pub const NEW_TIMELINE_ID: TimelineId =
3728 : TimelineId::from_array(hex!("AA223344556677881122334455667788"));
3729 :
3730 : /// Convenience function to create a page image with given string as the only content
3731 5028422 : pub fn test_img(s: &str) -> Bytes {
3732 5028422 : let mut buf = BytesMut::new();
3733 5028422 : buf.extend_from_slice(s.as_bytes());
3734 5028422 : buf.resize(64, 0);
3735 5028422 :
3736 5028422 : buf.freeze()
3737 5028422 : }
3738 :
3739 : impl From<TenantConf> for TenantConfOpt {
3740 168 : fn from(tenant_conf: TenantConf) -> Self {
3741 168 : Self {
3742 168 : checkpoint_distance: Some(tenant_conf.checkpoint_distance),
3743 168 : checkpoint_timeout: Some(tenant_conf.checkpoint_timeout),
3744 168 : compaction_target_size: Some(tenant_conf.compaction_target_size),
3745 168 : compaction_period: Some(tenant_conf.compaction_period),
3746 168 : compaction_threshold: Some(tenant_conf.compaction_threshold),
3747 168 : compaction_algorithm: Some(tenant_conf.compaction_algorithm),
3748 168 : gc_horizon: Some(tenant_conf.gc_horizon),
3749 168 : gc_period: Some(tenant_conf.gc_period),
3750 168 : image_creation_threshold: Some(tenant_conf.image_creation_threshold),
3751 168 : pitr_interval: Some(tenant_conf.pitr_interval),
3752 168 : walreceiver_connect_timeout: Some(tenant_conf.walreceiver_connect_timeout),
3753 168 : lagging_wal_timeout: Some(tenant_conf.lagging_wal_timeout),
3754 168 : max_lsn_wal_lag: Some(tenant_conf.max_lsn_wal_lag),
3755 168 : eviction_policy: Some(tenant_conf.eviction_policy),
3756 168 : min_resident_size_override: tenant_conf.min_resident_size_override,
3757 168 : evictions_low_residence_duration_metric_threshold: Some(
3758 168 : tenant_conf.evictions_low_residence_duration_metric_threshold,
3759 168 : ),
3760 168 : heatmap_period: Some(tenant_conf.heatmap_period),
3761 168 : lazy_slru_download: Some(tenant_conf.lazy_slru_download),
3762 168 : timeline_get_throttle: Some(tenant_conf.timeline_get_throttle),
3763 168 : image_layer_creation_check_threshold: Some(
3764 168 : tenant_conf.image_layer_creation_check_threshold,
3765 168 : ),
3766 168 : switch_aux_file_policy: Some(tenant_conf.switch_aux_file_policy),
3767 168 : lsn_lease_length: Some(tenant_conf.lsn_lease_length),
3768 168 : lsn_lease_length_for_ts: Some(tenant_conf.lsn_lease_length_for_ts),
3769 168 : }
3770 168 : }
3771 : }
3772 :
3773 : pub struct TenantHarness {
3774 : pub conf: &'static PageServerConf,
3775 : pub tenant_conf: TenantConf,
3776 : pub tenant_shard_id: TenantShardId,
3777 : pub generation: Generation,
3778 : pub shard: ShardIndex,
3779 : pub remote_storage: GenericRemoteStorage,
3780 : pub remote_fs_dir: Utf8PathBuf,
3781 : pub deletion_queue: MockDeletionQueue,
3782 : }
3783 :
3784 : static LOG_HANDLE: OnceCell<()> = OnceCell::new();
3785 :
3786 174 : pub(crate) fn setup_logging() {
3787 174 : LOG_HANDLE.get_or_init(|| {
3788 162 : logging::init(
3789 162 : logging::LogFormat::Test,
3790 162 : // enable it in case the tests exercise code paths that use
3791 162 : // debug_assert_current_span_has_tenant_and_timeline_id
3792 162 : logging::TracingErrorLayerEnablement::EnableWithRustLogFilter,
3793 162 : logging::Output::Stdout,
3794 162 : )
3795 162 : .expect("Failed to init test logging")
3796 174 : });
3797 174 : }
3798 :
3799 : impl TenantHarness {
3800 168 : pub async fn create_custom(
3801 168 : test_name: &'static str,
3802 168 : tenant_conf: TenantConf,
3803 168 : tenant_id: TenantId,
3804 168 : shard_identity: ShardIdentity,
3805 168 : generation: Generation,
3806 168 : ) -> anyhow::Result<Self> {
3807 168 : setup_logging();
3808 168 :
3809 168 : let repo_dir = PageServerConf::test_repo_dir(test_name);
3810 168 : let _ = fs::remove_dir_all(&repo_dir);
3811 168 : fs::create_dir_all(&repo_dir)?;
3812 :
3813 168 : let conf = PageServerConf::dummy_conf(repo_dir);
3814 168 : // Make a static copy of the config. This can never be free'd, but that's
3815 168 : // OK in a test.
3816 168 : let conf: &'static PageServerConf = Box::leak(Box::new(conf));
3817 168 :
3818 168 : let shard = shard_identity.shard_index();
3819 168 : let tenant_shard_id = TenantShardId {
3820 168 : tenant_id,
3821 168 : shard_number: shard.shard_number,
3822 168 : shard_count: shard.shard_count,
3823 168 : };
3824 168 : fs::create_dir_all(conf.tenant_path(&tenant_shard_id))?;
3825 168 : fs::create_dir_all(conf.timelines_path(&tenant_shard_id))?;
3826 :
3827 : use remote_storage::{RemoteStorageConfig, RemoteStorageKind};
3828 168 : let remote_fs_dir = conf.workdir.join("localfs");
3829 168 : std::fs::create_dir_all(&remote_fs_dir).unwrap();
3830 168 : let config = RemoteStorageConfig {
3831 168 : storage: RemoteStorageKind::LocalFs {
3832 168 : local_path: remote_fs_dir.clone(),
3833 168 : },
3834 168 : timeout: RemoteStorageConfig::DEFAULT_TIMEOUT,
3835 168 : };
3836 168 : let remote_storage = GenericRemoteStorage::from_config(&config).await.unwrap();
3837 168 : let deletion_queue = MockDeletionQueue::new(Some(remote_storage.clone()));
3838 168 :
3839 168 : Ok(Self {
3840 168 : conf,
3841 168 : tenant_conf,
3842 168 : tenant_shard_id,
3843 168 : generation,
3844 168 : shard,
3845 168 : remote_storage,
3846 168 : remote_fs_dir,
3847 168 : deletion_queue,
3848 168 : })
3849 168 : }
3850 :
3851 156 : pub async fn create(test_name: &'static str) -> anyhow::Result<Self> {
3852 156 : // Disable automatic GC and compaction to make the unit tests more deterministic.
3853 156 : // The tests perform them manually if needed.
3854 156 : let tenant_conf = TenantConf {
3855 156 : gc_period: Duration::ZERO,
3856 156 : compaction_period: Duration::ZERO,
3857 156 : ..TenantConf::default()
3858 156 : };
3859 156 : let tenant_id = TenantId::generate();
3860 156 : let shard = ShardIdentity::unsharded();
3861 156 : Self::create_custom(
3862 156 : test_name,
3863 156 : tenant_conf,
3864 156 : tenant_id,
3865 156 : shard,
3866 156 : Generation::new(0xdeadbeef),
3867 156 : )
3868 0 : .await
3869 156 : }
3870 :
3871 18 : pub fn span(&self) -> tracing::Span {
3872 18 : info_span!("TenantHarness", tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug())
3873 18 : }
3874 :
3875 168 : pub(crate) async fn load(&self) -> (Arc<Tenant>, RequestContext) {
3876 168 : let ctx = RequestContext::new(TaskKind::UnitTest, DownloadBehavior::Error);
3877 168 : (
3878 168 : self.do_try_load(&ctx)
3879 677 : .await
3880 168 : .expect("failed to load test tenant"),
3881 168 : ctx,
3882 168 : )
3883 168 : }
3884 :
3885 336 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
3886 : pub(crate) async fn do_try_load(
3887 : &self,
3888 : ctx: &RequestContext,
3889 : ) -> anyhow::Result<Arc<Tenant>> {
3890 : let walredo_mgr = Arc::new(WalRedoManager::from(TestRedoManager));
3891 :
3892 : let tenant = Arc::new(Tenant::new(
3893 : TenantState::Loading,
3894 : self.conf,
3895 : AttachedTenantConf::try_from(LocationConf::attached_single(
3896 : TenantConfOpt::from(self.tenant_conf.clone()),
3897 : self.generation,
3898 : &ShardParameters::default(),
3899 : ))
3900 : .unwrap(),
3901 : // This is a legacy/test code path: sharding isn't supported here.
3902 : ShardIdentity::unsharded(),
3903 : Some(walredo_mgr),
3904 : self.tenant_shard_id,
3905 : self.remote_storage.clone(),
3906 : self.deletion_queue.new_client(),
3907 : // TODO: ideally we should run all unit tests with both configs
3908 : L0FlushGlobalState::new(L0FlushConfig::default()),
3909 : ));
3910 :
3911 : let preload = tenant
3912 : .preload(&self.remote_storage, CancellationToken::new())
3913 : .await?;
3914 : tenant.attach(Some(preload), ctx).await?;
3915 :
3916 : tenant.state.send_replace(TenantState::Active);
3917 : for timeline in tenant.timelines.lock().unwrap().values() {
3918 : timeline.set_state(TimelineState::Active);
3919 : }
3920 : Ok(tenant)
3921 : }
3922 :
3923 2 : pub fn timeline_path(&self, timeline_id: &TimelineId) -> Utf8PathBuf {
3924 2 : self.conf.timeline_path(&self.tenant_shard_id, timeline_id)
3925 2 : }
3926 : }
3927 :
3928 : // Mock WAL redo manager that doesn't do much
3929 : pub(crate) struct TestRedoManager;
3930 :
3931 : impl TestRedoManager {
3932 : /// # Cancel-Safety
3933 : ///
3934 : /// This method is cancellation-safe.
3935 58 : pub async fn request_redo(
3936 58 : &self,
3937 58 : key: Key,
3938 58 : lsn: Lsn,
3939 58 : base_img: Option<(Lsn, Bytes)>,
3940 58 : records: Vec<(Lsn, NeonWalRecord)>,
3941 58 : _pg_version: u32,
3942 58 : ) -> Result<Bytes, walredo::Error> {
3943 76 : let records_neon = records.iter().all(|r| apply_neon::can_apply_in_neon(&r.1));
3944 58 : if records_neon {
3945 : // For Neon wal records, we can decode without spawning postgres, so do so.
3946 58 : let base_img = base_img.expect("Neon WAL redo requires base image").1;
3947 58 : let mut page = BytesMut::new();
3948 58 : page.extend_from_slice(&base_img);
3949 134 : for (record_lsn, record) in records {
3950 76 : apply_neon::apply_in_neon(&record, record_lsn, key, &mut page)?;
3951 : }
3952 58 : Ok(page.freeze())
3953 : } else {
3954 : // We never spawn a postgres walredo process in unit tests: just log what we might have done.
3955 0 : let s = format!(
3956 0 : "redo for {} to get to {}, with {} and {} records",
3957 0 : key,
3958 0 : lsn,
3959 0 : if base_img.is_some() {
3960 0 : "base image"
3961 : } else {
3962 0 : "no base image"
3963 : },
3964 0 : records.len()
3965 0 : );
3966 0 : println!("{s}");
3967 0 :
3968 0 : Ok(test_img(&s))
3969 : }
3970 58 : }
3971 : }
3972 : }
3973 :
3974 : #[cfg(test)]
3975 : mod tests {
3976 : use std::collections::BTreeMap;
3977 :
3978 : use super::*;
3979 : use crate::keyspace::KeySpaceAccum;
3980 : use crate::pgdatadir_mapping::AuxFilesDirectory;
3981 : use crate::repository::{Key, Value};
3982 : use crate::tenant::harness::*;
3983 : use crate::tenant::timeline::CompactFlags;
3984 : use crate::walrecord::NeonWalRecord;
3985 : use crate::DEFAULT_PG_VERSION;
3986 : use bytes::{Bytes, BytesMut};
3987 : use hex_literal::hex;
3988 : use itertools::Itertools;
3989 : use pageserver_api::key::{AUX_FILES_KEY, AUX_KEY_PREFIX, NON_INHERITED_RANGE};
3990 : use pageserver_api::keyspace::KeySpace;
3991 : use pageserver_api::models::{CompactionAlgorithm, CompactionAlgorithmSettings};
3992 : use rand::{thread_rng, Rng};
3993 : use storage_layer::PersistentLayerKey;
3994 : use tests::storage_layer::ValuesReconstructState;
3995 : use tests::timeline::{GetVectoredError, ShutdownMode};
3996 : use timeline::{DeltaLayerTestDesc, GcInfo};
3997 : use utils::bin_ser::BeSer;
3998 : use utils::id::TenantId;
3999 :
4000 : static TEST_KEY: Lazy<Key> =
4001 18 : Lazy::new(|| Key::from_slice(&hex!("010000000033333333444444445500000001")));
4002 :
4003 : #[tokio::test]
4004 2 : async fn test_basic() -> anyhow::Result<()> {
4005 8 : let (tenant, ctx) = TenantHarness::create("test_basic").await?.load().await;
4006 2 : let tline = tenant
4007 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
4008 6 : .await?;
4009 2 :
4010 2 : let mut writer = tline.writer().await;
4011 2 : writer
4012 2 : .put(
4013 2 : *TEST_KEY,
4014 2 : Lsn(0x10),
4015 2 : &Value::Image(test_img("foo at 0x10")),
4016 2 : &ctx,
4017 2 : )
4018 2 : .await?;
4019 2 : writer.finish_write(Lsn(0x10));
4020 2 : drop(writer);
4021 2 :
4022 2 : let mut writer = tline.writer().await;
4023 2 : writer
4024 2 : .put(
4025 2 : *TEST_KEY,
4026 2 : Lsn(0x20),
4027 2 : &Value::Image(test_img("foo at 0x20")),
4028 2 : &ctx,
4029 2 : )
4030 2 : .await?;
4031 2 : writer.finish_write(Lsn(0x20));
4032 2 : drop(writer);
4033 2 :
4034 2 : assert_eq!(
4035 2 : tline.get(*TEST_KEY, Lsn(0x10), &ctx).await?,
4036 2 : test_img("foo at 0x10")
4037 2 : );
4038 2 : assert_eq!(
4039 2 : tline.get(*TEST_KEY, Lsn(0x1f), &ctx).await?,
4040 2 : test_img("foo at 0x10")
4041 2 : );
4042 2 : assert_eq!(
4043 2 : tline.get(*TEST_KEY, Lsn(0x20), &ctx).await?,
4044 2 : test_img("foo at 0x20")
4045 2 : );
4046 2 :
4047 2 : Ok(())
4048 2 : }
4049 :
4050 : #[tokio::test]
4051 2 : async fn no_duplicate_timelines() -> anyhow::Result<()> {
4052 2 : let (tenant, ctx) = TenantHarness::create("no_duplicate_timelines")
4053 2 : .await?
4054 2 : .load()
4055 8 : .await;
4056 2 : let _ = tenant
4057 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
4058 5 : .await?;
4059 2 :
4060 2 : match tenant
4061 2 : .create_empty_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
4062 2 : .await
4063 2 : {
4064 2 : Ok(_) => panic!("duplicate timeline creation should fail"),
4065 2 : Err(e) => assert_eq!(e.to_string(), "Already exists".to_string()),
4066 2 : }
4067 2 :
4068 2 : Ok(())
4069 2 : }
4070 :
4071 : /// Convenience function to create a page image with given string as the only content
4072 10 : pub fn test_value(s: &str) -> Value {
4073 10 : let mut buf = BytesMut::new();
4074 10 : buf.extend_from_slice(s.as_bytes());
4075 10 : Value::Image(buf.freeze())
4076 10 : }
4077 :
4078 : ///
4079 : /// Test branch creation
4080 : ///
4081 : #[tokio::test]
4082 2 : async fn test_branch() -> anyhow::Result<()> {
4083 2 : use std::str::from_utf8;
4084 2 :
4085 8 : let (tenant, ctx) = TenantHarness::create("test_branch").await?.load().await;
4086 2 : let tline = tenant
4087 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
4088 6 : .await?;
4089 2 : let mut writer = tline.writer().await;
4090 2 :
4091 2 : #[allow(non_snake_case)]
4092 2 : let TEST_KEY_A: Key = Key::from_hex("110000000033333333444444445500000001").unwrap();
4093 2 : #[allow(non_snake_case)]
4094 2 : let TEST_KEY_B: Key = Key::from_hex("110000000033333333444444445500000002").unwrap();
4095 2 :
4096 2 : // Insert a value on the timeline
4097 2 : writer
4098 2 : .put(TEST_KEY_A, Lsn(0x20), &test_value("foo at 0x20"), &ctx)
4099 2 : .await?;
4100 2 : writer
4101 2 : .put(TEST_KEY_B, Lsn(0x20), &test_value("foobar at 0x20"), &ctx)
4102 2 : .await?;
4103 2 : writer.finish_write(Lsn(0x20));
4104 2 :
4105 2 : writer
4106 2 : .put(TEST_KEY_A, Lsn(0x30), &test_value("foo at 0x30"), &ctx)
4107 2 : .await?;
4108 2 : writer.finish_write(Lsn(0x30));
4109 2 : writer
4110 2 : .put(TEST_KEY_A, Lsn(0x40), &test_value("foo at 0x40"), &ctx)
4111 2 : .await?;
4112 2 : writer.finish_write(Lsn(0x40));
4113 2 :
4114 2 : //assert_current_logical_size(&tline, Lsn(0x40));
4115 2 :
4116 2 : // Branch the history, modify relation differently on the new timeline
4117 2 : tenant
4118 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x30)), &ctx)
4119 2 : .await?;
4120 2 : let newtline = tenant
4121 2 : .get_timeline(NEW_TIMELINE_ID, true)
4122 2 : .expect("Should have a local timeline");
4123 2 : let mut new_writer = newtline.writer().await;
4124 2 : new_writer
4125 2 : .put(TEST_KEY_A, Lsn(0x40), &test_value("bar at 0x40"), &ctx)
4126 2 : .await?;
4127 2 : new_writer.finish_write(Lsn(0x40));
4128 2 :
4129 2 : // Check page contents on both branches
4130 2 : assert_eq!(
4131 2 : from_utf8(&tline.get(TEST_KEY_A, Lsn(0x40), &ctx).await?)?,
4132 2 : "foo at 0x40"
4133 2 : );
4134 2 : assert_eq!(
4135 2 : from_utf8(&newtline.get(TEST_KEY_A, Lsn(0x40), &ctx).await?)?,
4136 2 : "bar at 0x40"
4137 2 : );
4138 2 : assert_eq!(
4139 2 : from_utf8(&newtline.get(TEST_KEY_B, Lsn(0x40), &ctx).await?)?,
4140 2 : "foobar at 0x20"
4141 2 : );
4142 2 :
4143 2 : //assert_current_logical_size(&tline, Lsn(0x40));
4144 2 :
4145 2 : Ok(())
4146 2 : }
4147 :
4148 20 : async fn make_some_layers(
4149 20 : tline: &Timeline,
4150 20 : start_lsn: Lsn,
4151 20 : ctx: &RequestContext,
4152 20 : ) -> anyhow::Result<()> {
4153 20 : let mut lsn = start_lsn;
4154 : {
4155 20 : let mut writer = tline.writer().await;
4156 : // Create a relation on the timeline
4157 20 : writer
4158 20 : .put(
4159 20 : *TEST_KEY,
4160 20 : lsn,
4161 20 : &Value::Image(test_img(&format!("foo at {}", lsn))),
4162 20 : ctx,
4163 20 : )
4164 10 : .await?;
4165 20 : writer.finish_write(lsn);
4166 20 : lsn += 0x10;
4167 20 : writer
4168 20 : .put(
4169 20 : *TEST_KEY,
4170 20 : lsn,
4171 20 : &Value::Image(test_img(&format!("foo at {}", lsn))),
4172 20 : ctx,
4173 20 : )
4174 0 : .await?;
4175 20 : writer.finish_write(lsn);
4176 20 : lsn += 0x10;
4177 20 : }
4178 20 : tline.freeze_and_flush().await?;
4179 : {
4180 20 : let mut writer = tline.writer().await;
4181 20 : writer
4182 20 : .put(
4183 20 : *TEST_KEY,
4184 20 : lsn,
4185 20 : &Value::Image(test_img(&format!("foo at {}", lsn))),
4186 20 : ctx,
4187 20 : )
4188 10 : .await?;
4189 20 : writer.finish_write(lsn);
4190 20 : lsn += 0x10;
4191 20 : writer
4192 20 : .put(
4193 20 : *TEST_KEY,
4194 20 : lsn,
4195 20 : &Value::Image(test_img(&format!("foo at {}", lsn))),
4196 20 : ctx,
4197 20 : )
4198 0 : .await?;
4199 20 : writer.finish_write(lsn);
4200 20 : }
4201 20 : tline.freeze_and_flush().await.map_err(|e| e.into())
4202 20 : }
4203 :
4204 : #[tokio::test]
4205 2 : async fn test_prohibit_branch_creation_on_garbage_collected_data() -> anyhow::Result<()> {
4206 2 : let (tenant, ctx) =
4207 2 : TenantHarness::create("test_prohibit_branch_creation_on_garbage_collected_data")
4208 2 : .await?
4209 2 : .load()
4210 8 : .await;
4211 2 : let tline = tenant
4212 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
4213 6 : .await?;
4214 6 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
4215 2 :
4216 2 : // this removes layers before lsn 40 (50 minus 10), so there are two remaining layers, image and delta for 31-50
4217 2 : // FIXME: this doesn't actually remove any layer currently, given how the flushing
4218 2 : // and compaction works. But it does set the 'cutoff' point so that the cross check
4219 2 : // below should fail.
4220 2 : tenant
4221 2 : .gc_iteration(
4222 2 : Some(TIMELINE_ID),
4223 2 : 0x10,
4224 2 : Duration::ZERO,
4225 2 : &CancellationToken::new(),
4226 2 : &ctx,
4227 2 : )
4228 2 : .await?;
4229 2 :
4230 2 : // try to branch at lsn 25, should fail because we already garbage collected the data
4231 2 : match tenant
4232 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x25)), &ctx)
4233 2 : .await
4234 2 : {
4235 2 : Ok(_) => panic!("branching should have failed"),
4236 2 : Err(err) => {
4237 2 : let CreateTimelineError::AncestorLsn(err) = err else {
4238 2 : panic!("wrong error type")
4239 2 : };
4240 2 : assert!(err.to_string().contains("invalid branch start lsn"));
4241 2 : assert!(err
4242 2 : .source()
4243 2 : .unwrap()
4244 2 : .to_string()
4245 2 : .contains("we might've already garbage collected needed data"))
4246 2 : }
4247 2 : }
4248 2 :
4249 2 : Ok(())
4250 2 : }
4251 :
4252 : #[tokio::test]
4253 2 : async fn test_prohibit_branch_creation_on_pre_initdb_lsn() -> anyhow::Result<()> {
4254 2 : let (tenant, ctx) =
4255 2 : TenantHarness::create("test_prohibit_branch_creation_on_pre_initdb_lsn")
4256 2 : .await?
4257 2 : .load()
4258 8 : .await;
4259 2 :
4260 2 : let tline = tenant
4261 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x50), DEFAULT_PG_VERSION, &ctx)
4262 6 : .await?;
4263 2 : // try to branch at lsn 0x25, should fail because initdb lsn is 0x50
4264 2 : match tenant
4265 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x25)), &ctx)
4266 2 : .await
4267 2 : {
4268 2 : Ok(_) => panic!("branching should have failed"),
4269 2 : Err(err) => {
4270 2 : let CreateTimelineError::AncestorLsn(err) = err else {
4271 2 : panic!("wrong error type");
4272 2 : };
4273 2 : assert!(&err.to_string().contains("invalid branch start lsn"));
4274 2 : assert!(&err
4275 2 : .source()
4276 2 : .unwrap()
4277 2 : .to_string()
4278 2 : .contains("is earlier than latest GC cutoff"));
4279 2 : }
4280 2 : }
4281 2 :
4282 2 : Ok(())
4283 2 : }
4284 :
4285 : /*
4286 : // FIXME: This currently fails to error out. Calling GC doesn't currently
4287 : // remove the old value, we'd need to work a little harder
4288 : #[tokio::test]
4289 : async fn test_prohibit_get_for_garbage_collected_data() -> anyhow::Result<()> {
4290 : let repo =
4291 : RepoHarness::create("test_prohibit_get_for_garbage_collected_data")?
4292 : .load();
4293 :
4294 : let tline = repo.create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION)?;
4295 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
4296 :
4297 : repo.gc_iteration(Some(TIMELINE_ID), 0x10, Duration::ZERO)?;
4298 : let latest_gc_cutoff_lsn = tline.get_latest_gc_cutoff_lsn();
4299 : assert!(*latest_gc_cutoff_lsn > Lsn(0x25));
4300 : match tline.get(*TEST_KEY, Lsn(0x25)) {
4301 : Ok(_) => panic!("request for page should have failed"),
4302 : Err(err) => assert!(err.to_string().contains("not found at")),
4303 : }
4304 : Ok(())
4305 : }
4306 : */
4307 :
4308 : #[tokio::test]
4309 2 : async fn test_get_branchpoints_from_an_inactive_timeline() -> anyhow::Result<()> {
4310 2 : let (tenant, ctx) =
4311 2 : TenantHarness::create("test_get_branchpoints_from_an_inactive_timeline")
4312 2 : .await?
4313 2 : .load()
4314 8 : .await;
4315 2 : let tline = tenant
4316 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
4317 6 : .await?;
4318 6 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
4319 2 :
4320 2 : tenant
4321 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
4322 2 : .await?;
4323 2 : let newtline = tenant
4324 2 : .get_timeline(NEW_TIMELINE_ID, true)
4325 2 : .expect("Should have a local timeline");
4326 2 :
4327 6 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
4328 2 :
4329 2 : tline.set_broken("test".to_owned());
4330 2 :
4331 2 : tenant
4332 2 : .gc_iteration(
4333 2 : Some(TIMELINE_ID),
4334 2 : 0x10,
4335 2 : Duration::ZERO,
4336 2 : &CancellationToken::new(),
4337 2 : &ctx,
4338 2 : )
4339 2 : .await?;
4340 2 :
4341 2 : // The branchpoints should contain all timelines, even ones marked
4342 2 : // as Broken.
4343 2 : {
4344 2 : let branchpoints = &tline.gc_info.read().unwrap().retain_lsns;
4345 2 : assert_eq!(branchpoints.len(), 1);
4346 2 : assert_eq!(branchpoints[0], Lsn(0x40));
4347 2 : }
4348 2 :
4349 2 : // You can read the key from the child branch even though the parent is
4350 2 : // Broken, as long as you don't need to access data from the parent.
4351 2 : assert_eq!(
4352 4 : newtline.get(*TEST_KEY, Lsn(0x70), &ctx).await?,
4353 2 : test_img(&format!("foo at {}", Lsn(0x70)))
4354 2 : );
4355 2 :
4356 2 : // This needs to traverse to the parent, and fails.
4357 2 : let err = newtline.get(*TEST_KEY, Lsn(0x50), &ctx).await.unwrap_err();
4358 2 : assert!(err.to_string().starts_with(&format!(
4359 2 : "Bad state on timeline {}: Broken",
4360 2 : tline.timeline_id
4361 2 : )));
4362 2 :
4363 2 : Ok(())
4364 2 : }
4365 :
4366 : #[tokio::test]
4367 2 : async fn test_retain_data_in_parent_which_is_needed_for_child() -> anyhow::Result<()> {
4368 2 : let (tenant, ctx) =
4369 2 : TenantHarness::create("test_retain_data_in_parent_which_is_needed_for_child")
4370 2 : .await?
4371 2 : .load()
4372 8 : .await;
4373 2 : let tline = tenant
4374 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
4375 6 : .await?;
4376 6 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
4377 2 :
4378 2 : tenant
4379 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
4380 2 : .await?;
4381 2 : let newtline = tenant
4382 2 : .get_timeline(NEW_TIMELINE_ID, true)
4383 2 : .expect("Should have a local timeline");
4384 2 : // this removes layers before lsn 40 (50 minus 10), so there are two remaining layers, image and delta for 31-50
4385 2 : tenant
4386 2 : .gc_iteration(
4387 2 : Some(TIMELINE_ID),
4388 2 : 0x10,
4389 2 : Duration::ZERO,
4390 2 : &CancellationToken::new(),
4391 2 : &ctx,
4392 2 : )
4393 2 : .await?;
4394 4 : assert!(newtline.get(*TEST_KEY, Lsn(0x25), &ctx).await.is_ok());
4395 2 :
4396 2 : Ok(())
4397 2 : }
4398 : #[tokio::test]
4399 2 : async fn test_parent_keeps_data_forever_after_branching() -> anyhow::Result<()> {
4400 2 : let (tenant, ctx) = TenantHarness::create("test_parent_keeps_data_forever_after_branching")
4401 2 : .await?
4402 2 : .load()
4403 8 : .await;
4404 2 : let tline = tenant
4405 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
4406 6 : .await?;
4407 6 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
4408 2 :
4409 2 : tenant
4410 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
4411 2 : .await?;
4412 2 : let newtline = tenant
4413 2 : .get_timeline(NEW_TIMELINE_ID, true)
4414 2 : .expect("Should have a local timeline");
4415 2 :
4416 6 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
4417 2 :
4418 2 : // run gc on parent
4419 2 : tenant
4420 2 : .gc_iteration(
4421 2 : Some(TIMELINE_ID),
4422 2 : 0x10,
4423 2 : Duration::ZERO,
4424 2 : &CancellationToken::new(),
4425 2 : &ctx,
4426 2 : )
4427 2 : .await?;
4428 2 :
4429 2 : // Check that the data is still accessible on the branch.
4430 2 : assert_eq!(
4431 7 : newtline.get(*TEST_KEY, Lsn(0x50), &ctx).await?,
4432 2 : test_img(&format!("foo at {}", Lsn(0x40)))
4433 2 : );
4434 2 :
4435 2 : Ok(())
4436 2 : }
4437 :
4438 : #[tokio::test]
4439 2 : async fn timeline_load() -> anyhow::Result<()> {
4440 2 : const TEST_NAME: &str = "timeline_load";
4441 2 : let harness = TenantHarness::create(TEST_NAME).await?;
4442 2 : {
4443 8 : let (tenant, ctx) = harness.load().await;
4444 2 : let tline = tenant
4445 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x7000), DEFAULT_PG_VERSION, &ctx)
4446 6 : .await?;
4447 6 : make_some_layers(tline.as_ref(), Lsn(0x8000), &ctx).await?;
4448 2 : // so that all uploads finish & we can call harness.load() below again
4449 2 : tenant
4450 2 : .shutdown(Default::default(), ShutdownMode::FreezeAndFlush)
4451 2 : .instrument(harness.span())
4452 2 : .await
4453 2 : .ok()
4454 2 : .unwrap();
4455 2 : }
4456 2 :
4457 8 : let (tenant, _ctx) = harness.load().await;
4458 2 : tenant
4459 2 : .get_timeline(TIMELINE_ID, true)
4460 2 : .expect("cannot load timeline");
4461 2 :
4462 2 : Ok(())
4463 2 : }
4464 :
4465 : #[tokio::test]
4466 2 : async fn timeline_load_with_ancestor() -> anyhow::Result<()> {
4467 2 : const TEST_NAME: &str = "timeline_load_with_ancestor";
4468 2 : let harness = TenantHarness::create(TEST_NAME).await?;
4469 2 : // create two timelines
4470 2 : {
4471 8 : let (tenant, ctx) = harness.load().await;
4472 2 : let tline = tenant
4473 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
4474 6 : .await?;
4475 2 :
4476 6 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
4477 2 :
4478 2 : let child_tline = tenant
4479 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
4480 2 : .await?;
4481 2 : child_tline.set_state(TimelineState::Active);
4482 2 :
4483 2 : let newtline = tenant
4484 2 : .get_timeline(NEW_TIMELINE_ID, true)
4485 2 : .expect("Should have a local timeline");
4486 2 :
4487 6 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
4488 2 :
4489 2 : // so that all uploads finish & we can call harness.load() below again
4490 2 : tenant
4491 2 : .shutdown(Default::default(), ShutdownMode::FreezeAndFlush)
4492 2 : .instrument(harness.span())
4493 4 : .await
4494 2 : .ok()
4495 2 : .unwrap();
4496 2 : }
4497 2 :
4498 2 : // check that both of them are initially unloaded
4499 13 : let (tenant, _ctx) = harness.load().await;
4500 2 :
4501 2 : // check that both, child and ancestor are loaded
4502 2 : let _child_tline = tenant
4503 2 : .get_timeline(NEW_TIMELINE_ID, true)
4504 2 : .expect("cannot get child timeline loaded");
4505 2 :
4506 2 : let _ancestor_tline = tenant
4507 2 : .get_timeline(TIMELINE_ID, true)
4508 2 : .expect("cannot get ancestor timeline loaded");
4509 2 :
4510 2 : Ok(())
4511 2 : }
4512 :
4513 : #[tokio::test]
4514 2 : async fn delta_layer_dumping() -> anyhow::Result<()> {
4515 2 : use storage_layer::AsLayerDesc;
4516 2 : let (tenant, ctx) = TenantHarness::create("test_layer_dumping")
4517 2 : .await?
4518 2 : .load()
4519 8 : .await;
4520 2 : let tline = tenant
4521 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
4522 5 : .await?;
4523 6 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
4524 2 :
4525 2 : let layer_map = tline.layers.read().await;
4526 2 : let level0_deltas = layer_map
4527 2 : .layer_map()
4528 2 : .get_level0_deltas()?
4529 2 : .into_iter()
4530 4 : .map(|desc| layer_map.get_from_desc(&desc))
4531 2 : .collect::<Vec<_>>();
4532 2 :
4533 2 : assert!(!level0_deltas.is_empty());
4534 2 :
4535 6 : for delta in level0_deltas {
4536 2 : // Ensure we are dumping a delta layer here
4537 4 : assert!(delta.layer_desc().is_delta);
4538 8 : delta.dump(true, &ctx).await.unwrap();
4539 2 : }
4540 2 :
4541 2 : Ok(())
4542 2 : }
4543 :
4544 : #[tokio::test]
4545 2 : async fn test_images() -> anyhow::Result<()> {
4546 8 : let (tenant, ctx) = TenantHarness::create("test_images").await?.load().await;
4547 2 : let tline = tenant
4548 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
4549 6 : .await?;
4550 2 :
4551 2 : let mut writer = tline.writer().await;
4552 2 : writer
4553 2 : .put(
4554 2 : *TEST_KEY,
4555 2 : Lsn(0x10),
4556 2 : &Value::Image(test_img("foo at 0x10")),
4557 2 : &ctx,
4558 2 : )
4559 2 : .await?;
4560 2 : writer.finish_write(Lsn(0x10));
4561 2 : drop(writer);
4562 2 :
4563 2 : tline.freeze_and_flush().await?;
4564 2 : tline
4565 2 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
4566 2 : .await?;
4567 2 :
4568 2 : let mut writer = tline.writer().await;
4569 2 : writer
4570 2 : .put(
4571 2 : *TEST_KEY,
4572 2 : Lsn(0x20),
4573 2 : &Value::Image(test_img("foo at 0x20")),
4574 2 : &ctx,
4575 2 : )
4576 2 : .await?;
4577 2 : writer.finish_write(Lsn(0x20));
4578 2 : drop(writer);
4579 2 :
4580 2 : tline.freeze_and_flush().await?;
4581 2 : tline
4582 2 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
4583 2 : .await?;
4584 2 :
4585 2 : let mut writer = tline.writer().await;
4586 2 : writer
4587 2 : .put(
4588 2 : *TEST_KEY,
4589 2 : Lsn(0x30),
4590 2 : &Value::Image(test_img("foo at 0x30")),
4591 2 : &ctx,
4592 2 : )
4593 2 : .await?;
4594 2 : writer.finish_write(Lsn(0x30));
4595 2 : drop(writer);
4596 2 :
4597 2 : tline.freeze_and_flush().await?;
4598 2 : tline
4599 2 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
4600 2 : .await?;
4601 2 :
4602 2 : let mut writer = tline.writer().await;
4603 2 : writer
4604 2 : .put(
4605 2 : *TEST_KEY,
4606 2 : Lsn(0x40),
4607 2 : &Value::Image(test_img("foo at 0x40")),
4608 2 : &ctx,
4609 2 : )
4610 2 : .await?;
4611 2 : writer.finish_write(Lsn(0x40));
4612 2 : drop(writer);
4613 2 :
4614 2 : tline.freeze_and_flush().await?;
4615 2 : tline
4616 2 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
4617 2 : .await?;
4618 2 :
4619 2 : assert_eq!(
4620 4 : tline.get(*TEST_KEY, Lsn(0x10), &ctx).await?,
4621 2 : test_img("foo at 0x10")
4622 2 : );
4623 2 : assert_eq!(
4624 3 : tline.get(*TEST_KEY, Lsn(0x1f), &ctx).await?,
4625 2 : test_img("foo at 0x10")
4626 2 : );
4627 2 : assert_eq!(
4628 2 : tline.get(*TEST_KEY, Lsn(0x20), &ctx).await?,
4629 2 : test_img("foo at 0x20")
4630 2 : );
4631 2 : assert_eq!(
4632 4 : tline.get(*TEST_KEY, Lsn(0x30), &ctx).await?,
4633 2 : test_img("foo at 0x30")
4634 2 : );
4635 2 : assert_eq!(
4636 4 : tline.get(*TEST_KEY, Lsn(0x40), &ctx).await?,
4637 2 : test_img("foo at 0x40")
4638 2 : );
4639 2 :
4640 2 : Ok(())
4641 2 : }
4642 :
4643 4 : async fn bulk_insert_compact_gc(
4644 4 : tenant: &Tenant,
4645 4 : timeline: &Arc<Timeline>,
4646 4 : ctx: &RequestContext,
4647 4 : lsn: Lsn,
4648 4 : repeat: usize,
4649 4 : key_count: usize,
4650 4 : ) -> anyhow::Result<()> {
4651 4 : let compact = true;
4652 72774 : bulk_insert_maybe_compact_gc(tenant, timeline, ctx, lsn, repeat, key_count, compact).await
4653 4 : }
4654 :
4655 8 : async fn bulk_insert_maybe_compact_gc(
4656 8 : tenant: &Tenant,
4657 8 : timeline: &Arc<Timeline>,
4658 8 : ctx: &RequestContext,
4659 8 : mut lsn: Lsn,
4660 8 : repeat: usize,
4661 8 : key_count: usize,
4662 8 : compact: bool,
4663 8 : ) -> anyhow::Result<()> {
4664 8 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
4665 8 : let mut blknum = 0;
4666 8 :
4667 8 : // Enforce that key range is monotonously increasing
4668 8 : let mut keyspace = KeySpaceAccum::new();
4669 8 :
4670 8 : let cancel = CancellationToken::new();
4671 8 :
4672 8 : for _ in 0..repeat {
4673 400 : for _ in 0..key_count {
4674 4000000 : test_key.field6 = blknum;
4675 4000000 : let mut writer = timeline.writer().await;
4676 4000000 : writer
4677 4000000 : .put(
4678 4000000 : test_key,
4679 4000000 : lsn,
4680 4000000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
4681 4000000 : ctx,
4682 4000000 : )
4683 2600 : .await?;
4684 4000000 : writer.finish_write(lsn);
4685 4000000 : drop(writer);
4686 4000000 :
4687 4000000 : keyspace.add_key(test_key);
4688 4000000 :
4689 4000000 : lsn = Lsn(lsn.0 + 0x10);
4690 4000000 : blknum += 1;
4691 : }
4692 :
4693 400 : timeline.freeze_and_flush().await?;
4694 400 : if compact {
4695 : // this requires timeline to be &Arc<Timeline>
4696 40174 : timeline.compact(&cancel, EnumSet::empty(), ctx).await?;
4697 200 : }
4698 :
4699 : // this doesn't really need to use the timeline_id target, but it is closer to what it
4700 : // originally was.
4701 400 : let res = tenant
4702 400 : .gc_iteration(Some(timeline.timeline_id), 0, Duration::ZERO, &cancel, ctx)
4703 400 : .await?;
4704 :
4705 400 : assert_eq!(res.layers_removed, 0, "this never removes anything");
4706 : }
4707 :
4708 8 : Ok(())
4709 8 : }
4710 :
4711 : //
4712 : // Insert 1000 key-value pairs with increasing keys, flush, compact, GC.
4713 : // Repeat 50 times.
4714 : //
4715 : #[tokio::test]
4716 2 : async fn test_bulk_insert() -> anyhow::Result<()> {
4717 2 : let harness = TenantHarness::create("test_bulk_insert").await?;
4718 8 : let (tenant, ctx) = harness.load().await;
4719 2 : let tline = tenant
4720 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
4721 6 : .await?;
4722 2 :
4723 2 : let lsn = Lsn(0x10);
4724 36387 : bulk_insert_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000).await?;
4725 2 :
4726 2 : Ok(())
4727 2 : }
4728 :
4729 : // Test the vectored get real implementation against a simple sequential implementation.
4730 : //
4731 : // The test generates a keyspace by repeatedly flushing the in-memory layer and compacting.
4732 : // Projected to 2D the key space looks like below. Lsn grows upwards on the Y axis and keys
4733 : // grow to the right on the X axis.
4734 : // [Delta]
4735 : // [Delta]
4736 : // [Delta]
4737 : // [Delta]
4738 : // ------------ Image ---------------
4739 : //
4740 : // After layer generation we pick the ranges to query as follows:
4741 : // 1. The beginning of each delta layer
4742 : // 2. At the seam between two adjacent delta layers
4743 : //
4744 : // There's one major downside to this test: delta layers only contains images,
4745 : // so the search can stop at the first delta layer and doesn't traverse any deeper.
4746 : #[tokio::test]
4747 2 : async fn test_get_vectored() -> anyhow::Result<()> {
4748 2 : let harness = TenantHarness::create("test_get_vectored").await?;
4749 8 : let (tenant, ctx) = harness.load().await;
4750 2 : let tline = tenant
4751 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
4752 6 : .await?;
4753 2 :
4754 2 : let lsn = Lsn(0x10);
4755 36387 : bulk_insert_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000).await?;
4756 2 :
4757 2 : let guard = tline.layers.read().await;
4758 2 : guard.layer_map().dump(true, &ctx).await?;
4759 2 :
4760 2 : let mut reads = Vec::new();
4761 2 : let mut prev = None;
4762 12 : guard.layer_map().iter_historic_layers().for_each(|desc| {
4763 12 : if !desc.is_delta() {
4764 2 : prev = Some(desc.clone());
4765 2 : return;
4766 10 : }
4767 10 :
4768 10 : let start = desc.key_range.start;
4769 10 : let end = desc
4770 10 : .key_range
4771 10 : .start
4772 10 : .add(Timeline::MAX_GET_VECTORED_KEYS.try_into().unwrap());
4773 10 : reads.push(KeySpace {
4774 10 : ranges: vec![start..end],
4775 10 : });
4776 2 :
4777 10 : if let Some(prev) = &prev {
4778 10 : if !prev.is_delta() {
4779 10 : return;
4780 2 : }
4781 0 :
4782 0 : let first_range = Key {
4783 0 : field6: prev.key_range.end.field6 - 4,
4784 0 : ..prev.key_range.end
4785 0 : }..prev.key_range.end;
4786 0 :
4787 0 : let second_range = desc.key_range.start..Key {
4788 0 : field6: desc.key_range.start.field6 + 4,
4789 0 : ..desc.key_range.start
4790 0 : };
4791 0 :
4792 0 : reads.push(KeySpace {
4793 0 : ranges: vec![first_range, second_range],
4794 0 : });
4795 2 : };
4796 2 :
4797 2 : prev = Some(desc.clone());
4798 12 : });
4799 2 :
4800 2 : drop(guard);
4801 2 :
4802 2 : // Pick a big LSN such that we query over all the changes.
4803 2 : let reads_lsn = Lsn(u64::MAX - 1);
4804 2 :
4805 12 : for read in reads {
4806 10 : info!("Doing vectored read on {:?}", read);
4807 2 :
4808 10 : let vectored_res = tline
4809 10 : .get_vectored_impl(
4810 10 : read.clone(),
4811 10 : reads_lsn,
4812 10 : &mut ValuesReconstructState::new(),
4813 10 : &ctx,
4814 10 : )
4815 25 : .await;
4816 10 : tline
4817 10 : .validate_get_vectored_impl(&vectored_res, read, reads_lsn, &ctx)
4818 19 : .await;
4819 2 : }
4820 2 :
4821 2 : Ok(())
4822 2 : }
4823 :
4824 : #[tokio::test]
4825 2 : async fn test_get_vectored_aux_files() -> anyhow::Result<()> {
4826 2 : let harness = TenantHarness::create("test_get_vectored_aux_files").await?;
4827 2 :
4828 8 : let (tenant, ctx) = harness.load().await;
4829 2 : let tline = tenant
4830 2 : .create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION, &ctx)
4831 2 : .await?;
4832 2 : let tline = tline.raw_timeline().unwrap();
4833 2 :
4834 2 : let mut modification = tline.begin_modification(Lsn(0x1000));
4835 2 : modification.put_file("foo/bar1", b"content1", &ctx).await?;
4836 2 : modification.set_lsn(Lsn(0x1008))?;
4837 2 : modification.put_file("foo/bar2", b"content2", &ctx).await?;
4838 2 : modification.commit(&ctx).await?;
4839 2 :
4840 2 : let child_timeline_id = TimelineId::generate();
4841 2 : tenant
4842 2 : .branch_timeline_test(
4843 2 : tline,
4844 2 : child_timeline_id,
4845 2 : Some(tline.get_last_record_lsn()),
4846 2 : &ctx,
4847 2 : )
4848 2 : .await?;
4849 2 :
4850 2 : let child_timeline = tenant
4851 2 : .get_timeline(child_timeline_id, true)
4852 2 : .expect("Should have the branched timeline");
4853 2 :
4854 2 : let aux_keyspace = KeySpace {
4855 2 : ranges: vec![NON_INHERITED_RANGE],
4856 2 : };
4857 2 : let read_lsn = child_timeline.get_last_record_lsn();
4858 2 :
4859 2 : let vectored_res = child_timeline
4860 2 : .get_vectored_impl(
4861 2 : aux_keyspace.clone(),
4862 2 : read_lsn,
4863 2 : &mut ValuesReconstructState::new(),
4864 2 : &ctx,
4865 2 : )
4866 2 : .await;
4867 2 :
4868 2 : child_timeline
4869 2 : .validate_get_vectored_impl(&vectored_res, aux_keyspace, read_lsn, &ctx)
4870 2 : .await;
4871 2 :
4872 2 : let images = vectored_res?;
4873 2 : assert!(images.is_empty());
4874 2 : Ok(())
4875 2 : }
4876 :
4877 : // Test that vectored get handles layer gaps correctly
4878 : // by advancing into the next ancestor timeline if required.
4879 : //
4880 : // The test generates timelines that look like the diagram below.
4881 : // We leave a gap in one of the L1 layers at `gap_at_key` (`/` in the diagram).
4882 : // The reconstruct data for that key lies in the ancestor timeline (`X` in the diagram).
4883 : //
4884 : // ```
4885 : //-------------------------------+
4886 : // ... |
4887 : // [ L1 ] |
4888 : // [ / L1 ] | Child Timeline
4889 : // ... |
4890 : // ------------------------------+
4891 : // [ X L1 ] | Parent Timeline
4892 : // ------------------------------+
4893 : // ```
4894 : #[tokio::test]
4895 2 : async fn test_get_vectored_key_gap() -> anyhow::Result<()> {
4896 2 : let tenant_conf = TenantConf {
4897 2 : // Make compaction deterministic
4898 2 : gc_period: Duration::ZERO,
4899 2 : compaction_period: Duration::ZERO,
4900 2 : // Encourage creation of L1 layers
4901 2 : checkpoint_distance: 16 * 1024,
4902 2 : compaction_target_size: 8 * 1024,
4903 2 : ..TenantConf::default()
4904 2 : };
4905 2 :
4906 2 : let harness = TenantHarness::create_custom(
4907 2 : "test_get_vectored_key_gap",
4908 2 : tenant_conf,
4909 2 : TenantId::generate(),
4910 2 : ShardIdentity::unsharded(),
4911 2 : Generation::new(0xdeadbeef),
4912 2 : )
4913 2 : .await?;
4914 8 : let (tenant, ctx) = harness.load().await;
4915 2 :
4916 2 : let mut current_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
4917 2 : let gap_at_key = current_key.add(100);
4918 2 : let mut current_lsn = Lsn(0x10);
4919 2 :
4920 2 : const KEY_COUNT: usize = 10_000;
4921 2 :
4922 2 : let timeline_id = TimelineId::generate();
4923 2 : let current_timeline = tenant
4924 2 : .create_test_timeline(timeline_id, current_lsn, DEFAULT_PG_VERSION, &ctx)
4925 5 : .await?;
4926 2 :
4927 2 : current_lsn += 0x100;
4928 2 :
4929 2 : let mut writer = current_timeline.writer().await;
4930 2 : writer
4931 2 : .put(
4932 2 : gap_at_key,
4933 2 : current_lsn,
4934 2 : &Value::Image(test_img(&format!("{} at {}", gap_at_key, current_lsn))),
4935 2 : &ctx,
4936 2 : )
4937 2 : .await?;
4938 2 : writer.finish_write(current_lsn);
4939 2 : drop(writer);
4940 2 :
4941 2 : let mut latest_lsns = HashMap::new();
4942 2 : latest_lsns.insert(gap_at_key, current_lsn);
4943 2 :
4944 2 : current_timeline.freeze_and_flush().await?;
4945 2 :
4946 2 : let child_timeline_id = TimelineId::generate();
4947 2 :
4948 2 : tenant
4949 2 : .branch_timeline_test(
4950 2 : ¤t_timeline,
4951 2 : child_timeline_id,
4952 2 : Some(current_lsn),
4953 2 : &ctx,
4954 2 : )
4955 2 : .await?;
4956 2 : let child_timeline = tenant
4957 2 : .get_timeline(child_timeline_id, true)
4958 2 : .expect("Should have the branched timeline");
4959 2 :
4960 20002 : for i in 0..KEY_COUNT {
4961 20000 : if current_key == gap_at_key {
4962 2 : current_key = current_key.next();
4963 2 : continue;
4964 19998 : }
4965 19998 :
4966 19998 : current_lsn += 0x10;
4967 2 :
4968 19998 : let mut writer = child_timeline.writer().await;
4969 19998 : writer
4970 19998 : .put(
4971 19998 : current_key,
4972 19998 : current_lsn,
4973 19998 : &Value::Image(test_img(&format!("{} at {}", current_key, current_lsn))),
4974 19998 : &ctx,
4975 19998 : )
4976 72 : .await?;
4977 19998 : writer.finish_write(current_lsn);
4978 19998 : drop(writer);
4979 19998 :
4980 19998 : latest_lsns.insert(current_key, current_lsn);
4981 19998 : current_key = current_key.next();
4982 19998 :
4983 19998 : // Flush every now and then to encourage layer file creation.
4984 19998 : if i % 500 == 0 {
4985 42 : child_timeline.freeze_and_flush().await?;
4986 19958 : }
4987 2 : }
4988 2 :
4989 2 : child_timeline.freeze_and_flush().await?;
4990 2 : let mut flags = EnumSet::new();
4991 2 : flags.insert(CompactFlags::ForceRepartition);
4992 2 : child_timeline
4993 2 : .compact(&CancellationToken::new(), flags, &ctx)
4994 2184 : .await?;
4995 2 :
4996 2 : let key_near_end = {
4997 2 : let mut tmp = current_key;
4998 2 : tmp.field6 -= 10;
4999 2 : tmp
5000 2 : };
5001 2 :
5002 2 : let key_near_gap = {
5003 2 : let mut tmp = gap_at_key;
5004 2 : tmp.field6 -= 10;
5005 2 : tmp
5006 2 : };
5007 2 :
5008 2 : let read = KeySpace {
5009 2 : ranges: vec![key_near_gap..gap_at_key.next(), key_near_end..current_key],
5010 2 : };
5011 2 : let results = child_timeline
5012 2 : .get_vectored_impl(
5013 2 : read.clone(),
5014 2 : current_lsn,
5015 2 : &mut ValuesReconstructState::new(),
5016 2 : &ctx,
5017 2 : )
5018 15 : .await?;
5019 2 :
5020 44 : for (key, img_res) in results {
5021 42 : let expected = test_img(&format!("{} at {}", key, latest_lsns[&key]));
5022 42 : assert_eq!(img_res?, expected);
5023 2 : }
5024 2 :
5025 2 : Ok(())
5026 2 : }
5027 :
5028 : // Test that vectored get descends into ancestor timelines correctly and
5029 : // does not return an image that's newer than requested.
5030 : //
5031 : // The diagram below ilustrates an interesting case. We have a parent timeline
5032 : // (top of the Lsn range) and a child timeline. The request key cannot be reconstructed
5033 : // from the child timeline, so the parent timeline must be visited. When advacing into
5034 : // the child timeline, the read path needs to remember what the requested Lsn was in
5035 : // order to avoid returning an image that's too new. The test below constructs such
5036 : // a timeline setup and does a few queries around the Lsn of each page image.
5037 : // ```
5038 : // LSN
5039 : // ^
5040 : // |
5041 : // |
5042 : // 500 | --------------------------------------> branch point
5043 : // 400 | X
5044 : // 300 | X
5045 : // 200 | --------------------------------------> requested lsn
5046 : // 100 | X
5047 : // |---------------------------------------> Key
5048 : // |
5049 : // ------> requested key
5050 : //
5051 : // Legend:
5052 : // * X - page images
5053 : // ```
5054 : #[tokio::test]
5055 2 : async fn test_get_vectored_ancestor_descent() -> anyhow::Result<()> {
5056 2 : let harness = TenantHarness::create("test_get_vectored_on_lsn_axis").await?;
5057 8 : let (tenant, ctx) = harness.load().await;
5058 2 :
5059 2 : let start_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
5060 2 : let end_key = start_key.add(1000);
5061 2 : let child_gap_at_key = start_key.add(500);
5062 2 : let mut parent_gap_lsns: BTreeMap<Lsn, String> = BTreeMap::new();
5063 2 :
5064 2 : let mut current_lsn = Lsn(0x10);
5065 2 :
5066 2 : let timeline_id = TimelineId::generate();
5067 2 : let parent_timeline = tenant
5068 2 : .create_test_timeline(timeline_id, current_lsn, DEFAULT_PG_VERSION, &ctx)
5069 6 : .await?;
5070 2 :
5071 2 : current_lsn += 0x100;
5072 2 :
5073 8 : for _ in 0..3 {
5074 6 : let mut key = start_key;
5075 6006 : while key < end_key {
5076 6000 : current_lsn += 0x10;
5077 6000 :
5078 6000 : let image_value = format!("{} at {}", child_gap_at_key, current_lsn);
5079 2 :
5080 6000 : let mut writer = parent_timeline.writer().await;
5081 6000 : writer
5082 6000 : .put(
5083 6000 : key,
5084 6000 : current_lsn,
5085 6000 : &Value::Image(test_img(&image_value)),
5086 6000 : &ctx,
5087 6000 : )
5088 6 : .await?;
5089 6000 : writer.finish_write(current_lsn);
5090 6000 :
5091 6000 : if key == child_gap_at_key {
5092 6 : parent_gap_lsns.insert(current_lsn, image_value);
5093 5994 : }
5094 2 :
5095 6000 : key = key.next();
5096 2 : }
5097 2 :
5098 6 : parent_timeline.freeze_and_flush().await?;
5099 2 : }
5100 2 :
5101 2 : let child_timeline_id = TimelineId::generate();
5102 2 :
5103 2 : let child_timeline = tenant
5104 2 : .branch_timeline_test(&parent_timeline, child_timeline_id, Some(current_lsn), &ctx)
5105 2 : .await?;
5106 2 :
5107 2 : let mut key = start_key;
5108 2002 : while key < end_key {
5109 2000 : if key == child_gap_at_key {
5110 2 : key = key.next();
5111 2 : continue;
5112 1998 : }
5113 1998 :
5114 1998 : current_lsn += 0x10;
5115 2 :
5116 1998 : let mut writer = child_timeline.writer().await;
5117 1998 : writer
5118 1998 : .put(
5119 1998 : key,
5120 1998 : current_lsn,
5121 1998 : &Value::Image(test_img(&format!("{} at {}", key, current_lsn))),
5122 1998 : &ctx,
5123 1998 : )
5124 17 : .await?;
5125 1998 : writer.finish_write(current_lsn);
5126 1998 :
5127 1998 : key = key.next();
5128 2 : }
5129 2 :
5130 2 : child_timeline.freeze_and_flush().await?;
5131 2 :
5132 2 : let lsn_offsets: [i64; 5] = [-10, -1, 0, 1, 10];
5133 2 : let mut query_lsns = Vec::new();
5134 6 : for image_lsn in parent_gap_lsns.keys().rev() {
5135 36 : for offset in lsn_offsets {
5136 30 : query_lsns.push(Lsn(image_lsn
5137 30 : .0
5138 30 : .checked_add_signed(offset)
5139 30 : .expect("Shouldn't overflow")));
5140 30 : }
5141 2 : }
5142 2 :
5143 32 : for query_lsn in query_lsns {
5144 30 : let results = child_timeline
5145 30 : .get_vectored_impl(
5146 30 : KeySpace {
5147 30 : ranges: vec![child_gap_at_key..child_gap_at_key.next()],
5148 30 : },
5149 30 : query_lsn,
5150 30 : &mut ValuesReconstructState::new(),
5151 30 : &ctx,
5152 30 : )
5153 29 : .await;
5154 2 :
5155 30 : let expected_item = parent_gap_lsns
5156 30 : .iter()
5157 30 : .rev()
5158 68 : .find(|(lsn, _)| **lsn <= query_lsn);
5159 30 :
5160 30 : info!(
5161 2 : "Doing vectored read at LSN {}. Expecting image to be: {:?}",
5162 2 : query_lsn, expected_item
5163 2 : );
5164 2 :
5165 30 : match expected_item {
5166 26 : Some((_, img_value)) => {
5167 26 : let key_results = results.expect("No vectored get error expected");
5168 26 : let key_result = &key_results[&child_gap_at_key];
5169 26 : let returned_img = key_result
5170 26 : .as_ref()
5171 26 : .expect("No page reconstruct error expected");
5172 26 :
5173 26 : info!(
5174 2 : "Vectored read at LSN {} returned image {}",
5175 0 : query_lsn,
5176 0 : std::str::from_utf8(returned_img)?
5177 2 : );
5178 26 : assert_eq!(*returned_img, test_img(img_value));
5179 2 : }
5180 2 : None => {
5181 4 : assert!(matches!(results, Err(GetVectoredError::MissingKey(_))));
5182 2 : }
5183 2 : }
5184 2 : }
5185 2 :
5186 2 : Ok(())
5187 2 : }
5188 :
5189 : #[tokio::test]
5190 2 : async fn test_random_updates() -> anyhow::Result<()> {
5191 2 : let names_algorithms = [
5192 2 : ("test_random_updates_legacy", CompactionAlgorithm::Legacy),
5193 2 : ("test_random_updates_tiered", CompactionAlgorithm::Tiered),
5194 2 : ];
5195 6 : for (name, algorithm) in names_algorithms {
5196 17875 : test_random_updates_algorithm(name, algorithm).await?;
5197 2 : }
5198 2 : Ok(())
5199 2 : }
5200 :
5201 4 : async fn test_random_updates_algorithm(
5202 4 : name: &'static str,
5203 4 : compaction_algorithm: CompactionAlgorithm,
5204 4 : ) -> anyhow::Result<()> {
5205 4 : let mut harness = TenantHarness::create(name).await?;
5206 4 : harness.tenant_conf.compaction_algorithm = CompactionAlgorithmSettings {
5207 4 : kind: compaction_algorithm,
5208 4 : };
5209 16 : let (tenant, ctx) = harness.load().await;
5210 4 : let tline = tenant
5211 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5212 11 : .await?;
5213 :
5214 : const NUM_KEYS: usize = 1000;
5215 4 : let cancel = CancellationToken::new();
5216 4 :
5217 4 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
5218 4 : let mut test_key_end = test_key;
5219 4 : test_key_end.field6 = NUM_KEYS as u32;
5220 4 : tline.add_extra_test_dense_keyspace(KeySpace::single(test_key..test_key_end));
5221 4 :
5222 4 : let mut keyspace = KeySpaceAccum::new();
5223 4 :
5224 4 : // Track when each page was last modified. Used to assert that
5225 4 : // a read sees the latest page version.
5226 4 : let mut updated = [Lsn(0); NUM_KEYS];
5227 4 :
5228 4 : let mut lsn = Lsn(0x10);
5229 : #[allow(clippy::needless_range_loop)]
5230 4004 : for blknum in 0..NUM_KEYS {
5231 4000 : lsn = Lsn(lsn.0 + 0x10);
5232 4000 : test_key.field6 = blknum as u32;
5233 4000 : let mut writer = tline.writer().await;
5234 4000 : writer
5235 4000 : .put(
5236 4000 : test_key,
5237 4000 : lsn,
5238 4000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
5239 4000 : &ctx,
5240 4000 : )
5241 4 : .await?;
5242 4000 : writer.finish_write(lsn);
5243 4000 : updated[blknum] = lsn;
5244 4000 : drop(writer);
5245 4000 :
5246 4000 : keyspace.add_key(test_key);
5247 : }
5248 :
5249 204 : for _ in 0..50 {
5250 200200 : for _ in 0..NUM_KEYS {
5251 200000 : lsn = Lsn(lsn.0 + 0x10);
5252 200000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
5253 200000 : test_key.field6 = blknum as u32;
5254 200000 : let mut writer = tline.writer().await;
5255 200000 : writer
5256 200000 : .put(
5257 200000 : test_key,
5258 200000 : lsn,
5259 200000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
5260 200000 : &ctx,
5261 200000 : )
5262 198 : .await?;
5263 200000 : writer.finish_write(lsn);
5264 200000 : drop(writer);
5265 200000 : updated[blknum] = lsn;
5266 : }
5267 :
5268 : // Read all the blocks
5269 200000 : for (blknum, last_lsn) in updated.iter().enumerate() {
5270 200000 : test_key.field6 = blknum as u32;
5271 200000 : assert_eq!(
5272 200000 : tline.get(test_key, lsn, &ctx).await?,
5273 200000 : test_img(&format!("{} at {}", blknum, last_lsn))
5274 : );
5275 : }
5276 :
5277 : // Perform a cycle of flush, and GC
5278 201 : tline.freeze_and_flush().await?;
5279 200 : tenant
5280 200 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
5281 196 : .await?;
5282 : }
5283 :
5284 4 : Ok(())
5285 4 : }
5286 :
5287 : #[tokio::test]
5288 2 : async fn test_traverse_branches() -> anyhow::Result<()> {
5289 2 : let (tenant, ctx) = TenantHarness::create("test_traverse_branches")
5290 2 : .await?
5291 2 : .load()
5292 8 : .await;
5293 2 : let mut tline = tenant
5294 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5295 6 : .await?;
5296 2 :
5297 2 : const NUM_KEYS: usize = 1000;
5298 2 :
5299 2 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
5300 2 :
5301 2 : let mut keyspace = KeySpaceAccum::new();
5302 2 :
5303 2 : let cancel = CancellationToken::new();
5304 2 :
5305 2 : // Track when each page was last modified. Used to assert that
5306 2 : // a read sees the latest page version.
5307 2 : let mut updated = [Lsn(0); NUM_KEYS];
5308 2 :
5309 2 : let mut lsn = Lsn(0x10);
5310 2 : #[allow(clippy::needless_range_loop)]
5311 2002 : for blknum in 0..NUM_KEYS {
5312 2000 : lsn = Lsn(lsn.0 + 0x10);
5313 2000 : test_key.field6 = blknum as u32;
5314 2000 : let mut writer = tline.writer().await;
5315 2000 : writer
5316 2000 : .put(
5317 2000 : test_key,
5318 2000 : lsn,
5319 2000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
5320 2000 : &ctx,
5321 2000 : )
5322 2 : .await?;
5323 2000 : writer.finish_write(lsn);
5324 2000 : updated[blknum] = lsn;
5325 2000 : drop(writer);
5326 2000 :
5327 2000 : keyspace.add_key(test_key);
5328 2 : }
5329 2 :
5330 102 : for _ in 0..50 {
5331 100 : let new_tline_id = TimelineId::generate();
5332 100 : tenant
5333 100 : .branch_timeline_test(&tline, new_tline_id, Some(lsn), &ctx)
5334 2 : .await?;
5335 100 : tline = tenant
5336 100 : .get_timeline(new_tline_id, true)
5337 100 : .expect("Should have the branched timeline");
5338 2 :
5339 100100 : for _ in 0..NUM_KEYS {
5340 100000 : lsn = Lsn(lsn.0 + 0x10);
5341 100000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
5342 100000 : test_key.field6 = blknum as u32;
5343 100000 : let mut writer = tline.writer().await;
5344 100000 : writer
5345 100000 : .put(
5346 100000 : test_key,
5347 100000 : lsn,
5348 100000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
5349 100000 : &ctx,
5350 100000 : )
5351 851 : .await?;
5352 100000 : println!("updating {} at {}", blknum, lsn);
5353 100000 : writer.finish_write(lsn);
5354 100000 : drop(writer);
5355 100000 : updated[blknum] = lsn;
5356 2 : }
5357 2 :
5358 2 : // Read all the blocks
5359 100000 : for (blknum, last_lsn) in updated.iter().enumerate() {
5360 100000 : test_key.field6 = blknum as u32;
5361 100000 : assert_eq!(
5362 100000 : tline.get(test_key, lsn, &ctx).await?,
5363 100000 : test_img(&format!("{} at {}", blknum, last_lsn))
5364 2 : );
5365 2 : }
5366 2 :
5367 2 : // Perform a cycle of flush, compact, and GC
5368 104 : tline.freeze_and_flush().await?;
5369 13257 : tline.compact(&cancel, EnumSet::empty(), &ctx).await?;
5370 100 : tenant
5371 100 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
5372 96 : .await?;
5373 2 : }
5374 2 :
5375 2 : Ok(())
5376 2 : }
5377 :
5378 : #[tokio::test]
5379 2 : async fn test_traverse_ancestors() -> anyhow::Result<()> {
5380 2 : let (tenant, ctx) = TenantHarness::create("test_traverse_ancestors")
5381 2 : .await?
5382 2 : .load()
5383 8 : .await;
5384 2 : let mut tline = tenant
5385 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5386 6 : .await?;
5387 2 :
5388 2 : const NUM_KEYS: usize = 100;
5389 2 : const NUM_TLINES: usize = 50;
5390 2 :
5391 2 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
5392 2 : // Track page mutation lsns across different timelines.
5393 2 : let mut updated = [[Lsn(0); NUM_KEYS]; NUM_TLINES];
5394 2 :
5395 2 : let mut lsn = Lsn(0x10);
5396 2 :
5397 2 : #[allow(clippy::needless_range_loop)]
5398 102 : for idx in 0..NUM_TLINES {
5399 100 : let new_tline_id = TimelineId::generate();
5400 100 : tenant
5401 100 : .branch_timeline_test(&tline, new_tline_id, Some(lsn), &ctx)
5402 2 : .await?;
5403 100 : tline = tenant
5404 100 : .get_timeline(new_tline_id, true)
5405 100 : .expect("Should have the branched timeline");
5406 2 :
5407 10100 : for _ in 0..NUM_KEYS {
5408 10000 : lsn = Lsn(lsn.0 + 0x10);
5409 10000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
5410 10000 : test_key.field6 = blknum as u32;
5411 10000 : let mut writer = tline.writer().await;
5412 10000 : writer
5413 10000 : .put(
5414 10000 : test_key,
5415 10000 : lsn,
5416 10000 : &Value::Image(test_img(&format!("{} {} at {}", idx, blknum, lsn))),
5417 10000 : &ctx,
5418 10000 : )
5419 88 : .await?;
5420 10000 : println!("updating [{}][{}] at {}", idx, blknum, lsn);
5421 10000 : writer.finish_write(lsn);
5422 10000 : drop(writer);
5423 10000 : updated[idx][blknum] = lsn;
5424 2 : }
5425 2 : }
5426 2 :
5427 2 : // Read pages from leaf timeline across all ancestors.
5428 100 : for (idx, lsns) in updated.iter().enumerate() {
5429 10000 : for (blknum, lsn) in lsns.iter().enumerate() {
5430 2 : // Skip empty mutations.
5431 10000 : if lsn.0 == 0 {
5432 3644 : continue;
5433 6356 : }
5434 6356 : println!("checking [{idx}][{blknum}] at {lsn}");
5435 6356 : test_key.field6 = blknum as u32;
5436 6356 : assert_eq!(
5437 6356 : tline.get(test_key, *lsn, &ctx).await?,
5438 6356 : test_img(&format!("{idx} {blknum} at {lsn}"))
5439 2 : );
5440 2 : }
5441 2 : }
5442 2 : Ok(())
5443 2 : }
5444 :
5445 : #[tokio::test]
5446 2 : async fn test_write_at_initdb_lsn_takes_optimization_code_path() -> anyhow::Result<()> {
5447 2 : let (tenant, ctx) = TenantHarness::create("test_empty_test_timeline_is_usable")
5448 2 : .await?
5449 2 : .load()
5450 8 : .await;
5451 2 :
5452 2 : let initdb_lsn = Lsn(0x20);
5453 2 : let utline = tenant
5454 2 : .create_empty_timeline(TIMELINE_ID, initdb_lsn, DEFAULT_PG_VERSION, &ctx)
5455 2 : .await?;
5456 2 : let tline = utline.raw_timeline().unwrap();
5457 2 :
5458 2 : // Spawn flush loop now so that we can set the `expect_initdb_optimization`
5459 2 : tline.maybe_spawn_flush_loop();
5460 2 :
5461 2 : // Make sure the timeline has the minimum set of required keys for operation.
5462 2 : // The only operation you can always do on an empty timeline is to `put` new data.
5463 2 : // Except if you `put` at `initdb_lsn`.
5464 2 : // In that case, there's an optimization to directly create image layers instead of delta layers.
5465 2 : // It uses `repartition()`, which assumes some keys to be present.
5466 2 : // Let's make sure the test timeline can handle that case.
5467 2 : {
5468 2 : let mut state = tline.flush_loop_state.lock().unwrap();
5469 2 : assert_eq!(
5470 2 : timeline::FlushLoopState::Running {
5471 2 : expect_initdb_optimization: false,
5472 2 : initdb_optimization_count: 0,
5473 2 : },
5474 2 : *state
5475 2 : );
5476 2 : *state = timeline::FlushLoopState::Running {
5477 2 : expect_initdb_optimization: true,
5478 2 : initdb_optimization_count: 0,
5479 2 : };
5480 2 : }
5481 2 :
5482 2 : // Make writes at the initdb_lsn. When we flush it below, it should be handled by the optimization.
5483 2 : // As explained above, the optimization requires some keys to be present.
5484 2 : // As per `create_empty_timeline` documentation, use init_empty to set them.
5485 2 : // This is what `create_test_timeline` does, by the way.
5486 2 : let mut modification = tline.begin_modification(initdb_lsn);
5487 2 : modification
5488 2 : .init_empty_test_timeline()
5489 2 : .context("init_empty_test_timeline")?;
5490 2 : modification
5491 2 : .commit(&ctx)
5492 2 : .await
5493 2 : .context("commit init_empty_test_timeline modification")?;
5494 2 :
5495 2 : // Do the flush. The flush code will check the expectations that we set above.
5496 2 : tline.freeze_and_flush().await?;
5497 2 :
5498 2 : // assert freeze_and_flush exercised the initdb optimization
5499 2 : {
5500 2 : let state = tline.flush_loop_state.lock().unwrap();
5501 2 : let timeline::FlushLoopState::Running {
5502 2 : expect_initdb_optimization,
5503 2 : initdb_optimization_count,
5504 2 : } = *state
5505 2 : else {
5506 2 : panic!("unexpected state: {:?}", *state);
5507 2 : };
5508 2 : assert!(expect_initdb_optimization);
5509 2 : assert!(initdb_optimization_count > 0);
5510 2 : }
5511 2 : Ok(())
5512 2 : }
5513 :
5514 : #[tokio::test]
5515 2 : async fn test_create_guard_crash() -> anyhow::Result<()> {
5516 2 : let name = "test_create_guard_crash";
5517 2 : let harness = TenantHarness::create(name).await?;
5518 2 : {
5519 8 : let (tenant, ctx) = harness.load().await;
5520 2 : let tline = tenant
5521 2 : .create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION, &ctx)
5522 2 : .await?;
5523 2 : // Leave the timeline ID in [`Tenant::timelines_creating`] to exclude attempting to create it again
5524 2 : let raw_tline = tline.raw_timeline().unwrap();
5525 2 : raw_tline
5526 2 : .shutdown(super::timeline::ShutdownMode::Hard)
5527 2 : .instrument(info_span!("test_shutdown", tenant_id=%raw_tline.tenant_shard_id, shard_id=%raw_tline.tenant_shard_id.shard_slug(), timeline_id=%TIMELINE_ID))
5528 2 : .await;
5529 2 : std::mem::forget(tline);
5530 2 : }
5531 2 :
5532 8 : let (tenant, _) = harness.load().await;
5533 2 : match tenant.get_timeline(TIMELINE_ID, false) {
5534 2 : Ok(_) => panic!("timeline should've been removed during load"),
5535 2 : Err(e) => {
5536 2 : assert_eq!(
5537 2 : e,
5538 2 : GetTimelineError::NotFound {
5539 2 : tenant_id: tenant.tenant_shard_id,
5540 2 : timeline_id: TIMELINE_ID,
5541 2 : }
5542 2 : )
5543 2 : }
5544 2 : }
5545 2 :
5546 2 : assert!(!harness
5547 2 : .conf
5548 2 : .timeline_path(&tenant.tenant_shard_id, &TIMELINE_ID)
5549 2 : .exists());
5550 2 :
5551 2 : Ok(())
5552 2 : }
5553 :
5554 : #[tokio::test]
5555 2 : async fn test_read_at_max_lsn() -> anyhow::Result<()> {
5556 2 : let names_algorithms = [
5557 2 : ("test_read_at_max_lsn_legacy", CompactionAlgorithm::Legacy),
5558 2 : ("test_read_at_max_lsn_tiered", CompactionAlgorithm::Tiered),
5559 2 : ];
5560 6 : for (name, algorithm) in names_algorithms {
5561 32937 : test_read_at_max_lsn_algorithm(name, algorithm).await?;
5562 2 : }
5563 2 : Ok(())
5564 2 : }
5565 :
5566 4 : async fn test_read_at_max_lsn_algorithm(
5567 4 : name: &'static str,
5568 4 : compaction_algorithm: CompactionAlgorithm,
5569 4 : ) -> anyhow::Result<()> {
5570 4 : let mut harness = TenantHarness::create(name).await?;
5571 4 : harness.tenant_conf.compaction_algorithm = CompactionAlgorithmSettings {
5572 4 : kind: compaction_algorithm,
5573 4 : };
5574 16 : let (tenant, ctx) = harness.load().await;
5575 4 : let tline = tenant
5576 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
5577 11 : .await?;
5578 :
5579 4 : let lsn = Lsn(0x10);
5580 4 : let compact = false;
5581 32600 : bulk_insert_maybe_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000, compact).await?;
5582 :
5583 4 : let test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
5584 4 : let read_lsn = Lsn(u64::MAX - 1);
5585 :
5586 310 : let result = tline.get(test_key, read_lsn, &ctx).await;
5587 4 : assert!(result.is_ok(), "result is not Ok: {}", result.unwrap_err());
5588 :
5589 4 : Ok(())
5590 4 : }
5591 :
5592 : #[tokio::test]
5593 2 : async fn test_metadata_scan() -> anyhow::Result<()> {
5594 2 : let harness = TenantHarness::create("test_metadata_scan").await?;
5595 8 : let (tenant, ctx) = harness.load().await;
5596 2 : let tline = tenant
5597 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5598 6 : .await?;
5599 2 :
5600 2 : const NUM_KEYS: usize = 1000;
5601 2 : const STEP: usize = 10000; // random update + scan base_key + idx * STEP
5602 2 :
5603 2 : let cancel = CancellationToken::new();
5604 2 :
5605 2 : let mut base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
5606 2 : base_key.field1 = AUX_KEY_PREFIX;
5607 2 : let mut test_key = base_key;
5608 2 :
5609 2 : // Track when each page was last modified. Used to assert that
5610 2 : // a read sees the latest page version.
5611 2 : let mut updated = [Lsn(0); NUM_KEYS];
5612 2 :
5613 2 : let mut lsn = Lsn(0x10);
5614 2 : #[allow(clippy::needless_range_loop)]
5615 2002 : for blknum in 0..NUM_KEYS {
5616 2000 : lsn = Lsn(lsn.0 + 0x10);
5617 2000 : test_key.field6 = (blknum * STEP) as u32;
5618 2000 : let mut writer = tline.writer().await;
5619 2000 : writer
5620 2000 : .put(
5621 2000 : test_key,
5622 2000 : lsn,
5623 2000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
5624 2000 : &ctx,
5625 2000 : )
5626 2 : .await?;
5627 2000 : writer.finish_write(lsn);
5628 2000 : updated[blknum] = lsn;
5629 2000 : drop(writer);
5630 2 : }
5631 2 :
5632 2 : let keyspace = KeySpace::single(base_key..base_key.add((NUM_KEYS * STEP) as u32));
5633 2 :
5634 24 : for iter in 0..=10 {
5635 2 : // Read all the blocks
5636 22000 : for (blknum, last_lsn) in updated.iter().enumerate() {
5637 22000 : test_key.field6 = (blknum * STEP) as u32;
5638 22000 : assert_eq!(
5639 22000 : tline.get(test_key, lsn, &ctx).await?,
5640 22000 : test_img(&format!("{} at {}", blknum, last_lsn))
5641 2 : );
5642 2 : }
5643 2 :
5644 22 : let mut cnt = 0;
5645 22000 : for (key, value) in tline
5646 22 : .get_vectored_impl(
5647 22 : keyspace.clone(),
5648 22 : lsn,
5649 22 : &mut ValuesReconstructState::default(),
5650 22 : &ctx,
5651 22 : )
5652 5599 : .await?
5653 2 : {
5654 22000 : let blknum = key.field6 as usize;
5655 22000 : let value = value?;
5656 22000 : assert!(blknum % STEP == 0);
5657 22000 : let blknum = blknum / STEP;
5658 22000 : assert_eq!(
5659 22000 : value,
5660 22000 : test_img(&format!("{} at {}", blknum, updated[blknum]))
5661 22000 : );
5662 22000 : cnt += 1;
5663 2 : }
5664 2 :
5665 22 : assert_eq!(cnt, NUM_KEYS);
5666 2 :
5667 22022 : for _ in 0..NUM_KEYS {
5668 22000 : lsn = Lsn(lsn.0 + 0x10);
5669 22000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
5670 22000 : test_key.field6 = (blknum * STEP) as u32;
5671 22000 : let mut writer = tline.writer().await;
5672 22000 : writer
5673 22000 : .put(
5674 22000 : test_key,
5675 22000 : lsn,
5676 22000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
5677 22000 : &ctx,
5678 22000 : )
5679 129 : .await?;
5680 22000 : writer.finish_write(lsn);
5681 22000 : drop(writer);
5682 22000 : updated[blknum] = lsn;
5683 2 : }
5684 2 :
5685 2 : // Perform two cycles of flush, compact, and GC
5686 66 : for round in 0..2 {
5687 44 : tline.freeze_and_flush().await?;
5688 44 : tline
5689 44 : .compact(
5690 44 : &cancel,
5691 44 : if iter % 5 == 0 && round == 0 {
5692 6 : let mut flags = EnumSet::new();
5693 6 : flags.insert(CompactFlags::ForceImageLayerCreation);
5694 6 : flags.insert(CompactFlags::ForceRepartition);
5695 6 : flags
5696 2 : } else {
5697 38 : EnumSet::empty()
5698 2 : },
5699 44 : &ctx,
5700 2 : )
5701 8873 : .await?;
5702 44 : tenant
5703 44 : .gc_iteration(Some(tline.timeline_id), 0, Duration::ZERO, &cancel, &ctx)
5704 42 : .await?;
5705 2 : }
5706 2 : }
5707 2 :
5708 2 : Ok(())
5709 2 : }
5710 :
5711 : #[tokio::test]
5712 2 : async fn test_metadata_compaction_trigger() -> anyhow::Result<()> {
5713 2 : let harness = TenantHarness::create("test_metadata_compaction_trigger").await?;
5714 8 : let (tenant, ctx) = harness.load().await;
5715 2 : let tline = tenant
5716 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
5717 6 : .await?;
5718 2 :
5719 2 : let cancel = CancellationToken::new();
5720 2 :
5721 2 : let mut base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
5722 2 : base_key.field1 = AUX_KEY_PREFIX;
5723 2 : let test_key = base_key;
5724 2 : let mut lsn = Lsn(0x10);
5725 2 :
5726 42 : for _ in 0..20 {
5727 40 : lsn = Lsn(lsn.0 + 0x10);
5728 40 : let mut writer = tline.writer().await;
5729 40 : writer
5730 40 : .put(
5731 40 : test_key,
5732 40 : lsn,
5733 40 : &Value::Image(test_img(&format!("{} at {}", 0, lsn))),
5734 40 : &ctx,
5735 40 : )
5736 20 : .await?;
5737 40 : writer.finish_write(lsn);
5738 40 : drop(writer);
5739 40 : tline.freeze_and_flush().await?; // force create a delta layer
5740 2 : }
5741 2 :
5742 2 : let before_num_l0_delta_files = tline
5743 2 : .layers
5744 2 : .read()
5745 2 : .await
5746 2 : .layer_map()
5747 2 : .get_level0_deltas()?
5748 2 : .len();
5749 2 :
5750 110 : tline.compact(&cancel, EnumSet::empty(), &ctx).await?;
5751 2 :
5752 2 : let after_num_l0_delta_files = tline
5753 2 : .layers
5754 2 : .read()
5755 2 : .await
5756 2 : .layer_map()
5757 2 : .get_level0_deltas()?
5758 2 : .len();
5759 2 :
5760 2 : assert!(after_num_l0_delta_files < before_num_l0_delta_files, "after_num_l0_delta_files={after_num_l0_delta_files}, before_num_l0_delta_files={before_num_l0_delta_files}");
5761 2 :
5762 2 : assert_eq!(
5763 4 : tline.get(test_key, lsn, &ctx).await?,
5764 2 : test_img(&format!("{} at {}", 0, lsn))
5765 2 : );
5766 2 :
5767 2 : Ok(())
5768 2 : }
5769 :
5770 : #[tokio::test]
5771 2 : async fn test_branch_copies_dirty_aux_file_flag() {
5772 2 : let harness = TenantHarness::create("test_branch_copies_dirty_aux_file_flag")
5773 2 : .await
5774 2 : .unwrap();
5775 2 :
5776 2 : // the default aux file policy to switch is v1 if not set by the admins
5777 2 : assert_eq!(
5778 2 : harness.tenant_conf.switch_aux_file_policy,
5779 2 : AuxFilePolicy::V1
5780 2 : );
5781 8 : let (tenant, ctx) = harness.load().await;
5782 2 :
5783 2 : let mut lsn = Lsn(0x08);
5784 2 :
5785 2 : let tline: Arc<Timeline> = tenant
5786 2 : .create_test_timeline(TIMELINE_ID, lsn, DEFAULT_PG_VERSION, &ctx)
5787 6 : .await
5788 2 : .unwrap();
5789 2 :
5790 2 : // no aux file is written at this point, so the persistent flag should be unset
5791 2 : assert_eq!(tline.last_aux_file_policy.load(), None);
5792 2 :
5793 2 : {
5794 2 : lsn += 8;
5795 2 : let mut modification = tline.begin_modification(lsn);
5796 2 : modification
5797 2 : .put_file("pg_logical/mappings/test1", b"first", &ctx)
5798 4 : .await
5799 2 : .unwrap();
5800 2 : modification.commit(&ctx).await.unwrap();
5801 2 : }
5802 2 :
5803 2 : // there is no tenant manager to pass the configuration through, so lets mimic it
5804 2 : tenant.set_new_location_config(
5805 2 : AttachedTenantConf::try_from(LocationConf::attached_single(
5806 2 : TenantConfOpt {
5807 2 : switch_aux_file_policy: Some(AuxFilePolicy::V2),
5808 2 : ..Default::default()
5809 2 : },
5810 2 : tenant.generation,
5811 2 : &pageserver_api::models::ShardParameters::default(),
5812 2 : ))
5813 2 : .unwrap(),
5814 2 : );
5815 2 :
5816 2 : assert_eq!(
5817 2 : tline.get_switch_aux_file_policy(),
5818 2 : AuxFilePolicy::V2,
5819 2 : "wanted state has been updated"
5820 2 : );
5821 2 : assert_eq!(
5822 2 : tline.last_aux_file_policy.load(),
5823 2 : Some(AuxFilePolicy::V1),
5824 2 : "aux file is written with switch_aux_file_policy unset (which is v1), so we should keep v1"
5825 2 : );
5826 2 :
5827 2 : // we can read everything from the storage
5828 2 : let files = tline.list_aux_files(lsn, &ctx).await.unwrap();
5829 2 : assert_eq!(
5830 2 : files.get("pg_logical/mappings/test1"),
5831 2 : Some(&bytes::Bytes::from_static(b"first"))
5832 2 : );
5833 2 :
5834 2 : {
5835 2 : lsn += 8;
5836 2 : let mut modification = tline.begin_modification(lsn);
5837 2 : modification
5838 2 : .put_file("pg_logical/mappings/test2", b"second", &ctx)
5839 2 : .await
5840 2 : .unwrap();
5841 2 : modification.commit(&ctx).await.unwrap();
5842 2 : }
5843 2 :
5844 2 : assert_eq!(
5845 2 : tline.last_aux_file_policy.load(),
5846 2 : Some(AuxFilePolicy::V1),
5847 2 : "keep v1 storage format when new files are written"
5848 2 : );
5849 2 :
5850 2 : let files = tline.list_aux_files(lsn, &ctx).await.unwrap();
5851 2 : assert_eq!(
5852 2 : files.get("pg_logical/mappings/test2"),
5853 2 : Some(&bytes::Bytes::from_static(b"second"))
5854 2 : );
5855 2 :
5856 2 : let child = tenant
5857 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(lsn), &ctx)
5858 2 : .await
5859 2 : .unwrap();
5860 2 :
5861 2 : // child copies the last flag even if that is not on remote storage yet
5862 2 : assert_eq!(child.get_switch_aux_file_policy(), AuxFilePolicy::V2);
5863 2 : assert_eq!(child.last_aux_file_policy.load(), Some(AuxFilePolicy::V1));
5864 2 :
5865 2 : let files = child.list_aux_files(lsn, &ctx).await.unwrap();
5866 2 : assert_eq!(files.get("pg_logical/mappings/test1"), None);
5867 2 : assert_eq!(files.get("pg_logical/mappings/test2"), None);
5868 2 :
5869 2 : // even if we crash here without flushing parent timeline with it's new
5870 2 : // last_aux_file_policy we are safe, because child was never meant to access ancestor's
5871 2 : // files. the ancestor can even switch back to V1 because of a migration safely.
5872 2 : }
5873 :
5874 : #[tokio::test]
5875 2 : async fn aux_file_policy_switch() {
5876 2 : let mut harness = TenantHarness::create("aux_file_policy_switch")
5877 2 : .await
5878 2 : .unwrap();
5879 2 : harness.tenant_conf.switch_aux_file_policy = AuxFilePolicy::CrossValidation; // set to cross-validation mode
5880 8 : let (tenant, ctx) = harness.load().await;
5881 2 :
5882 2 : let mut lsn = Lsn(0x08);
5883 2 :
5884 2 : let tline: Arc<Timeline> = tenant
5885 2 : .create_test_timeline(TIMELINE_ID, lsn, DEFAULT_PG_VERSION, &ctx)
5886 6 : .await
5887 2 : .unwrap();
5888 2 :
5889 2 : assert_eq!(
5890 2 : tline.last_aux_file_policy.load(),
5891 2 : None,
5892 2 : "no aux file is written so it should be unset"
5893 2 : );
5894 2 :
5895 2 : {
5896 2 : lsn += 8;
5897 2 : let mut modification = tline.begin_modification(lsn);
5898 2 : modification
5899 2 : .put_file("pg_logical/mappings/test1", b"first", &ctx)
5900 4 : .await
5901 2 : .unwrap();
5902 2 : modification.commit(&ctx).await.unwrap();
5903 2 : }
5904 2 :
5905 2 : // there is no tenant manager to pass the configuration through, so lets mimic it
5906 2 : tenant.set_new_location_config(
5907 2 : AttachedTenantConf::try_from(LocationConf::attached_single(
5908 2 : TenantConfOpt {
5909 2 : switch_aux_file_policy: Some(AuxFilePolicy::V2),
5910 2 : ..Default::default()
5911 2 : },
5912 2 : tenant.generation,
5913 2 : &pageserver_api::models::ShardParameters::default(),
5914 2 : ))
5915 2 : .unwrap(),
5916 2 : );
5917 2 :
5918 2 : assert_eq!(
5919 2 : tline.get_switch_aux_file_policy(),
5920 2 : AuxFilePolicy::V2,
5921 2 : "wanted state has been updated"
5922 2 : );
5923 2 : assert_eq!(
5924 2 : tline.last_aux_file_policy.load(),
5925 2 : Some(AuxFilePolicy::CrossValidation),
5926 2 : "dirty index_part.json reflected state is yet to be updated"
5927 2 : );
5928 2 :
5929 2 : // we can still read the auxfile v1 before we ingest anything new
5930 2 : let files = tline.list_aux_files(lsn, &ctx).await.unwrap();
5931 2 : assert_eq!(
5932 2 : files.get("pg_logical/mappings/test1"),
5933 2 : Some(&bytes::Bytes::from_static(b"first"))
5934 2 : );
5935 2 :
5936 2 : {
5937 2 : lsn += 8;
5938 2 : let mut modification = tline.begin_modification(lsn);
5939 2 : modification
5940 2 : .put_file("pg_logical/mappings/test2", b"second", &ctx)
5941 2 : .await
5942 2 : .unwrap();
5943 2 : modification.commit(&ctx).await.unwrap();
5944 2 : }
5945 2 :
5946 2 : assert_eq!(
5947 2 : tline.last_aux_file_policy.load(),
5948 2 : Some(AuxFilePolicy::V2),
5949 2 : "ingesting a file should apply the wanted switch state when applicable"
5950 2 : );
5951 2 :
5952 2 : let files = tline.list_aux_files(lsn, &ctx).await.unwrap();
5953 2 : assert_eq!(
5954 2 : files.get("pg_logical/mappings/test1"),
5955 2 : Some(&bytes::Bytes::from_static(b"first")),
5956 2 : "cross validation writes to both v1 and v2 so this should be available in v2"
5957 2 : );
5958 2 : assert_eq!(
5959 2 : files.get("pg_logical/mappings/test2"),
5960 2 : Some(&bytes::Bytes::from_static(b"second"))
5961 2 : );
5962 2 :
5963 2 : // mimic again by trying to flip it from V2 to V1 (not switched to while ingesting a file)
5964 2 : tenant.set_new_location_config(
5965 2 : AttachedTenantConf::try_from(LocationConf::attached_single(
5966 2 : TenantConfOpt {
5967 2 : switch_aux_file_policy: Some(AuxFilePolicy::V1),
5968 2 : ..Default::default()
5969 2 : },
5970 2 : tenant.generation,
5971 2 : &pageserver_api::models::ShardParameters::default(),
5972 2 : ))
5973 2 : .unwrap(),
5974 2 : );
5975 2 :
5976 2 : {
5977 2 : lsn += 8;
5978 2 : let mut modification = tline.begin_modification(lsn);
5979 2 : modification
5980 2 : .put_file("pg_logical/mappings/test2", b"third", &ctx)
5981 2 : .await
5982 2 : .unwrap();
5983 2 : modification.commit(&ctx).await.unwrap();
5984 2 : }
5985 2 :
5986 2 : assert_eq!(
5987 2 : tline.get_switch_aux_file_policy(),
5988 2 : AuxFilePolicy::V1,
5989 2 : "wanted state has been updated again, even if invalid request"
5990 2 : );
5991 2 :
5992 2 : assert_eq!(
5993 2 : tline.last_aux_file_policy.load(),
5994 2 : Some(AuxFilePolicy::V2),
5995 2 : "ingesting a file should apply the wanted switch state when applicable"
5996 2 : );
5997 2 :
5998 2 : let files = tline.list_aux_files(lsn, &ctx).await.unwrap();
5999 2 : assert_eq!(
6000 2 : files.get("pg_logical/mappings/test1"),
6001 2 : Some(&bytes::Bytes::from_static(b"first"))
6002 2 : );
6003 2 : assert_eq!(
6004 2 : files.get("pg_logical/mappings/test2"),
6005 2 : Some(&bytes::Bytes::from_static(b"third"))
6006 2 : );
6007 2 :
6008 2 : // mimic again by trying to flip it from from V1 to V2 (not switched to while ingesting a file)
6009 2 : tenant.set_new_location_config(
6010 2 : AttachedTenantConf::try_from(LocationConf::attached_single(
6011 2 : TenantConfOpt {
6012 2 : switch_aux_file_policy: Some(AuxFilePolicy::V2),
6013 2 : ..Default::default()
6014 2 : },
6015 2 : tenant.generation,
6016 2 : &pageserver_api::models::ShardParameters::default(),
6017 2 : ))
6018 2 : .unwrap(),
6019 2 : );
6020 2 :
6021 2 : {
6022 2 : lsn += 8;
6023 2 : let mut modification = tline.begin_modification(lsn);
6024 2 : modification
6025 2 : .put_file("pg_logical/mappings/test3", b"last", &ctx)
6026 2 : .await
6027 2 : .unwrap();
6028 2 : modification.commit(&ctx).await.unwrap();
6029 2 : }
6030 2 :
6031 2 : assert_eq!(tline.get_switch_aux_file_policy(), AuxFilePolicy::V2);
6032 2 :
6033 2 : assert_eq!(tline.last_aux_file_policy.load(), Some(AuxFilePolicy::V2));
6034 2 :
6035 2 : let files = tline.list_aux_files(lsn, &ctx).await.unwrap();
6036 2 : assert_eq!(
6037 2 : files.get("pg_logical/mappings/test1"),
6038 2 : Some(&bytes::Bytes::from_static(b"first"))
6039 2 : );
6040 2 : assert_eq!(
6041 2 : files.get("pg_logical/mappings/test2"),
6042 2 : Some(&bytes::Bytes::from_static(b"third"))
6043 2 : );
6044 2 : assert_eq!(
6045 2 : files.get("pg_logical/mappings/test3"),
6046 2 : Some(&bytes::Bytes::from_static(b"last"))
6047 2 : );
6048 2 : }
6049 :
6050 : #[tokio::test]
6051 2 : async fn aux_file_policy_force_switch() {
6052 2 : let mut harness = TenantHarness::create("aux_file_policy_force_switch")
6053 2 : .await
6054 2 : .unwrap();
6055 2 : harness.tenant_conf.switch_aux_file_policy = AuxFilePolicy::V1;
6056 8 : let (tenant, ctx) = harness.load().await;
6057 2 :
6058 2 : let mut lsn = Lsn(0x08);
6059 2 :
6060 2 : let tline: Arc<Timeline> = tenant
6061 2 : .create_test_timeline(TIMELINE_ID, lsn, DEFAULT_PG_VERSION, &ctx)
6062 6 : .await
6063 2 : .unwrap();
6064 2 :
6065 2 : assert_eq!(
6066 2 : tline.last_aux_file_policy.load(),
6067 2 : None,
6068 2 : "no aux file is written so it should be unset"
6069 2 : );
6070 2 :
6071 2 : {
6072 2 : lsn += 8;
6073 2 : let mut modification = tline.begin_modification(lsn);
6074 2 : modification
6075 2 : .put_file("pg_logical/mappings/test1", b"first", &ctx)
6076 4 : .await
6077 2 : .unwrap();
6078 2 : modification.commit(&ctx).await.unwrap();
6079 2 : }
6080 2 :
6081 2 : tline.do_switch_aux_policy(AuxFilePolicy::V2).unwrap();
6082 2 :
6083 2 : assert_eq!(
6084 2 : tline.last_aux_file_policy.load(),
6085 2 : Some(AuxFilePolicy::V2),
6086 2 : "dirty index_part.json reflected state is yet to be updated"
6087 2 : );
6088 2 :
6089 2 : // lose all data from v1
6090 2 : let files = tline.list_aux_files(lsn, &ctx).await.unwrap();
6091 2 : assert_eq!(files.get("pg_logical/mappings/test1"), None);
6092 2 :
6093 2 : {
6094 2 : lsn += 8;
6095 2 : let mut modification = tline.begin_modification(lsn);
6096 2 : modification
6097 2 : .put_file("pg_logical/mappings/test2", b"second", &ctx)
6098 2 : .await
6099 2 : .unwrap();
6100 2 : modification.commit(&ctx).await.unwrap();
6101 2 : }
6102 2 :
6103 2 : // read data ingested in v2
6104 2 : let files = tline.list_aux_files(lsn, &ctx).await.unwrap();
6105 2 : assert_eq!(
6106 2 : files.get("pg_logical/mappings/test2"),
6107 2 : Some(&bytes::Bytes::from_static(b"second"))
6108 2 : );
6109 2 : // lose all data from v1
6110 2 : assert_eq!(files.get("pg_logical/mappings/test1"), None);
6111 2 : }
6112 :
6113 : #[tokio::test]
6114 2 : async fn aux_file_policy_auto_detect() {
6115 2 : let mut harness = TenantHarness::create("aux_file_policy_auto_detect")
6116 2 : .await
6117 2 : .unwrap();
6118 2 : harness.tenant_conf.switch_aux_file_policy = AuxFilePolicy::V2; // set to cross-validation mode
6119 8 : let (tenant, ctx) = harness.load().await;
6120 2 :
6121 2 : let mut lsn = Lsn(0x08);
6122 2 :
6123 2 : let tline: Arc<Timeline> = tenant
6124 2 : .create_test_timeline(TIMELINE_ID, lsn, DEFAULT_PG_VERSION, &ctx)
6125 6 : .await
6126 2 : .unwrap();
6127 2 :
6128 2 : assert_eq!(
6129 2 : tline.last_aux_file_policy.load(),
6130 2 : None,
6131 2 : "no aux file is written so it should be unset"
6132 2 : );
6133 2 :
6134 2 : {
6135 2 : lsn += 8;
6136 2 : let mut modification = tline.begin_modification(lsn);
6137 2 : let buf = AuxFilesDirectory::ser(&AuxFilesDirectory {
6138 2 : files: vec![(
6139 2 : "test_file".to_string(),
6140 2 : Bytes::copy_from_slice(b"test_file"),
6141 2 : )]
6142 2 : .into_iter()
6143 2 : .collect(),
6144 2 : })
6145 2 : .unwrap();
6146 2 : modification.put_for_test(AUX_FILES_KEY, Value::Image(Bytes::from(buf)));
6147 2 : modification.commit(&ctx).await.unwrap();
6148 2 : }
6149 2 :
6150 2 : {
6151 2 : lsn += 8;
6152 2 : let mut modification = tline.begin_modification(lsn);
6153 2 : modification
6154 2 : .put_file("pg_logical/mappings/test1", b"first", &ctx)
6155 2 : .await
6156 2 : .unwrap();
6157 2 : modification.commit(&ctx).await.unwrap();
6158 2 : }
6159 2 :
6160 2 : assert_eq!(
6161 2 : tline.last_aux_file_policy.load(),
6162 2 : Some(AuxFilePolicy::V1),
6163 2 : "keep using v1 because there are aux files writting with v1"
6164 2 : );
6165 2 :
6166 2 : // we can still read the auxfile v1
6167 2 : let files = tline.list_aux_files(lsn, &ctx).await.unwrap();
6168 2 : assert_eq!(
6169 2 : files.get("pg_logical/mappings/test1"),
6170 2 : Some(&bytes::Bytes::from_static(b"first"))
6171 2 : );
6172 2 : assert_eq!(
6173 2 : files.get("test_file"),
6174 2 : Some(&bytes::Bytes::from_static(b"test_file"))
6175 2 : );
6176 2 : }
6177 :
6178 : #[tokio::test]
6179 2 : async fn test_metadata_image_creation() -> anyhow::Result<()> {
6180 2 : let harness = TenantHarness::create("test_metadata_image_creation").await?;
6181 8 : let (tenant, ctx) = harness.load().await;
6182 2 : let tline = tenant
6183 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
6184 6 : .await?;
6185 2 :
6186 2 : const NUM_KEYS: usize = 1000;
6187 2 : const STEP: usize = 10000; // random update + scan base_key + idx * STEP
6188 2 :
6189 2 : let cancel = CancellationToken::new();
6190 2 :
6191 2 : let base_key = Key::from_hex("620000000033333333444444445500000000").unwrap();
6192 2 : assert_eq!(base_key.field1, AUX_KEY_PREFIX); // in case someone accidentally changed the prefix...
6193 2 : let mut test_key = base_key;
6194 2 : let mut lsn = Lsn(0x10);
6195 2 :
6196 8 : async fn scan_with_statistics(
6197 8 : tline: &Timeline,
6198 8 : keyspace: &KeySpace,
6199 8 : lsn: Lsn,
6200 8 : ctx: &RequestContext,
6201 8 : ) -> anyhow::Result<(BTreeMap<Key, Result<Bytes, PageReconstructError>>, usize)> {
6202 8 : let mut reconstruct_state = ValuesReconstructState::default();
6203 8 : let res = tline
6204 8 : .get_vectored_impl(keyspace.clone(), lsn, &mut reconstruct_state, ctx)
6205 1680 : .await?;
6206 8 : Ok((res, reconstruct_state.get_delta_layers_visited() as usize))
6207 8 : }
6208 2 :
6209 2 : #[allow(clippy::needless_range_loop)]
6210 2002 : for blknum in 0..NUM_KEYS {
6211 2000 : lsn = Lsn(lsn.0 + 0x10);
6212 2000 : test_key.field6 = (blknum * STEP) as u32;
6213 2000 : let mut writer = tline.writer().await;
6214 2000 : writer
6215 2000 : .put(
6216 2000 : test_key,
6217 2000 : lsn,
6218 2000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
6219 2000 : &ctx,
6220 2000 : )
6221 2 : .await?;
6222 2000 : writer.finish_write(lsn);
6223 2000 : drop(writer);
6224 2 : }
6225 2 :
6226 2 : let keyspace = KeySpace::single(base_key..base_key.add((NUM_KEYS * STEP) as u32));
6227 2 :
6228 22 : for iter in 1..=10 {
6229 20020 : for _ in 0..NUM_KEYS {
6230 20000 : lsn = Lsn(lsn.0 + 0x10);
6231 20000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
6232 20000 : test_key.field6 = (blknum * STEP) as u32;
6233 20000 : let mut writer = tline.writer().await;
6234 20000 : writer
6235 20000 : .put(
6236 20000 : test_key,
6237 20000 : lsn,
6238 20000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
6239 20000 : &ctx,
6240 20000 : )
6241 35 : .await?;
6242 20000 : writer.finish_write(lsn);
6243 20000 : drop(writer);
6244 2 : }
6245 2 :
6246 20 : tline.freeze_and_flush().await?;
6247 2 :
6248 20 : if iter % 5 == 0 {
6249 4 : let (_, before_delta_file_accessed) =
6250 1672 : scan_with_statistics(&tline, &keyspace, lsn, &ctx).await?;
6251 4 : tline
6252 4 : .compact(
6253 4 : &cancel,
6254 4 : {
6255 4 : let mut flags = EnumSet::new();
6256 4 : flags.insert(CompactFlags::ForceImageLayerCreation);
6257 4 : flags.insert(CompactFlags::ForceRepartition);
6258 4 : flags
6259 4 : },
6260 4 : &ctx,
6261 4 : )
6262 6513 : .await?;
6263 4 : let (_, after_delta_file_accessed) =
6264 8 : scan_with_statistics(&tline, &keyspace, lsn, &ctx).await?;
6265 4 : assert!(after_delta_file_accessed < before_delta_file_accessed, "after_delta_file_accessed={after_delta_file_accessed}, before_delta_file_accessed={before_delta_file_accessed}");
6266 2 : // Given that we already produced an image layer, there should be no delta layer needed for the scan, but still setting a low threshold there for unforeseen circumstances.
6267 4 : assert!(
6268 4 : after_delta_file_accessed <= 2,
6269 2 : "after_delta_file_accessed={after_delta_file_accessed}"
6270 2 : );
6271 16 : }
6272 2 : }
6273 2 :
6274 2 : Ok(())
6275 2 : }
6276 :
6277 : #[tokio::test]
6278 2 : async fn test_vectored_missing_data_key_reads() -> anyhow::Result<()> {
6279 2 : let harness = TenantHarness::create("test_vectored_missing_data_key_reads").await?;
6280 8 : let (tenant, ctx) = harness.load().await;
6281 2 :
6282 2 : let base_key = Key::from_hex("000000000033333333444444445500000000").unwrap();
6283 2 : let base_key_child = Key::from_hex("000000000033333333444444445500000001").unwrap();
6284 2 : let base_key_nonexist = Key::from_hex("000000000033333333444444445500000002").unwrap();
6285 2 :
6286 2 : let tline = tenant
6287 2 : .create_test_timeline_with_layers(
6288 2 : TIMELINE_ID,
6289 2 : Lsn(0x10),
6290 2 : DEFAULT_PG_VERSION,
6291 2 : &ctx,
6292 2 : Vec::new(), // delta layers
6293 2 : vec![(Lsn(0x20), vec![(base_key, test_img("data key 1"))])], // image layers
6294 2 : Lsn(0x20), // it's fine to not advance LSN to 0x30 while using 0x30 to get below because `get_vectored_impl` does not wait for LSN
6295 2 : )
6296 13 : .await?;
6297 2 : tline.add_extra_test_dense_keyspace(KeySpace::single(base_key..(base_key_nonexist.next())));
6298 2 :
6299 2 : let child = tenant
6300 2 : .branch_timeline_test_with_layers(
6301 2 : &tline,
6302 2 : NEW_TIMELINE_ID,
6303 2 : Some(Lsn(0x20)),
6304 2 : &ctx,
6305 2 : Vec::new(), // delta layers
6306 2 : vec![(Lsn(0x30), vec![(base_key_child, test_img("data key 2"))])], // image layers
6307 2 : Lsn(0x30),
6308 2 : )
6309 7 : .await
6310 2 : .unwrap();
6311 2 :
6312 2 : let lsn = Lsn(0x30);
6313 2 :
6314 2 : // test vectored get on parent timeline
6315 2 : assert_eq!(
6316 4 : get_vectored_impl_wrapper(&tline, base_key, lsn, &ctx).await?,
6317 2 : Some(test_img("data key 1"))
6318 2 : );
6319 2 : assert!(get_vectored_impl_wrapper(&tline, base_key_child, lsn, &ctx)
6320 3 : .await
6321 2 : .unwrap_err()
6322 2 : .is_missing_key_error());
6323 2 : assert!(
6324 2 : get_vectored_impl_wrapper(&tline, base_key_nonexist, lsn, &ctx)
6325 2 : .await
6326 2 : .unwrap_err()
6327 2 : .is_missing_key_error()
6328 2 : );
6329 2 :
6330 2 : // test vectored get on child timeline
6331 2 : assert_eq!(
6332 2 : get_vectored_impl_wrapper(&child, base_key, lsn, &ctx).await?,
6333 2 : Some(test_img("data key 1"))
6334 2 : );
6335 2 : assert_eq!(
6336 4 : get_vectored_impl_wrapper(&child, base_key_child, lsn, &ctx).await?,
6337 2 : Some(test_img("data key 2"))
6338 2 : );
6339 2 : assert!(
6340 2 : get_vectored_impl_wrapper(&child, base_key_nonexist, lsn, &ctx)
6341 2 : .await
6342 2 : .unwrap_err()
6343 2 : .is_missing_key_error()
6344 2 : );
6345 2 :
6346 2 : Ok(())
6347 2 : }
6348 :
6349 : #[tokio::test]
6350 2 : async fn test_vectored_missing_metadata_key_reads() -> anyhow::Result<()> {
6351 2 : let harness = TenantHarness::create("test_vectored_missing_metadata_key_reads").await?;
6352 8 : let (tenant, ctx) = harness.load().await;
6353 2 :
6354 2 : let base_key = Key::from_hex("620000000033333333444444445500000000").unwrap();
6355 2 : let base_key_child = Key::from_hex("620000000033333333444444445500000001").unwrap();
6356 2 : let base_key_nonexist = Key::from_hex("620000000033333333444444445500000002").unwrap();
6357 2 : assert_eq!(base_key.field1, AUX_KEY_PREFIX); // in case someone accidentally changed the prefix...
6358 2 :
6359 2 : let tline = tenant
6360 2 : .create_test_timeline_with_layers(
6361 2 : TIMELINE_ID,
6362 2 : Lsn(0x10),
6363 2 : DEFAULT_PG_VERSION,
6364 2 : &ctx,
6365 2 : Vec::new(), // delta layers
6366 2 : vec![(Lsn(0x20), vec![(base_key, test_img("metadata key 1"))])], // image layers
6367 2 : Lsn(0x20), // it's fine to not advance LSN to 0x30 while using 0x30 to get below because `get_vectored_impl` does not wait for LSN
6368 2 : )
6369 13 : .await?;
6370 2 :
6371 2 : let child = tenant
6372 2 : .branch_timeline_test_with_layers(
6373 2 : &tline,
6374 2 : NEW_TIMELINE_ID,
6375 2 : Some(Lsn(0x20)),
6376 2 : &ctx,
6377 2 : Vec::new(), // delta layers
6378 2 : vec![(
6379 2 : Lsn(0x30),
6380 2 : vec![(base_key_child, test_img("metadata key 2"))],
6381 2 : )], // image layers
6382 2 : Lsn(0x30),
6383 2 : )
6384 7 : .await
6385 2 : .unwrap();
6386 2 :
6387 2 : let lsn = Lsn(0x30);
6388 2 :
6389 2 : // test vectored get on parent timeline
6390 2 : assert_eq!(
6391 4 : get_vectored_impl_wrapper(&tline, base_key, lsn, &ctx).await?,
6392 2 : Some(test_img("metadata key 1"))
6393 2 : );
6394 2 : assert_eq!(
6395 2 : get_vectored_impl_wrapper(&tline, base_key_child, lsn, &ctx).await?,
6396 2 : None
6397 2 : );
6398 2 : assert_eq!(
6399 2 : get_vectored_impl_wrapper(&tline, base_key_nonexist, lsn, &ctx).await?,
6400 2 : None
6401 2 : );
6402 2 :
6403 2 : // test vectored get on child timeline
6404 2 : assert_eq!(
6405 2 : get_vectored_impl_wrapper(&child, base_key, lsn, &ctx).await?,
6406 2 : None
6407 2 : );
6408 2 : assert_eq!(
6409 4 : get_vectored_impl_wrapper(&child, base_key_child, lsn, &ctx).await?,
6410 2 : Some(test_img("metadata key 2"))
6411 2 : );
6412 2 : assert_eq!(
6413 2 : get_vectored_impl_wrapper(&child, base_key_nonexist, lsn, &ctx).await?,
6414 2 : None
6415 2 : );
6416 2 :
6417 2 : Ok(())
6418 2 : }
6419 :
6420 36 : async fn get_vectored_impl_wrapper(
6421 36 : tline: &Arc<Timeline>,
6422 36 : key: Key,
6423 36 : lsn: Lsn,
6424 36 : ctx: &RequestContext,
6425 36 : ) -> Result<Option<Bytes>, GetVectoredError> {
6426 36 : let mut reconstruct_state = ValuesReconstructState::new();
6427 36 : let mut res = tline
6428 36 : .get_vectored_impl(
6429 36 : KeySpace::single(key..key.next()),
6430 36 : lsn,
6431 36 : &mut reconstruct_state,
6432 36 : ctx,
6433 36 : )
6434 33 : .await?;
6435 30 : Ok(res.pop_last().map(|(k, v)| {
6436 18 : assert_eq!(k, key);
6437 18 : v.unwrap()
6438 30 : }))
6439 36 : }
6440 :
6441 : #[tokio::test]
6442 2 : async fn test_metadata_tombstone_reads() -> anyhow::Result<()> {
6443 2 : let harness = TenantHarness::create("test_metadata_tombstone_reads").await?;
6444 8 : let (tenant, ctx) = harness.load().await;
6445 2 : let key0 = Key::from_hex("620000000033333333444444445500000000").unwrap();
6446 2 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
6447 2 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
6448 2 : let key3 = Key::from_hex("620000000033333333444444445500000003").unwrap();
6449 2 :
6450 2 : // We emulate the situation that the compaction algorithm creates an image layer that removes the tombstones
6451 2 : // Lsn 0x30 key0, key3, no key1+key2
6452 2 : // Lsn 0x20 key1+key2 tomestones
6453 2 : // Lsn 0x10 key1 in image, key2 in delta
6454 2 : let tline = tenant
6455 2 : .create_test_timeline_with_layers(
6456 2 : TIMELINE_ID,
6457 2 : Lsn(0x10),
6458 2 : DEFAULT_PG_VERSION,
6459 2 : &ctx,
6460 2 : // delta layers
6461 2 : vec![
6462 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
6463 2 : Lsn(0x10)..Lsn(0x20),
6464 2 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
6465 2 : ),
6466 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
6467 2 : Lsn(0x20)..Lsn(0x30),
6468 2 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
6469 2 : ),
6470 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
6471 2 : Lsn(0x20)..Lsn(0x30),
6472 2 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
6473 2 : ),
6474 2 : ],
6475 2 : // image layers
6476 2 : vec![
6477 2 : (Lsn(0x10), vec![(key1, test_img("metadata key 1"))]),
6478 2 : (
6479 2 : Lsn(0x30),
6480 2 : vec![
6481 2 : (key0, test_img("metadata key 0")),
6482 2 : (key3, test_img("metadata key 3")),
6483 2 : ],
6484 2 : ),
6485 2 : ],
6486 2 : Lsn(0x30),
6487 2 : )
6488 40 : .await?;
6489 2 :
6490 2 : let lsn = Lsn(0x30);
6491 2 : let old_lsn = Lsn(0x20);
6492 2 :
6493 2 : assert_eq!(
6494 4 : get_vectored_impl_wrapper(&tline, key0, lsn, &ctx).await?,
6495 2 : Some(test_img("metadata key 0"))
6496 2 : );
6497 2 : assert_eq!(
6498 2 : get_vectored_impl_wrapper(&tline, key1, lsn, &ctx).await?,
6499 2 : None,
6500 2 : );
6501 2 : assert_eq!(
6502 2 : get_vectored_impl_wrapper(&tline, key2, lsn, &ctx).await?,
6503 2 : None,
6504 2 : );
6505 2 : assert_eq!(
6506 4 : get_vectored_impl_wrapper(&tline, key1, old_lsn, &ctx).await?,
6507 2 : Some(Bytes::new()),
6508 2 : );
6509 2 : assert_eq!(
6510 4 : get_vectored_impl_wrapper(&tline, key2, old_lsn, &ctx).await?,
6511 2 : Some(Bytes::new()),
6512 2 : );
6513 2 : assert_eq!(
6514 2 : get_vectored_impl_wrapper(&tline, key3, lsn, &ctx).await?,
6515 2 : Some(test_img("metadata key 3"))
6516 2 : );
6517 2 :
6518 2 : Ok(())
6519 2 : }
6520 :
6521 : #[tokio::test]
6522 2 : async fn test_metadata_tombstone_image_creation() {
6523 2 : let harness = TenantHarness::create("test_metadata_tombstone_image_creation")
6524 2 : .await
6525 2 : .unwrap();
6526 8 : let (tenant, ctx) = harness.load().await;
6527 2 :
6528 2 : let key0 = Key::from_hex("620000000033333333444444445500000000").unwrap();
6529 2 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
6530 2 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
6531 2 : let key3 = Key::from_hex("620000000033333333444444445500000003").unwrap();
6532 2 :
6533 2 : let tline = tenant
6534 2 : .create_test_timeline_with_layers(
6535 2 : TIMELINE_ID,
6536 2 : Lsn(0x10),
6537 2 : DEFAULT_PG_VERSION,
6538 2 : &ctx,
6539 2 : // delta layers
6540 2 : vec![
6541 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
6542 2 : Lsn(0x10)..Lsn(0x20),
6543 2 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
6544 2 : ),
6545 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
6546 2 : Lsn(0x20)..Lsn(0x30),
6547 2 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
6548 2 : ),
6549 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
6550 2 : Lsn(0x20)..Lsn(0x30),
6551 2 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
6552 2 : ),
6553 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
6554 2 : Lsn(0x30)..Lsn(0x40),
6555 2 : vec![
6556 2 : (key0, Lsn(0x30), Value::Image(test_img("metadata key 0"))),
6557 2 : (key3, Lsn(0x30), Value::Image(test_img("metadata key 3"))),
6558 2 : ],
6559 2 : ),
6560 2 : ],
6561 2 : // image layers
6562 2 : vec![(Lsn(0x10), vec![(key1, test_img("metadata key 1"))])],
6563 2 : Lsn(0x40),
6564 2 : )
6565 37 : .await
6566 2 : .unwrap();
6567 2 :
6568 2 : let cancel = CancellationToken::new();
6569 2 :
6570 2 : tline
6571 2 : .compact(
6572 2 : &cancel,
6573 2 : {
6574 2 : let mut flags = EnumSet::new();
6575 2 : flags.insert(CompactFlags::ForceImageLayerCreation);
6576 2 : flags.insert(CompactFlags::ForceRepartition);
6577 2 : flags
6578 2 : },
6579 2 : &ctx,
6580 2 : )
6581 49 : .await
6582 2 : .unwrap();
6583 2 :
6584 2 : // Image layers are created at last_record_lsn
6585 2 : let images = tline
6586 2 : .inspect_image_layers(Lsn(0x40), &ctx)
6587 8 : .await
6588 2 : .unwrap()
6589 2 : .into_iter()
6590 20 : .filter(|(k, _)| k.is_metadata_key())
6591 2 : .collect::<Vec<_>>();
6592 2 : assert_eq!(images.len(), 2); // the image layer should only contain two existing keys, tombstones should be removed.
6593 2 : }
6594 :
6595 : #[tokio::test]
6596 2 : async fn test_metadata_tombstone_empty_image_creation() {
6597 2 : let harness = TenantHarness::create("test_metadata_tombstone_empty_image_creation")
6598 2 : .await
6599 2 : .unwrap();
6600 8 : let (tenant, ctx) = harness.load().await;
6601 2 :
6602 2 : let key1 = Key::from_hex("620000000033333333444444445500000001").unwrap();
6603 2 : let key2 = Key::from_hex("620000000033333333444444445500000002").unwrap();
6604 2 :
6605 2 : let tline = tenant
6606 2 : .create_test_timeline_with_layers(
6607 2 : TIMELINE_ID,
6608 2 : Lsn(0x10),
6609 2 : DEFAULT_PG_VERSION,
6610 2 : &ctx,
6611 2 : // delta layers
6612 2 : vec![
6613 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
6614 2 : Lsn(0x10)..Lsn(0x20),
6615 2 : vec![(key2, Lsn(0x10), Value::Image(test_img("metadata key 2")))],
6616 2 : ),
6617 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
6618 2 : Lsn(0x20)..Lsn(0x30),
6619 2 : vec![(key1, Lsn(0x20), Value::Image(Bytes::new()))],
6620 2 : ),
6621 2 : DeltaLayerTestDesc::new_with_inferred_key_range(
6622 2 : Lsn(0x20)..Lsn(0x30),
6623 2 : vec![(key2, Lsn(0x20), Value::Image(Bytes::new()))],
6624 2 : ),
6625 2 : ],
6626 2 : // image layers
6627 2 : vec![(Lsn(0x10), vec![(key1, test_img("metadata key 1"))])],
6628 2 : Lsn(0x30),
6629 2 : )
6630 31 : .await
6631 2 : .unwrap();
6632 2 :
6633 2 : let cancel = CancellationToken::new();
6634 2 :
6635 2 : tline
6636 2 : .compact(
6637 2 : &cancel,
6638 2 : {
6639 2 : let mut flags = EnumSet::new();
6640 2 : flags.insert(CompactFlags::ForceImageLayerCreation);
6641 2 : flags.insert(CompactFlags::ForceRepartition);
6642 2 : flags
6643 2 : },
6644 2 : &ctx,
6645 2 : )
6646 37 : .await
6647 2 : .unwrap();
6648 2 :
6649 2 : // Image layers are created at last_record_lsn
6650 2 : let images = tline
6651 2 : .inspect_image_layers(Lsn(0x30), &ctx)
6652 4 : .await
6653 2 : .unwrap()
6654 2 : .into_iter()
6655 16 : .filter(|(k, _)| k.is_metadata_key())
6656 2 : .collect::<Vec<_>>();
6657 2 : assert_eq!(images.len(), 0); // the image layer should not contain tombstones, or it is not created
6658 2 : }
6659 :
6660 : #[tokio::test]
6661 2 : async fn test_simple_bottom_most_compaction_images() -> anyhow::Result<()> {
6662 2 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_images").await?;
6663 8 : let (tenant, ctx) = harness.load().await;
6664 2 :
6665 102 : fn get_key(id: u32) -> Key {
6666 102 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
6667 102 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
6668 102 : key.field6 = id;
6669 102 : key
6670 102 : }
6671 2 :
6672 2 : // We create
6673 2 : // - one bottom-most image layer,
6674 2 : // - a delta layer D1 crossing the GC horizon with data below and above the horizon,
6675 2 : // - a delta layer D2 crossing the GC horizon with data only below the horizon,
6676 2 : // - a delta layer D3 above the horizon.
6677 2 : //
6678 2 : // | D3 |
6679 2 : // | D1 |
6680 2 : // -| |-- gc horizon -----------------
6681 2 : // | | | D2 |
6682 2 : // --------- img layer ------------------
6683 2 : //
6684 2 : // What we should expact from this compaction is:
6685 2 : // | D3 |
6686 2 : // | Part of D1 |
6687 2 : // --------- img layer with D1+D2 at GC horizon------------------
6688 2 :
6689 2 : // img layer at 0x10
6690 2 : let img_layer = (0..10)
6691 20 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
6692 2 : .collect_vec();
6693 2 :
6694 2 : let delta1 = vec![
6695 2 : (
6696 2 : get_key(1),
6697 2 : Lsn(0x20),
6698 2 : Value::Image(Bytes::from("value 1@0x20")),
6699 2 : ),
6700 2 : (
6701 2 : get_key(2),
6702 2 : Lsn(0x30),
6703 2 : Value::Image(Bytes::from("value 2@0x30")),
6704 2 : ),
6705 2 : (
6706 2 : get_key(3),
6707 2 : Lsn(0x40),
6708 2 : Value::Image(Bytes::from("value 3@0x40")),
6709 2 : ),
6710 2 : ];
6711 2 : let delta2 = vec![
6712 2 : (
6713 2 : get_key(5),
6714 2 : Lsn(0x20),
6715 2 : Value::Image(Bytes::from("value 5@0x20")),
6716 2 : ),
6717 2 : (
6718 2 : get_key(6),
6719 2 : Lsn(0x20),
6720 2 : Value::Image(Bytes::from("value 6@0x20")),
6721 2 : ),
6722 2 : ];
6723 2 : let delta3 = vec![
6724 2 : (
6725 2 : get_key(8),
6726 2 : Lsn(0x48),
6727 2 : Value::Image(Bytes::from("value 8@0x48")),
6728 2 : ),
6729 2 : (
6730 2 : get_key(9),
6731 2 : Lsn(0x48),
6732 2 : Value::Image(Bytes::from("value 9@0x48")),
6733 2 : ),
6734 2 : ];
6735 2 :
6736 2 : let tline = tenant
6737 2 : .create_test_timeline_with_layers(
6738 2 : TIMELINE_ID,
6739 2 : Lsn(0x10),
6740 2 : DEFAULT_PG_VERSION,
6741 2 : &ctx,
6742 2 : vec![
6743 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta1),
6744 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x20)..Lsn(0x48), delta2),
6745 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
6746 2 : ], // delta layers
6747 2 : vec![(Lsn(0x10), img_layer)], // image layers
6748 2 : Lsn(0x50),
6749 2 : )
6750 49 : .await?;
6751 2 : {
6752 2 : // Update GC info
6753 2 : let mut guard = tline.gc_info.write().unwrap();
6754 2 : guard.cutoffs.time = Lsn(0x30);
6755 2 : guard.cutoffs.space = Lsn(0x30);
6756 2 : }
6757 2 :
6758 2 : let expected_result = [
6759 2 : Bytes::from_static(b"value 0@0x10"),
6760 2 : Bytes::from_static(b"value 1@0x20"),
6761 2 : Bytes::from_static(b"value 2@0x30"),
6762 2 : Bytes::from_static(b"value 3@0x40"),
6763 2 : Bytes::from_static(b"value 4@0x10"),
6764 2 : Bytes::from_static(b"value 5@0x20"),
6765 2 : Bytes::from_static(b"value 6@0x20"),
6766 2 : Bytes::from_static(b"value 7@0x10"),
6767 2 : Bytes::from_static(b"value 8@0x48"),
6768 2 : Bytes::from_static(b"value 9@0x48"),
6769 2 : ];
6770 2 :
6771 20 : for (idx, expected) in expected_result.iter().enumerate() {
6772 20 : assert_eq!(
6773 20 : tline
6774 20 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
6775 17 : .await
6776 20 : .unwrap(),
6777 2 : expected
6778 2 : );
6779 2 : }
6780 2 :
6781 2 : let cancel = CancellationToken::new();
6782 55 : tline.compact_with_gc(&cancel, &ctx).await.unwrap();
6783 2 :
6784 20 : for (idx, expected) in expected_result.iter().enumerate() {
6785 20 : assert_eq!(
6786 20 : tline
6787 20 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
6788 8 : .await
6789 20 : .unwrap(),
6790 2 : expected
6791 2 : );
6792 2 : }
6793 2 :
6794 2 : // Check if the image layer at the GC horizon contains exactly what we want
6795 2 : let image_at_gc_horizon = tline
6796 2 : .inspect_image_layers(Lsn(0x30), &ctx)
6797 2 : .await
6798 2 : .unwrap()
6799 2 : .into_iter()
6800 36 : .filter(|(k, _)| k.is_metadata_key())
6801 2 : .collect::<Vec<_>>();
6802 2 :
6803 2 : assert_eq!(image_at_gc_horizon.len(), 10);
6804 2 : let expected_result = [
6805 2 : Bytes::from_static(b"value 0@0x10"),
6806 2 : Bytes::from_static(b"value 1@0x20"),
6807 2 : Bytes::from_static(b"value 2@0x30"),
6808 2 : Bytes::from_static(b"value 3@0x10"),
6809 2 : Bytes::from_static(b"value 4@0x10"),
6810 2 : Bytes::from_static(b"value 5@0x20"),
6811 2 : Bytes::from_static(b"value 6@0x20"),
6812 2 : Bytes::from_static(b"value 7@0x10"),
6813 2 : Bytes::from_static(b"value 8@0x10"),
6814 2 : Bytes::from_static(b"value 9@0x10"),
6815 2 : ];
6816 22 : for idx in 0..10 {
6817 20 : assert_eq!(
6818 20 : image_at_gc_horizon[idx],
6819 20 : (get_key(idx as u32), expected_result[idx].clone())
6820 20 : );
6821 2 : }
6822 2 :
6823 2 : // Check if old layers are removed / new layers have the expected LSN
6824 2 : let mut all_layers = tline.inspect_historic_layers().await.unwrap();
6825 4 : all_layers.sort_by(|k1, k2| {
6826 4 : (
6827 4 : k1.is_delta,
6828 4 : k1.key_range.start,
6829 4 : k1.key_range.end,
6830 4 : k1.lsn_range.start,
6831 4 : k1.lsn_range.end,
6832 4 : )
6833 4 : .cmp(&(
6834 4 : k2.is_delta,
6835 4 : k2.key_range.start,
6836 4 : k2.key_range.end,
6837 4 : k2.lsn_range.start,
6838 4 : k2.lsn_range.end,
6839 4 : ))
6840 4 : });
6841 2 : assert_eq!(
6842 2 : all_layers,
6843 2 : vec![
6844 2 : // Image layer at GC horizon
6845 2 : PersistentLayerKey {
6846 2 : key_range: Key::MIN..Key::MAX,
6847 2 : lsn_range: Lsn(0x30)..Lsn(0x31),
6848 2 : is_delta: false
6849 2 : },
6850 2 : // The delta layer that is cut in the middle
6851 2 : PersistentLayerKey {
6852 2 : key_range: get_key(3)..get_key(4),
6853 2 : lsn_range: Lsn(0x30)..Lsn(0x41),
6854 2 : is_delta: true
6855 2 : },
6856 2 : // The delta3 layer that should not be picked for the compaction
6857 2 : PersistentLayerKey {
6858 2 : key_range: get_key(8)..get_key(10),
6859 2 : lsn_range: Lsn(0x48)..Lsn(0x50),
6860 2 : is_delta: true
6861 2 : }
6862 2 : ]
6863 2 : );
6864 2 :
6865 2 : Ok(())
6866 2 : }
6867 :
6868 : #[tokio::test]
6869 2 : async fn test_neon_test_record() -> anyhow::Result<()> {
6870 2 : let harness = TenantHarness::create("test_neon_test_record").await?;
6871 8 : let (tenant, ctx) = harness.load().await;
6872 2 :
6873 24 : fn get_key(id: u32) -> Key {
6874 24 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
6875 24 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
6876 24 : key.field6 = id;
6877 24 : key
6878 24 : }
6879 2 :
6880 2 : let delta1 = vec![
6881 2 : (
6882 2 : get_key(1),
6883 2 : Lsn(0x20),
6884 2 : Value::WalRecord(NeonWalRecord::wal_append(",0x20")),
6885 2 : ),
6886 2 : (
6887 2 : get_key(1),
6888 2 : Lsn(0x30),
6889 2 : Value::WalRecord(NeonWalRecord::wal_append(",0x30")),
6890 2 : ),
6891 2 : (get_key(2), Lsn(0x10), Value::Image("0x10".into())),
6892 2 : (
6893 2 : get_key(2),
6894 2 : Lsn(0x20),
6895 2 : Value::WalRecord(NeonWalRecord::wal_append(",0x20")),
6896 2 : ),
6897 2 : (
6898 2 : get_key(2),
6899 2 : Lsn(0x30),
6900 2 : Value::WalRecord(NeonWalRecord::wal_append(",0x30")),
6901 2 : ),
6902 2 : (get_key(3), Lsn(0x10), Value::Image("0x10".into())),
6903 2 : (
6904 2 : get_key(3),
6905 2 : Lsn(0x20),
6906 2 : Value::WalRecord(NeonWalRecord::wal_clear()),
6907 2 : ),
6908 2 : (get_key(4), Lsn(0x10), Value::Image("0x10".into())),
6909 2 : (
6910 2 : get_key(4),
6911 2 : Lsn(0x20),
6912 2 : Value::WalRecord(NeonWalRecord::wal_init()),
6913 2 : ),
6914 2 : ];
6915 2 : let image1 = vec![(get_key(1), "0x10".into())];
6916 2 :
6917 2 : let tline = tenant
6918 2 : .create_test_timeline_with_layers(
6919 2 : TIMELINE_ID,
6920 2 : Lsn(0x10),
6921 2 : DEFAULT_PG_VERSION,
6922 2 : &ctx,
6923 2 : vec![DeltaLayerTestDesc::new_with_inferred_key_range(
6924 2 : Lsn(0x10)..Lsn(0x40),
6925 2 : delta1,
6926 2 : )], // delta layers
6927 2 : vec![(Lsn(0x10), image1)], // image layers
6928 2 : Lsn(0x50),
6929 2 : )
6930 19 : .await?;
6931 2 :
6932 2 : assert_eq!(
6933 8 : tline.get(get_key(1), Lsn(0x50), &ctx).await?,
6934 2 : Bytes::from_static(b"0x10,0x20,0x30")
6935 2 : );
6936 2 : assert_eq!(
6937 2 : tline.get(get_key(2), Lsn(0x50), &ctx).await?,
6938 2 : Bytes::from_static(b"0x10,0x20,0x30")
6939 2 : );
6940 2 :
6941 2 : // Need to remove the limit of "Neon WAL redo requires base image".
6942 2 :
6943 2 : // assert_eq!(tline.get(get_key(3), Lsn(0x50), &ctx).await?, Bytes::new());
6944 2 : // assert_eq!(tline.get(get_key(4), Lsn(0x50), &ctx).await?, Bytes::new());
6945 2 :
6946 2 : Ok(())
6947 2 : }
6948 :
6949 : #[tokio::test]
6950 2 : async fn test_lsn_lease() -> anyhow::Result<()> {
6951 8 : let (tenant, ctx) = TenantHarness::create("test_lsn_lease").await?.load().await;
6952 2 : let key = Key::from_hex("010000000033333333444444445500000000").unwrap();
6953 2 :
6954 2 : let end_lsn = Lsn(0x100);
6955 2 : let image_layers = (0x20..=0x90)
6956 2 : .step_by(0x10)
6957 16 : .map(|n| {
6958 16 : (
6959 16 : Lsn(n),
6960 16 : vec![(key, test_img(&format!("data key at {:x}", n)))],
6961 16 : )
6962 16 : })
6963 2 : .collect();
6964 2 :
6965 2 : let timeline = tenant
6966 2 : .create_test_timeline_with_layers(
6967 2 : TIMELINE_ID,
6968 2 : Lsn(0x10),
6969 2 : DEFAULT_PG_VERSION,
6970 2 : &ctx,
6971 2 : Vec::new(),
6972 2 : image_layers,
6973 2 : end_lsn,
6974 2 : )
6975 62 : .await?;
6976 2 :
6977 2 : let leased_lsns = [0x30, 0x50, 0x70];
6978 2 : let mut leases = Vec::new();
6979 6 : let _: anyhow::Result<_> = leased_lsns.iter().try_for_each(|n| {
6980 6 : leases.push(timeline.make_lsn_lease(Lsn(*n), timeline.get_lsn_lease_length(), &ctx)?);
6981 6 : Ok(())
6982 6 : });
6983 2 :
6984 2 : // Renewing with shorter lease should not change the lease.
6985 2 : let updated_lease_0 =
6986 2 : timeline.make_lsn_lease(Lsn(leased_lsns[0]), Duration::from_secs(0), &ctx)?;
6987 2 : assert_eq!(updated_lease_0.valid_until, leases[0].valid_until);
6988 2 :
6989 2 : // Renewing with a long lease should renew lease with later expiration time.
6990 2 : let updated_lease_1 = timeline.make_lsn_lease(
6991 2 : Lsn(leased_lsns[1]),
6992 2 : timeline.get_lsn_lease_length() * 2,
6993 2 : &ctx,
6994 2 : )?;
6995 2 :
6996 2 : assert!(updated_lease_1.valid_until > leases[1].valid_until);
6997 2 :
6998 2 : // Force set disk consistent lsn so we can get the cutoff at `end_lsn`.
6999 2 : info!(
7000 2 : "latest_gc_cutoff_lsn: {}",
7001 0 : *timeline.get_latest_gc_cutoff_lsn()
7002 2 : );
7003 2 : timeline.force_set_disk_consistent_lsn(end_lsn);
7004 2 :
7005 2 : let res = tenant
7006 2 : .gc_iteration(
7007 2 : Some(TIMELINE_ID),
7008 2 : 0,
7009 2 : Duration::ZERO,
7010 2 : &CancellationToken::new(),
7011 2 : &ctx,
7012 2 : )
7013 2 : .await?;
7014 2 :
7015 2 : // Keeping everything <= Lsn(0x80) b/c leases:
7016 2 : // 0/10: initdb layer
7017 2 : // (0/20..=0/70).step_by(0x10): image layers added when creating the timeline.
7018 2 : assert_eq!(res.layers_needed_by_leases, 7);
7019 2 : // Keeping 0/90 b/c it is the latest layer.
7020 2 : assert_eq!(res.layers_not_updated, 1);
7021 2 : // Removed 0/80.
7022 2 : assert_eq!(res.layers_removed, 1);
7023 2 :
7024 2 : // Make lease on a already GC-ed LSN.
7025 2 : // 0/80 does not have a valid lease + is below latest_gc_cutoff
7026 2 : assert!(Lsn(0x80) < *timeline.get_latest_gc_cutoff_lsn());
7027 2 : let res = timeline.make_lsn_lease(Lsn(0x80), timeline.get_lsn_lease_length(), &ctx);
7028 2 : assert!(res.is_err());
7029 2 :
7030 2 : // Should still be able to renew a currently valid lease
7031 2 : // Assumption: original lease to is still valid for 0/50.
7032 2 : let _ =
7033 2 : timeline.make_lsn_lease(Lsn(leased_lsns[1]), timeline.get_lsn_lease_length(), &ctx)?;
7034 2 :
7035 2 : Ok(())
7036 2 : }
7037 :
7038 : #[tokio::test]
7039 2 : async fn test_simple_bottom_most_compaction_deltas() -> anyhow::Result<()> {
7040 2 : let harness = TenantHarness::create("test_simple_bottom_most_compaction_deltas").await?;
7041 8 : let (tenant, ctx) = harness.load().await;
7042 2 :
7043 118 : fn get_key(id: u32) -> Key {
7044 118 : // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
7045 118 : let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
7046 118 : key.field6 = id;
7047 118 : key
7048 118 : }
7049 2 :
7050 2 : // We create
7051 2 : // - one bottom-most image layer,
7052 2 : // - a delta layer D1 crossing the GC horizon with data below and above the horizon,
7053 2 : // - a delta layer D2 crossing the GC horizon with data only below the horizon,
7054 2 : // - a delta layer D3 above the horizon.
7055 2 : //
7056 2 : // | D3 |
7057 2 : // | D1 |
7058 2 : // -| |-- gc horizon -----------------
7059 2 : // | | | D2 |
7060 2 : // --------- img layer ------------------
7061 2 : //
7062 2 : // What we should expact from this compaction is:
7063 2 : // | D3 |
7064 2 : // | Part of D1 |
7065 2 : // --------- img layer with D1+D2 at GC horizon------------------
7066 2 :
7067 2 : // img layer at 0x10
7068 2 : let img_layer = (0..10)
7069 20 : .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
7070 2 : .collect_vec();
7071 2 :
7072 2 : let delta1 = vec![
7073 2 : (
7074 2 : get_key(1),
7075 2 : Lsn(0x20),
7076 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
7077 2 : ),
7078 2 : (
7079 2 : get_key(2),
7080 2 : Lsn(0x30),
7081 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
7082 2 : ),
7083 2 : (
7084 2 : get_key(3),
7085 2 : Lsn(0x28),
7086 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x28")),
7087 2 : ),
7088 2 : (
7089 2 : get_key(3),
7090 2 : Lsn(0x30),
7091 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x30")),
7092 2 : ),
7093 2 : (
7094 2 : get_key(3),
7095 2 : Lsn(0x40),
7096 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x40")),
7097 2 : ),
7098 2 : ];
7099 2 : let delta2 = vec![
7100 2 : (
7101 2 : get_key(5),
7102 2 : Lsn(0x20),
7103 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
7104 2 : ),
7105 2 : (
7106 2 : get_key(6),
7107 2 : Lsn(0x20),
7108 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
7109 2 : ),
7110 2 : ];
7111 2 : let delta3 = vec![
7112 2 : (
7113 2 : get_key(8),
7114 2 : Lsn(0x48),
7115 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
7116 2 : ),
7117 2 : (
7118 2 : get_key(9),
7119 2 : Lsn(0x48),
7120 2 : Value::WalRecord(NeonWalRecord::wal_append("@0x48")),
7121 2 : ),
7122 2 : ];
7123 2 :
7124 2 : let tline = tenant
7125 2 : .create_test_timeline_with_layers(
7126 2 : TIMELINE_ID,
7127 2 : Lsn(0x10),
7128 2 : DEFAULT_PG_VERSION,
7129 2 : &ctx,
7130 2 : vec![
7131 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x48), delta1),
7132 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x10)..Lsn(0x48), delta2),
7133 2 : DeltaLayerTestDesc::new_with_inferred_key_range(Lsn(0x48)..Lsn(0x50), delta3),
7134 2 : ], // delta layers
7135 2 : vec![(Lsn(0x10), img_layer)], // image layers
7136 2 : Lsn(0x50),
7137 2 : )
7138 49 : .await?;
7139 2 : {
7140 2 : // Update GC info
7141 2 : let mut guard = tline.gc_info.write().unwrap();
7142 2 : *guard = GcInfo {
7143 2 : retain_lsns: vec![],
7144 2 : cutoffs: GcCutoffs {
7145 2 : time: Lsn(0x30),
7146 2 : space: Lsn(0x30),
7147 2 : },
7148 2 : leases: Default::default(),
7149 2 : within_ancestor_pitr: false,
7150 2 : };
7151 2 : }
7152 2 :
7153 2 : let expected_result = [
7154 2 : Bytes::from_static(b"value 0@0x10"),
7155 2 : Bytes::from_static(b"value 1@0x10@0x20"),
7156 2 : Bytes::from_static(b"value 2@0x10@0x30"),
7157 2 : Bytes::from_static(b"value 3@0x10@0x28@0x30@0x40"),
7158 2 : Bytes::from_static(b"value 4@0x10"),
7159 2 : Bytes::from_static(b"value 5@0x10@0x20"),
7160 2 : Bytes::from_static(b"value 6@0x10@0x20"),
7161 2 : Bytes::from_static(b"value 7@0x10"),
7162 2 : Bytes::from_static(b"value 8@0x10@0x48"),
7163 2 : Bytes::from_static(b"value 9@0x10@0x48"),
7164 2 : ];
7165 2 :
7166 2 : let expected_result_at_gc_horizon = [
7167 2 : Bytes::from_static(b"value 0@0x10"),
7168 2 : Bytes::from_static(b"value 1@0x10@0x20"),
7169 2 : Bytes::from_static(b"value 2@0x10@0x30"),
7170 2 : Bytes::from_static(b"value 3@0x10@0x28@0x30"),
7171 2 : Bytes::from_static(b"value 4@0x10"),
7172 2 : Bytes::from_static(b"value 5@0x10@0x20"),
7173 2 : Bytes::from_static(b"value 6@0x10@0x20"),
7174 2 : Bytes::from_static(b"value 7@0x10"),
7175 2 : Bytes::from_static(b"value 8@0x10"),
7176 2 : Bytes::from_static(b"value 9@0x10"),
7177 2 : ];
7178 2 :
7179 22 : for idx in 0..10 {
7180 20 : assert_eq!(
7181 20 : tline
7182 20 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
7183 17 : .await
7184 20 : .unwrap(),
7185 20 : &expected_result[idx]
7186 2 : );
7187 20 : assert_eq!(
7188 20 : tline
7189 20 : .get(get_key(idx as u32), Lsn(0x30), &ctx)
7190 2 : .await
7191 20 : .unwrap(),
7192 20 : &expected_result_at_gc_horizon[idx]
7193 2 : );
7194 2 : }
7195 2 :
7196 2 : let cancel = CancellationToken::new();
7197 55 : tline.compact_with_gc(&cancel, &ctx).await.unwrap();
7198 2 :
7199 22 : for idx in 0..10 {
7200 20 : assert_eq!(
7201 20 : tline
7202 20 : .get(get_key(idx as u32), Lsn(0x50), &ctx)
7203 8 : .await
7204 20 : .unwrap(),
7205 20 : &expected_result[idx]
7206 2 : );
7207 20 : assert_eq!(
7208 20 : tline
7209 20 : .get(get_key(idx as u32), Lsn(0x30), &ctx)
7210 2 : .await
7211 20 : .unwrap(),
7212 20 : &expected_result_at_gc_horizon[idx]
7213 2 : );
7214 2 : }
7215 2 :
7216 2 : Ok(())
7217 2 : }
7218 : }
|
