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 camino::Utf8Path;
16 : use camino::Utf8PathBuf;
17 : use enumset::EnumSet;
18 : use futures::stream::FuturesUnordered;
19 : use futures::FutureExt;
20 : use futures::StreamExt;
21 : use pageserver_api::models;
22 : use pageserver_api::models::TimelineState;
23 : use pageserver_api::models::WalRedoManagerStatus;
24 : use pageserver_api::shard::ShardIdentity;
25 : use pageserver_api::shard::TenantShardId;
26 : use remote_storage::DownloadError;
27 : use remote_storage::GenericRemoteStorage;
28 : use remote_storage::TimeoutOrCancel;
29 : use std::fmt;
30 : use storage_broker::BrokerClientChannel;
31 : use tokio::io::BufReader;
32 : use tokio::sync::watch;
33 : use tokio::task::JoinSet;
34 : use tokio_util::sync::CancellationToken;
35 : use tracing::*;
36 : use utils::backoff;
37 : use utils::completion;
38 : use utils::crashsafe::path_with_suffix_extension;
39 : use utils::failpoint_support;
40 : use utils::fs_ext;
41 : use utils::sync::gate::Gate;
42 : use utils::sync::gate::GateGuard;
43 : use utils::timeout::timeout_cancellable;
44 : use utils::timeout::TimeoutCancellableError;
45 :
46 : use self::config::AttachedLocationConfig;
47 : use self::config::AttachmentMode;
48 : use self::config::LocationConf;
49 : use self::config::TenantConf;
50 : use self::delete::DeleteTenantFlow;
51 : use self::metadata::TimelineMetadata;
52 : use self::mgr::GetActiveTenantError;
53 : use self::mgr::GetTenantError;
54 : use self::mgr::TenantsMap;
55 : use self::remote_timeline_client::upload::upload_index_part;
56 : use self::remote_timeline_client::RemoteTimelineClient;
57 : use self::timeline::uninit::TimelineExclusionError;
58 : use self::timeline::uninit::TimelineUninitMark;
59 : use self::timeline::uninit::UninitializedTimeline;
60 : use self::timeline::EvictionTaskTenantState;
61 : use self::timeline::TimelineResources;
62 : use self::timeline::WaitLsnError;
63 : use crate::config::PageServerConf;
64 : use crate::context::{DownloadBehavior, RequestContext};
65 : use crate::deletion_queue::DeletionQueueClient;
66 : use crate::deletion_queue::DeletionQueueError;
67 : use crate::import_datadir;
68 : use crate::is_uninit_mark;
69 : use crate::metrics::TENANT;
70 : use crate::metrics::{
71 : remove_tenant_metrics, BROKEN_TENANTS_SET, TENANT_STATE_METRIC, TENANT_SYNTHETIC_SIZE_METRIC,
72 : };
73 : use crate::repository::GcResult;
74 : use crate::task_mgr;
75 : use crate::task_mgr::TaskKind;
76 : use crate::tenant::config::LocationMode;
77 : use crate::tenant::config::TenantConfOpt;
78 : pub use crate::tenant::remote_timeline_client::index::IndexPart;
79 : use crate::tenant::remote_timeline_client::remote_initdb_archive_path;
80 : use crate::tenant::remote_timeline_client::MaybeDeletedIndexPart;
81 : use crate::tenant::remote_timeline_client::INITDB_PATH;
82 : use crate::tenant::storage_layer::DeltaLayer;
83 : use crate::tenant::storage_layer::ImageLayer;
84 : use crate::InitializationOrder;
85 : use std::cmp::min;
86 : use std::collections::hash_map::Entry;
87 : use std::collections::BTreeSet;
88 : use std::collections::HashMap;
89 : use std::collections::HashSet;
90 : use std::fmt::Debug;
91 : use std::fmt::Display;
92 : use std::fs;
93 : use std::fs::File;
94 : use std::ops::Bound::Included;
95 : use std::sync::atomic::AtomicU64;
96 : use std::sync::atomic::Ordering;
97 : use std::sync::Arc;
98 : use std::sync::{Mutex, RwLock};
99 : use std::time::{Duration, Instant};
100 :
101 : use crate::span;
102 : use crate::tenant::timeline::delete::DeleteTimelineFlow;
103 : use crate::tenant::timeline::uninit::cleanup_timeline_directory;
104 : use crate::virtual_file::VirtualFile;
105 : use crate::walredo::PostgresRedoManager;
106 : use crate::TEMP_FILE_SUFFIX;
107 : use once_cell::sync::Lazy;
108 : pub use pageserver_api::models::TenantState;
109 : use tokio::sync::Semaphore;
110 :
111 0 : static INIT_DB_SEMAPHORE: Lazy<Semaphore> = Lazy::new(|| Semaphore::new(8));
112 : use toml_edit;
113 : use utils::{
114 : crashsafe,
115 : generation::Generation,
116 : id::TimelineId,
117 : lsn::{Lsn, RecordLsn},
118 : };
119 :
120 : /// Declare a failpoint that can use the `pause` failpoint action.
121 : /// We don't want to block the executor thread, hence, spawn_blocking + await.
122 : macro_rules! pausable_failpoint {
123 : ($name:literal) => {
124 : if cfg!(feature = "testing") {
125 : tokio::task::spawn_blocking({
126 : let current = tracing::Span::current();
127 : move || {
128 : let _entered = current.entered();
129 : tracing::info!("at failpoint {}", $name);
130 : fail::fail_point!($name);
131 : }
132 : })
133 : .await
134 : .expect("spawn_blocking");
135 : }
136 : };
137 : ($name:literal, $cond:expr) => {
138 : if cfg!(feature = "testing") {
139 : if $cond {
140 : pausable_failpoint!($name)
141 : }
142 : }
143 : };
144 : }
145 :
146 : pub mod blob_io;
147 : pub mod block_io;
148 : pub mod vectored_blob_io;
149 :
150 : pub mod disk_btree;
151 : pub(crate) mod ephemeral_file;
152 : pub mod layer_map;
153 :
154 : pub mod metadata;
155 : mod par_fsync;
156 : pub mod remote_timeline_client;
157 : pub mod storage_layer;
158 :
159 : pub mod config;
160 : pub mod delete;
161 : pub mod mgr;
162 : pub mod secondary;
163 : pub mod tasks;
164 : pub mod upload_queue;
165 :
166 : pub(crate) mod timeline;
167 :
168 : pub mod size;
169 :
170 : pub(crate) mod throttle;
171 :
172 : pub(crate) use crate::span::debug_assert_current_span_has_tenant_and_timeline_id;
173 : pub(crate) use timeline::{LogicalSizeCalculationCause, PageReconstructError, Timeline};
174 :
175 : // re-export for use in walreceiver
176 : pub use crate::tenant::timeline::WalReceiverInfo;
177 :
178 : /// The "tenants" part of `tenants/<tenant>/timelines...`
179 : pub const TENANTS_SEGMENT_NAME: &str = "tenants";
180 :
181 : /// Parts of the `.neon/tenants/<tenant_id>/timelines/<timeline_id>` directory prefix.
182 : pub const TIMELINES_SEGMENT_NAME: &str = "timelines";
183 :
184 : pub const TENANT_DELETED_MARKER_FILE_NAME: &str = "deleted";
185 :
186 : /// References to shared objects that are passed into each tenant, such
187 : /// as the shared remote storage client and process initialization state.
188 0 : #[derive(Clone)]
189 : pub struct TenantSharedResources {
190 : pub broker_client: storage_broker::BrokerClientChannel,
191 : pub remote_storage: Option<GenericRemoteStorage>,
192 : pub deletion_queue_client: DeletionQueueClient,
193 : }
194 :
195 : /// A [`Tenant`] is really an _attached_ tenant. The configuration
196 : /// for an attached tenant is a subset of the [`LocationConf`], represented
197 : /// in this struct.
198 : pub(super) struct AttachedTenantConf {
199 : tenant_conf: TenantConfOpt,
200 : location: AttachedLocationConfig,
201 : }
202 :
203 : impl AttachedTenantConf {
204 88 : fn try_from(location_conf: LocationConf) -> anyhow::Result<Self> {
205 88 : match &location_conf.mode {
206 88 : LocationMode::Attached(attach_conf) => Ok(Self {
207 88 : tenant_conf: location_conf.tenant_conf,
208 88 : location: *attach_conf,
209 88 : }),
210 : LocationMode::Secondary(_) => {
211 0 : anyhow::bail!("Attempted to construct AttachedTenantConf from a LocationConf in secondary mode")
212 : }
213 : }
214 88 : }
215 : }
216 : struct TimelinePreload {
217 : timeline_id: TimelineId,
218 : client: RemoteTimelineClient,
219 : index_part: Result<MaybeDeletedIndexPart, DownloadError>,
220 : }
221 :
222 : pub(crate) struct TenantPreload {
223 : deleting: bool,
224 : timelines: HashMap<TimelineId, TimelinePreload>,
225 : }
226 :
227 : /// When we spawn a tenant, there is a special mode for tenant creation that
228 : /// avoids trying to read anything from remote storage.
229 : pub(crate) enum SpawnMode {
230 : Normal,
231 : Create,
232 : }
233 :
234 : ///
235 : /// Tenant consists of multiple timelines. Keep them in a hash table.
236 : ///
237 : pub struct Tenant {
238 : // Global pageserver config parameters
239 : pub conf: &'static PageServerConf,
240 :
241 : /// The value creation timestamp, used to measure activation delay, see:
242 : /// <https://github.com/neondatabase/neon/issues/4025>
243 : constructed_at: Instant,
244 :
245 : state: watch::Sender<TenantState>,
246 :
247 : // Overridden tenant-specific config parameters.
248 : // We keep TenantConfOpt sturct here to preserve the information
249 : // about parameters that are not set.
250 : // This is necessary to allow global config updates.
251 : tenant_conf: Arc<RwLock<AttachedTenantConf>>,
252 :
253 : tenant_shard_id: TenantShardId,
254 :
255 : // The detailed sharding information, beyond the number/count in tenant_shard_id
256 : shard_identity: ShardIdentity,
257 :
258 : /// The remote storage generation, used to protect S3 objects from split-brain.
259 : /// Does not change over the lifetime of the [`Tenant`] object.
260 : ///
261 : /// This duplicates the generation stored in LocationConf, but that structure is mutable:
262 : /// this copy enforces the invariant that generatio doesn't change during a Tenant's lifetime.
263 : generation: Generation,
264 :
265 : timelines: Mutex<HashMap<TimelineId, Arc<Timeline>>>,
266 :
267 : /// During timeline creation, we first insert the TimelineId to the
268 : /// creating map, then `timelines`, then remove it from the creating map.
269 : /// **Lock order**: if acquring both, acquire`timelines` before `timelines_creating`
270 : timelines_creating: std::sync::Mutex<HashSet<TimelineId>>,
271 :
272 : // This mutex prevents creation of new timelines during GC.
273 : // Adding yet another mutex (in addition to `timelines`) is needed because holding
274 : // `timelines` mutex during all GC iteration
275 : // may block for a long time `get_timeline`, `get_timelines_state`,... and other operations
276 : // with timelines, which in turn may cause dropping replication connection, expiration of wait_for_lsn
277 : // timeout...
278 : gc_cs: tokio::sync::Mutex<()>,
279 : walredo_mgr: Option<Arc<WalRedoManager>>,
280 :
281 : // provides access to timeline data sitting in the remote storage
282 : pub(crate) remote_storage: Option<GenericRemoteStorage>,
283 :
284 : // Access to global deletion queue for when this tenant wants to schedule a deletion
285 : deletion_queue_client: DeletionQueueClient,
286 :
287 : /// Cached logical sizes updated updated on each [`Tenant::gather_size_inputs`].
288 : cached_logical_sizes: tokio::sync::Mutex<HashMap<(TimelineId, Lsn), u64>>,
289 : cached_synthetic_tenant_size: Arc<AtomicU64>,
290 :
291 : eviction_task_tenant_state: tokio::sync::Mutex<EvictionTaskTenantState>,
292 :
293 : /// If the tenant is in Activating state, notify this to encourage it
294 : /// to proceed to Active as soon as possible, rather than waiting for lazy
295 : /// background warmup.
296 : pub(crate) activate_now_sem: tokio::sync::Semaphore,
297 :
298 : pub(crate) delete_progress: Arc<tokio::sync::Mutex<DeleteTenantFlow>>,
299 :
300 : // Cancellation token fires when we have entered shutdown(). This is a parent of
301 : // Timelines' cancellation token.
302 : pub(crate) cancel: CancellationToken,
303 :
304 : // Users of the Tenant such as the page service must take this Gate to avoid
305 : // trying to use a Tenant which is shutting down.
306 : pub(crate) gate: Gate,
307 :
308 : /// Throttle applied at the top of [`Timeline::get`].
309 : /// All [`Tenant::timelines`] of a given [`Tenant`] instance share the same [`throttle::Throttle`] instance.
310 : pub(crate) timeline_get_throttle:
311 : Arc<throttle::Throttle<&'static crate::metrics::tenant_throttling::TimelineGet>>,
312 : }
313 :
314 : impl std::fmt::Debug for Tenant {
315 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
316 0 : write!(f, "{} ({})", self.tenant_shard_id, self.current_state())
317 0 : }
318 : }
319 :
320 : pub(crate) enum WalRedoManager {
321 : Prod(PostgresRedoManager),
322 : #[cfg(test)]
323 : Test(harness::TestRedoManager),
324 : }
325 :
326 : impl From<PostgresRedoManager> for WalRedoManager {
327 0 : fn from(mgr: PostgresRedoManager) -> Self {
328 0 : Self::Prod(mgr)
329 0 : }
330 : }
331 :
332 : #[cfg(test)]
333 : impl From<harness::TestRedoManager> for WalRedoManager {
334 88 : fn from(mgr: harness::TestRedoManager) -> Self {
335 88 : Self::Test(mgr)
336 88 : }
337 : }
338 :
339 : impl WalRedoManager {
340 0 : pub(crate) fn maybe_quiesce(&self, idle_timeout: Duration) {
341 0 : match self {
342 0 : Self::Prod(mgr) => mgr.maybe_quiesce(idle_timeout),
343 0 : #[cfg(test)]
344 0 : Self::Test(_) => {
345 0 : // Not applicable to test redo manager
346 0 : }
347 0 : }
348 0 : }
349 :
350 : /// # Cancel-Safety
351 : ///
352 : /// This method is cancellation-safe.
353 6 : pub async fn request_redo(
354 6 : &self,
355 6 : key: crate::repository::Key,
356 6 : lsn: Lsn,
357 6 : base_img: Option<(Lsn, bytes::Bytes)>,
358 6 : records: Vec<(Lsn, crate::walrecord::NeonWalRecord)>,
359 6 : pg_version: u32,
360 6 : ) -> anyhow::Result<bytes::Bytes> {
361 6 : match self {
362 0 : Self::Prod(mgr) => {
363 0 : mgr.request_redo(key, lsn, base_img, records, pg_version)
364 0 : .await
365 : }
366 : #[cfg(test)]
367 6 : Self::Test(mgr) => {
368 6 : mgr.request_redo(key, lsn, base_img, records, pg_version)
369 0 : .await
370 : }
371 : }
372 6 : }
373 :
374 0 : pub(crate) fn status(&self) -> Option<WalRedoManagerStatus> {
375 0 : match self {
376 0 : WalRedoManager::Prod(m) => m.status(),
377 0 : #[cfg(test)]
378 0 : WalRedoManager::Test(_) => None,
379 0 : }
380 0 : }
381 : }
382 :
383 2 : #[derive(Debug, thiserror::Error, PartialEq, Eq)]
384 : pub enum GetTimelineError {
385 : #[error("Timeline {tenant_id}/{timeline_id} is not active, state: {state:?}")]
386 : NotActive {
387 : tenant_id: TenantShardId,
388 : timeline_id: TimelineId,
389 : state: TimelineState,
390 : },
391 : #[error("Timeline {tenant_id}/{timeline_id} was not found")]
392 : NotFound {
393 : tenant_id: TenantShardId,
394 : timeline_id: TimelineId,
395 : },
396 : }
397 :
398 0 : #[derive(Debug, thiserror::Error)]
399 : pub enum LoadLocalTimelineError {
400 : #[error("FailedToLoad")]
401 : Load(#[source] anyhow::Error),
402 : #[error("FailedToResumeDeletion")]
403 : ResumeDeletion(#[source] anyhow::Error),
404 : }
405 :
406 0 : #[derive(thiserror::Error)]
407 : pub enum DeleteTimelineError {
408 : #[error("NotFound")]
409 : NotFound,
410 :
411 : #[error("HasChildren")]
412 : HasChildren(Vec<TimelineId>),
413 :
414 : #[error("Timeline deletion is already in progress")]
415 : AlreadyInProgress(Arc<tokio::sync::Mutex<DeleteTimelineFlow>>),
416 :
417 : #[error(transparent)]
418 : Other(#[from] anyhow::Error),
419 : }
420 :
421 : impl Debug for DeleteTimelineError {
422 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
423 0 : match self {
424 0 : Self::NotFound => write!(f, "NotFound"),
425 0 : Self::HasChildren(c) => f.debug_tuple("HasChildren").field(c).finish(),
426 0 : Self::AlreadyInProgress(_) => f.debug_tuple("AlreadyInProgress").finish(),
427 0 : Self::Other(e) => f.debug_tuple("Other").field(e).finish(),
428 : }
429 0 : }
430 : }
431 :
432 : pub enum SetStoppingError {
433 : AlreadyStopping(completion::Barrier),
434 : Broken,
435 : }
436 :
437 : impl Debug for SetStoppingError {
438 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
439 0 : match self {
440 0 : Self::AlreadyStopping(_) => f.debug_tuple("AlreadyStopping").finish(),
441 0 : Self::Broken => write!(f, "Broken"),
442 : }
443 0 : }
444 : }
445 :
446 0 : #[derive(thiserror::Error, Debug)]
447 : pub enum CreateTimelineError {
448 : #[error("creation of timeline with the given ID is in progress")]
449 : AlreadyCreating,
450 : #[error("timeline already exists with different parameters")]
451 : Conflict,
452 : #[error(transparent)]
453 : AncestorLsn(anyhow::Error),
454 : #[error("ancestor timeline is not active")]
455 : AncestorNotActive,
456 : #[error("tenant shutting down")]
457 : ShuttingDown,
458 : #[error(transparent)]
459 : Other(#[from] anyhow::Error),
460 : }
461 :
462 0 : #[derive(thiserror::Error, Debug)]
463 : enum InitdbError {
464 : Other(anyhow::Error),
465 : Cancelled,
466 : Spawn(std::io::Result<()>),
467 : Failed(std::process::ExitStatus, Vec<u8>),
468 : }
469 :
470 : impl fmt::Display for InitdbError {
471 0 : fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
472 0 : match self {
473 0 : InitdbError::Cancelled => write!(f, "Operation was cancelled"),
474 0 : InitdbError::Spawn(e) => write!(f, "Spawn error: {:?}", e),
475 0 : InitdbError::Failed(status, stderr) => write!(
476 0 : f,
477 0 : "Command failed with status {:?}: {}",
478 0 : status,
479 0 : String::from_utf8_lossy(stderr)
480 0 : ),
481 0 : InitdbError::Other(e) => write!(f, "Error: {:?}", e),
482 : }
483 0 : }
484 : }
485 :
486 : impl From<std::io::Error> for InitdbError {
487 0 : fn from(error: std::io::Error) -> Self {
488 0 : InitdbError::Spawn(Err(error))
489 0 : }
490 : }
491 :
492 : enum CreateTimelineCause {
493 : Load,
494 : Delete,
495 : }
496 :
497 : impl Tenant {
498 : /// Yet another helper for timeline initialization.
499 : ///
500 : /// - Initializes the Timeline struct and inserts it into the tenant's hash map
501 : /// - Scans the local timeline directory for layer files and builds the layer map
502 : /// - Downloads remote index file and adds remote files to the layer map
503 : /// - Schedules remote upload tasks for any files that are present locally but missing from remote storage.
504 : ///
505 : /// If the operation fails, the timeline is left in the tenant's hash map in Broken state. On success,
506 : /// it is marked as Active.
507 : #[allow(clippy::too_many_arguments)]
508 6 : async fn timeline_init_and_sync(
509 6 : &self,
510 6 : timeline_id: TimelineId,
511 6 : resources: TimelineResources,
512 6 : index_part: Option<IndexPart>,
513 6 : metadata: TimelineMetadata,
514 6 : ancestor: Option<Arc<Timeline>>,
515 6 : _ctx: &RequestContext,
516 6 : ) -> anyhow::Result<()> {
517 6 : let tenant_id = self.tenant_shard_id;
518 :
519 6 : let timeline = self.create_timeline_struct(
520 6 : timeline_id,
521 6 : &metadata,
522 6 : ancestor.clone(),
523 6 : resources,
524 6 : CreateTimelineCause::Load,
525 6 : )?;
526 6 : let disk_consistent_lsn = timeline.get_disk_consistent_lsn();
527 6 : anyhow::ensure!(
528 6 : disk_consistent_lsn.is_valid(),
529 0 : "Timeline {tenant_id}/{timeline_id} has invalid disk_consistent_lsn"
530 : );
531 6 : assert_eq!(
532 6 : disk_consistent_lsn,
533 6 : metadata.disk_consistent_lsn(),
534 0 : "these are used interchangeably"
535 : );
536 :
537 6 : if let Some(index_part) = index_part.as_ref() {
538 6 : timeline
539 6 : .remote_client
540 6 : .as_ref()
541 6 : .unwrap()
542 6 : .init_upload_queue(index_part)?;
543 0 : } else if self.remote_storage.is_some() {
544 : // No data on the remote storage, but we have local metadata file. We can end up
545 : // here with timeline_create being interrupted before finishing index part upload.
546 : // By doing what we do here, the index part upload is retried.
547 : // If control plane retries timeline creation in the meantime, the mgmt API handler
548 : // for timeline creation will coalesce on the upload we queue here.
549 0 : let rtc = timeline.remote_client.as_ref().unwrap();
550 0 : rtc.init_upload_queue_for_empty_remote(&metadata)?;
551 0 : rtc.schedule_index_upload_for_metadata_update(&metadata)?;
552 0 : }
553 :
554 6 : timeline
555 6 : .load_layer_map(disk_consistent_lsn, index_part)
556 6 : .await
557 6 : .with_context(|| {
558 0 : format!("Failed to load layermap for timeline {tenant_id}/{timeline_id}")
559 6 : })?;
560 :
561 : {
562 : // avoiding holding it across awaits
563 6 : let mut timelines_accessor = self.timelines.lock().unwrap();
564 6 : match timelines_accessor.entry(timeline_id) {
565 : Entry::Occupied(_) => {
566 : // The uninit mark file acts as a lock that prevents another task from
567 : // initializing the timeline at the same time.
568 0 : unreachable!(
569 0 : "Timeline {tenant_id}/{timeline_id} already exists in the tenant map"
570 0 : );
571 : }
572 6 : Entry::Vacant(v) => {
573 6 : v.insert(Arc::clone(&timeline));
574 6 : timeline.maybe_spawn_flush_loop();
575 6 : }
576 6 : }
577 6 : };
578 6 :
579 6 : // Sanity check: a timeline should have some content.
580 6 : anyhow::ensure!(
581 6 : ancestor.is_some()
582 4 : || timeline
583 4 : .layers
584 4 : .read()
585 0 : .await
586 4 : .layer_map()
587 4 : .iter_historic_layers()
588 4 : .next()
589 4 : .is_some(),
590 0 : "Timeline has no ancestor and no layer files"
591 : );
592 :
593 6 : Ok(())
594 6 : }
595 :
596 : /// Attach a tenant that's available in cloud storage.
597 : ///
598 : /// This returns quickly, after just creating the in-memory object
599 : /// Tenant struct and launching a background task to download
600 : /// the remote index files. On return, the tenant is most likely still in
601 : /// Attaching state, and it will become Active once the background task
602 : /// finishes. You can use wait_until_active() to wait for the task to
603 : /// complete.
604 : ///
605 : #[allow(clippy::too_many_arguments)]
606 0 : pub(crate) fn spawn(
607 0 : conf: &'static PageServerConf,
608 0 : tenant_shard_id: TenantShardId,
609 0 : resources: TenantSharedResources,
610 0 : attached_conf: AttachedTenantConf,
611 0 : shard_identity: ShardIdentity,
612 0 : init_order: Option<InitializationOrder>,
613 0 : tenants: &'static std::sync::RwLock<TenantsMap>,
614 0 : mode: SpawnMode,
615 0 : ctx: &RequestContext,
616 0 : ) -> anyhow::Result<Arc<Tenant>> {
617 0 : let wal_redo_manager = Arc::new(WalRedoManager::from(PostgresRedoManager::new(
618 0 : conf,
619 0 : tenant_shard_id,
620 0 : )));
621 0 :
622 0 : let TenantSharedResources {
623 0 : broker_client,
624 0 : remote_storage,
625 0 : deletion_queue_client,
626 0 : } = resources;
627 0 :
628 0 : let attach_mode = attached_conf.location.attach_mode;
629 0 : let generation = attached_conf.location.generation;
630 0 :
631 0 : let tenant = Arc::new(Tenant::new(
632 0 : TenantState::Attaching,
633 0 : conf,
634 0 : attached_conf,
635 0 : shard_identity,
636 0 : Some(wal_redo_manager),
637 0 : tenant_shard_id,
638 0 : remote_storage.clone(),
639 0 : deletion_queue_client,
640 0 : ));
641 0 :
642 0 : // The attach task will carry a GateGuard, so that shutdown() reliably waits for it to drop out if
643 0 : // we shut down while attaching.
644 0 : let attach_gate_guard = tenant
645 0 : .gate
646 0 : .enter()
647 0 : .expect("We just created the Tenant: nothing else can have shut it down yet");
648 0 :
649 0 : // Do all the hard work in the background
650 0 : let tenant_clone = Arc::clone(&tenant);
651 0 : let ctx = ctx.detached_child(TaskKind::Attach, DownloadBehavior::Warn);
652 0 : task_mgr::spawn(
653 0 : &tokio::runtime::Handle::current(),
654 0 : TaskKind::Attach,
655 0 : Some(tenant_shard_id),
656 0 : None,
657 0 : "attach tenant",
658 : false,
659 0 : async move {
660 0 :
661 0 : info!(
662 0 : ?attach_mode,
663 0 : "Attaching tenant"
664 0 : );
665 :
666 0 : let _gate_guard = attach_gate_guard;
667 0 :
668 0 : // Is this tenant being spawned as part of process startup?
669 0 : let starting_up = init_order.is_some();
670 0 : scopeguard::defer! {
671 0 : if starting_up {
672 0 : TENANT.startup_complete.inc();
673 0 : }
674 : }
675 :
676 : // Ideally we should use Tenant::set_broken_no_wait, but it is not supposed to be used when tenant is in loading state.
677 0 : let make_broken =
678 0 : |t: &Tenant, err: anyhow::Error| {
679 0 : error!("attach failed, setting tenant state to Broken: {err:?}");
680 0 : t.state.send_modify(|state| {
681 0 : // The Stopping case is for when we have passed control on to DeleteTenantFlow:
682 0 : // if it errors, we will call make_broken when tenant is already in Stopping.
683 0 : assert!(
684 0 : matches!(*state, TenantState::Attaching | TenantState::Stopping { .. }),
685 0 : "the attach task owns the tenant state until activation is complete"
686 : );
687 :
688 0 : *state = TenantState::broken_from_reason(err.to_string());
689 0 : });
690 0 : };
691 :
692 0 : let mut init_order = init_order;
693 0 : // take the completion because initial tenant loading will complete when all of
694 0 : // these tasks complete.
695 0 : let _completion = init_order
696 0 : .as_mut()
697 0 : .and_then(|x| x.initial_tenant_load.take());
698 0 : let remote_load_completion = init_order
699 0 : .as_mut()
700 0 : .and_then(|x| x.initial_tenant_load_remote.take());
701 :
702 : enum AttachType<'a> {
703 : // During pageserver startup, we are attaching this tenant lazily in the background
704 : Warmup(tokio::sync::SemaphorePermit<'a>),
705 : // During pageserver startup, we are attaching this tenant as soon as we can,
706 : // because a client tried to access it.
707 : OnDemand,
708 : // During normal operations after startup, we are attaching a tenant.
709 : Normal,
710 : }
711 :
712 : // Before doing any I/O, wait for either or:
713 : // - A client to attempt to access to this tenant (on-demand loading)
714 : // - A permit to become available in the warmup semaphore (background warmup)
715 : //
716 : // Some-ness of init_order is how we know if we're attaching during startup or later
717 : // in process lifetime.
718 0 : let attach_type = if init_order.is_some() {
719 0 : tokio::select!(
720 : _ = tenant_clone.activate_now_sem.acquire() => {
721 0 : tracing::info!("Activating tenant (on-demand)");
722 : AttachType::OnDemand
723 : },
724 0 : permit_result = conf.concurrent_tenant_warmup.inner().acquire() => {
725 : match permit_result {
726 : Ok(p) => {
727 0 : tracing::info!("Activating tenant (warmup)");
728 : AttachType::Warmup(p)
729 : }
730 : Err(_) => {
731 : // This is unexpected: the warmup semaphore should stay alive
732 : // for the lifetime of init_order. Log a warning and proceed.
733 0 : tracing::warn!("warmup_limit semaphore unexpectedly closed");
734 : AttachType::Normal
735 : }
736 : }
737 :
738 : }
739 : _ = tenant_clone.cancel.cancelled() => {
740 : // This is safe, but should be pretty rare: it is interesting if a tenant
741 : // stayed in Activating for such a long time that shutdown found it in
742 : // that state.
743 0 : tracing::info!(state=%tenant_clone.current_state(), "Tenant shut down before activation");
744 : // Make the tenant broken so that set_stopping will not hang waiting for it to leave
745 : // the Attaching state. This is an over-reaction (nothing really broke, the tenant is
746 : // just shutting down), but ensures progress.
747 : make_broken(&tenant_clone, anyhow::anyhow!("Shut down while Attaching"));
748 : return Ok(());
749 : },
750 : )
751 : } else {
752 0 : AttachType::Normal
753 : };
754 :
755 0 : let preload = match (&mode, &remote_storage) {
756 : (SpawnMode::Create, _) => {
757 0 : None
758 : },
759 0 : (SpawnMode::Normal, Some(remote_storage)) => {
760 0 : let _preload_timer = TENANT.preload.start_timer();
761 0 : let res = tenant_clone
762 0 : .preload(remote_storage, task_mgr::shutdown_token())
763 0 : .await;
764 0 : match res {
765 0 : Ok(p) => Some(p),
766 0 : Err(e) => {
767 0 : make_broken(&tenant_clone, anyhow::anyhow!(e));
768 0 : return Ok(());
769 : }
770 : }
771 : }
772 : (SpawnMode::Normal, None) => {
773 0 : let _preload_timer = TENANT.preload.start_timer();
774 0 : None
775 : }
776 : };
777 :
778 : // Remote preload is complete.
779 0 : drop(remote_load_completion);
780 :
781 0 : let pending_deletion = {
782 0 : match DeleteTenantFlow::should_resume_deletion(
783 0 : conf,
784 0 : preload.as_ref().map(|p| p.deleting).unwrap_or(false),
785 0 : &tenant_clone,
786 0 : )
787 0 : .await
788 : {
789 0 : Ok(should_resume_deletion) => should_resume_deletion,
790 0 : Err(err) => {
791 0 : make_broken(&tenant_clone, anyhow::anyhow!(err));
792 0 : return Ok(());
793 : }
794 : }
795 : };
796 :
797 0 : info!("pending_deletion {}", pending_deletion.is_some());
798 :
799 0 : if let Some(deletion) = pending_deletion {
800 : // as we are no longer loading, signal completion by dropping
801 : // the completion while we resume deletion
802 0 : drop(_completion);
803 0 : let background_jobs_can_start =
804 0 : init_order.as_ref().map(|x| &x.background_jobs_can_start);
805 0 : if let Some(background) = background_jobs_can_start {
806 0 : info!("waiting for backgound jobs barrier");
807 0 : background.clone().wait().await;
808 0 : info!("ready for backgound jobs barrier");
809 0 : }
810 :
811 0 : let deleted = DeleteTenantFlow::resume_from_attach(
812 0 : deletion,
813 0 : &tenant_clone,
814 0 : preload,
815 0 : tenants,
816 0 : &ctx,
817 0 : )
818 0 : .await;
819 :
820 0 : if let Err(e) = deleted {
821 0 : make_broken(&tenant_clone, anyhow::anyhow!(e));
822 0 : }
823 :
824 0 : return Ok(());
825 0 : }
826 :
827 : // We will time the duration of the attach phase unless this is a creation (attach will do no work)
828 0 : let attached = {
829 0 : let _attach_timer = match mode {
830 0 : SpawnMode::Create => None,
831 0 : SpawnMode::Normal => {Some(TENANT.attach.start_timer())}
832 : };
833 0 : tenant_clone.attach(preload, mode, &ctx).await
834 : };
835 :
836 0 : match attached {
837 : Ok(()) => {
838 0 : info!("attach finished, activating");
839 0 : tenant_clone.activate(broker_client, None, &ctx);
840 : }
841 0 : Err(e) => {
842 0 : make_broken(&tenant_clone, anyhow::anyhow!(e));
843 0 : }
844 : }
845 :
846 : // If we are doing an opportunistic warmup attachment at startup, initialize
847 : // logical size at the same time. This is better than starting a bunch of idle tenants
848 : // with cold caches and then coming back later to initialize their logical sizes.
849 : //
850 : // It also prevents the warmup proccess competing with the concurrency limit on
851 : // logical size calculations: if logical size calculation semaphore is saturated,
852 : // then warmup will wait for that before proceeding to the next tenant.
853 0 : if let AttachType::Warmup(_permit) = attach_type {
854 0 : let mut futs: FuturesUnordered<_> = tenant_clone.timelines.lock().unwrap().values().cloned().map(|t| t.await_initial_logical_size()).collect();
855 0 : tracing::info!("Waiting for initial logical sizes while warming up...");
856 0 : while futs.next().await.is_some() {}
857 0 : tracing::info!("Warm-up complete");
858 0 : }
859 :
860 0 : Ok(())
861 0 : }
862 0 : .instrument(tracing::info_span!(parent: None, "attach", tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), gen=?generation)),
863 : );
864 0 : Ok(tenant)
865 0 : }
866 :
867 176 : #[instrument(skip_all)]
868 : pub(crate) async fn preload(
869 : self: &Arc<Tenant>,
870 : remote_storage: &GenericRemoteStorage,
871 : cancel: CancellationToken,
872 : ) -> anyhow::Result<TenantPreload> {
873 : span::debug_assert_current_span_has_tenant_id();
874 : // Get list of remote timelines
875 : // download index files for every tenant timeline
876 88 : info!("listing remote timelines");
877 : let (remote_timeline_ids, other_keys) = remote_timeline_client::list_remote_timelines(
878 : remote_storage,
879 : self.tenant_shard_id,
880 : cancel.clone(),
881 : )
882 : .await?;
883 :
884 : let deleting = other_keys.contains(TENANT_DELETED_MARKER_FILE_NAME);
885 88 : info!(
886 88 : "found {} timelines, deleting={}",
887 88 : remote_timeline_ids.len(),
888 88 : deleting
889 88 : );
890 :
891 : for k in other_keys {
892 : if k != TENANT_DELETED_MARKER_FILE_NAME {
893 0 : warn!("Unexpected non timeline key {k}");
894 : }
895 : }
896 :
897 : Ok(TenantPreload {
898 : deleting,
899 : timelines: self
900 : .load_timeline_metadata(remote_timeline_ids, remote_storage, cancel)
901 : .await?,
902 : })
903 : }
904 :
905 : ///
906 : /// Background task that downloads all data for a tenant and brings it to Active state.
907 : ///
908 : /// No background tasks are started as part of this routine.
909 : ///
910 88 : async fn attach(
911 88 : self: &Arc<Tenant>,
912 88 : preload: Option<TenantPreload>,
913 88 : mode: SpawnMode,
914 88 : ctx: &RequestContext,
915 88 : ) -> anyhow::Result<()> {
916 88 : span::debug_assert_current_span_has_tenant_id();
917 :
918 0 : failpoint_support::sleep_millis_async!("before-attaching-tenant");
919 :
920 88 : let preload = match (preload, mode) {
921 88 : (Some(p), _) => p,
922 0 : (None, SpawnMode::Create) => TenantPreload {
923 0 : deleting: false,
924 0 : timelines: HashMap::new(),
925 0 : },
926 : (None, SpawnMode::Normal) => {
927 0 : anyhow::bail!("local-only deployment is no longer supported, https://github.com/neondatabase/neon/issues/5624");
928 : }
929 : };
930 :
931 88 : let mut timelines_to_resume_deletions = vec![];
932 88 :
933 88 : let mut remote_index_and_client = HashMap::new();
934 88 : let mut timeline_ancestors = HashMap::new();
935 88 : let mut existent_timelines = HashSet::new();
936 94 : for (timeline_id, preload) in preload.timelines {
937 6 : let index_part = match preload.index_part {
938 6 : Ok(i) => {
939 0 : debug!("remote index part exists for timeline {timeline_id}");
940 : // We found index_part on the remote, this is the standard case.
941 6 : existent_timelines.insert(timeline_id);
942 6 : i
943 : }
944 : Err(DownloadError::NotFound) => {
945 : // There is no index_part on the remote. We only get here
946 : // if there is some prefix for the timeline in the remote storage.
947 : // This can e.g. be the initdb.tar.zst archive, maybe a
948 : // remnant from a prior incomplete creation or deletion attempt.
949 : // Delete the local directory as the deciding criterion for a
950 : // timeline's existence is presence of index_part.
951 0 : info!(%timeline_id, "index_part not found on remote");
952 0 : continue;
953 : }
954 0 : Err(e) => {
955 : // Some (possibly ephemeral) error happened during index_part download.
956 : // Pretend the timeline exists to not delete the timeline directory,
957 : // as it might be a temporary issue and we don't want to re-download
958 : // everything after it resolves.
959 0 : warn!(%timeline_id, "Failed to load index_part from remote storage, failed creation? ({e})");
960 :
961 0 : existent_timelines.insert(timeline_id);
962 0 : continue;
963 : }
964 : };
965 6 : match index_part {
966 6 : MaybeDeletedIndexPart::IndexPart(index_part) => {
967 6 : timeline_ancestors.insert(timeline_id, index_part.metadata.clone());
968 6 : remote_index_and_client.insert(timeline_id, (index_part, preload.client));
969 6 : }
970 0 : MaybeDeletedIndexPart::Deleted(index_part) => {
971 0 : info!(
972 0 : "timeline {} is deleted, picking to resume deletion",
973 0 : timeline_id
974 0 : );
975 0 : timelines_to_resume_deletions.push((timeline_id, index_part, preload.client));
976 : }
977 : }
978 : }
979 :
980 : // For every timeline, download the metadata file, scan the local directory,
981 : // and build a layer map that contains an entry for each remote and local
982 : // layer file.
983 88 : let sorted_timelines = tree_sort_timelines(timeline_ancestors, |m| m.ancestor_timeline())?;
984 94 : for (timeline_id, remote_metadata) in sorted_timelines {
985 6 : let (index_part, remote_client) = remote_index_and_client
986 6 : .remove(&timeline_id)
987 6 : .expect("just put it in above");
988 6 :
989 6 : // TODO again handle early failure
990 6 : self.load_remote_timeline(
991 6 : timeline_id,
992 6 : index_part,
993 6 : remote_metadata,
994 6 : TimelineResources {
995 6 : remote_client: Some(remote_client),
996 6 : deletion_queue_client: self.deletion_queue_client.clone(),
997 6 : timeline_get_throttle: self.timeline_get_throttle.clone(),
998 6 : },
999 6 : ctx,
1000 6 : )
1001 12 : .await
1002 6 : .with_context(|| {
1003 0 : format!(
1004 0 : "failed to load remote timeline {} for tenant {}",
1005 0 : timeline_id, self.tenant_shard_id
1006 0 : )
1007 6 : })?;
1008 : }
1009 :
1010 : // Walk through deleted timelines, resume deletion
1011 88 : for (timeline_id, index_part, remote_timeline_client) in timelines_to_resume_deletions {
1012 0 : remote_timeline_client
1013 0 : .init_upload_queue_stopped_to_continue_deletion(&index_part)
1014 0 : .context("init queue stopped")
1015 0 : .map_err(LoadLocalTimelineError::ResumeDeletion)?;
1016 :
1017 0 : DeleteTimelineFlow::resume_deletion(
1018 0 : Arc::clone(self),
1019 0 : timeline_id,
1020 0 : &index_part.metadata,
1021 0 : Some(remote_timeline_client),
1022 0 : self.deletion_queue_client.clone(),
1023 0 : )
1024 0 : .instrument(tracing::info_span!("timeline_delete", %timeline_id))
1025 0 : .await
1026 0 : .context("resume_deletion")
1027 0 : .map_err(LoadLocalTimelineError::ResumeDeletion)?;
1028 : }
1029 :
1030 : // The local filesystem contents are a cache of what's in the remote IndexPart;
1031 : // IndexPart is the source of truth.
1032 88 : self.clean_up_timelines(&existent_timelines)?;
1033 :
1034 88 : fail::fail_point!("attach-before-activate", |_| {
1035 0 : anyhow::bail!("attach-before-activate");
1036 88 : });
1037 0 : failpoint_support::sleep_millis_async!("attach-before-activate-sleep", &self.cancel);
1038 :
1039 88 : info!("Done");
1040 :
1041 88 : Ok(())
1042 88 : }
1043 :
1044 : /// Check for any local timeline directories that are temporary, or do not correspond to a
1045 : /// timeline that still exists: this can happen if we crashed during a deletion/creation, or
1046 : /// if a timeline was deleted while the tenant was attached to a different pageserver.
1047 88 : fn clean_up_timelines(&self, existent_timelines: &HashSet<TimelineId>) -> anyhow::Result<()> {
1048 88 : let timelines_dir = self.conf.timelines_path(&self.tenant_shard_id);
1049 :
1050 88 : let entries = match timelines_dir.read_dir_utf8() {
1051 88 : Ok(d) => d,
1052 0 : Err(e) => {
1053 0 : if e.kind() == std::io::ErrorKind::NotFound {
1054 0 : return Ok(());
1055 : } else {
1056 0 : return Err(e).context("list timelines directory for tenant");
1057 : }
1058 : }
1059 : };
1060 :
1061 98 : for entry in entries {
1062 10 : let entry = entry.context("read timeline dir entry")?;
1063 10 : let entry_path = entry.path();
1064 :
1065 10 : let purge = if crate::is_temporary(entry_path)
1066 : // TODO: uninit_mark isn't needed any more, since uninitialized timelines are already
1067 : // covered by the check that the timeline must exist in remote storage.
1068 10 : || is_uninit_mark(entry_path)
1069 8 : || crate::is_delete_mark(entry_path)
1070 : {
1071 2 : true
1072 : } else {
1073 8 : match TimelineId::try_from(entry_path.file_name()) {
1074 8 : Ok(i) => {
1075 8 : // Purge if the timeline ID does not exist in remote storage: remote storage is the authority.
1076 8 : !existent_timelines.contains(&i)
1077 : }
1078 0 : Err(e) => {
1079 0 : tracing::warn!(
1080 0 : "Unparseable directory in timelines directory: {entry_path}, ignoring ({e})"
1081 0 : );
1082 : // Do not purge junk: if we don't recognize it, be cautious and leave it for a human.
1083 0 : false
1084 : }
1085 : }
1086 : };
1087 :
1088 10 : if purge {
1089 4 : tracing::info!("Purging stale timeline dentry {entry_path}");
1090 4 : if let Err(e) = match entry.file_type() {
1091 4 : Ok(t) => if t.is_dir() {
1092 2 : std::fs::remove_dir_all(entry_path)
1093 : } else {
1094 2 : std::fs::remove_file(entry_path)
1095 : }
1096 4 : .or_else(fs_ext::ignore_not_found),
1097 0 : Err(e) => Err(e),
1098 : } {
1099 0 : tracing::warn!("Failed to purge stale timeline dentry {entry_path}: {e}");
1100 4 : }
1101 6 : }
1102 : }
1103 :
1104 88 : Ok(())
1105 88 : }
1106 :
1107 : /// Get sum of all remote timelines sizes
1108 : ///
1109 : /// This function relies on the index_part instead of listing the remote storage
1110 0 : pub fn remote_size(&self) -> u64 {
1111 0 : let mut size = 0;
1112 :
1113 0 : for timeline in self.list_timelines() {
1114 0 : if let Some(remote_client) = &timeline.remote_client {
1115 0 : size += remote_client.get_remote_physical_size();
1116 0 : }
1117 : }
1118 :
1119 0 : size
1120 0 : }
1121 :
1122 12 : #[instrument(skip_all, fields(timeline_id=%timeline_id))]
1123 : async fn load_remote_timeline(
1124 : &self,
1125 : timeline_id: TimelineId,
1126 : index_part: IndexPart,
1127 : remote_metadata: TimelineMetadata,
1128 : resources: TimelineResources,
1129 : ctx: &RequestContext,
1130 : ) -> anyhow::Result<()> {
1131 : span::debug_assert_current_span_has_tenant_id();
1132 :
1133 6 : info!("downloading index file for timeline {}", timeline_id);
1134 : tokio::fs::create_dir_all(self.conf.timeline_path(&self.tenant_shard_id, &timeline_id))
1135 : .await
1136 : .context("Failed to create new timeline directory")?;
1137 :
1138 : let ancestor = if let Some(ancestor_id) = remote_metadata.ancestor_timeline() {
1139 : let timelines = self.timelines.lock().unwrap();
1140 : Some(Arc::clone(timelines.get(&ancestor_id).ok_or_else(
1141 0 : || {
1142 0 : anyhow::anyhow!(
1143 0 : "cannot find ancestor timeline {ancestor_id} for timeline {timeline_id}"
1144 0 : )
1145 0 : },
1146 : )?))
1147 : } else {
1148 : None
1149 : };
1150 :
1151 : self.timeline_init_and_sync(
1152 : timeline_id,
1153 : resources,
1154 : Some(index_part),
1155 : remote_metadata,
1156 : ancestor,
1157 : ctx,
1158 : )
1159 : .await
1160 : }
1161 :
1162 : /// Create a placeholder Tenant object for a broken tenant
1163 0 : pub fn create_broken_tenant(
1164 0 : conf: &'static PageServerConf,
1165 0 : tenant_shard_id: TenantShardId,
1166 0 : reason: String,
1167 0 : ) -> Arc<Tenant> {
1168 0 : Arc::new(Tenant::new(
1169 0 : TenantState::Broken {
1170 0 : reason,
1171 0 : backtrace: String::new(),
1172 0 : },
1173 0 : conf,
1174 0 : AttachedTenantConf::try_from(LocationConf::default()).unwrap(),
1175 0 : // Shard identity isn't meaningful for a broken tenant: it's just a placeholder
1176 0 : // to occupy the slot for this TenantShardId.
1177 0 : ShardIdentity::broken(tenant_shard_id.shard_number, tenant_shard_id.shard_count),
1178 0 : None,
1179 0 : tenant_shard_id,
1180 0 : None,
1181 0 : DeletionQueueClient::broken(),
1182 0 : ))
1183 0 : }
1184 :
1185 88 : async fn load_timeline_metadata(
1186 88 : self: &Arc<Tenant>,
1187 88 : timeline_ids: HashSet<TimelineId>,
1188 88 : remote_storage: &GenericRemoteStorage,
1189 88 : cancel: CancellationToken,
1190 88 : ) -> anyhow::Result<HashMap<TimelineId, TimelinePreload>> {
1191 88 : let mut part_downloads = JoinSet::new();
1192 94 : for timeline_id in timeline_ids {
1193 6 : let client = RemoteTimelineClient::new(
1194 6 : remote_storage.clone(),
1195 6 : self.deletion_queue_client.clone(),
1196 6 : self.conf,
1197 6 : self.tenant_shard_id,
1198 6 : timeline_id,
1199 6 : self.generation,
1200 6 : );
1201 6 : let cancel_clone = cancel.clone();
1202 6 : part_downloads.spawn(
1203 6 : async move {
1204 6 : debug!("starting index part download");
1205 :
1206 18 : let index_part = client.download_index_file(&cancel_clone).await;
1207 :
1208 6 : debug!("finished index part download");
1209 :
1210 6 : Result::<_, anyhow::Error>::Ok(TimelinePreload {
1211 6 : client,
1212 6 : timeline_id,
1213 6 : index_part,
1214 6 : })
1215 6 : }
1216 6 : .map(move |res| {
1217 6 : res.with_context(|| format!("download index part for timeline {timeline_id}"))
1218 6 : })
1219 6 : .instrument(info_span!("download_index_part", %timeline_id)),
1220 : );
1221 : }
1222 :
1223 88 : let mut timeline_preloads: HashMap<TimelineId, TimelinePreload> = HashMap::new();
1224 :
1225 94 : loop {
1226 98 : tokio::select!(
1227 94 : next = part_downloads.join_next() => {
1228 : match next {
1229 : Some(result) => {
1230 : let preload_result = result.context("join preload task")?;
1231 : let preload = preload_result?;
1232 : timeline_preloads.insert(preload.timeline_id, preload);
1233 : },
1234 : None => {
1235 : break;
1236 : }
1237 : }
1238 : },
1239 : _ = cancel.cancelled() => {
1240 : anyhow::bail!("Cancelled while waiting for remote index download")
1241 : }
1242 94 : )
1243 94 : }
1244 :
1245 88 : Ok(timeline_preloads)
1246 88 : }
1247 :
1248 4 : pub(crate) fn tenant_shard_id(&self) -> TenantShardId {
1249 4 : self.tenant_shard_id
1250 4 : }
1251 :
1252 : /// Get Timeline handle for given Neon timeline ID.
1253 : /// This function is idempotent. It doesn't change internal state in any way.
1254 232 : pub fn get_timeline(
1255 232 : &self,
1256 232 : timeline_id: TimelineId,
1257 232 : active_only: bool,
1258 232 : ) -> Result<Arc<Timeline>, GetTimelineError> {
1259 232 : let timelines_accessor = self.timelines.lock().unwrap();
1260 232 : let timeline = timelines_accessor
1261 232 : .get(&timeline_id)
1262 232 : .ok_or(GetTimelineError::NotFound {
1263 232 : tenant_id: self.tenant_shard_id,
1264 232 : timeline_id,
1265 232 : })?;
1266 :
1267 230 : if active_only && !timeline.is_active() {
1268 0 : Err(GetTimelineError::NotActive {
1269 0 : tenant_id: self.tenant_shard_id,
1270 0 : timeline_id,
1271 0 : state: timeline.current_state(),
1272 0 : })
1273 : } else {
1274 230 : Ok(Arc::clone(timeline))
1275 : }
1276 232 : }
1277 :
1278 : /// Lists timelines the tenant contains.
1279 : /// Up to tenant's implementation to omit certain timelines that ar not considered ready for use.
1280 0 : pub fn list_timelines(&self) -> Vec<Arc<Timeline>> {
1281 0 : self.timelines
1282 0 : .lock()
1283 0 : .unwrap()
1284 0 : .values()
1285 0 : .map(Arc::clone)
1286 0 : .collect()
1287 0 : }
1288 :
1289 0 : pub fn list_timeline_ids(&self) -> Vec<TimelineId> {
1290 0 : self.timelines.lock().unwrap().keys().cloned().collect()
1291 0 : }
1292 :
1293 : /// This is used to create the initial 'main' timeline during bootstrapping,
1294 : /// or when importing a new base backup. The caller is expected to load an
1295 : /// initial image of the datadir to the new timeline after this.
1296 : ///
1297 : /// Until that happens, the on-disk state is invalid (disk_consistent_lsn=Lsn(0))
1298 : /// and the timeline will fail to load at a restart.
1299 : ///
1300 : /// That's why we add an uninit mark file, and wrap it together witht the Timeline
1301 : /// in-memory object into UninitializedTimeline.
1302 : /// Once the caller is done setting up the timeline, they should call
1303 : /// `UninitializedTimeline::initialize_with_lock` to remove the uninit mark.
1304 : ///
1305 : /// For tests, use `DatadirModification::init_empty_test_timeline` + `commit` to setup the
1306 : /// minimum amount of keys required to get a writable timeline.
1307 : /// (Without it, `put` might fail due to `repartition` failing.)
1308 80 : pub(crate) async fn create_empty_timeline(
1309 80 : &self,
1310 80 : new_timeline_id: TimelineId,
1311 80 : initdb_lsn: Lsn,
1312 80 : pg_version: u32,
1313 80 : _ctx: &RequestContext,
1314 80 : ) -> anyhow::Result<UninitializedTimeline> {
1315 80 : anyhow::ensure!(
1316 80 : self.is_active(),
1317 0 : "Cannot create empty timelines on inactive tenant"
1318 : );
1319 :
1320 80 : let timeline_uninit_mark = self.create_timeline_uninit_mark(new_timeline_id)?;
1321 78 : let new_metadata = TimelineMetadata::new(
1322 78 : // Initialize disk_consistent LSN to 0, The caller must import some data to
1323 78 : // make it valid, before calling finish_creation()
1324 78 : Lsn(0),
1325 78 : None,
1326 78 : None,
1327 78 : Lsn(0),
1328 78 : initdb_lsn,
1329 78 : initdb_lsn,
1330 78 : pg_version,
1331 78 : );
1332 78 : self.prepare_new_timeline(
1333 78 : new_timeline_id,
1334 78 : &new_metadata,
1335 78 : timeline_uninit_mark,
1336 78 : initdb_lsn,
1337 78 : None,
1338 78 : )
1339 0 : .await
1340 80 : }
1341 :
1342 : /// Helper for unit tests to create an empty timeline.
1343 : ///
1344 : /// The timeline is has state value `Active` but its background loops are not running.
1345 : // This makes the various functions which anyhow::ensure! for Active state work in tests.
1346 : // Our current tests don't need the background loops.
1347 : #[cfg(test)]
1348 72 : pub async fn create_test_timeline(
1349 72 : &self,
1350 72 : new_timeline_id: TimelineId,
1351 72 : initdb_lsn: Lsn,
1352 72 : pg_version: u32,
1353 72 : ctx: &RequestContext,
1354 72 : ) -> anyhow::Result<Arc<Timeline>> {
1355 72 : let uninit_tl = self
1356 72 : .create_empty_timeline(new_timeline_id, initdb_lsn, pg_version, ctx)
1357 0 : .await?;
1358 72 : let tline = uninit_tl.raw_timeline().expect("we just created it");
1359 72 : assert_eq!(tline.get_last_record_lsn(), Lsn(0));
1360 :
1361 : // Setup minimum keys required for the timeline to be usable.
1362 72 : let mut modification = tline.begin_modification(initdb_lsn);
1363 72 : modification
1364 72 : .init_empty_test_timeline()
1365 72 : .context("init_empty_test_timeline")?;
1366 72 : modification
1367 72 : .commit(ctx)
1368 72 : .await
1369 72 : .context("commit init_empty_test_timeline modification")?;
1370 :
1371 : // Flush to disk so that uninit_tl's check for valid disk_consistent_lsn passes.
1372 72 : tline.maybe_spawn_flush_loop();
1373 72 : tline.freeze_and_flush().await.context("freeze_and_flush")?;
1374 :
1375 : // Make sure the freeze_and_flush reaches remote storage.
1376 72 : tline
1377 72 : .remote_client
1378 72 : .as_ref()
1379 72 : .unwrap()
1380 72 : .wait_completion()
1381 70 : .await
1382 72 : .unwrap();
1383 :
1384 72 : let tl = uninit_tl.finish_creation()?;
1385 : // The non-test code would call tl.activate() here.
1386 72 : tl.set_state(TimelineState::Active);
1387 72 : Ok(tl)
1388 72 : }
1389 :
1390 : /// Create a new timeline.
1391 : ///
1392 : /// Returns the new timeline ID and reference to its Timeline object.
1393 : ///
1394 : /// If the caller specified the timeline ID to use (`new_timeline_id`), and timeline with
1395 : /// the same timeline ID already exists, returns CreateTimelineError::AlreadyExists.
1396 : #[allow(clippy::too_many_arguments)]
1397 0 : pub(crate) async fn create_timeline(
1398 0 : &self,
1399 0 : new_timeline_id: TimelineId,
1400 0 : ancestor_timeline_id: Option<TimelineId>,
1401 0 : mut ancestor_start_lsn: Option<Lsn>,
1402 0 : pg_version: u32,
1403 0 : load_existing_initdb: Option<TimelineId>,
1404 0 : broker_client: storage_broker::BrokerClientChannel,
1405 0 : ctx: &RequestContext,
1406 0 : ) -> Result<Arc<Timeline>, CreateTimelineError> {
1407 0 : if !self.is_active() {
1408 0 : if matches!(self.current_state(), TenantState::Stopping { .. }) {
1409 0 : return Err(CreateTimelineError::ShuttingDown);
1410 : } else {
1411 0 : return Err(CreateTimelineError::Other(anyhow::anyhow!(
1412 0 : "Cannot create timelines on inactive tenant"
1413 0 : )));
1414 : }
1415 0 : }
1416 :
1417 0 : let _gate = self
1418 0 : .gate
1419 0 : .enter()
1420 0 : .map_err(|_| CreateTimelineError::ShuttingDown)?;
1421 :
1422 : // Get exclusive access to the timeline ID: this ensures that it does not already exist,
1423 : // and that no other creation attempts will be allowed in while we are working. The
1424 : // uninit_mark is a guard.
1425 0 : let uninit_mark = match self.create_timeline_uninit_mark(new_timeline_id) {
1426 0 : Ok(m) => m,
1427 : Err(TimelineExclusionError::AlreadyCreating) => {
1428 : // Creation is in progress, we cannot create it again, and we cannot
1429 : // check if this request matches the existing one, so caller must try
1430 : // again later.
1431 0 : return Err(CreateTimelineError::AlreadyCreating);
1432 : }
1433 0 : Err(TimelineExclusionError::Other(e)) => {
1434 0 : return Err(CreateTimelineError::Other(e));
1435 : }
1436 0 : Err(TimelineExclusionError::AlreadyExists(existing)) => {
1437 0 : debug!("timeline {new_timeline_id} already exists");
1438 :
1439 : // Idempotency: creating the same timeline twice is not an error, unless
1440 : // the second creation has different parameters.
1441 0 : if existing.get_ancestor_timeline_id() != ancestor_timeline_id
1442 0 : || existing.pg_version != pg_version
1443 0 : || (ancestor_start_lsn.is_some()
1444 0 : && ancestor_start_lsn != Some(existing.get_ancestor_lsn()))
1445 : {
1446 0 : return Err(CreateTimelineError::Conflict);
1447 0 : }
1448 :
1449 0 : if let Some(remote_client) = existing.remote_client.as_ref() {
1450 : // Wait for uploads to complete, so that when we return Ok, the timeline
1451 : // is known to be durable on remote storage. Just like we do at the end of
1452 : // this function, after we have created the timeline ourselves.
1453 : //
1454 : // We only really care that the initial version of `index_part.json` has
1455 : // been uploaded. That's enough to remember that the timeline
1456 : // exists. However, there is no function to wait specifically for that so
1457 : // we just wait for all in-progress uploads to finish.
1458 0 : remote_client
1459 0 : .wait_completion()
1460 0 : .await
1461 0 : .context("wait for timeline uploads to complete")?;
1462 0 : }
1463 :
1464 0 : return Ok(existing);
1465 : }
1466 : };
1467 :
1468 0 : let loaded_timeline = match ancestor_timeline_id {
1469 0 : Some(ancestor_timeline_id) => {
1470 0 : let ancestor_timeline = self
1471 0 : .get_timeline(ancestor_timeline_id, false)
1472 0 : .context("Cannot branch off the timeline that's not present in pageserver")?;
1473 :
1474 : // instead of waiting around, just deny the request because ancestor is not yet
1475 : // ready for other purposes either.
1476 0 : if !ancestor_timeline.is_active() {
1477 0 : return Err(CreateTimelineError::AncestorNotActive);
1478 0 : }
1479 :
1480 0 : if let Some(lsn) = ancestor_start_lsn.as_mut() {
1481 0 : *lsn = lsn.align();
1482 0 :
1483 0 : let ancestor_ancestor_lsn = ancestor_timeline.get_ancestor_lsn();
1484 0 : if ancestor_ancestor_lsn > *lsn {
1485 : // can we safely just branch from the ancestor instead?
1486 0 : return Err(CreateTimelineError::AncestorLsn(anyhow::anyhow!(
1487 0 : "invalid start lsn {} for ancestor timeline {}: less than timeline ancestor lsn {}",
1488 0 : lsn,
1489 0 : ancestor_timeline_id,
1490 0 : ancestor_ancestor_lsn,
1491 0 : )));
1492 0 : }
1493 0 :
1494 0 : // Wait for the WAL to arrive and be processed on the parent branch up
1495 0 : // to the requested branch point. The repository code itself doesn't
1496 0 : // require it, but if we start to receive WAL on the new timeline,
1497 0 : // decoding the new WAL might need to look up previous pages, relation
1498 0 : // sizes etc. and that would get confused if the previous page versions
1499 0 : // are not in the repository yet.
1500 0 : ancestor_timeline
1501 0 : .wait_lsn(*lsn, ctx)
1502 0 : .await
1503 0 : .map_err(|e| match e {
1504 0 : e @ (WaitLsnError::Timeout(_) | WaitLsnError::BadState) => {
1505 0 : CreateTimelineError::AncestorLsn(anyhow::anyhow!(e))
1506 : }
1507 0 : WaitLsnError::Shutdown => CreateTimelineError::ShuttingDown,
1508 0 : })?;
1509 0 : }
1510 :
1511 0 : self.branch_timeline(
1512 0 : &ancestor_timeline,
1513 0 : new_timeline_id,
1514 0 : ancestor_start_lsn,
1515 0 : uninit_mark,
1516 0 : ctx,
1517 0 : )
1518 0 : .await?
1519 : }
1520 : None => {
1521 0 : self.bootstrap_timeline(
1522 0 : new_timeline_id,
1523 0 : pg_version,
1524 0 : load_existing_initdb,
1525 0 : uninit_mark,
1526 0 : ctx,
1527 0 : )
1528 0 : .await?
1529 : }
1530 : };
1531 :
1532 : // At this point we have dropped our guard on [`Self::timelines_creating`], and
1533 : // the timeline is visible in [`Self::timelines`], but it is _not_ durable yet. We must
1534 : // not send a success to the caller until it is. The same applies to handling retries,
1535 : // see the handling of [`TimelineExclusionError::AlreadyExists`] above.
1536 0 : if let Some(remote_client) = loaded_timeline.remote_client.as_ref() {
1537 0 : let kind = ancestor_timeline_id
1538 0 : .map(|_| "branched")
1539 0 : .unwrap_or("bootstrapped");
1540 0 : remote_client.wait_completion().await.with_context(|| {
1541 0 : format!("wait for {} timeline initial uploads to complete", kind)
1542 0 : })?;
1543 0 : }
1544 :
1545 0 : loaded_timeline.activate(broker_client, None, ctx);
1546 0 :
1547 0 : Ok(loaded_timeline)
1548 0 : }
1549 :
1550 0 : pub(crate) async fn delete_timeline(
1551 0 : self: Arc<Self>,
1552 0 : timeline_id: TimelineId,
1553 0 : ) -> Result<(), DeleteTimelineError> {
1554 0 : DeleteTimelineFlow::run(&self, timeline_id, false).await?;
1555 :
1556 0 : Ok(())
1557 0 : }
1558 :
1559 : /// perform one garbage collection iteration, removing old data files from disk.
1560 : /// this function is periodically called by gc task.
1561 : /// also it can be explicitly requested through page server api 'do_gc' command.
1562 : ///
1563 : /// `target_timeline_id` specifies the timeline to GC, or None for all.
1564 : ///
1565 : /// The `horizon` an `pitr` parameters determine how much WAL history needs to be retained.
1566 : /// Also known as the retention period, or the GC cutoff point. `horizon` specifies
1567 : /// the amount of history, as LSN difference from current latest LSN on each timeline.
1568 : /// `pitr` specifies the same as a time difference from the current time. The effective
1569 : /// GC cutoff point is determined conservatively by either `horizon` and `pitr`, whichever
1570 : /// requires more history to be retained.
1571 : //
1572 8 : pub async fn gc_iteration(
1573 8 : &self,
1574 8 : target_timeline_id: Option<TimelineId>,
1575 8 : horizon: u64,
1576 8 : pitr: Duration,
1577 8 : cancel: &CancellationToken,
1578 8 : ctx: &RequestContext,
1579 8 : ) -> anyhow::Result<GcResult> {
1580 8 : // Don't start doing work during shutdown
1581 8 : if let TenantState::Stopping { .. } = self.current_state() {
1582 0 : return Ok(GcResult::default());
1583 8 : }
1584 8 :
1585 8 : // there is a global allowed_error for this
1586 8 : anyhow::ensure!(
1587 8 : self.is_active(),
1588 0 : "Cannot run GC iteration on inactive tenant"
1589 : );
1590 :
1591 : {
1592 8 : let conf = self.tenant_conf.read().unwrap();
1593 8 :
1594 8 : if !conf.location.may_delete_layers_hint() {
1595 0 : info!("Skipping GC in location state {:?}", conf.location);
1596 0 : return Ok(GcResult::default());
1597 8 : }
1598 8 : }
1599 8 :
1600 8 : self.gc_iteration_internal(target_timeline_id, horizon, pitr, cancel, ctx)
1601 0 : .await
1602 8 : }
1603 :
1604 : /// Perform one compaction iteration.
1605 : /// This function is periodically called by compactor task.
1606 : /// Also it can be explicitly requested per timeline through page server
1607 : /// api's 'compact' command.
1608 0 : async fn compaction_iteration(
1609 0 : &self,
1610 0 : cancel: &CancellationToken,
1611 0 : ctx: &RequestContext,
1612 0 : ) -> anyhow::Result<(), timeline::CompactionError> {
1613 0 : // Don't start doing work during shutdown, or when broken, we do not need those in the logs
1614 0 : if !self.is_active() {
1615 0 : return Ok(());
1616 0 : }
1617 0 :
1618 0 : {
1619 0 : let conf = self.tenant_conf.read().unwrap();
1620 0 : if !conf.location.may_delete_layers_hint() || !conf.location.may_upload_layers_hint() {
1621 0 : info!("Skipping compaction in location state {:?}", conf.location);
1622 0 : return Ok(());
1623 0 : }
1624 0 : }
1625 0 :
1626 0 : // Scan through the hashmap and collect a list of all the timelines,
1627 0 : // while holding the lock. Then drop the lock and actually perform the
1628 0 : // compactions. We don't want to block everything else while the
1629 0 : // compaction runs.
1630 0 : let timelines_to_compact = {
1631 0 : let timelines = self.timelines.lock().unwrap();
1632 0 : let timelines_to_compact = timelines
1633 0 : .iter()
1634 0 : .filter_map(|(timeline_id, timeline)| {
1635 0 : if timeline.is_active() {
1636 0 : Some((*timeline_id, timeline.clone()))
1637 : } else {
1638 0 : None
1639 : }
1640 0 : })
1641 0 : .collect::<Vec<_>>();
1642 0 : drop(timelines);
1643 0 : timelines_to_compact
1644 : };
1645 :
1646 0 : for (timeline_id, timeline) in &timelines_to_compact {
1647 0 : timeline
1648 0 : .compact(cancel, EnumSet::empty(), ctx)
1649 0 : .instrument(info_span!("compact_timeline", %timeline_id))
1650 0 : .await?;
1651 : }
1652 :
1653 0 : Ok(())
1654 0 : }
1655 :
1656 102 : pub fn current_state(&self) -> TenantState {
1657 102 : self.state.borrow().clone()
1658 102 : }
1659 :
1660 88 : pub fn is_active(&self) -> bool {
1661 88 : self.current_state() == TenantState::Active
1662 88 : }
1663 :
1664 0 : pub fn generation(&self) -> Generation {
1665 0 : self.generation
1666 0 : }
1667 :
1668 0 : pub(crate) fn wal_redo_manager_status(&self) -> Option<WalRedoManagerStatus> {
1669 0 : self.walredo_mgr.as_ref().and_then(|mgr| mgr.status())
1670 0 : }
1671 :
1672 : /// Changes tenant status to active, unless shutdown was already requested.
1673 : ///
1674 : /// `background_jobs_can_start` is an optional barrier set to a value during pageserver startup
1675 : /// to delay background jobs. Background jobs can be started right away when None is given.
1676 0 : fn activate(
1677 0 : self: &Arc<Self>,
1678 0 : broker_client: BrokerClientChannel,
1679 0 : background_jobs_can_start: Option<&completion::Barrier>,
1680 0 : ctx: &RequestContext,
1681 0 : ) {
1682 0 : span::debug_assert_current_span_has_tenant_id();
1683 0 :
1684 0 : let mut activating = false;
1685 0 : self.state.send_modify(|current_state| {
1686 0 : use pageserver_api::models::ActivatingFrom;
1687 0 : match &*current_state {
1688 : TenantState::Activating(_) | TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => {
1689 0 : panic!("caller is responsible for calling activate() only on Loading / Attaching tenants, got {state:?}", state = current_state);
1690 : }
1691 0 : TenantState::Loading => {
1692 0 : *current_state = TenantState::Activating(ActivatingFrom::Loading);
1693 0 : }
1694 0 : TenantState::Attaching => {
1695 0 : *current_state = TenantState::Activating(ActivatingFrom::Attaching);
1696 0 : }
1697 : }
1698 0 : debug!(tenant_id = %self.tenant_shard_id.tenant_id, shard_id = %self.tenant_shard_id.shard_slug(), "Activating tenant");
1699 0 : activating = true;
1700 0 : // Continue outside the closure. We need to grab timelines.lock()
1701 0 : // and we plan to turn it into a tokio::sync::Mutex in a future patch.
1702 0 : });
1703 0 :
1704 0 : if activating {
1705 0 : let timelines_accessor = self.timelines.lock().unwrap();
1706 0 : let timelines_to_activate = timelines_accessor
1707 0 : .values()
1708 0 : .filter(|timeline| !(timeline.is_broken() || timeline.is_stopping()));
1709 0 :
1710 0 : // Spawn gc and compaction loops. The loops will shut themselves
1711 0 : // down when they notice that the tenant is inactive.
1712 0 : tasks::start_background_loops(self, background_jobs_can_start);
1713 0 :
1714 0 : let mut activated_timelines = 0;
1715 :
1716 0 : for timeline in timelines_to_activate {
1717 0 : timeline.activate(broker_client.clone(), background_jobs_can_start, ctx);
1718 0 : activated_timelines += 1;
1719 0 : }
1720 :
1721 0 : self.state.send_modify(move |current_state| {
1722 0 : assert!(
1723 0 : matches!(current_state, TenantState::Activating(_)),
1724 0 : "set_stopping and set_broken wait for us to leave Activating state",
1725 : );
1726 0 : *current_state = TenantState::Active;
1727 0 :
1728 0 : let elapsed = self.constructed_at.elapsed();
1729 0 : let total_timelines = timelines_accessor.len();
1730 0 :
1731 0 : // log a lot of stuff, because some tenants sometimes suffer from user-visible
1732 0 : // times to activate. see https://github.com/neondatabase/neon/issues/4025
1733 0 : info!(
1734 0 : since_creation_millis = elapsed.as_millis(),
1735 0 : tenant_id = %self.tenant_shard_id.tenant_id,
1736 0 : shard_id = %self.tenant_shard_id.shard_slug(),
1737 0 : activated_timelines,
1738 0 : total_timelines,
1739 0 : post_state = <&'static str>::from(&*current_state),
1740 0 : "activation attempt finished"
1741 0 : );
1742 :
1743 0 : TENANT.activation.observe(elapsed.as_secs_f64());
1744 0 : });
1745 0 : }
1746 0 : }
1747 :
1748 : /// Shutdown the tenant and join all of the spawned tasks.
1749 : ///
1750 : /// The method caters for all use-cases:
1751 : /// - pageserver shutdown (freeze_and_flush == true)
1752 : /// - detach + ignore (freeze_and_flush == false)
1753 : ///
1754 : /// This will attempt to shutdown even if tenant is broken.
1755 : ///
1756 : /// `shutdown_progress` is a [`completion::Barrier`] for the shutdown initiated by this call.
1757 : /// If the tenant is already shutting down, we return a clone of the first shutdown call's
1758 : /// `Barrier` as an `Err`. This not-first caller can use the returned barrier to join with
1759 : /// the ongoing shutdown.
1760 6 : async fn shutdown(
1761 6 : &self,
1762 6 : shutdown_progress: completion::Barrier,
1763 6 : freeze_and_flush: bool,
1764 6 : ) -> Result<(), completion::Barrier> {
1765 6 : span::debug_assert_current_span_has_tenant_id();
1766 6 :
1767 6 : // Set tenant (and its timlines) to Stoppping state.
1768 6 : //
1769 6 : // Since we can only transition into Stopping state after activation is complete,
1770 6 : // run it in a JoinSet so all tenants have a chance to stop before we get SIGKILLed.
1771 6 : //
1772 6 : // Transitioning tenants to Stopping state has a couple of non-obvious side effects:
1773 6 : // 1. Lock out any new requests to the tenants.
1774 6 : // 2. Signal cancellation to WAL receivers (we wait on it below).
1775 6 : // 3. Signal cancellation for other tenant background loops.
1776 6 : // 4. ???
1777 6 : //
1778 6 : // The waiting for the cancellation is not done uniformly.
1779 6 : // We certainly wait for WAL receivers to shut down.
1780 6 : // That is necessary so that no new data comes in before the freeze_and_flush.
1781 6 : // But the tenant background loops are joined-on in our caller.
1782 6 : // It's mesed up.
1783 6 : // we just ignore the failure to stop
1784 6 :
1785 6 : // If we're still attaching, fire the cancellation token early to drop out: this
1786 6 : // will prevent us flushing, but ensures timely shutdown if some I/O during attach
1787 6 : // is very slow.
1788 6 : if matches!(self.current_state(), TenantState::Attaching) {
1789 0 : self.cancel.cancel();
1790 6 : }
1791 :
1792 6 : match self.set_stopping(shutdown_progress, false, false).await {
1793 6 : Ok(()) => {}
1794 0 : Err(SetStoppingError::Broken) => {
1795 0 : // assume that this is acceptable
1796 0 : }
1797 0 : Err(SetStoppingError::AlreadyStopping(other)) => {
1798 : // give caller the option to wait for this this shutdown
1799 0 : info!("Tenant::shutdown: AlreadyStopping");
1800 0 : return Err(other);
1801 : }
1802 : };
1803 :
1804 6 : let mut js = tokio::task::JoinSet::new();
1805 6 : {
1806 6 : let timelines = self.timelines.lock().unwrap();
1807 6 : timelines.values().for_each(|timeline| {
1808 6 : let timeline = Arc::clone(timeline);
1809 6 : let timeline_id = timeline.timeline_id;
1810 :
1811 6 : let span =
1812 6 : tracing::info_span!("timeline_shutdown", %timeline_id, ?freeze_and_flush);
1813 6 : js.spawn(async move {
1814 6 : if freeze_and_flush {
1815 16 : timeline.flush_and_shutdown().instrument(span).await
1816 : } else {
1817 0 : timeline.shutdown().instrument(span).await
1818 : }
1819 6 : });
1820 6 : })
1821 : };
1822 : // test_long_timeline_create_then_tenant_delete is leaning on this message
1823 6 : tracing::info!("Waiting for timelines...");
1824 12 : while let Some(res) = js.join_next().await {
1825 0 : match res {
1826 6 : Ok(()) => {}
1827 0 : Err(je) if je.is_cancelled() => unreachable!("no cancelling used"),
1828 0 : Err(je) if je.is_panic() => { /* logged already */ }
1829 0 : Err(je) => warn!("unexpected JoinError: {je:?}"),
1830 : }
1831 : }
1832 :
1833 : // We cancel the Tenant's cancellation token _after_ the timelines have all shut down. This permits
1834 : // them to continue to do work during their shutdown methods, e.g. flushing data.
1835 0 : tracing::debug!("Cancelling CancellationToken");
1836 6 : self.cancel.cancel();
1837 :
1838 : // shutdown all tenant and timeline tasks: gc, compaction, page service
1839 : // No new tasks will be started for this tenant because it's in `Stopping` state.
1840 : //
1841 : // this will additionally shutdown and await all timeline tasks.
1842 0 : tracing::debug!("Waiting for tasks...");
1843 6 : task_mgr::shutdown_tasks(None, Some(self.tenant_shard_id), None).await;
1844 :
1845 : // Wait for any in-flight operations to complete
1846 6 : self.gate.close().await;
1847 :
1848 6 : Ok(())
1849 6 : }
1850 :
1851 : /// Change tenant status to Stopping, to mark that it is being shut down.
1852 : ///
1853 : /// This function waits for the tenant to become active if it isn't already, before transitioning it into Stopping state.
1854 : ///
1855 : /// This function is not cancel-safe!
1856 : ///
1857 : /// `allow_transition_from_loading` is needed for the special case of loading task deleting the tenant.
1858 : /// `allow_transition_from_attaching` is needed for the special case of attaching deleted tenant.
1859 6 : async fn set_stopping(
1860 6 : &self,
1861 6 : progress: completion::Barrier,
1862 6 : allow_transition_from_loading: bool,
1863 6 : allow_transition_from_attaching: bool,
1864 6 : ) -> Result<(), SetStoppingError> {
1865 6 : let mut rx = self.state.subscribe();
1866 6 :
1867 6 : // cannot stop before we're done activating, so wait out until we're done activating
1868 6 : rx.wait_for(|state| match state {
1869 0 : TenantState::Attaching if allow_transition_from_attaching => true,
1870 : TenantState::Activating(_) | TenantState::Attaching => {
1871 0 : info!(
1872 0 : "waiting for {} to turn Active|Broken|Stopping",
1873 0 : <&'static str>::from(state)
1874 0 : );
1875 0 : false
1876 : }
1877 0 : TenantState::Loading => allow_transition_from_loading,
1878 6 : TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => true,
1879 6 : })
1880 0 : .await
1881 6 : .expect("cannot drop self.state while on a &self method");
1882 6 :
1883 6 : // we now know we're done activating, let's see whether this task is the winner to transition into Stopping
1884 6 : let mut err = None;
1885 6 : let stopping = self.state.send_if_modified(|current_state| match current_state {
1886 : TenantState::Activating(_) => {
1887 0 : unreachable!("1we ensured above that we're done with activation, and, there is no re-activation")
1888 : }
1889 : TenantState::Attaching => {
1890 0 : if !allow_transition_from_attaching {
1891 0 : unreachable!("2we ensured above that we're done with activation, and, there is no re-activation")
1892 0 : };
1893 0 : *current_state = TenantState::Stopping { progress };
1894 0 : true
1895 : }
1896 : TenantState::Loading => {
1897 0 : if !allow_transition_from_loading {
1898 0 : unreachable!("3we ensured above that we're done with activation, and, there is no re-activation")
1899 0 : };
1900 0 : *current_state = TenantState::Stopping { progress };
1901 0 : true
1902 : }
1903 : TenantState::Active => {
1904 : // FIXME: due to time-of-check vs time-of-use issues, it can happen that new timelines
1905 : // are created after the transition to Stopping. That's harmless, as the Timelines
1906 : // won't be accessible to anyone afterwards, because the Tenant is in Stopping state.
1907 6 : *current_state = TenantState::Stopping { progress };
1908 6 : // Continue stopping outside the closure. We need to grab timelines.lock()
1909 6 : // and we plan to turn it into a tokio::sync::Mutex in a future patch.
1910 6 : true
1911 : }
1912 0 : TenantState::Broken { reason, .. } => {
1913 0 : info!(
1914 0 : "Cannot set tenant to Stopping state, it is in Broken state due to: {reason}"
1915 0 : );
1916 0 : err = Some(SetStoppingError::Broken);
1917 0 : false
1918 : }
1919 0 : TenantState::Stopping { progress } => {
1920 0 : info!("Tenant is already in Stopping state");
1921 0 : err = Some(SetStoppingError::AlreadyStopping(progress.clone()));
1922 0 : false
1923 : }
1924 6 : });
1925 6 : match (stopping, err) {
1926 6 : (true, None) => {} // continue
1927 0 : (false, Some(err)) => return Err(err),
1928 0 : (true, Some(_)) => unreachable!(
1929 0 : "send_if_modified closure must error out if not transitioning to Stopping"
1930 0 : ),
1931 0 : (false, None) => unreachable!(
1932 0 : "send_if_modified closure must return true if transitioning to Stopping"
1933 0 : ),
1934 : }
1935 :
1936 6 : let timelines_accessor = self.timelines.lock().unwrap();
1937 6 : let not_broken_timelines = timelines_accessor
1938 6 : .values()
1939 6 : .filter(|timeline| !timeline.is_broken());
1940 12 : for timeline in not_broken_timelines {
1941 6 : timeline.set_state(TimelineState::Stopping);
1942 6 : }
1943 6 : Ok(())
1944 6 : }
1945 :
1946 : /// Method for tenant::mgr to transition us into Broken state in case of a late failure in
1947 : /// `remove_tenant_from_memory`
1948 : ///
1949 : /// This function waits for the tenant to become active if it isn't already, before transitioning it into Stopping state.
1950 : ///
1951 : /// In tests, we also use this to set tenants to Broken state on purpose.
1952 0 : pub(crate) async fn set_broken(&self, reason: String) {
1953 0 : let mut rx = self.state.subscribe();
1954 0 :
1955 0 : // The load & attach routines own the tenant state until it has reached `Active`.
1956 0 : // So, wait until it's done.
1957 0 : rx.wait_for(|state| match state {
1958 : TenantState::Activating(_) | TenantState::Loading | TenantState::Attaching => {
1959 0 : info!(
1960 0 : "waiting for {} to turn Active|Broken|Stopping",
1961 0 : <&'static str>::from(state)
1962 0 : );
1963 0 : false
1964 : }
1965 0 : TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => true,
1966 0 : })
1967 0 : .await
1968 0 : .expect("cannot drop self.state while on a &self method");
1969 0 :
1970 0 : // we now know we're done activating, let's see whether this task is the winner to transition into Broken
1971 0 : self.set_broken_no_wait(reason)
1972 0 : }
1973 :
1974 0 : pub(crate) fn set_broken_no_wait(&self, reason: impl Display) {
1975 0 : let reason = reason.to_string();
1976 0 : self.state.send_modify(|current_state| {
1977 0 : match *current_state {
1978 : TenantState::Activating(_) | TenantState::Loading | TenantState::Attaching => {
1979 0 : unreachable!("we ensured above that we're done with activation, and, there is no re-activation")
1980 : }
1981 : TenantState::Active => {
1982 0 : if cfg!(feature = "testing") {
1983 0 : warn!("Changing Active tenant to Broken state, reason: {}", reason);
1984 0 : *current_state = TenantState::broken_from_reason(reason);
1985 : } else {
1986 0 : unreachable!("not allowed to call set_broken on Active tenants in non-testing builds")
1987 : }
1988 : }
1989 : TenantState::Broken { .. } => {
1990 0 : warn!("Tenant is already in Broken state");
1991 : }
1992 : // This is the only "expected" path, any other path is a bug.
1993 : TenantState::Stopping { .. } => {
1994 0 : warn!(
1995 0 : "Marking Stopping tenant as Broken state, reason: {}",
1996 0 : reason
1997 0 : );
1998 0 : *current_state = TenantState::broken_from_reason(reason);
1999 : }
2000 : }
2001 0 : });
2002 0 : }
2003 :
2004 0 : pub fn subscribe_for_state_updates(&self) -> watch::Receiver<TenantState> {
2005 0 : self.state.subscribe()
2006 0 : }
2007 :
2008 : /// The activate_now semaphore is initialized with zero units. As soon as
2009 : /// we add a unit, waiters will be able to acquire a unit and proceed.
2010 0 : pub(crate) fn activate_now(&self) {
2011 0 : self.activate_now_sem.add_permits(1);
2012 0 : }
2013 :
2014 0 : pub(crate) async fn wait_to_become_active(
2015 0 : &self,
2016 0 : timeout: Duration,
2017 0 : ) -> Result<(), GetActiveTenantError> {
2018 0 : let mut receiver = self.state.subscribe();
2019 0 : loop {
2020 0 : let current_state = receiver.borrow_and_update().clone();
2021 0 : match current_state {
2022 : TenantState::Loading | TenantState::Attaching | TenantState::Activating(_) => {
2023 : // in these states, there's a chance that we can reach ::Active
2024 0 : self.activate_now();
2025 0 : match timeout_cancellable(timeout, &self.cancel, receiver.changed()).await {
2026 0 : Ok(r) => {
2027 0 : r.map_err(
2028 0 : |_e: tokio::sync::watch::error::RecvError|
2029 : // Tenant existed but was dropped: report it as non-existent
2030 0 : GetActiveTenantError::NotFound(GetTenantError::NotFound(self.tenant_shard_id.tenant_id))
2031 0 : )?
2032 : }
2033 : Err(TimeoutCancellableError::Cancelled) => {
2034 0 : return Err(GetActiveTenantError::Cancelled);
2035 : }
2036 : Err(TimeoutCancellableError::Timeout) => {
2037 0 : return Err(GetActiveTenantError::WaitForActiveTimeout {
2038 0 : latest_state: Some(self.current_state()),
2039 0 : wait_time: timeout,
2040 0 : });
2041 : }
2042 : }
2043 : }
2044 : TenantState::Active { .. } => {
2045 0 : return Ok(());
2046 : }
2047 : TenantState::Broken { .. } | TenantState::Stopping { .. } => {
2048 : // There's no chance the tenant can transition back into ::Active
2049 0 : return Err(GetActiveTenantError::WillNotBecomeActive(current_state));
2050 : }
2051 : }
2052 : }
2053 0 : }
2054 :
2055 0 : pub(crate) fn get_attach_mode(&self) -> AttachmentMode {
2056 0 : self.tenant_conf.read().unwrap().location.attach_mode
2057 0 : }
2058 :
2059 : /// For API access: generate a LocationConfig equivalent to the one that would be used to
2060 : /// create a Tenant in the same state. Do not use this in hot paths: it's for relatively
2061 : /// rare external API calls, like a reconciliation at startup.
2062 0 : pub(crate) fn get_location_conf(&self) -> models::LocationConfig {
2063 0 : let conf = self.tenant_conf.read().unwrap();
2064 :
2065 0 : let location_config_mode = match conf.location.attach_mode {
2066 0 : AttachmentMode::Single => models::LocationConfigMode::AttachedSingle,
2067 0 : AttachmentMode::Multi => models::LocationConfigMode::AttachedMulti,
2068 0 : AttachmentMode::Stale => models::LocationConfigMode::AttachedStale,
2069 : };
2070 :
2071 : // We have a pageserver TenantConf, we need the API-facing TenantConfig.
2072 0 : let tenant_config: models::TenantConfig = conf.tenant_conf.clone().into();
2073 0 :
2074 0 : models::LocationConfig {
2075 0 : mode: location_config_mode,
2076 0 : generation: self.generation.into(),
2077 0 : secondary_conf: None,
2078 0 : shard_number: self.shard_identity.number.0,
2079 0 : shard_count: self.shard_identity.count.literal(),
2080 0 : shard_stripe_size: self.shard_identity.stripe_size.0,
2081 0 : tenant_conf: tenant_config,
2082 0 : }
2083 0 : }
2084 :
2085 0 : pub(crate) fn get_tenant_shard_id(&self) -> &TenantShardId {
2086 0 : &self.tenant_shard_id
2087 0 : }
2088 :
2089 0 : pub(crate) fn get_generation(&self) -> Generation {
2090 0 : self.generation
2091 0 : }
2092 :
2093 : /// This function partially shuts down the tenant (it shuts down the Timelines) and is fallible,
2094 : /// and can leave the tenant in a bad state if it fails. The caller is responsible for
2095 : /// resetting this tenant to a valid state if we fail.
2096 0 : pub(crate) async fn split_prepare(
2097 0 : &self,
2098 0 : child_shards: &Vec<TenantShardId>,
2099 0 : ) -> anyhow::Result<()> {
2100 0 : let timelines = self.timelines.lock().unwrap().clone();
2101 0 : for timeline in timelines.values() {
2102 0 : let Some(tl_client) = &timeline.remote_client else {
2103 0 : anyhow::bail!("Remote storage is mandatory");
2104 : };
2105 :
2106 0 : let Some(remote_storage) = &self.remote_storage else {
2107 0 : anyhow::bail!("Remote storage is mandatory");
2108 : };
2109 :
2110 : // We do not block timeline creation/deletion during splits inside the pageserver: it is up to higher levels
2111 : // to ensure that they do not start a split if currently in the process of doing these.
2112 :
2113 : // Upload an index from the parent: this is partly to provide freshness for the
2114 : // child tenants that will copy it, and partly for general ease-of-debugging: there will
2115 : // always be a parent shard index in the same generation as we wrote the child shard index.
2116 0 : tl_client.schedule_index_upload_for_file_changes()?;
2117 0 : tl_client.wait_completion().await?;
2118 :
2119 : // Shut down the timeline's remote client: this means that the indices we write
2120 : // for child shards will not be invalidated by the parent shard deleting layers.
2121 0 : tl_client.shutdown().await?;
2122 :
2123 : // Download methods can still be used after shutdown, as they don't flow through the remote client's
2124 : // queue. In principal the RemoteTimelineClient could provide this without downloading it, but this
2125 : // operation is rare, so it's simpler to just download it (and robustly guarantees that the index
2126 : // we use here really is the remotely persistent one).
2127 0 : let result = tl_client
2128 0 : .download_index_file(&self.cancel)
2129 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))
2130 0 : .await?;
2131 0 : let index_part = match result {
2132 : MaybeDeletedIndexPart::Deleted(_) => {
2133 0 : anyhow::bail!("Timeline deletion happened concurrently with split")
2134 : }
2135 0 : MaybeDeletedIndexPart::IndexPart(p) => p,
2136 : };
2137 :
2138 0 : for child_shard in child_shards {
2139 0 : upload_index_part(
2140 0 : remote_storage,
2141 0 : child_shard,
2142 0 : &timeline.timeline_id,
2143 0 : self.generation,
2144 0 : &index_part,
2145 0 : &self.cancel,
2146 0 : )
2147 0 : .await?;
2148 : }
2149 : }
2150 :
2151 0 : Ok(())
2152 0 : }
2153 : }
2154 :
2155 : /// Given a Vec of timelines and their ancestors (timeline_id, ancestor_id),
2156 : /// perform a topological sort, so that the parent of each timeline comes
2157 : /// before the children.
2158 : /// E extracts the ancestor from T
2159 : /// This allows for T to be different. It can be TimelineMetadata, can be Timeline itself, etc.
2160 88 : fn tree_sort_timelines<T, E>(
2161 88 : timelines: HashMap<TimelineId, T>,
2162 88 : extractor: E,
2163 88 : ) -> anyhow::Result<Vec<(TimelineId, T)>>
2164 88 : where
2165 88 : E: Fn(&T) -> Option<TimelineId>,
2166 88 : {
2167 88 : let mut result = Vec::with_capacity(timelines.len());
2168 88 :
2169 88 : let mut now = Vec::with_capacity(timelines.len());
2170 88 : // (ancestor, children)
2171 88 : let mut later: HashMap<TimelineId, Vec<(TimelineId, T)>> =
2172 88 : HashMap::with_capacity(timelines.len());
2173 :
2174 94 : for (timeline_id, value) in timelines {
2175 6 : if let Some(ancestor_id) = extractor(&value) {
2176 2 : let children = later.entry(ancestor_id).or_default();
2177 2 : children.push((timeline_id, value));
2178 4 : } else {
2179 4 : now.push((timeline_id, value));
2180 4 : }
2181 : }
2182 :
2183 94 : while let Some((timeline_id, metadata)) = now.pop() {
2184 6 : result.push((timeline_id, metadata));
2185 : // All children of this can be loaded now
2186 6 : if let Some(mut children) = later.remove(&timeline_id) {
2187 2 : now.append(&mut children);
2188 4 : }
2189 : }
2190 :
2191 : // All timelines should be visited now. Unless there were timelines with missing ancestors.
2192 88 : if !later.is_empty() {
2193 0 : for (missing_id, orphan_ids) in later {
2194 0 : for (orphan_id, _) in orphan_ids {
2195 0 : error!("could not load timeline {orphan_id} because its ancestor timeline {missing_id} could not be loaded");
2196 : }
2197 : }
2198 0 : bail!("could not load tenant because some timelines are missing ancestors");
2199 88 : }
2200 88 :
2201 88 : Ok(result)
2202 88 : }
2203 :
2204 : impl Tenant {
2205 0 : pub fn tenant_specific_overrides(&self) -> TenantConfOpt {
2206 0 : self.tenant_conf.read().unwrap().tenant_conf.clone()
2207 0 : }
2208 :
2209 0 : pub fn effective_config(&self) -> TenantConf {
2210 0 : self.tenant_specific_overrides()
2211 0 : .merge(self.conf.default_tenant_conf.clone())
2212 0 : }
2213 :
2214 0 : pub fn get_checkpoint_distance(&self) -> u64 {
2215 0 : let tenant_conf = self.tenant_conf.read().unwrap().tenant_conf.clone();
2216 0 : tenant_conf
2217 0 : .checkpoint_distance
2218 0 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_distance)
2219 0 : }
2220 :
2221 0 : pub fn get_checkpoint_timeout(&self) -> Duration {
2222 0 : let tenant_conf = self.tenant_conf.read().unwrap().tenant_conf.clone();
2223 0 : tenant_conf
2224 0 : .checkpoint_timeout
2225 0 : .unwrap_or(self.conf.default_tenant_conf.checkpoint_timeout)
2226 0 : }
2227 :
2228 0 : pub fn get_compaction_target_size(&self) -> u64 {
2229 0 : let tenant_conf = self.tenant_conf.read().unwrap().tenant_conf.clone();
2230 0 : tenant_conf
2231 0 : .compaction_target_size
2232 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_target_size)
2233 0 : }
2234 :
2235 0 : pub fn get_compaction_period(&self) -> Duration {
2236 0 : let tenant_conf = self.tenant_conf.read().unwrap().tenant_conf.clone();
2237 0 : tenant_conf
2238 0 : .compaction_period
2239 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_period)
2240 0 : }
2241 :
2242 0 : pub fn get_compaction_threshold(&self) -> usize {
2243 0 : let tenant_conf = self.tenant_conf.read().unwrap().tenant_conf.clone();
2244 0 : tenant_conf
2245 0 : .compaction_threshold
2246 0 : .unwrap_or(self.conf.default_tenant_conf.compaction_threshold)
2247 0 : }
2248 :
2249 0 : pub fn get_gc_horizon(&self) -> u64 {
2250 0 : let tenant_conf = self.tenant_conf.read().unwrap().tenant_conf.clone();
2251 0 : tenant_conf
2252 0 : .gc_horizon
2253 0 : .unwrap_or(self.conf.default_tenant_conf.gc_horizon)
2254 0 : }
2255 :
2256 0 : pub fn get_gc_period(&self) -> Duration {
2257 0 : let tenant_conf = self.tenant_conf.read().unwrap().tenant_conf.clone();
2258 0 : tenant_conf
2259 0 : .gc_period
2260 0 : .unwrap_or(self.conf.default_tenant_conf.gc_period)
2261 0 : }
2262 :
2263 0 : pub fn get_image_creation_threshold(&self) -> usize {
2264 0 : let tenant_conf = self.tenant_conf.read().unwrap().tenant_conf.clone();
2265 0 : tenant_conf
2266 0 : .image_creation_threshold
2267 0 : .unwrap_or(self.conf.default_tenant_conf.image_creation_threshold)
2268 0 : }
2269 :
2270 0 : pub fn get_pitr_interval(&self) -> Duration {
2271 0 : let tenant_conf = self.tenant_conf.read().unwrap().tenant_conf.clone();
2272 0 : tenant_conf
2273 0 : .pitr_interval
2274 0 : .unwrap_or(self.conf.default_tenant_conf.pitr_interval)
2275 0 : }
2276 :
2277 0 : pub fn get_trace_read_requests(&self) -> bool {
2278 0 : let tenant_conf = self.tenant_conf.read().unwrap().tenant_conf.clone();
2279 0 : tenant_conf
2280 0 : .trace_read_requests
2281 0 : .unwrap_or(self.conf.default_tenant_conf.trace_read_requests)
2282 0 : }
2283 :
2284 0 : pub fn get_min_resident_size_override(&self) -> Option<u64> {
2285 0 : let tenant_conf = self.tenant_conf.read().unwrap().tenant_conf.clone();
2286 0 : tenant_conf
2287 0 : .min_resident_size_override
2288 0 : .or(self.conf.default_tenant_conf.min_resident_size_override)
2289 0 : }
2290 :
2291 0 : pub fn get_heatmap_period(&self) -> Option<Duration> {
2292 0 : let tenant_conf = self.tenant_conf.read().unwrap().tenant_conf.clone();
2293 0 : let heatmap_period = tenant_conf
2294 0 : .heatmap_period
2295 0 : .unwrap_or(self.conf.default_tenant_conf.heatmap_period);
2296 0 : if heatmap_period.is_zero() {
2297 0 : None
2298 : } else {
2299 0 : Some(heatmap_period)
2300 : }
2301 0 : }
2302 :
2303 0 : pub fn set_new_tenant_config(&self, new_tenant_conf: TenantConfOpt) {
2304 0 : self.tenant_conf.write().unwrap().tenant_conf = new_tenant_conf;
2305 0 : self.tenant_conf_updated();
2306 0 : // Don't hold self.timelines.lock() during the notifies.
2307 0 : // There's no risk of deadlock right now, but there could be if we consolidate
2308 0 : // mutexes in struct Timeline in the future.
2309 0 : let timelines = self.list_timelines();
2310 0 : for timeline in timelines {
2311 0 : timeline.tenant_conf_updated();
2312 0 : }
2313 0 : }
2314 :
2315 0 : pub(crate) fn set_new_location_config(&self, new_conf: AttachedTenantConf) {
2316 0 : *self.tenant_conf.write().unwrap() = new_conf;
2317 0 : self.tenant_conf_updated();
2318 0 : // Don't hold self.timelines.lock() during the notifies.
2319 0 : // There's no risk of deadlock right now, but there could be if we consolidate
2320 0 : // mutexes in struct Timeline in the future.
2321 0 : let timelines = self.list_timelines();
2322 0 : for timeline in timelines {
2323 0 : timeline.tenant_conf_updated();
2324 0 : }
2325 0 : }
2326 :
2327 88 : fn get_timeline_get_throttle_config(
2328 88 : psconf: &'static PageServerConf,
2329 88 : overrides: &TenantConfOpt,
2330 88 : ) -> throttle::Config {
2331 88 : overrides
2332 88 : .timeline_get_throttle
2333 88 : .clone()
2334 88 : .unwrap_or(psconf.default_tenant_conf.timeline_get_throttle.clone())
2335 88 : }
2336 :
2337 0 : pub(crate) fn tenant_conf_updated(&self) {
2338 0 : let conf = {
2339 0 : let guard = self.tenant_conf.read().unwrap();
2340 0 : Self::get_timeline_get_throttle_config(self.conf, &guard.tenant_conf)
2341 0 : };
2342 0 : self.timeline_get_throttle.reconfigure(conf)
2343 0 : }
2344 :
2345 : /// Helper function to create a new Timeline struct.
2346 : ///
2347 : /// The returned Timeline is in Loading state. The caller is responsible for
2348 : /// initializing any on-disk state, and for inserting the Timeline to the 'timelines'
2349 : /// map.
2350 : ///
2351 : /// `validate_ancestor == false` is used when a timeline is created for deletion
2352 : /// and we might not have the ancestor present anymore which is fine for to be
2353 : /// deleted timelines.
2354 296 : fn create_timeline_struct(
2355 296 : &self,
2356 296 : new_timeline_id: TimelineId,
2357 296 : new_metadata: &TimelineMetadata,
2358 296 : ancestor: Option<Arc<Timeline>>,
2359 296 : resources: TimelineResources,
2360 296 : cause: CreateTimelineCause,
2361 296 : ) -> anyhow::Result<Arc<Timeline>> {
2362 296 : let state = match cause {
2363 : CreateTimelineCause::Load => {
2364 296 : let ancestor_id = new_metadata.ancestor_timeline();
2365 296 : anyhow::ensure!(
2366 296 : ancestor_id == ancestor.as_ref().map(|t| t.timeline_id),
2367 0 : "Timeline's {new_timeline_id} ancestor {ancestor_id:?} was not found"
2368 : );
2369 296 : TimelineState::Loading
2370 : }
2371 0 : CreateTimelineCause::Delete => TimelineState::Stopping,
2372 : };
2373 :
2374 296 : let pg_version = new_metadata.pg_version();
2375 296 :
2376 296 : let timeline = Timeline::new(
2377 296 : self.conf,
2378 296 : Arc::clone(&self.tenant_conf),
2379 296 : new_metadata,
2380 296 : ancestor,
2381 296 : new_timeline_id,
2382 296 : self.tenant_shard_id,
2383 296 : self.generation,
2384 296 : self.shard_identity,
2385 296 : self.walredo_mgr.as_ref().map(Arc::clone),
2386 296 : resources,
2387 296 : pg_version,
2388 296 : state,
2389 296 : self.cancel.child_token(),
2390 296 : );
2391 296 :
2392 296 : Ok(timeline)
2393 296 : }
2394 :
2395 : // Allow too_many_arguments because a constructor's argument list naturally grows with the
2396 : // number of attributes in the struct: breaking these out into a builder wouldn't be helpful.
2397 : #[allow(clippy::too_many_arguments)]
2398 88 : fn new(
2399 88 : state: TenantState,
2400 88 : conf: &'static PageServerConf,
2401 88 : attached_conf: AttachedTenantConf,
2402 88 : shard_identity: ShardIdentity,
2403 88 : walredo_mgr: Option<Arc<WalRedoManager>>,
2404 88 : tenant_shard_id: TenantShardId,
2405 88 : remote_storage: Option<GenericRemoteStorage>,
2406 88 : deletion_queue_client: DeletionQueueClient,
2407 88 : ) -> Tenant {
2408 88 : let (state, mut rx) = watch::channel(state);
2409 88 :
2410 88 : tokio::spawn(async move {
2411 83 : // reflect tenant state in metrics:
2412 83 : // - global per tenant state: TENANT_STATE_METRIC
2413 83 : // - "set" of broken tenants: BROKEN_TENANTS_SET
2414 83 : //
2415 83 : // set of broken tenants should not have zero counts so that it remains accessible for
2416 83 : // alerting.
2417 83 :
2418 83 : let tid = tenant_shard_id.to_string();
2419 83 : let shard_id = tenant_shard_id.shard_slug().to_string();
2420 83 : let set_key = &[tid.as_str(), shard_id.as_str()][..];
2421 83 :
2422 89 : fn inspect_state(state: &TenantState) -> ([&'static str; 1], bool) {
2423 89 : ([state.into()], matches!(state, TenantState::Broken { .. }))
2424 89 : }
2425 83 :
2426 83 : let mut tuple = inspect_state(&rx.borrow_and_update());
2427 83 :
2428 83 : let is_broken = tuple.1;
2429 83 : let mut counted_broken = if is_broken {
2430 : // add the id to the set right away, there should not be any updates on the channel
2431 : // after before tenant is removed, if ever
2432 0 : BROKEN_TENANTS_SET.with_label_values(set_key).set(1);
2433 0 : true
2434 : } else {
2435 83 : false
2436 : };
2437 :
2438 89 : loop {
2439 89 : let labels = &tuple.0;
2440 89 : let current = TENANT_STATE_METRIC.with_label_values(labels);
2441 89 : current.inc();
2442 89 :
2443 89 : if rx.changed().await.is_err() {
2444 : // tenant has been dropped
2445 4 : current.dec();
2446 4 : drop(BROKEN_TENANTS_SET.remove_label_values(set_key));
2447 4 : break;
2448 6 : }
2449 6 :
2450 6 : current.dec();
2451 6 : tuple = inspect_state(&rx.borrow_and_update());
2452 6 :
2453 6 : let is_broken = tuple.1;
2454 6 : if is_broken && !counted_broken {
2455 0 : counted_broken = true;
2456 0 : // insert the tenant_id (back) into the set while avoiding needless counter
2457 0 : // access
2458 0 : BROKEN_TENANTS_SET.with_label_values(set_key).set(1);
2459 6 : }
2460 : }
2461 88 : });
2462 88 :
2463 88 : Tenant {
2464 88 : tenant_shard_id,
2465 88 : shard_identity,
2466 88 : generation: attached_conf.location.generation,
2467 88 : conf,
2468 88 : // using now here is good enough approximation to catch tenants with really long
2469 88 : // activation times.
2470 88 : constructed_at: Instant::now(),
2471 88 : timelines: Mutex::new(HashMap::new()),
2472 88 : timelines_creating: Mutex::new(HashSet::new()),
2473 88 : gc_cs: tokio::sync::Mutex::new(()),
2474 88 : walredo_mgr,
2475 88 : remote_storage,
2476 88 : deletion_queue_client,
2477 88 : state,
2478 88 : cached_logical_sizes: tokio::sync::Mutex::new(HashMap::new()),
2479 88 : cached_synthetic_tenant_size: Arc::new(AtomicU64::new(0)),
2480 88 : eviction_task_tenant_state: tokio::sync::Mutex::new(EvictionTaskTenantState::default()),
2481 88 : activate_now_sem: tokio::sync::Semaphore::new(0),
2482 88 : delete_progress: Arc::new(tokio::sync::Mutex::new(DeleteTenantFlow::default())),
2483 88 : cancel: CancellationToken::default(),
2484 88 : gate: Gate::default(),
2485 88 : timeline_get_throttle: Arc::new(throttle::Throttle::new(
2486 88 : Tenant::get_timeline_get_throttle_config(conf, &attached_conf.tenant_conf),
2487 88 : &crate::metrics::tenant_throttling::TIMELINE_GET,
2488 88 : )),
2489 88 : tenant_conf: Arc::new(RwLock::new(attached_conf)),
2490 88 : }
2491 88 : }
2492 :
2493 : /// Locate and load config
2494 0 : pub(super) fn load_tenant_config(
2495 0 : conf: &'static PageServerConf,
2496 0 : tenant_shard_id: &TenantShardId,
2497 0 : ) -> anyhow::Result<LocationConf> {
2498 0 : let legacy_config_path = conf.tenant_config_path(tenant_shard_id);
2499 0 : let config_path = conf.tenant_location_config_path(tenant_shard_id);
2500 0 :
2501 0 : if config_path.exists() {
2502 : // New-style config takes precedence
2503 0 : let deserialized = Self::read_config(&config_path)?;
2504 0 : Ok(toml_edit::de::from_document::<LocationConf>(deserialized)?)
2505 0 : } else if legacy_config_path.exists() {
2506 : // Upgrade path: found an old-style configuration only
2507 0 : let deserialized = Self::read_config(&legacy_config_path)?;
2508 :
2509 0 : let mut tenant_conf = TenantConfOpt::default();
2510 0 : for (key, item) in deserialized.iter() {
2511 0 : match key {
2512 0 : "tenant_config" => {
2513 0 : tenant_conf = TenantConfOpt::try_from(item.to_owned()).context(format!("Failed to parse config from file '{legacy_config_path}' as pageserver config"))?;
2514 : }
2515 0 : _ => bail!(
2516 0 : "config file {legacy_config_path} has unrecognized pageserver option '{key}'"
2517 0 : ),
2518 : }
2519 : }
2520 :
2521 : // Legacy configs are implicitly in attached state, and do not support sharding
2522 0 : Ok(LocationConf::attached_single(
2523 0 : tenant_conf,
2524 0 : Generation::none(),
2525 0 : &models::ShardParameters::default(),
2526 0 : ))
2527 : } else {
2528 : // FIXME If the config file is not found, assume that we're attaching
2529 : // a detached tenant and config is passed via attach command.
2530 : // https://github.com/neondatabase/neon/issues/1555
2531 : // OR: we're loading after incomplete deletion that managed to remove config.
2532 0 : info!(
2533 0 : "tenant config not found in {} or {}",
2534 0 : config_path, legacy_config_path
2535 0 : );
2536 0 : Ok(LocationConf::default())
2537 : }
2538 0 : }
2539 :
2540 0 : fn read_config(path: &Utf8Path) -> anyhow::Result<toml_edit::Document> {
2541 0 : info!("loading tenant configuration from {path}");
2542 :
2543 : // load and parse file
2544 0 : let config = fs::read_to_string(path)
2545 0 : .with_context(|| format!("Failed to load config from path '{path}'"))?;
2546 :
2547 0 : config
2548 0 : .parse::<toml_edit::Document>()
2549 0 : .with_context(|| format!("Failed to parse config from file '{path}' as toml file"))
2550 0 : }
2551 :
2552 0 : #[tracing::instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug()))]
2553 : pub(super) async fn persist_tenant_config(
2554 : conf: &'static PageServerConf,
2555 : tenant_shard_id: &TenantShardId,
2556 : location_conf: &LocationConf,
2557 : ) -> anyhow::Result<()> {
2558 : let legacy_config_path = conf.tenant_config_path(tenant_shard_id);
2559 : let config_path = conf.tenant_location_config_path(tenant_shard_id);
2560 :
2561 : Self::persist_tenant_config_at(
2562 : tenant_shard_id,
2563 : &config_path,
2564 : &legacy_config_path,
2565 : location_conf,
2566 : )
2567 : .await
2568 : }
2569 :
2570 0 : #[tracing::instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug()))]
2571 : pub(super) async fn persist_tenant_config_at(
2572 : tenant_shard_id: &TenantShardId,
2573 : config_path: &Utf8Path,
2574 : legacy_config_path: &Utf8Path,
2575 : location_conf: &LocationConf,
2576 : ) -> anyhow::Result<()> {
2577 : if let LocationMode::Attached(attach_conf) = &location_conf.mode {
2578 : // The modern-style LocationConf config file requires a generation to be set. In case someone
2579 : // is running a pageserver without the infrastructure to set generations, write out the legacy-style
2580 : // config file that only contains TenantConf.
2581 : //
2582 : // This will eventually be removed in https://github.com/neondatabase/neon/issues/5388
2583 :
2584 : if attach_conf.generation.is_none() {
2585 0 : tracing::info!(
2586 0 : "Running without generations, writing legacy-style tenant config file"
2587 0 : );
2588 : Self::persist_tenant_config_legacy(
2589 : tenant_shard_id,
2590 : legacy_config_path,
2591 : &location_conf.tenant_conf,
2592 : )
2593 : .await?;
2594 :
2595 : return Ok(());
2596 : }
2597 : }
2598 :
2599 0 : debug!("persisting tenantconf to {config_path}");
2600 :
2601 : let mut conf_content = r#"# This file contains a specific per-tenant's config.
2602 : # It is read in case of pageserver restart.
2603 : "#
2604 : .to_string();
2605 :
2606 0 : fail::fail_point!("tenant-config-before-write", |_| {
2607 0 : anyhow::bail!("tenant-config-before-write");
2608 0 : });
2609 :
2610 : // Convert the config to a toml file.
2611 : conf_content += &toml_edit::ser::to_string_pretty(&location_conf)?;
2612 :
2613 : let temp_path = path_with_suffix_extension(config_path, TEMP_FILE_SUFFIX);
2614 :
2615 : let tenant_shard_id = *tenant_shard_id;
2616 : let config_path = config_path.to_owned();
2617 : let conf_content = conf_content.into_bytes();
2618 : VirtualFile::crashsafe_overwrite(config_path.clone(), temp_path, conf_content)
2619 : .await
2620 0 : .with_context(|| format!("write tenant {tenant_shard_id} config to {config_path}"))?;
2621 :
2622 : Ok(())
2623 : }
2624 :
2625 0 : #[tracing::instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug()))]
2626 : async fn persist_tenant_config_legacy(
2627 : tenant_shard_id: &TenantShardId,
2628 : target_config_path: &Utf8Path,
2629 : tenant_conf: &TenantConfOpt,
2630 : ) -> anyhow::Result<()> {
2631 0 : debug!("persisting tenantconf to {target_config_path}");
2632 :
2633 : let mut conf_content = r#"# This file contains a specific per-tenant's config.
2634 : # It is read in case of pageserver restart.
2635 :
2636 : [tenant_config]
2637 : "#
2638 : .to_string();
2639 :
2640 : // Convert the config to a toml file.
2641 : conf_content += &toml_edit::ser::to_string(&tenant_conf)?;
2642 :
2643 : let temp_path = path_with_suffix_extension(target_config_path, TEMP_FILE_SUFFIX);
2644 :
2645 : let tenant_shard_id = *tenant_shard_id;
2646 : let target_config_path = target_config_path.to_owned();
2647 : let conf_content = conf_content.into_bytes();
2648 : VirtualFile::crashsafe_overwrite(target_config_path.clone(), temp_path, conf_content)
2649 : .await
2650 0 : .with_context(|| {
2651 0 : format!("write tenant {tenant_shard_id} config to {target_config_path}")
2652 0 : })?;
2653 : Ok(())
2654 : }
2655 :
2656 : //
2657 : // How garbage collection works:
2658 : //
2659 : // +--bar------------->
2660 : // /
2661 : // +----+-----foo---------------->
2662 : // /
2663 : // ----main--+-------------------------->
2664 : // \
2665 : // +-----baz-------->
2666 : //
2667 : //
2668 : // 1. Grab 'gc_cs' mutex to prevent new timelines from being created while Timeline's
2669 : // `gc_infos` are being refreshed
2670 : // 2. Scan collected timelines, and on each timeline, make note of the
2671 : // all the points where other timelines have been branched off.
2672 : // We will refrain from removing page versions at those LSNs.
2673 : // 3. For each timeline, scan all layer files on the timeline.
2674 : // Remove all files for which a newer file exists and which
2675 : // don't cover any branch point LSNs.
2676 : //
2677 : // TODO:
2678 : // - if a relation has a non-incremental persistent layer on a child branch, then we
2679 : // don't need to keep that in the parent anymore. But currently
2680 : // we do.
2681 8 : async fn gc_iteration_internal(
2682 8 : &self,
2683 8 : target_timeline_id: Option<TimelineId>,
2684 8 : horizon: u64,
2685 8 : pitr: Duration,
2686 8 : cancel: &CancellationToken,
2687 8 : ctx: &RequestContext,
2688 8 : ) -> anyhow::Result<GcResult> {
2689 8 : let mut totals: GcResult = Default::default();
2690 8 : let now = Instant::now();
2691 :
2692 8 : let gc_timelines = match self
2693 8 : .refresh_gc_info_internal(target_timeline_id, horizon, pitr, cancel, ctx)
2694 0 : .await
2695 : {
2696 8 : Ok(result) => result,
2697 0 : Err(e) => {
2698 0 : if let Some(PageReconstructError::Cancelled) =
2699 0 : e.downcast_ref::<PageReconstructError>()
2700 : {
2701 : // Handle cancellation
2702 0 : totals.elapsed = now.elapsed();
2703 0 : return Ok(totals);
2704 : } else {
2705 : // Propagate other errors
2706 0 : return Err(e);
2707 : }
2708 : }
2709 : };
2710 :
2711 0 : failpoint_support::sleep_millis_async!("gc_iteration_internal_after_getting_gc_timelines");
2712 :
2713 : // If there is nothing to GC, we don't want any messages in the INFO log.
2714 8 : if !gc_timelines.is_empty() {
2715 8 : info!("{} timelines need GC", gc_timelines.len());
2716 : } else {
2717 0 : debug!("{} timelines need GC", gc_timelines.len());
2718 : }
2719 :
2720 : // Perform GC for each timeline.
2721 : //
2722 : // Note that we don't hold the `Tenant::gc_cs` lock here because we don't want to delay the
2723 : // branch creation task, which requires the GC lock. A GC iteration can run concurrently
2724 : // with branch creation.
2725 : //
2726 : // See comments in [`Tenant::branch_timeline`] for more information about why branch
2727 : // creation task can run concurrently with timeline's GC iteration.
2728 16 : for timeline in gc_timelines {
2729 8 : if task_mgr::is_shutdown_requested() || cancel.is_cancelled() {
2730 : // We were requested to shut down. Stop and return with the progress we
2731 : // made.
2732 0 : break;
2733 8 : }
2734 8 : let result = timeline.gc().await?;
2735 8 : totals += result;
2736 : }
2737 :
2738 8 : totals.elapsed = now.elapsed();
2739 8 : Ok(totals)
2740 8 : }
2741 :
2742 : /// Refreshes the Timeline::gc_info for all timelines, returning the
2743 : /// vector of timelines which have [`Timeline::get_last_record_lsn`] past
2744 : /// [`Tenant::get_gc_horizon`].
2745 : ///
2746 : /// This is usually executed as part of periodic gc, but can now be triggered more often.
2747 0 : pub async fn refresh_gc_info(
2748 0 : &self,
2749 0 : cancel: &CancellationToken,
2750 0 : ctx: &RequestContext,
2751 0 : ) -> anyhow::Result<Vec<Arc<Timeline>>> {
2752 0 : // since this method can now be called at different rates than the configured gc loop, it
2753 0 : // might be that these configuration values get applied faster than what it was previously,
2754 0 : // since these were only read from the gc task.
2755 0 : let horizon = self.get_gc_horizon();
2756 0 : let pitr = self.get_pitr_interval();
2757 0 :
2758 0 : // refresh all timelines
2759 0 : let target_timeline_id = None;
2760 0 :
2761 0 : self.refresh_gc_info_internal(target_timeline_id, horizon, pitr, cancel, ctx)
2762 0 : .await
2763 0 : }
2764 :
2765 8 : async fn refresh_gc_info_internal(
2766 8 : &self,
2767 8 : target_timeline_id: Option<TimelineId>,
2768 8 : horizon: u64,
2769 8 : pitr: Duration,
2770 8 : cancel: &CancellationToken,
2771 8 : ctx: &RequestContext,
2772 8 : ) -> anyhow::Result<Vec<Arc<Timeline>>> {
2773 : // grab mutex to prevent new timelines from being created here.
2774 8 : let gc_cs = self.gc_cs.lock().await;
2775 :
2776 : // Scan all timelines. For each timeline, remember the timeline ID and
2777 : // the branch point where it was created.
2778 8 : let (all_branchpoints, timeline_ids): (BTreeSet<(TimelineId, Lsn)>, _) = {
2779 8 : let timelines = self.timelines.lock().unwrap();
2780 8 : let mut all_branchpoints = BTreeSet::new();
2781 8 : let timeline_ids = {
2782 8 : if let Some(target_timeline_id) = target_timeline_id.as_ref() {
2783 8 : if timelines.get(target_timeline_id).is_none() {
2784 0 : bail!("gc target timeline does not exist")
2785 8 : }
2786 0 : };
2787 :
2788 8 : timelines
2789 8 : .iter()
2790 14 : .map(|(timeline_id, timeline_entry)| {
2791 6 : if let Some(ancestor_timeline_id) =
2792 14 : &timeline_entry.get_ancestor_timeline_id()
2793 : {
2794 : // If target_timeline is specified, we only need to know branchpoints of its children
2795 6 : if let Some(timeline_id) = target_timeline_id {
2796 6 : if ancestor_timeline_id == &timeline_id {
2797 6 : all_branchpoints.insert((
2798 6 : *ancestor_timeline_id,
2799 6 : timeline_entry.get_ancestor_lsn(),
2800 6 : ));
2801 6 : }
2802 : }
2803 : // Collect branchpoints for all timelines
2804 0 : else {
2805 0 : all_branchpoints.insert((
2806 0 : *ancestor_timeline_id,
2807 0 : timeline_entry.get_ancestor_lsn(),
2808 0 : ));
2809 0 : }
2810 8 : }
2811 :
2812 14 : *timeline_id
2813 14 : })
2814 8 : .collect::<Vec<_>>()
2815 8 : };
2816 8 : (all_branchpoints, timeline_ids)
2817 8 : };
2818 8 :
2819 8 : // Ok, we now know all the branch points.
2820 8 : // Update the GC information for each timeline.
2821 8 : let mut gc_timelines = Vec::with_capacity(timeline_ids.len());
2822 22 : for timeline_id in timeline_ids {
2823 : // Timeline is known to be local and loaded.
2824 14 : let timeline = self
2825 14 : .get_timeline(timeline_id, false)
2826 14 : .with_context(|| format!("Timeline {timeline_id} was not found"))?;
2827 :
2828 : // If target_timeline is specified, ignore all other timelines
2829 14 : if let Some(target_timeline_id) = target_timeline_id {
2830 14 : if timeline_id != target_timeline_id {
2831 6 : continue;
2832 8 : }
2833 0 : }
2834 :
2835 8 : if let Some(cutoff) = timeline.get_last_record_lsn().checked_sub(horizon) {
2836 8 : let branchpoints: Vec<Lsn> = all_branchpoints
2837 8 : .range((
2838 8 : Included((timeline_id, Lsn(0))),
2839 8 : Included((timeline_id, Lsn(u64::MAX))),
2840 8 : ))
2841 8 : .map(|&x| x.1)
2842 8 : .collect();
2843 8 : timeline
2844 8 : .update_gc_info(branchpoints, cutoff, pitr, cancel, ctx)
2845 0 : .await?;
2846 :
2847 8 : gc_timelines.push(timeline);
2848 0 : }
2849 : }
2850 8 : drop(gc_cs);
2851 8 : Ok(gc_timelines)
2852 8 : }
2853 :
2854 : /// A substitute for `branch_timeline` for use in unit tests.
2855 : /// The returned timeline will have state value `Active` to make various `anyhow::ensure!()`
2856 : /// calls pass, but, we do not actually call `.activate()` under the hood. So, none of the
2857 : /// timeline background tasks are launched, except the flush loop.
2858 : #[cfg(test)]
2859 214 : async fn branch_timeline_test(
2860 214 : &self,
2861 214 : src_timeline: &Arc<Timeline>,
2862 214 : dst_id: TimelineId,
2863 214 : start_lsn: Option<Lsn>,
2864 214 : ctx: &RequestContext,
2865 214 : ) -> Result<Arc<Timeline>, CreateTimelineError> {
2866 214 : let uninit_mark = self.create_timeline_uninit_mark(dst_id).unwrap();
2867 214 : let tl = self
2868 214 : .branch_timeline_impl(src_timeline, dst_id, start_lsn, uninit_mark, ctx)
2869 4 : .await?;
2870 210 : tl.set_state(TimelineState::Active);
2871 210 : Ok(tl)
2872 214 : }
2873 :
2874 : /// Branch an existing timeline.
2875 : ///
2876 : /// The caller is responsible for activating the returned timeline.
2877 0 : async fn branch_timeline(
2878 0 : &self,
2879 0 : src_timeline: &Arc<Timeline>,
2880 0 : dst_id: TimelineId,
2881 0 : start_lsn: Option<Lsn>,
2882 0 : timeline_uninit_mark: TimelineUninitMark<'_>,
2883 0 : ctx: &RequestContext,
2884 0 : ) -> Result<Arc<Timeline>, CreateTimelineError> {
2885 0 : self.branch_timeline_impl(src_timeline, dst_id, start_lsn, timeline_uninit_mark, ctx)
2886 0 : .await
2887 0 : }
2888 :
2889 214 : async fn branch_timeline_impl(
2890 214 : &self,
2891 214 : src_timeline: &Arc<Timeline>,
2892 214 : dst_id: TimelineId,
2893 214 : start_lsn: Option<Lsn>,
2894 214 : timeline_uninit_mark: TimelineUninitMark<'_>,
2895 214 : _ctx: &RequestContext,
2896 214 : ) -> Result<Arc<Timeline>, CreateTimelineError> {
2897 214 : let src_id = src_timeline.timeline_id;
2898 :
2899 : // We will validate our ancestor LSN in this function. Acquire the GC lock so that
2900 : // this check cannot race with GC, and the ancestor LSN is guaranteed to remain
2901 : // valid while we are creating the branch.
2902 214 : let _gc_cs = self.gc_cs.lock().await;
2903 :
2904 : // If no start LSN is specified, we branch the new timeline from the source timeline's last record LSN
2905 214 : let start_lsn = start_lsn.unwrap_or_else(|| {
2906 0 : let lsn = src_timeline.get_last_record_lsn();
2907 0 : info!("branching timeline {dst_id} from timeline {src_id} at last record LSN: {lsn}");
2908 0 : lsn
2909 214 : });
2910 214 :
2911 214 : // Ensure that `start_lsn` is valid, i.e. the LSN is within the PITR
2912 214 : // horizon on the source timeline
2913 214 : //
2914 214 : // We check it against both the planned GC cutoff stored in 'gc_info',
2915 214 : // and the 'latest_gc_cutoff' of the last GC that was performed. The
2916 214 : // planned GC cutoff in 'gc_info' is normally larger than
2917 214 : // 'latest_gc_cutoff_lsn', but beware of corner cases like if you just
2918 214 : // changed the GC settings for the tenant to make the PITR window
2919 214 : // larger, but some of the data was already removed by an earlier GC
2920 214 : // iteration.
2921 214 :
2922 214 : // check against last actual 'latest_gc_cutoff' first
2923 214 : let latest_gc_cutoff_lsn = src_timeline.get_latest_gc_cutoff_lsn();
2924 214 : src_timeline
2925 214 : .check_lsn_is_in_scope(start_lsn, &latest_gc_cutoff_lsn)
2926 214 : .context(format!(
2927 214 : "invalid branch start lsn: less than latest GC cutoff {}",
2928 214 : *latest_gc_cutoff_lsn,
2929 214 : ))
2930 214 : .map_err(CreateTimelineError::AncestorLsn)?;
2931 :
2932 : // and then the planned GC cutoff
2933 : {
2934 210 : let gc_info = src_timeline.gc_info.read().unwrap();
2935 210 : let cutoff = min(gc_info.pitr_cutoff, gc_info.horizon_cutoff);
2936 210 : if start_lsn < cutoff {
2937 0 : return Err(CreateTimelineError::AncestorLsn(anyhow::anyhow!(
2938 0 : "invalid branch start lsn: less than planned GC cutoff {cutoff}"
2939 0 : )));
2940 210 : }
2941 210 : }
2942 210 :
2943 210 : //
2944 210 : // The branch point is valid, and we are still holding the 'gc_cs' lock
2945 210 : // so that GC cannot advance the GC cutoff until we are finished.
2946 210 : // Proceed with the branch creation.
2947 210 : //
2948 210 :
2949 210 : // Determine prev-LSN for the new timeline. We can only determine it if
2950 210 : // the timeline was branched at the current end of the source timeline.
2951 210 : let RecordLsn {
2952 210 : last: src_last,
2953 210 : prev: src_prev,
2954 210 : } = src_timeline.get_last_record_rlsn();
2955 210 : let dst_prev = if src_last == start_lsn {
2956 200 : Some(src_prev)
2957 : } else {
2958 10 : None
2959 : };
2960 :
2961 : // Create the metadata file, noting the ancestor of the new timeline.
2962 : // There is initially no data in it, but all the read-calls know to look
2963 : // into the ancestor.
2964 210 : let metadata = TimelineMetadata::new(
2965 210 : start_lsn,
2966 210 : dst_prev,
2967 210 : Some(src_id),
2968 210 : start_lsn,
2969 210 : *src_timeline.latest_gc_cutoff_lsn.read(), // FIXME: should we hold onto this guard longer?
2970 210 : src_timeline.initdb_lsn,
2971 210 : src_timeline.pg_version,
2972 210 : );
2973 :
2974 210 : let uninitialized_timeline = self
2975 210 : .prepare_new_timeline(
2976 210 : dst_id,
2977 210 : &metadata,
2978 210 : timeline_uninit_mark,
2979 210 : start_lsn + 1,
2980 210 : Some(Arc::clone(src_timeline)),
2981 210 : )
2982 0 : .await?;
2983 :
2984 210 : let new_timeline = uninitialized_timeline.finish_creation()?;
2985 :
2986 : // Root timeline gets its layers during creation and uploads them along with the metadata.
2987 : // A branch timeline though, when created, can get no writes for some time, hence won't get any layers created.
2988 : // We still need to upload its metadata eagerly: if other nodes `attach` the tenant and miss this timeline, their GC
2989 : // could get incorrect information and remove more layers, than needed.
2990 : // See also https://github.com/neondatabase/neon/issues/3865
2991 210 : if let Some(remote_client) = new_timeline.remote_client.as_ref() {
2992 210 : remote_client
2993 210 : .schedule_index_upload_for_metadata_update(&metadata)
2994 210 : .context("branch initial metadata upload")?;
2995 0 : }
2996 :
2997 210 : Ok(new_timeline)
2998 214 : }
2999 :
3000 : /// For unit tests, make this visible so that other modules can directly create timelines
3001 : #[cfg(test)]
3002 4 : #[tracing::instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug(), %timeline_id))]
3003 : pub(crate) async fn bootstrap_timeline_test(
3004 : &self,
3005 : timeline_id: TimelineId,
3006 : pg_version: u32,
3007 : load_existing_initdb: Option<TimelineId>,
3008 : ctx: &RequestContext,
3009 : ) -> anyhow::Result<Arc<Timeline>> {
3010 : let uninit_mark = self.create_timeline_uninit_mark(timeline_id).unwrap();
3011 : self.bootstrap_timeline(
3012 : timeline_id,
3013 : pg_version,
3014 : load_existing_initdb,
3015 : uninit_mark,
3016 : ctx,
3017 : )
3018 : .await
3019 : }
3020 :
3021 0 : async fn upload_initdb(
3022 0 : &self,
3023 0 : timelines_path: &Utf8PathBuf,
3024 0 : pgdata_path: &Utf8PathBuf,
3025 0 : timeline_id: &TimelineId,
3026 0 : ) -> anyhow::Result<()> {
3027 0 : let Some(storage) = &self.remote_storage else {
3028 : // No remote storage? No upload.
3029 0 : return Ok(());
3030 : };
3031 :
3032 0 : let temp_path = timelines_path.join(format!(
3033 0 : "{INITDB_PATH}.upload-{timeline_id}.{TEMP_FILE_SUFFIX}"
3034 0 : ));
3035 :
3036 0 : scopeguard::defer! {
3037 0 : if let Err(e) = fs::remove_file(&temp_path) {
3038 0 : error!("Failed to remove temporary initdb archive '{temp_path}': {e}");
3039 0 : }
3040 : }
3041 :
3042 0 : let (pgdata_zstd, tar_zst_size) =
3043 0 : import_datadir::create_tar_zst(pgdata_path, &temp_path).await?;
3044 :
3045 0 : pausable_failpoint!("before-initdb-upload");
3046 :
3047 0 : backoff::retry(
3048 0 : || async {
3049 0 : self::remote_timeline_client::upload_initdb_dir(
3050 0 : storage,
3051 0 : &self.tenant_shard_id.tenant_id,
3052 0 : timeline_id,
3053 0 : pgdata_zstd.try_clone().await?,
3054 0 : tar_zst_size,
3055 0 : &self.cancel,
3056 : )
3057 0 : .await
3058 0 : },
3059 0 : |_| false,
3060 0 : 3,
3061 0 : u32::MAX,
3062 0 : "persist_initdb_tar_zst",
3063 0 : &self.cancel,
3064 0 : )
3065 0 : .await
3066 0 : .ok_or_else(|| anyhow::Error::new(TimeoutOrCancel::Cancel))
3067 0 : .and_then(|x| x)
3068 0 : }
3069 :
3070 : /// - run initdb to init temporary instance and get bootstrap data
3071 : /// - after initialization completes, tar up the temp dir and upload it to S3.
3072 : ///
3073 : /// The caller is responsible for activating the returned timeline.
3074 2 : async fn bootstrap_timeline(
3075 2 : &self,
3076 2 : timeline_id: TimelineId,
3077 2 : pg_version: u32,
3078 2 : load_existing_initdb: Option<TimelineId>,
3079 2 : timeline_uninit_mark: TimelineUninitMark<'_>,
3080 2 : ctx: &RequestContext,
3081 2 : ) -> anyhow::Result<Arc<Timeline>> {
3082 2 : // create a `tenant/{tenant_id}/timelines/basebackup-{timeline_id}.{TEMP_FILE_SUFFIX}/`
3083 2 : // temporary directory for basebackup files for the given timeline.
3084 2 :
3085 2 : let timelines_path = self.conf.timelines_path(&self.tenant_shard_id);
3086 2 : let pgdata_path = path_with_suffix_extension(
3087 2 : timelines_path.join(format!("basebackup-{timeline_id}")),
3088 2 : TEMP_FILE_SUFFIX,
3089 2 : );
3090 2 :
3091 2 : // an uninit mark was placed before, nothing else can access this timeline files
3092 2 : // current initdb was not run yet, so remove whatever was left from the previous runs
3093 2 : if pgdata_path.exists() {
3094 0 : fs::remove_dir_all(&pgdata_path).with_context(|| {
3095 0 : format!("Failed to remove already existing initdb directory: {pgdata_path}")
3096 0 : })?;
3097 2 : }
3098 : // this new directory is very temporary, set to remove it immediately after bootstrap, we don't need it
3099 2 : scopeguard::defer! {
3100 2 : if let Err(e) = fs::remove_dir_all(&pgdata_path) {
3101 : // this is unlikely, but we will remove the directory on pageserver restart or another bootstrap call
3102 0 : error!("Failed to remove temporary initdb directory '{pgdata_path}': {e}");
3103 2 : }
3104 : }
3105 2 : if let Some(existing_initdb_timeline_id) = load_existing_initdb {
3106 2 : let Some(storage) = &self.remote_storage else {
3107 0 : bail!("no storage configured but load_existing_initdb set to {existing_initdb_timeline_id}");
3108 : };
3109 2 : if existing_initdb_timeline_id != timeline_id {
3110 0 : let source_path = &remote_initdb_archive_path(
3111 0 : &self.tenant_shard_id.tenant_id,
3112 0 : &existing_initdb_timeline_id,
3113 0 : );
3114 0 : let dest_path =
3115 0 : &remote_initdb_archive_path(&self.tenant_shard_id.tenant_id, &timeline_id);
3116 0 :
3117 0 : // if this fails, it will get retried by retried control plane requests
3118 0 : storage
3119 0 : .copy_object(source_path, dest_path, &self.cancel)
3120 0 : .await
3121 0 : .context("copy initdb tar")?;
3122 2 : }
3123 2 : let (initdb_tar_zst_path, initdb_tar_zst) =
3124 2 : self::remote_timeline_client::download_initdb_tar_zst(
3125 2 : self.conf,
3126 2 : storage,
3127 2 : &self.tenant_shard_id,
3128 2 : &existing_initdb_timeline_id,
3129 2 : &self.cancel,
3130 2 : )
3131 714 : .await
3132 2 : .context("download initdb tar")?;
3133 :
3134 2 : scopeguard::defer! {
3135 2 : if let Err(e) = fs::remove_file(&initdb_tar_zst_path) {
3136 0 : error!("Failed to remove temporary initdb archive '{initdb_tar_zst_path}': {e}");
3137 2 : }
3138 : }
3139 :
3140 2 : let buf_read =
3141 2 : BufReader::with_capacity(remote_timeline_client::BUFFER_SIZE, initdb_tar_zst);
3142 2 : import_datadir::extract_tar_zst(&pgdata_path, buf_read)
3143 10182 : .await
3144 2 : .context("extract initdb tar")?;
3145 : } else {
3146 : // Init temporarily repo to get bootstrap data, this creates a directory in the `pgdata_path` path
3147 0 : run_initdb(self.conf, &pgdata_path, pg_version, &self.cancel).await?;
3148 :
3149 : // Upload the created data dir to S3
3150 0 : if self.tenant_shard_id().is_zero() {
3151 0 : self.upload_initdb(&timelines_path, &pgdata_path, &timeline_id)
3152 0 : .await?;
3153 0 : }
3154 : }
3155 2 : let pgdata_lsn = import_datadir::get_lsn_from_controlfile(&pgdata_path)?.align();
3156 2 :
3157 2 : // Import the contents of the data directory at the initial checkpoint
3158 2 : // LSN, and any WAL after that.
3159 2 : // Initdb lsn will be equal to last_record_lsn which will be set after import.
3160 2 : // Because we know it upfront avoid having an option or dummy zero value by passing it to the metadata.
3161 2 : let new_metadata = TimelineMetadata::new(
3162 2 : Lsn(0),
3163 2 : None,
3164 2 : None,
3165 2 : Lsn(0),
3166 2 : pgdata_lsn,
3167 2 : pgdata_lsn,
3168 2 : pg_version,
3169 2 : );
3170 2 : let raw_timeline = self
3171 2 : .prepare_new_timeline(
3172 2 : timeline_id,
3173 2 : &new_metadata,
3174 2 : timeline_uninit_mark,
3175 2 : pgdata_lsn,
3176 2 : None,
3177 2 : )
3178 0 : .await?;
3179 :
3180 2 : let tenant_shard_id = raw_timeline.owning_tenant.tenant_shard_id;
3181 2 : let unfinished_timeline = raw_timeline.raw_timeline()?;
3182 :
3183 2 : import_datadir::import_timeline_from_postgres_datadir(
3184 2 : unfinished_timeline,
3185 2 : &pgdata_path,
3186 2 : pgdata_lsn,
3187 2 : ctx,
3188 2 : )
3189 9318 : .await
3190 2 : .with_context(|| {
3191 0 : format!("Failed to import pgdatadir for timeline {tenant_shard_id}/{timeline_id}")
3192 2 : })?;
3193 :
3194 : // Flush the new layer files to disk, before we make the timeline as available to
3195 : // the outside world.
3196 : //
3197 : // Flush loop needs to be spawned in order to be able to flush.
3198 2 : unfinished_timeline.maybe_spawn_flush_loop();
3199 2 :
3200 2 : fail::fail_point!("before-checkpoint-new-timeline", |_| {
3201 0 : anyhow::bail!("failpoint before-checkpoint-new-timeline");
3202 2 : });
3203 :
3204 2 : unfinished_timeline
3205 2 : .freeze_and_flush()
3206 2 : .await
3207 2 : .with_context(|| {
3208 0 : format!(
3209 0 : "Failed to flush after pgdatadir import for timeline {tenant_shard_id}/{timeline_id}"
3210 0 : )
3211 2 : })?;
3212 :
3213 : // All done!
3214 2 : let timeline = raw_timeline.finish_creation()?;
3215 :
3216 2 : Ok(timeline)
3217 2 : }
3218 :
3219 : /// Call this before constructing a timeline, to build its required structures
3220 290 : fn build_timeline_resources(&self, timeline_id: TimelineId) -> TimelineResources {
3221 290 : let remote_client = if let Some(remote_storage) = self.remote_storage.as_ref() {
3222 290 : let remote_client = RemoteTimelineClient::new(
3223 290 : remote_storage.clone(),
3224 290 : self.deletion_queue_client.clone(),
3225 290 : self.conf,
3226 290 : self.tenant_shard_id,
3227 290 : timeline_id,
3228 290 : self.generation,
3229 290 : );
3230 290 : Some(remote_client)
3231 : } else {
3232 0 : None
3233 : };
3234 :
3235 290 : TimelineResources {
3236 290 : remote_client,
3237 290 : deletion_queue_client: self.deletion_queue_client.clone(),
3238 290 : timeline_get_throttle: self.timeline_get_throttle.clone(),
3239 290 : }
3240 290 : }
3241 :
3242 : /// Creates intermediate timeline structure and its files.
3243 : ///
3244 : /// An empty layer map is initialized, and new data and WAL can be imported starting
3245 : /// at 'disk_consistent_lsn'. After any initial data has been imported, call
3246 : /// `finish_creation` to insert the Timeline into the timelines map and to remove the
3247 : /// uninit mark file.
3248 290 : async fn prepare_new_timeline<'a>(
3249 290 : &'a self,
3250 290 : new_timeline_id: TimelineId,
3251 290 : new_metadata: &TimelineMetadata,
3252 290 : uninit_mark: TimelineUninitMark<'a>,
3253 290 : start_lsn: Lsn,
3254 290 : ancestor: Option<Arc<Timeline>>,
3255 290 : ) -> anyhow::Result<UninitializedTimeline> {
3256 290 : let tenant_shard_id = self.tenant_shard_id;
3257 290 :
3258 290 : let resources = self.build_timeline_resources(new_timeline_id);
3259 290 : if let Some(remote_client) = &resources.remote_client {
3260 290 : remote_client.init_upload_queue_for_empty_remote(new_metadata)?;
3261 0 : }
3262 :
3263 290 : let timeline_struct = self
3264 290 : .create_timeline_struct(
3265 290 : new_timeline_id,
3266 290 : new_metadata,
3267 290 : ancestor,
3268 290 : resources,
3269 290 : CreateTimelineCause::Load,
3270 290 : )
3271 290 : .context("Failed to create timeline data structure")?;
3272 :
3273 290 : timeline_struct.init_empty_layer_map(start_lsn);
3274 :
3275 290 : if let Err(e) = self.create_timeline_files(&uninit_mark.timeline_path).await {
3276 0 : error!("Failed to create initial files for timeline {tenant_shard_id}/{new_timeline_id}, cleaning up: {e:?}");
3277 0 : cleanup_timeline_directory(uninit_mark);
3278 0 : return Err(e);
3279 290 : }
3280 :
3281 0 : debug!(
3282 0 : "Successfully created initial files for timeline {tenant_shard_id}/{new_timeline_id}"
3283 0 : );
3284 :
3285 290 : Ok(UninitializedTimeline::new(
3286 290 : self,
3287 290 : new_timeline_id,
3288 290 : Some((timeline_struct, uninit_mark)),
3289 290 : ))
3290 290 : }
3291 :
3292 290 : async fn create_timeline_files(&self, timeline_path: &Utf8Path) -> anyhow::Result<()> {
3293 290 : crashsafe::create_dir(timeline_path).context("Failed to create timeline directory")?;
3294 :
3295 290 : fail::fail_point!("after-timeline-uninit-mark-creation", |_| {
3296 0 : anyhow::bail!("failpoint after-timeline-uninit-mark-creation");
3297 290 : });
3298 :
3299 290 : Ok(())
3300 290 : }
3301 :
3302 : /// Attempts to create an uninit mark file for the timeline initialization.
3303 : /// Bails, if the timeline is already loaded into the memory (i.e. initialized before), or the uninit mark file already exists.
3304 : ///
3305 : /// This way, we need to hold the timelines lock only for small amount of time during the mark check/creation per timeline init.
3306 296 : fn create_timeline_uninit_mark(
3307 296 : &self,
3308 296 : timeline_id: TimelineId,
3309 296 : ) -> Result<TimelineUninitMark, TimelineExclusionError> {
3310 296 : let tenant_shard_id = self.tenant_shard_id;
3311 296 :
3312 296 : let uninit_mark_path = self
3313 296 : .conf
3314 296 : .timeline_uninit_mark_file_path(tenant_shard_id, timeline_id);
3315 296 : let timeline_path = self.conf.timeline_path(&tenant_shard_id, &timeline_id);
3316 :
3317 296 : let uninit_mark = TimelineUninitMark::new(
3318 296 : self,
3319 296 : timeline_id,
3320 296 : uninit_mark_path.clone(),
3321 296 : timeline_path.clone(),
3322 296 : )?;
3323 :
3324 : // At this stage, we have got exclusive access to in-memory state for this timeline ID
3325 : // for creation.
3326 : // A timeline directory should never exist on disk already:
3327 : // - a previous failed creation would have cleaned up after itself
3328 : // - a pageserver restart would clean up timeline directories that don't have valid remote state
3329 : //
3330 : // Therefore it is an unexpected internal error to encounter a timeline directory already existing here,
3331 : // this error may indicate a bug in cleanup on failed creations.
3332 294 : if timeline_path.exists() {
3333 0 : return Err(TimelineExclusionError::Other(anyhow::anyhow!(
3334 0 : "Timeline directory already exists! This is a bug."
3335 0 : )));
3336 294 : }
3337 294 :
3338 294 : // Create the on-disk uninit mark _after_ the in-memory acquisition of the tenant ID: guarantees
3339 294 : // that during process runtime, colliding creations will be caught in-memory without getting
3340 294 : // as far as failing to write a file.
3341 294 : fs::OpenOptions::new()
3342 294 : .write(true)
3343 294 : .create_new(true)
3344 294 : .open(&uninit_mark_path)
3345 294 : .context("Failed to create uninit mark file")
3346 294 : .and_then(|_| {
3347 294 : crashsafe::fsync_file_and_parent(&uninit_mark_path)
3348 294 : .context("Failed to fsync uninit mark file")
3349 294 : })
3350 294 : .with_context(|| {
3351 0 : format!("Failed to crate uninit mark for timeline {tenant_shard_id}/{timeline_id}")
3352 294 : })?;
3353 :
3354 294 : Ok(uninit_mark)
3355 296 : }
3356 :
3357 : /// Gathers inputs from all of the timelines to produce a sizing model input.
3358 : ///
3359 : /// Future is cancellation safe. Only one calculation can be running at once per tenant.
3360 0 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
3361 : pub async fn gather_size_inputs(
3362 : &self,
3363 : // `max_retention_period` overrides the cutoff that is used to calculate the size
3364 : // (only if it is shorter than the real cutoff).
3365 : max_retention_period: Option<u64>,
3366 : cause: LogicalSizeCalculationCause,
3367 : cancel: &CancellationToken,
3368 : ctx: &RequestContext,
3369 : ) -> anyhow::Result<size::ModelInputs> {
3370 : let logical_sizes_at_once = self
3371 : .conf
3372 : .concurrent_tenant_size_logical_size_queries
3373 : .inner();
3374 :
3375 : // TODO: Having a single mutex block concurrent reads is not great for performance.
3376 : //
3377 : // But the only case where we need to run multiple of these at once is when we
3378 : // request a size for a tenant manually via API, while another background calculation
3379 : // is in progress (which is not a common case).
3380 : //
3381 : // See more for on the issue #2748 condenced out of the initial PR review.
3382 : let mut shared_cache = self.cached_logical_sizes.lock().await;
3383 :
3384 : size::gather_inputs(
3385 : self,
3386 : logical_sizes_at_once,
3387 : max_retention_period,
3388 : &mut shared_cache,
3389 : cause,
3390 : cancel,
3391 : ctx,
3392 : )
3393 : .await
3394 : }
3395 :
3396 : /// Calculate synthetic tenant size and cache the result.
3397 : /// This is periodically called by background worker.
3398 : /// result is cached in tenant struct
3399 0 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
3400 : pub async fn calculate_synthetic_size(
3401 : &self,
3402 : cause: LogicalSizeCalculationCause,
3403 : cancel: &CancellationToken,
3404 : ctx: &RequestContext,
3405 : ) -> anyhow::Result<u64> {
3406 : let inputs = self.gather_size_inputs(None, cause, cancel, ctx).await?;
3407 :
3408 : let size = inputs.calculate()?;
3409 :
3410 : self.set_cached_synthetic_size(size);
3411 :
3412 : Ok(size)
3413 : }
3414 :
3415 : /// Cache given synthetic size and update the metric value
3416 0 : pub fn set_cached_synthetic_size(&self, size: u64) {
3417 0 : self.cached_synthetic_tenant_size
3418 0 : .store(size, Ordering::Relaxed);
3419 :
3420 : // Only shard zero should be calculating synthetic sizes
3421 0 : debug_assert!(self.shard_identity.is_zero());
3422 :
3423 0 : TENANT_SYNTHETIC_SIZE_METRIC
3424 0 : .get_metric_with_label_values(&[&self.tenant_shard_id.tenant_id.to_string()])
3425 0 : .unwrap()
3426 0 : .set(size);
3427 0 : }
3428 :
3429 0 : pub fn cached_synthetic_size(&self) -> u64 {
3430 0 : self.cached_synthetic_tenant_size.load(Ordering::Relaxed)
3431 0 : }
3432 :
3433 : /// Flush any in-progress layers, schedule uploads, and wait for uploads to complete.
3434 : ///
3435 : /// This function can take a long time: callers should wrap it in a timeout if calling
3436 : /// from an external API handler.
3437 : ///
3438 : /// Cancel-safety: cancelling this function may leave I/O running, but such I/O is
3439 : /// still bounded by tenant/timeline shutdown.
3440 0 : #[tracing::instrument(skip_all)]
3441 : pub(crate) async fn flush_remote(&self) -> anyhow::Result<()> {
3442 : let timelines = self.timelines.lock().unwrap().clone();
3443 :
3444 0 : async fn flush_timeline(_gate: GateGuard, timeline: Arc<Timeline>) -> anyhow::Result<()> {
3445 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Flushing...");
3446 0 : timeline.freeze_and_flush().await?;
3447 0 : tracing::info!(timeline_id=%timeline.timeline_id, "Waiting for uploads...");
3448 0 : if let Some(client) = &timeline.remote_client {
3449 0 : client.wait_completion().await?;
3450 0 : }
3451 :
3452 0 : Ok(())
3453 0 : }
3454 :
3455 : // We do not use a JoinSet for these tasks, because we don't want them to be
3456 : // aborted when this function's future is cancelled: they should stay alive
3457 : // holding their GateGuard until they complete, to ensure their I/Os complete
3458 : // before Timeline shutdown completes.
3459 : let mut results = FuturesUnordered::new();
3460 :
3461 : for (_timeline_id, timeline) in timelines {
3462 : // Run each timeline's flush in a task holding the timeline's gate: this
3463 : // means that if this function's future is cancelled, the Timeline shutdown
3464 : // will still wait for any I/O in here to complete.
3465 : let Ok(gate) = timeline.gate.enter() else {
3466 : continue;
3467 : };
3468 0 : let jh = tokio::task::spawn(async move { flush_timeline(gate, timeline).await });
3469 : results.push(jh);
3470 : }
3471 :
3472 : while let Some(r) = results.next().await {
3473 : if let Err(e) = r {
3474 : if !e.is_cancelled() && !e.is_panic() {
3475 0 : tracing::error!("unexpected join error: {e:?}");
3476 : }
3477 : }
3478 : }
3479 :
3480 : // The flushes we did above were just writes, but the Tenant might have had
3481 : // pending deletions as well from recent compaction/gc: we want to flush those
3482 : // as well. This requires flushing the global delete queue. This is cheap
3483 : // because it's typically a no-op.
3484 : match self.deletion_queue_client.flush_execute().await {
3485 : Ok(_) => {}
3486 : Err(DeletionQueueError::ShuttingDown) => {}
3487 : }
3488 :
3489 : Ok(())
3490 : }
3491 :
3492 0 : pub(crate) fn get_tenant_conf(&self) -> TenantConfOpt {
3493 0 : self.tenant_conf.read().unwrap().tenant_conf.clone()
3494 0 : }
3495 : }
3496 :
3497 : /// Create the cluster temporarily in 'initdbpath' directory inside the repository
3498 : /// to get bootstrap data for timeline initialization.
3499 0 : async fn run_initdb(
3500 0 : conf: &'static PageServerConf,
3501 0 : initdb_target_dir: &Utf8Path,
3502 0 : pg_version: u32,
3503 0 : cancel: &CancellationToken,
3504 0 : ) -> Result<(), InitdbError> {
3505 0 : let initdb_bin_path = conf
3506 0 : .pg_bin_dir(pg_version)
3507 0 : .map_err(InitdbError::Other)?
3508 0 : .join("initdb");
3509 0 : let initdb_lib_dir = conf.pg_lib_dir(pg_version).map_err(InitdbError::Other)?;
3510 0 : info!(
3511 0 : "running {} in {}, libdir: {}",
3512 0 : initdb_bin_path, initdb_target_dir, initdb_lib_dir,
3513 0 : );
3514 :
3515 0 : let _permit = INIT_DB_SEMAPHORE.acquire().await;
3516 :
3517 0 : let initdb_command = tokio::process::Command::new(&initdb_bin_path)
3518 0 : .args(["-D", initdb_target_dir.as_ref()])
3519 0 : .args(["-U", &conf.superuser])
3520 0 : .args(["-E", "utf8"])
3521 0 : .arg("--no-instructions")
3522 0 : .arg("--no-sync")
3523 0 : .env_clear()
3524 0 : .env("LD_LIBRARY_PATH", &initdb_lib_dir)
3525 0 : .env("DYLD_LIBRARY_PATH", &initdb_lib_dir)
3526 0 : .stdin(std::process::Stdio::null())
3527 0 : // stdout invocation produces the same output every time, we don't need it
3528 0 : .stdout(std::process::Stdio::null())
3529 0 : // we would be interested in the stderr output, if there was any
3530 0 : .stderr(std::process::Stdio::piped())
3531 0 : .spawn()?;
3532 :
3533 : // Ideally we'd select here with the cancellation token, but the problem is that
3534 : // we can't safely terminate initdb: it launches processes of its own, and killing
3535 : // initdb doesn't kill them. After we return from this function, we want the target
3536 : // directory to be able to be cleaned up.
3537 : // See https://github.com/neondatabase/neon/issues/6385
3538 0 : let initdb_output = initdb_command.wait_with_output().await?;
3539 0 : if !initdb_output.status.success() {
3540 0 : return Err(InitdbError::Failed(
3541 0 : initdb_output.status,
3542 0 : initdb_output.stderr,
3543 0 : ));
3544 0 : }
3545 0 :
3546 0 : // This isn't true cancellation support, see above. Still return an error to
3547 0 : // excercise the cancellation code path.
3548 0 : if cancel.is_cancelled() {
3549 0 : return Err(InitdbError::Cancelled);
3550 0 : }
3551 0 :
3552 0 : Ok(())
3553 0 : }
3554 :
3555 : impl Drop for Tenant {
3556 88 : fn drop(&mut self) {
3557 88 : remove_tenant_metrics(&self.tenant_shard_id);
3558 88 : }
3559 : }
3560 : /// Dump contents of a layer file to stdout.
3561 0 : pub async fn dump_layerfile_from_path(
3562 0 : path: &Utf8Path,
3563 0 : verbose: bool,
3564 0 : ctx: &RequestContext,
3565 0 : ) -> anyhow::Result<()> {
3566 : use std::os::unix::fs::FileExt;
3567 :
3568 : // All layer files start with a two-byte "magic" value, to identify the kind of
3569 : // file.
3570 0 : let file = File::open(path)?;
3571 0 : let mut header_buf = [0u8; 2];
3572 0 : file.read_exact_at(&mut header_buf, 0)?;
3573 :
3574 0 : match u16::from_be_bytes(header_buf) {
3575 : crate::IMAGE_FILE_MAGIC => {
3576 0 : ImageLayer::new_for_path(path, file)?
3577 0 : .dump(verbose, ctx)
3578 0 : .await?
3579 : }
3580 : crate::DELTA_FILE_MAGIC => {
3581 0 : DeltaLayer::new_for_path(path, file)?
3582 0 : .dump(verbose, ctx)
3583 0 : .await?
3584 : }
3585 0 : magic => bail!("unrecognized magic identifier: {:?}", magic),
3586 : }
3587 :
3588 0 : Ok(())
3589 0 : }
3590 :
3591 : #[cfg(test)]
3592 : pub(crate) mod harness {
3593 : use bytes::{Bytes, BytesMut};
3594 : use camino::Utf8PathBuf;
3595 : use once_cell::sync::OnceCell;
3596 : use pageserver_api::models::ShardParameters;
3597 : use pageserver_api::shard::ShardIndex;
3598 : use std::fs;
3599 : use std::sync::Arc;
3600 : use utils::logging;
3601 : use utils::lsn::Lsn;
3602 :
3603 : use crate::deletion_queue::mock::MockDeletionQueue;
3604 : use crate::walredo::apply_neon;
3605 : use crate::{
3606 : config::PageServerConf, repository::Key, tenant::Tenant, walrecord::NeonWalRecord,
3607 : };
3608 :
3609 : use super::*;
3610 : use crate::tenant::config::{TenantConf, TenantConfOpt};
3611 : use hex_literal::hex;
3612 : use utils::id::{TenantId, TimelineId};
3613 :
3614 : pub const TIMELINE_ID: TimelineId =
3615 : TimelineId::from_array(hex!("11223344556677881122334455667788"));
3616 : pub const NEW_TIMELINE_ID: TimelineId =
3617 : TimelineId::from_array(hex!("AA223344556677881122334455667788"));
3618 :
3619 : /// Convenience function to create a page image with given string as the only content
3620 2708254 : pub fn test_img(s: &str) -> Bytes {
3621 2708254 : let mut buf = BytesMut::new();
3622 2708254 : buf.extend_from_slice(s.as_bytes());
3623 2708254 : buf.resize(64, 0);
3624 2708254 :
3625 2708254 : buf.freeze()
3626 2708254 : }
3627 :
3628 : impl From<TenantConf> for TenantConfOpt {
3629 88 : fn from(tenant_conf: TenantConf) -> Self {
3630 88 : Self {
3631 88 : checkpoint_distance: Some(tenant_conf.checkpoint_distance),
3632 88 : checkpoint_timeout: Some(tenant_conf.checkpoint_timeout),
3633 88 : compaction_target_size: Some(tenant_conf.compaction_target_size),
3634 88 : compaction_period: Some(tenant_conf.compaction_period),
3635 88 : compaction_threshold: Some(tenant_conf.compaction_threshold),
3636 88 : compaction_algorithm: Some(tenant_conf.compaction_algorithm),
3637 88 : gc_horizon: Some(tenant_conf.gc_horizon),
3638 88 : gc_period: Some(tenant_conf.gc_period),
3639 88 : image_creation_threshold: Some(tenant_conf.image_creation_threshold),
3640 88 : pitr_interval: Some(tenant_conf.pitr_interval),
3641 88 : walreceiver_connect_timeout: Some(tenant_conf.walreceiver_connect_timeout),
3642 88 : lagging_wal_timeout: Some(tenant_conf.lagging_wal_timeout),
3643 88 : max_lsn_wal_lag: Some(tenant_conf.max_lsn_wal_lag),
3644 88 : trace_read_requests: Some(tenant_conf.trace_read_requests),
3645 88 : eviction_policy: Some(tenant_conf.eviction_policy),
3646 88 : min_resident_size_override: tenant_conf.min_resident_size_override,
3647 88 : evictions_low_residence_duration_metric_threshold: Some(
3648 88 : tenant_conf.evictions_low_residence_duration_metric_threshold,
3649 88 : ),
3650 88 : heatmap_period: Some(tenant_conf.heatmap_period),
3651 88 : lazy_slru_download: Some(tenant_conf.lazy_slru_download),
3652 88 : timeline_get_throttle: Some(tenant_conf.timeline_get_throttle),
3653 88 : }
3654 88 : }
3655 : }
3656 :
3657 : pub struct TenantHarness {
3658 : pub conf: &'static PageServerConf,
3659 : pub tenant_conf: TenantConf,
3660 : pub tenant_shard_id: TenantShardId,
3661 : pub generation: Generation,
3662 : pub shard: ShardIndex,
3663 : pub remote_storage: GenericRemoteStorage,
3664 : pub remote_fs_dir: Utf8PathBuf,
3665 : pub deletion_queue: MockDeletionQueue,
3666 : }
3667 :
3668 : static LOG_HANDLE: OnceCell<()> = OnceCell::new();
3669 :
3670 95 : pub(crate) fn setup_logging() {
3671 95 : LOG_HANDLE.get_or_init(|| {
3672 95 : logging::init(
3673 95 : logging::LogFormat::Test,
3674 95 : // enable it in case the tests exercise code paths that use
3675 95 : // debug_assert_current_span_has_tenant_and_timeline_id
3676 95 : logging::TracingErrorLayerEnablement::EnableWithRustLogFilter,
3677 95 : logging::Output::Stdout,
3678 95 : )
3679 95 : .expect("Failed to init test logging")
3680 95 : });
3681 95 : }
3682 :
3683 : impl TenantHarness {
3684 89 : pub fn create(test_name: &'static str) -> anyhow::Result<Self> {
3685 89 : setup_logging();
3686 89 :
3687 89 : let repo_dir = PageServerConf::test_repo_dir(test_name);
3688 89 : let _ = fs::remove_dir_all(&repo_dir);
3689 89 : fs::create_dir_all(&repo_dir)?;
3690 :
3691 89 : let conf = PageServerConf::dummy_conf(repo_dir);
3692 89 : // Make a static copy of the config. This can never be free'd, but that's
3693 89 : // OK in a test.
3694 89 : let conf: &'static PageServerConf = Box::leak(Box::new(conf));
3695 89 :
3696 89 : // Disable automatic GC and compaction to make the unit tests more deterministic.
3697 89 : // The tests perform them manually if needed.
3698 89 : let tenant_conf = TenantConf {
3699 89 : gc_period: Duration::ZERO,
3700 89 : compaction_period: Duration::ZERO,
3701 89 : ..TenantConf::default()
3702 89 : };
3703 89 :
3704 89 : let tenant_id = TenantId::generate();
3705 89 : let tenant_shard_id = TenantShardId::unsharded(tenant_id);
3706 89 : fs::create_dir_all(conf.tenant_path(&tenant_shard_id))?;
3707 89 : fs::create_dir_all(conf.timelines_path(&tenant_shard_id))?;
3708 :
3709 : use remote_storage::{RemoteStorageConfig, RemoteStorageKind};
3710 89 : let remote_fs_dir = conf.workdir.join("localfs");
3711 89 : std::fs::create_dir_all(&remote_fs_dir).unwrap();
3712 89 : let config = RemoteStorageConfig {
3713 89 : storage: RemoteStorageKind::LocalFs(remote_fs_dir.clone()),
3714 89 : timeout: RemoteStorageConfig::DEFAULT_TIMEOUT,
3715 89 : };
3716 89 : let remote_storage = GenericRemoteStorage::from_config(&config).unwrap();
3717 89 : let deletion_queue = MockDeletionQueue::new(Some(remote_storage.clone()));
3718 89 :
3719 89 : Ok(Self {
3720 89 : conf,
3721 89 : tenant_conf,
3722 89 : tenant_shard_id,
3723 89 : generation: Generation::new(0xdeadbeef),
3724 89 : shard: ShardIndex::unsharded(),
3725 89 : remote_storage,
3726 89 : remote_fs_dir,
3727 89 : deletion_queue,
3728 89 : })
3729 89 : }
3730 :
3731 8 : pub fn span(&self) -> tracing::Span {
3732 8 : info_span!("TenantHarness", tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug())
3733 8 : }
3734 :
3735 86 : pub(crate) async fn load(&self) -> (Arc<Tenant>, RequestContext) {
3736 86 : let ctx = RequestContext::new(TaskKind::UnitTest, DownloadBehavior::Error);
3737 86 : (
3738 86 : self.do_try_load(&ctx)
3739 26 : .await
3740 86 : .expect("failed to load test tenant"),
3741 86 : ctx,
3742 86 : )
3743 86 : }
3744 :
3745 176 : #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))]
3746 : pub(crate) async fn do_try_load(
3747 : &self,
3748 : ctx: &RequestContext,
3749 : ) -> anyhow::Result<Arc<Tenant>> {
3750 : let walredo_mgr = Arc::new(WalRedoManager::from(TestRedoManager));
3751 :
3752 : let tenant = Arc::new(Tenant::new(
3753 : TenantState::Loading,
3754 : self.conf,
3755 : AttachedTenantConf::try_from(LocationConf::attached_single(
3756 : TenantConfOpt::from(self.tenant_conf.clone()),
3757 : self.generation,
3758 : &ShardParameters::default(),
3759 : ))
3760 : .unwrap(),
3761 : // This is a legacy/test code path: sharding isn't supported here.
3762 : ShardIdentity::unsharded(),
3763 : Some(walredo_mgr),
3764 : self.tenant_shard_id,
3765 : Some(self.remote_storage.clone()),
3766 : self.deletion_queue.new_client(),
3767 : ));
3768 :
3769 : let preload = tenant
3770 : .preload(&self.remote_storage, CancellationToken::new())
3771 : .await?;
3772 : tenant.attach(Some(preload), SpawnMode::Normal, ctx).await?;
3773 :
3774 : tenant.state.send_replace(TenantState::Active);
3775 : for timeline in tenant.timelines.lock().unwrap().values() {
3776 : timeline.set_state(TimelineState::Active);
3777 : }
3778 : Ok(tenant)
3779 : }
3780 :
3781 4 : pub fn timeline_path(&self, timeline_id: &TimelineId) -> Utf8PathBuf {
3782 4 : self.conf.timeline_path(&self.tenant_shard_id, timeline_id)
3783 4 : }
3784 : }
3785 :
3786 : // Mock WAL redo manager that doesn't do much
3787 : pub(crate) struct TestRedoManager;
3788 :
3789 : impl TestRedoManager {
3790 : /// # Cancel-Safety
3791 : ///
3792 : /// This method is cancellation-safe.
3793 6 : pub async fn request_redo(
3794 6 : &self,
3795 6 : key: Key,
3796 6 : lsn: Lsn,
3797 6 : base_img: Option<(Lsn, Bytes)>,
3798 6 : records: Vec<(Lsn, NeonWalRecord)>,
3799 6 : _pg_version: u32,
3800 6 : ) -> anyhow::Result<Bytes> {
3801 10 : let records_neon = records.iter().all(|r| apply_neon::can_apply_in_neon(&r.1));
3802 6 : if records_neon {
3803 : // For Neon wal records, we can decode without spawning postgres, so do so.
3804 6 : let base_img = base_img.expect("Neon WAL redo requires base image").1;
3805 6 : let mut page = BytesMut::new();
3806 6 : page.extend_from_slice(&base_img);
3807 16 : for (_record_lsn, record) in records {
3808 10 : apply_neon::apply_in_neon(&record, key, &mut page)?;
3809 : }
3810 6 : Ok(page.freeze())
3811 : } else {
3812 : // We never spawn a postgres walredo process in unit tests: just log what we might have done.
3813 0 : let s = format!(
3814 0 : "redo for {} to get to {}, with {} and {} records",
3815 0 : key,
3816 0 : lsn,
3817 0 : if base_img.is_some() {
3818 0 : "base image"
3819 : } else {
3820 0 : "no base image"
3821 : },
3822 0 : records.len()
3823 0 : );
3824 0 : println!("{s}");
3825 0 :
3826 0 : Ok(test_img(&s))
3827 : }
3828 6 : }
3829 : }
3830 : }
3831 :
3832 : #[cfg(test)]
3833 : mod tests {
3834 : use super::*;
3835 : use crate::keyspace::KeySpaceAccum;
3836 : use crate::repository::{Key, Value};
3837 : use crate::tenant::harness::*;
3838 : use crate::DEFAULT_PG_VERSION;
3839 : use bytes::BytesMut;
3840 : use hex_literal::hex;
3841 : use once_cell::sync::Lazy;
3842 : use pageserver_api::keyspace::KeySpace;
3843 : use rand::{thread_rng, Rng};
3844 : use tokio_util::sync::CancellationToken;
3845 :
3846 : static TEST_KEY: Lazy<Key> =
3847 18 : Lazy::new(|| Key::from_slice(&hex!("010000000033333333444444445500000001")));
3848 :
3849 2 : #[tokio::test]
3850 2 : async fn test_basic() -> anyhow::Result<()> {
3851 2 : let (tenant, ctx) = TenantHarness::create("test_basic")?.load().await;
3852 2 : let tline = tenant
3853 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
3854 6 : .await?;
3855 2 :
3856 2 : let writer = tline.writer().await;
3857 2 : writer
3858 2 : .put(
3859 2 : *TEST_KEY,
3860 2 : Lsn(0x10),
3861 2 : &Value::Image(test_img("foo at 0x10")),
3862 2 : &ctx,
3863 2 : )
3864 2 : .await?;
3865 2 : writer.finish_write(Lsn(0x10));
3866 2 : drop(writer);
3867 2 :
3868 2 : let writer = tline.writer().await;
3869 2 : writer
3870 2 : .put(
3871 2 : *TEST_KEY,
3872 2 : Lsn(0x20),
3873 2 : &Value::Image(test_img("foo at 0x20")),
3874 2 : &ctx,
3875 2 : )
3876 2 : .await?;
3877 2 : writer.finish_write(Lsn(0x20));
3878 2 : drop(writer);
3879 2 :
3880 2 : assert_eq!(
3881 2 : tline.get(*TEST_KEY, Lsn(0x10), &ctx).await?,
3882 2 : test_img("foo at 0x10")
3883 2 : );
3884 2 : assert_eq!(
3885 2 : tline.get(*TEST_KEY, Lsn(0x1f), &ctx).await?,
3886 2 : test_img("foo at 0x10")
3887 2 : );
3888 2 : assert_eq!(
3889 2 : tline.get(*TEST_KEY, Lsn(0x20), &ctx).await?,
3890 2 : test_img("foo at 0x20")
3891 2 : );
3892 2 :
3893 2 : Ok(())
3894 2 : }
3895 :
3896 2 : #[tokio::test]
3897 2 : async fn no_duplicate_timelines() -> anyhow::Result<()> {
3898 2 : let (tenant, ctx) = TenantHarness::create("no_duplicate_timelines")?
3899 2 : .load()
3900 2 : .await;
3901 2 : let _ = tenant
3902 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
3903 6 : .await?;
3904 2 :
3905 2 : match tenant
3906 2 : .create_empty_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
3907 2 : .await
3908 2 : {
3909 2 : Ok(_) => panic!("duplicate timeline creation should fail"),
3910 2 : Err(e) => assert_eq!(e.to_string(), "Already exists".to_string()),
3911 2 : }
3912 2 :
3913 2 : Ok(())
3914 2 : }
3915 :
3916 : /// Convenience function to create a page image with given string as the only content
3917 10 : pub fn test_value(s: &str) -> Value {
3918 10 : let mut buf = BytesMut::new();
3919 10 : buf.extend_from_slice(s.as_bytes());
3920 10 : Value::Image(buf.freeze())
3921 10 : }
3922 :
3923 : ///
3924 : /// Test branch creation
3925 : ///
3926 2 : #[tokio::test]
3927 2 : async fn test_branch() -> anyhow::Result<()> {
3928 2 : use std::str::from_utf8;
3929 2 :
3930 2 : let (tenant, ctx) = TenantHarness::create("test_branch")?.load().await;
3931 2 : let tline = tenant
3932 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
3933 6 : .await?;
3934 2 : let writer = tline.writer().await;
3935 2 :
3936 2 : #[allow(non_snake_case)]
3937 2 : let TEST_KEY_A: Key = Key::from_hex("110000000033333333444444445500000001").unwrap();
3938 2 : #[allow(non_snake_case)]
3939 2 : let TEST_KEY_B: Key = Key::from_hex("110000000033333333444444445500000002").unwrap();
3940 2 :
3941 2 : // Insert a value on the timeline
3942 2 : writer
3943 2 : .put(TEST_KEY_A, Lsn(0x20), &test_value("foo at 0x20"), &ctx)
3944 2 : .await?;
3945 2 : writer
3946 2 : .put(TEST_KEY_B, Lsn(0x20), &test_value("foobar at 0x20"), &ctx)
3947 2 : .await?;
3948 2 : writer.finish_write(Lsn(0x20));
3949 2 :
3950 2 : writer
3951 2 : .put(TEST_KEY_A, Lsn(0x30), &test_value("foo at 0x30"), &ctx)
3952 2 : .await?;
3953 2 : writer.finish_write(Lsn(0x30));
3954 2 : writer
3955 2 : .put(TEST_KEY_A, Lsn(0x40), &test_value("foo at 0x40"), &ctx)
3956 2 : .await?;
3957 2 : writer.finish_write(Lsn(0x40));
3958 2 :
3959 2 : //assert_current_logical_size(&tline, Lsn(0x40));
3960 2 :
3961 2 : // Branch the history, modify relation differently on the new timeline
3962 2 : tenant
3963 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x30)), &ctx)
3964 2 : .await?;
3965 2 : let newtline = tenant
3966 2 : .get_timeline(NEW_TIMELINE_ID, true)
3967 2 : .expect("Should have a local timeline");
3968 2 : let new_writer = newtline.writer().await;
3969 2 : new_writer
3970 2 : .put(TEST_KEY_A, Lsn(0x40), &test_value("bar at 0x40"), &ctx)
3971 2 : .await?;
3972 2 : new_writer.finish_write(Lsn(0x40));
3973 2 :
3974 2 : // Check page contents on both branches
3975 2 : assert_eq!(
3976 2 : from_utf8(&tline.get(TEST_KEY_A, Lsn(0x40), &ctx).await?)?,
3977 2 : "foo at 0x40"
3978 2 : );
3979 2 : assert_eq!(
3980 2 : from_utf8(&newtline.get(TEST_KEY_A, Lsn(0x40), &ctx).await?)?,
3981 2 : "bar at 0x40"
3982 2 : );
3983 2 : assert_eq!(
3984 2 : from_utf8(&newtline.get(TEST_KEY_B, Lsn(0x40), &ctx).await?)?,
3985 2 : "foobar at 0x20"
3986 2 : );
3987 2 :
3988 2 : //assert_current_logical_size(&tline, Lsn(0x40));
3989 2 :
3990 2 : Ok(())
3991 2 : }
3992 :
3993 20 : async fn make_some_layers(
3994 20 : tline: &Timeline,
3995 20 : start_lsn: Lsn,
3996 20 : ctx: &RequestContext,
3997 20 : ) -> anyhow::Result<()> {
3998 20 : let mut lsn = start_lsn;
3999 : {
4000 20 : let writer = tline.writer().await;
4001 : // Create a relation on the timeline
4002 20 : writer
4003 20 : .put(
4004 20 : *TEST_KEY,
4005 20 : lsn,
4006 20 : &Value::Image(test_img(&format!("foo at {}", lsn))),
4007 20 : ctx,
4008 20 : )
4009 10 : .await?;
4010 20 : writer.finish_write(lsn);
4011 20 : lsn += 0x10;
4012 20 : writer
4013 20 : .put(
4014 20 : *TEST_KEY,
4015 20 : lsn,
4016 20 : &Value::Image(test_img(&format!("foo at {}", lsn))),
4017 20 : ctx,
4018 20 : )
4019 0 : .await?;
4020 20 : writer.finish_write(lsn);
4021 20 : lsn += 0x10;
4022 20 : }
4023 20 : tline.freeze_and_flush().await?;
4024 : {
4025 20 : let writer = tline.writer().await;
4026 20 : writer
4027 20 : .put(
4028 20 : *TEST_KEY,
4029 20 : lsn,
4030 20 : &Value::Image(test_img(&format!("foo at {}", lsn))),
4031 20 : ctx,
4032 20 : )
4033 10 : .await?;
4034 20 : writer.finish_write(lsn);
4035 20 : lsn += 0x10;
4036 20 : writer
4037 20 : .put(
4038 20 : *TEST_KEY,
4039 20 : lsn,
4040 20 : &Value::Image(test_img(&format!("foo at {}", lsn))),
4041 20 : ctx,
4042 20 : )
4043 0 : .await?;
4044 20 : writer.finish_write(lsn);
4045 20 : }
4046 20 : tline.freeze_and_flush().await
4047 20 : }
4048 :
4049 2 : #[tokio::test]
4050 2 : async fn test_prohibit_branch_creation_on_garbage_collected_data() -> anyhow::Result<()> {
4051 2 : let (tenant, ctx) =
4052 2 : TenantHarness::create("test_prohibit_branch_creation_on_garbage_collected_data")?
4053 2 : .load()
4054 2 : .await;
4055 2 : let tline = tenant
4056 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
4057 6 : .await?;
4058 6 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
4059 2 :
4060 2 : // this removes layers before lsn 40 (50 minus 10), so there are two remaining layers, image and delta for 31-50
4061 2 : // FIXME: this doesn't actually remove any layer currently, given how the flushing
4062 2 : // and compaction works. But it does set the 'cutoff' point so that the cross check
4063 2 : // below should fail.
4064 2 : tenant
4065 2 : .gc_iteration(
4066 2 : Some(TIMELINE_ID),
4067 2 : 0x10,
4068 2 : Duration::ZERO,
4069 2 : &CancellationToken::new(),
4070 2 : &ctx,
4071 2 : )
4072 2 : .await?;
4073 2 :
4074 2 : // try to branch at lsn 25, should fail because we already garbage collected the data
4075 2 : match tenant
4076 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x25)), &ctx)
4077 2 : .await
4078 2 : {
4079 2 : Ok(_) => panic!("branching should have failed"),
4080 2 : Err(err) => {
4081 2 : let CreateTimelineError::AncestorLsn(err) = err else {
4082 2 : panic!("wrong error type")
4083 2 : };
4084 2 : assert!(err.to_string().contains("invalid branch start lsn"));
4085 2 : assert!(err
4086 2 : .source()
4087 2 : .unwrap()
4088 2 : .to_string()
4089 2 : .contains("we might've already garbage collected needed data"))
4090 2 : }
4091 2 : }
4092 2 :
4093 2 : Ok(())
4094 2 : }
4095 :
4096 2 : #[tokio::test]
4097 2 : async fn test_prohibit_branch_creation_on_pre_initdb_lsn() -> anyhow::Result<()> {
4098 2 : let (tenant, ctx) =
4099 2 : TenantHarness::create("test_prohibit_branch_creation_on_pre_initdb_lsn")?
4100 2 : .load()
4101 2 : .await;
4102 2 :
4103 2 : let tline = tenant
4104 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x50), DEFAULT_PG_VERSION, &ctx)
4105 6 : .await?;
4106 2 : // try to branch at lsn 0x25, should fail because initdb lsn is 0x50
4107 2 : match tenant
4108 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x25)), &ctx)
4109 2 : .await
4110 2 : {
4111 2 : Ok(_) => panic!("branching should have failed"),
4112 2 : Err(err) => {
4113 2 : let CreateTimelineError::AncestorLsn(err) = err else {
4114 2 : panic!("wrong error type");
4115 2 : };
4116 2 : assert!(&err.to_string().contains("invalid branch start lsn"));
4117 2 : assert!(&err
4118 2 : .source()
4119 2 : .unwrap()
4120 2 : .to_string()
4121 2 : .contains("is earlier than latest GC horizon"));
4122 2 : }
4123 2 : }
4124 2 :
4125 2 : Ok(())
4126 2 : }
4127 :
4128 : /*
4129 : // FIXME: This currently fails to error out. Calling GC doesn't currently
4130 : // remove the old value, we'd need to work a little harder
4131 : #[tokio::test]
4132 : async fn test_prohibit_get_for_garbage_collected_data() -> anyhow::Result<()> {
4133 : let repo =
4134 : RepoHarness::create("test_prohibit_get_for_garbage_collected_data")?
4135 : .load();
4136 :
4137 : let tline = repo.create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION)?;
4138 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
4139 :
4140 : repo.gc_iteration(Some(TIMELINE_ID), 0x10, Duration::ZERO)?;
4141 : let latest_gc_cutoff_lsn = tline.get_latest_gc_cutoff_lsn();
4142 : assert!(*latest_gc_cutoff_lsn > Lsn(0x25));
4143 : match tline.get(*TEST_KEY, Lsn(0x25)) {
4144 : Ok(_) => panic!("request for page should have failed"),
4145 : Err(err) => assert!(err.to_string().contains("not found at")),
4146 : }
4147 : Ok(())
4148 : }
4149 : */
4150 :
4151 2 : #[tokio::test]
4152 2 : async fn test_get_branchpoints_from_an_inactive_timeline() -> anyhow::Result<()> {
4153 2 : let (tenant, ctx) =
4154 2 : TenantHarness::create("test_get_branchpoints_from_an_inactive_timeline")?
4155 2 : .load()
4156 2 : .await;
4157 2 : let tline = tenant
4158 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
4159 6 : .await?;
4160 6 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
4161 2 :
4162 2 : tenant
4163 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
4164 2 : .await?;
4165 2 : let newtline = tenant
4166 2 : .get_timeline(NEW_TIMELINE_ID, true)
4167 2 : .expect("Should have a local timeline");
4168 2 :
4169 6 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
4170 2 :
4171 2 : tline.set_broken("test".to_owned());
4172 2 :
4173 2 : tenant
4174 2 : .gc_iteration(
4175 2 : Some(TIMELINE_ID),
4176 2 : 0x10,
4177 2 : Duration::ZERO,
4178 2 : &CancellationToken::new(),
4179 2 : &ctx,
4180 2 : )
4181 2 : .await?;
4182 2 :
4183 2 : // The branchpoints should contain all timelines, even ones marked
4184 2 : // as Broken.
4185 2 : {
4186 2 : let branchpoints = &tline.gc_info.read().unwrap().retain_lsns;
4187 2 : assert_eq!(branchpoints.len(), 1);
4188 2 : assert_eq!(branchpoints[0], Lsn(0x40));
4189 2 : }
4190 2 :
4191 2 : // You can read the key from the child branch even though the parent is
4192 2 : // Broken, as long as you don't need to access data from the parent.
4193 2 : assert_eq!(
4194 4 : newtline.get(*TEST_KEY, Lsn(0x70), &ctx).await?,
4195 2 : test_img(&format!("foo at {}", Lsn(0x70)))
4196 2 : );
4197 2 :
4198 2 : // This needs to traverse to the parent, and fails.
4199 2 : let err = newtline.get(*TEST_KEY, Lsn(0x50), &ctx).await.unwrap_err();
4200 2 : assert!(err
4201 2 : .to_string()
4202 2 : .contains("will not become active. Current state: Broken"));
4203 2 :
4204 2 : Ok(())
4205 2 : }
4206 :
4207 2 : #[tokio::test]
4208 2 : async fn test_retain_data_in_parent_which_is_needed_for_child() -> anyhow::Result<()> {
4209 2 : let (tenant, ctx) =
4210 2 : TenantHarness::create("test_retain_data_in_parent_which_is_needed_for_child")?
4211 2 : .load()
4212 2 : .await;
4213 2 : let tline = tenant
4214 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
4215 6 : .await?;
4216 6 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
4217 2 :
4218 2 : tenant
4219 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
4220 2 : .await?;
4221 2 : let newtline = tenant
4222 2 : .get_timeline(NEW_TIMELINE_ID, true)
4223 2 : .expect("Should have a local timeline");
4224 2 : // this removes layers before lsn 40 (50 minus 10), so there are two remaining layers, image and delta for 31-50
4225 2 : tenant
4226 2 : .gc_iteration(
4227 2 : Some(TIMELINE_ID),
4228 2 : 0x10,
4229 2 : Duration::ZERO,
4230 2 : &CancellationToken::new(),
4231 2 : &ctx,
4232 2 : )
4233 2 : .await?;
4234 4 : assert!(newtline.get(*TEST_KEY, Lsn(0x25), &ctx).await.is_ok());
4235 2 :
4236 2 : Ok(())
4237 2 : }
4238 2 : #[tokio::test]
4239 2 : async fn test_parent_keeps_data_forever_after_branching() -> anyhow::Result<()> {
4240 2 : let (tenant, ctx) =
4241 2 : TenantHarness::create("test_parent_keeps_data_forever_after_branching")?
4242 2 : .load()
4243 2 : .await;
4244 2 : let tline = tenant
4245 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
4246 6 : .await?;
4247 6 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
4248 2 :
4249 2 : tenant
4250 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
4251 2 : .await?;
4252 2 : let newtline = tenant
4253 2 : .get_timeline(NEW_TIMELINE_ID, true)
4254 2 : .expect("Should have a local timeline");
4255 2 :
4256 6 : make_some_layers(newtline.as_ref(), Lsn(0x60), &ctx).await?;
4257 2 :
4258 2 : // run gc on parent
4259 2 : tenant
4260 2 : .gc_iteration(
4261 2 : Some(TIMELINE_ID),
4262 2 : 0x10,
4263 2 : Duration::ZERO,
4264 2 : &CancellationToken::new(),
4265 2 : &ctx,
4266 2 : )
4267 2 : .await?;
4268 2 :
4269 2 : // Check that the data is still accessible on the branch.
4270 2 : assert_eq!(
4271 7 : newtline.get(*TEST_KEY, Lsn(0x50), &ctx).await?,
4272 2 : test_img(&format!("foo at {}", Lsn(0x40)))
4273 2 : );
4274 2 :
4275 2 : Ok(())
4276 2 : }
4277 :
4278 2 : #[tokio::test]
4279 2 : async fn timeline_load() -> anyhow::Result<()> {
4280 2 : const TEST_NAME: &str = "timeline_load";
4281 2 : let harness = TenantHarness::create(TEST_NAME)?;
4282 2 : {
4283 2 : let (tenant, ctx) = harness.load().await;
4284 2 : let tline = tenant
4285 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x7000), DEFAULT_PG_VERSION, &ctx)
4286 6 : .await?;
4287 6 : make_some_layers(tline.as_ref(), Lsn(0x8000), &ctx).await?;
4288 2 : // so that all uploads finish & we can call harness.load() below again
4289 2 : tenant
4290 2 : .shutdown(Default::default(), true)
4291 2 : .instrument(harness.span())
4292 2 : .await
4293 2 : .ok()
4294 2 : .unwrap();
4295 2 : }
4296 2 :
4297 10 : let (tenant, _ctx) = harness.load().await;
4298 2 : tenant
4299 2 : .get_timeline(TIMELINE_ID, true)
4300 2 : .expect("cannot load timeline");
4301 2 :
4302 2 : Ok(())
4303 2 : }
4304 :
4305 2 : #[tokio::test]
4306 2 : async fn timeline_load_with_ancestor() -> anyhow::Result<()> {
4307 2 : const TEST_NAME: &str = "timeline_load_with_ancestor";
4308 2 : let harness = TenantHarness::create(TEST_NAME)?;
4309 2 : // create two timelines
4310 2 : {
4311 2 : let (tenant, ctx) = harness.load().await;
4312 2 : let tline = tenant
4313 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
4314 6 : .await?;
4315 2 :
4316 6 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
4317 2 :
4318 2 : let child_tline = tenant
4319 2 : .branch_timeline_test(&tline, NEW_TIMELINE_ID, Some(Lsn(0x40)), &ctx)
4320 2 : .await?;
4321 2 : child_tline.set_state(TimelineState::Active);
4322 2 :
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 : // so that all uploads finish & we can call harness.load() below again
4330 2 : tenant
4331 2 : .shutdown(Default::default(), true)
4332 2 : .instrument(harness.span())
4333 4 : .await
4334 2 : .ok()
4335 2 : .unwrap();
4336 2 : }
4337 2 :
4338 2 : // check that both of them are initially unloaded
4339 16 : let (tenant, _ctx) = harness.load().await;
4340 2 :
4341 2 : // check that both, child and ancestor are loaded
4342 2 : let _child_tline = tenant
4343 2 : .get_timeline(NEW_TIMELINE_ID, true)
4344 2 : .expect("cannot get child timeline loaded");
4345 2 :
4346 2 : let _ancestor_tline = tenant
4347 2 : .get_timeline(TIMELINE_ID, true)
4348 2 : .expect("cannot get ancestor timeline loaded");
4349 2 :
4350 2 : Ok(())
4351 2 : }
4352 :
4353 2 : #[tokio::test]
4354 2 : async fn delta_layer_dumping() -> anyhow::Result<()> {
4355 2 : use storage_layer::AsLayerDesc;
4356 2 : let (tenant, ctx) = TenantHarness::create("test_layer_dumping")?.load().await;
4357 2 : let tline = tenant
4358 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
4359 6 : .await?;
4360 6 : make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
4361 2 :
4362 2 : let layer_map = tline.layers.read().await;
4363 2 : let level0_deltas = layer_map
4364 2 : .layer_map()
4365 2 : .get_level0_deltas()?
4366 2 : .into_iter()
4367 4 : .map(|desc| layer_map.get_from_desc(&desc))
4368 2 : .collect::<Vec<_>>();
4369 2 :
4370 2 : assert!(!level0_deltas.is_empty());
4371 2 :
4372 6 : for delta in level0_deltas {
4373 2 : // Ensure we are dumping a delta layer here
4374 4 : assert!(delta.layer_desc().is_delta);
4375 8 : delta.dump(true, &ctx).await.unwrap();
4376 2 : }
4377 2 :
4378 2 : Ok(())
4379 2 : }
4380 :
4381 2 : #[tokio::test]
4382 2 : async fn test_images() -> anyhow::Result<()> {
4383 2 : let (tenant, ctx) = TenantHarness::create("test_images")?.load().await;
4384 2 : let tline = tenant
4385 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
4386 6 : .await?;
4387 2 :
4388 2 : let writer = tline.writer().await;
4389 2 : writer
4390 2 : .put(
4391 2 : *TEST_KEY,
4392 2 : Lsn(0x10),
4393 2 : &Value::Image(test_img("foo at 0x10")),
4394 2 : &ctx,
4395 2 : )
4396 2 : .await?;
4397 2 : writer.finish_write(Lsn(0x10));
4398 2 : drop(writer);
4399 2 :
4400 2 : tline.freeze_and_flush().await?;
4401 2 : tline
4402 2 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
4403 2 : .await?;
4404 2 :
4405 2 : let writer = tline.writer().await;
4406 2 : writer
4407 2 : .put(
4408 2 : *TEST_KEY,
4409 2 : Lsn(0x20),
4410 2 : &Value::Image(test_img("foo at 0x20")),
4411 2 : &ctx,
4412 2 : )
4413 2 : .await?;
4414 2 : writer.finish_write(Lsn(0x20));
4415 2 : drop(writer);
4416 2 :
4417 2 : tline.freeze_and_flush().await?;
4418 2 : tline
4419 2 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
4420 2 : .await?;
4421 2 :
4422 2 : let writer = tline.writer().await;
4423 2 : writer
4424 2 : .put(
4425 2 : *TEST_KEY,
4426 2 : Lsn(0x30),
4427 2 : &Value::Image(test_img("foo at 0x30")),
4428 2 : &ctx,
4429 2 : )
4430 2 : .await?;
4431 2 : writer.finish_write(Lsn(0x30));
4432 2 : drop(writer);
4433 2 :
4434 2 : tline.freeze_and_flush().await?;
4435 2 : tline
4436 2 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
4437 2 : .await?;
4438 2 :
4439 2 : let writer = tline.writer().await;
4440 2 : writer
4441 2 : .put(
4442 2 : *TEST_KEY,
4443 2 : Lsn(0x40),
4444 2 : &Value::Image(test_img("foo at 0x40")),
4445 2 : &ctx,
4446 2 : )
4447 2 : .await?;
4448 2 : writer.finish_write(Lsn(0x40));
4449 2 : drop(writer);
4450 2 :
4451 2 : tline.freeze_and_flush().await?;
4452 2 : tline
4453 2 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
4454 2 : .await?;
4455 2 :
4456 2 : assert_eq!(
4457 4 : tline.get(*TEST_KEY, Lsn(0x10), &ctx).await?,
4458 2 : test_img("foo at 0x10")
4459 2 : );
4460 2 : assert_eq!(
4461 3 : tline.get(*TEST_KEY, Lsn(0x1f), &ctx).await?,
4462 2 : test_img("foo at 0x10")
4463 2 : );
4464 2 : assert_eq!(
4465 2 : tline.get(*TEST_KEY, Lsn(0x20), &ctx).await?,
4466 2 : test_img("foo at 0x20")
4467 2 : );
4468 2 : assert_eq!(
4469 4 : tline.get(*TEST_KEY, Lsn(0x30), &ctx).await?,
4470 2 : test_img("foo at 0x30")
4471 2 : );
4472 2 : assert_eq!(
4473 4 : tline.get(*TEST_KEY, Lsn(0x40), &ctx).await?,
4474 2 : test_img("foo at 0x40")
4475 2 : );
4476 2 :
4477 2 : Ok(())
4478 2 : }
4479 :
4480 4 : async fn bulk_insert_compact_gc(
4481 4 : timeline: Arc<Timeline>,
4482 4 : ctx: &RequestContext,
4483 4 : mut lsn: Lsn,
4484 4 : repeat: usize,
4485 4 : key_count: usize,
4486 4 : ) -> anyhow::Result<()> {
4487 4 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
4488 4 : let mut blknum = 0;
4489 4 :
4490 4 : // Enforce that key range is monotonously increasing
4491 4 : let mut keyspace = KeySpaceAccum::new();
4492 4 :
4493 4 : for _ in 0..repeat {
4494 200 : for _ in 0..key_count {
4495 2000000 : test_key.field6 = blknum;
4496 2000000 : let writer = timeline.writer().await;
4497 2000000 : writer
4498 2000000 : .put(
4499 2000000 : test_key,
4500 2000000 : lsn,
4501 2000000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
4502 2000000 : ctx,
4503 2000000 : )
4504 31708 : .await?;
4505 2000000 : writer.finish_write(lsn);
4506 2000000 : drop(writer);
4507 2000000 :
4508 2000000 : keyspace.add_key(test_key);
4509 2000000 :
4510 2000000 : lsn = Lsn(lsn.0 + 0x10);
4511 2000000 : blknum += 1;
4512 : }
4513 :
4514 200 : let cutoff = timeline.get_last_record_lsn();
4515 200 :
4516 200 : timeline
4517 200 : .update_gc_info(
4518 200 : Vec::new(),
4519 200 : cutoff,
4520 200 : Duration::ZERO,
4521 200 : &CancellationToken::new(),
4522 200 : ctx,
4523 200 : )
4524 0 : .await?;
4525 200 : timeline.freeze_and_flush().await?;
4526 200 : timeline
4527 200 : .compact(&CancellationToken::new(), EnumSet::empty(), ctx)
4528 37280 : .await?;
4529 200 : timeline.gc().await?;
4530 : }
4531 :
4532 4 : Ok(())
4533 4 : }
4534 :
4535 : //
4536 : // Insert 1000 key-value pairs with increasing keys, flush, compact, GC.
4537 : // Repeat 50 times.
4538 : //
4539 2 : #[tokio::test]
4540 2 : async fn test_bulk_insert() -> anyhow::Result<()> {
4541 2 : let harness = TenantHarness::create("test_bulk_insert")?;
4542 2 : let (tenant, ctx) = harness.load().await;
4543 2 : let tline = tenant
4544 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
4545 6 : .await?;
4546 2 :
4547 2 : let lsn = Lsn(0x10);
4548 42390 : bulk_insert_compact_gc(tline.clone(), &ctx, lsn, 50, 10000).await?;
4549 2 :
4550 2 : Ok(())
4551 2 : }
4552 :
4553 : // Test the vectored get real implementation against a simple sequential implementation.
4554 : //
4555 : // The test generates a keyspace by repeatedly flushing the in-memory layer and compacting.
4556 : // Projected to 2D the key space looks like below. Lsn grows upwards on the Y axis and keys
4557 : // grow to the right on the X axis.
4558 : // [Delta]
4559 : // [Delta]
4560 : // [Delta]
4561 : // [Delta]
4562 : // ------------ Image ---------------
4563 : //
4564 : // After layer generation we pick the ranges to query as follows:
4565 : // 1. The beginning of each delta layer
4566 : // 2. At the seam between two adjacent delta layers
4567 : //
4568 : // There's one major downside to this test: delta layers only contains images,
4569 : // so the search can stop at the first delta layer and doesn't traverse any deeper.
4570 2 : #[tokio::test]
4571 2 : async fn test_get_vectored() -> anyhow::Result<()> {
4572 2 : let harness = TenantHarness::create("test_get_vectored")?;
4573 2 : let (tenant, ctx) = harness.load().await;
4574 2 : let tline = tenant
4575 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x08), DEFAULT_PG_VERSION, &ctx)
4576 6 : .await?;
4577 2 :
4578 2 : let lsn = Lsn(0x10);
4579 42390 : bulk_insert_compact_gc(tline.clone(), &ctx, lsn, 50, 10000).await?;
4580 2 :
4581 2 : let guard = tline.layers.read().await;
4582 2 : guard.layer_map().dump(true, &ctx).await?;
4583 2 :
4584 2 : let mut reads = Vec::new();
4585 2 : let mut prev = None;
4586 12 : guard.layer_map().iter_historic_layers().for_each(|desc| {
4587 12 : if !desc.is_delta() {
4588 2 : prev = Some(desc.clone());
4589 2 : return;
4590 10 : }
4591 10 :
4592 10 : let start = desc.key_range.start;
4593 10 : let end = desc
4594 10 : .key_range
4595 10 : .start
4596 10 : .add(Timeline::MAX_GET_VECTORED_KEYS.try_into().unwrap());
4597 10 : reads.push(KeySpace {
4598 10 : ranges: vec![start..end],
4599 10 : });
4600 2 :
4601 10 : if let Some(prev) = &prev {
4602 10 : if !prev.is_delta() {
4603 10 : return;
4604 2 : }
4605 0 :
4606 0 : let first_range = Key {
4607 0 : field6: prev.key_range.end.field6 - 4,
4608 0 : ..prev.key_range.end
4609 0 : }..prev.key_range.end;
4610 0 :
4611 0 : let second_range = desc.key_range.start..Key {
4612 0 : field6: desc.key_range.start.field6 + 4,
4613 0 : ..desc.key_range.start
4614 0 : };
4615 0 :
4616 0 : reads.push(KeySpace {
4617 0 : ranges: vec![first_range, second_range],
4618 0 : });
4619 2 : };
4620 2 :
4621 2 : prev = Some(desc.clone());
4622 12 : });
4623 2 :
4624 2 : drop(guard);
4625 2 :
4626 2 : // Pick a big LSN such that we query over all the changes.
4627 2 : // Technically, u64::MAX - 1 is the largest LSN supported by the read path,
4628 2 : // but there seems to be a bug on the non-vectored search path which surfaces
4629 2 : // in that case.
4630 2 : let reads_lsn = Lsn(u64::MAX - 1000);
4631 2 :
4632 12 : for read in reads {
4633 10 : info!("Doing vectored read on {:?}", read);
4634 2 :
4635 25 : let vectored_res = tline.get_vectored_impl(read.clone(), reads_lsn, &ctx).await;
4636 10 : tline
4637 10 : .validate_get_vectored_impl(&vectored_res, read, reads_lsn, &ctx)
4638 20 : .await;
4639 2 : }
4640 2 :
4641 2 : Ok(())
4642 2 : }
4643 :
4644 2 : #[tokio::test]
4645 2 : async fn test_random_updates() -> anyhow::Result<()> {
4646 2 : let harness = TenantHarness::create("test_random_updates")?;
4647 2 : let (tenant, ctx) = harness.load().await;
4648 2 : let tline = tenant
4649 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
4650 6 : .await?;
4651 2 :
4652 2 : const NUM_KEYS: usize = 1000;
4653 2 :
4654 2 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
4655 2 :
4656 2 : let mut keyspace = KeySpaceAccum::new();
4657 2 :
4658 2 : // Track when each page was last modified. Used to assert that
4659 2 : // a read sees the latest page version.
4660 2 : let mut updated = [Lsn(0); NUM_KEYS];
4661 2 :
4662 2 : let mut lsn = Lsn(0x10);
4663 2 : #[allow(clippy::needless_range_loop)]
4664 2002 : for blknum in 0..NUM_KEYS {
4665 2000 : lsn = Lsn(lsn.0 + 0x10);
4666 2000 : test_key.field6 = blknum as u32;
4667 2000 : let writer = tline.writer().await;
4668 2000 : writer
4669 2000 : .put(
4670 2000 : test_key,
4671 2000 : lsn,
4672 2000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
4673 2000 : &ctx,
4674 2000 : )
4675 33 : .await?;
4676 2000 : writer.finish_write(lsn);
4677 2000 : updated[blknum] = lsn;
4678 2000 : drop(writer);
4679 2000 :
4680 2000 : keyspace.add_key(test_key);
4681 2 : }
4682 2 :
4683 102 : for _ in 0..50 {
4684 100100 : for _ in 0..NUM_KEYS {
4685 100000 : lsn = Lsn(lsn.0 + 0x10);
4686 100000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
4687 100000 : test_key.field6 = blknum as u32;
4688 100000 : let writer = tline.writer().await;
4689 100000 : writer
4690 100000 : .put(
4691 100000 : test_key,
4692 100000 : lsn,
4693 100000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
4694 100000 : &ctx,
4695 100000 : )
4696 1604 : .await?;
4697 100000 : writer.finish_write(lsn);
4698 100000 : drop(writer);
4699 100000 : updated[blknum] = lsn;
4700 2 : }
4701 2 :
4702 2 : // Read all the blocks
4703 100000 : for (blknum, last_lsn) in updated.iter().enumerate() {
4704 100000 : test_key.field6 = blknum as u32;
4705 100000 : assert_eq!(
4706 100000 : tline.get(test_key, lsn, &ctx).await?,
4707 100000 : test_img(&format!("{} at {}", blknum, last_lsn))
4708 2 : );
4709 2 : }
4710 2 :
4711 2 : // Perform a cycle of flush, compact, and GC
4712 100 : let cutoff = tline.get_last_record_lsn();
4713 100 : tline
4714 100 : .update_gc_info(
4715 100 : Vec::new(),
4716 100 : cutoff,
4717 100 : Duration::ZERO,
4718 100 : &CancellationToken::new(),
4719 100 : &ctx,
4720 100 : )
4721 2 : .await?;
4722 105 : tline.freeze_and_flush().await?;
4723 100 : tline
4724 100 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
4725 2280 : .await?;
4726 100 : tline.gc().await?;
4727 2 : }
4728 2 :
4729 2 : Ok(())
4730 2 : }
4731 :
4732 2 : #[tokio::test]
4733 2 : async fn test_traverse_branches() -> anyhow::Result<()> {
4734 2 : let (tenant, ctx) = TenantHarness::create("test_traverse_branches")?
4735 2 : .load()
4736 2 : .await;
4737 2 : let mut tline = tenant
4738 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
4739 5 : .await?;
4740 2 :
4741 2 : const NUM_KEYS: usize = 1000;
4742 2 :
4743 2 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
4744 2 :
4745 2 : let mut keyspace = KeySpaceAccum::new();
4746 2 :
4747 2 : // Track when each page was last modified. Used to assert that
4748 2 : // a read sees the latest page version.
4749 2 : let mut updated = [Lsn(0); NUM_KEYS];
4750 2 :
4751 2 : let mut lsn = Lsn(0x10);
4752 2 : #[allow(clippy::needless_range_loop)]
4753 2002 : for blknum in 0..NUM_KEYS {
4754 2000 : lsn = Lsn(lsn.0 + 0x10);
4755 2000 : test_key.field6 = blknum as u32;
4756 2000 : let writer = tline.writer().await;
4757 2000 : writer
4758 2000 : .put(
4759 2000 : test_key,
4760 2000 : lsn,
4761 2000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
4762 2000 : &ctx,
4763 2000 : )
4764 33 : .await?;
4765 2000 : writer.finish_write(lsn);
4766 2000 : updated[blknum] = lsn;
4767 2000 : drop(writer);
4768 2000 :
4769 2000 : keyspace.add_key(test_key);
4770 2 : }
4771 2 :
4772 102 : for _ in 0..50 {
4773 100 : let new_tline_id = TimelineId::generate();
4774 100 : tenant
4775 100 : .branch_timeline_test(&tline, new_tline_id, Some(lsn), &ctx)
4776 2 : .await?;
4777 100 : tline = tenant
4778 100 : .get_timeline(new_tline_id, true)
4779 100 : .expect("Should have the branched timeline");
4780 2 :
4781 100100 : for _ in 0..NUM_KEYS {
4782 100000 : lsn = Lsn(lsn.0 + 0x10);
4783 100000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
4784 100000 : test_key.field6 = blknum as u32;
4785 100000 : let writer = tline.writer().await;
4786 100000 : writer
4787 100000 : .put(
4788 100000 : test_key,
4789 100000 : lsn,
4790 100000 : &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
4791 100000 : &ctx,
4792 100000 : )
4793 1642 : .await?;
4794 100000 : println!("updating {} at {}", blknum, lsn);
4795 100000 : writer.finish_write(lsn);
4796 100000 : drop(writer);
4797 100000 : updated[blknum] = lsn;
4798 2 : }
4799 2 :
4800 2 : // Read all the blocks
4801 100000 : for (blknum, last_lsn) in updated.iter().enumerate() {
4802 100000 : test_key.field6 = blknum as u32;
4803 100000 : assert_eq!(
4804 100000 : tline.get(test_key, lsn, &ctx).await?,
4805 100000 : test_img(&format!("{} at {}", blknum, last_lsn))
4806 2 : );
4807 2 : }
4808 2 :
4809 2 : // Perform a cycle of flush, compact, and GC
4810 100 : let cutoff = tline.get_last_record_lsn();
4811 100 : tline
4812 100 : .update_gc_info(
4813 100 : Vec::new(),
4814 100 : cutoff,
4815 100 : Duration::ZERO,
4816 100 : &CancellationToken::new(),
4817 100 : &ctx,
4818 100 : )
4819 2 : .await?;
4820 104 : tline.freeze_and_flush().await?;
4821 100 : tline
4822 100 : .compact(&CancellationToken::new(), EnumSet::empty(), &ctx)
4823 13314 : .await?;
4824 100 : tline.gc().await?;
4825 2 : }
4826 2 :
4827 2 : Ok(())
4828 2 : }
4829 :
4830 2 : #[tokio::test]
4831 2 : async fn test_traverse_ancestors() -> anyhow::Result<()> {
4832 2 : let (tenant, ctx) = TenantHarness::create("test_traverse_ancestors")?
4833 2 : .load()
4834 2 : .await;
4835 2 : let mut tline = tenant
4836 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
4837 6 : .await?;
4838 2 :
4839 2 : const NUM_KEYS: usize = 100;
4840 2 : const NUM_TLINES: usize = 50;
4841 2 :
4842 2 : let mut test_key = Key::from_hex("010000000033333333444444445500000000").unwrap();
4843 2 : // Track page mutation lsns across different timelines.
4844 2 : let mut updated = [[Lsn(0); NUM_KEYS]; NUM_TLINES];
4845 2 :
4846 2 : let mut lsn = Lsn(0x10);
4847 2 :
4848 2 : #[allow(clippy::needless_range_loop)]
4849 102 : for idx in 0..NUM_TLINES {
4850 100 : let new_tline_id = TimelineId::generate();
4851 100 : tenant
4852 100 : .branch_timeline_test(&tline, new_tline_id, Some(lsn), &ctx)
4853 2 : .await?;
4854 100 : tline = tenant
4855 100 : .get_timeline(new_tline_id, true)
4856 100 : .expect("Should have the branched timeline");
4857 2 :
4858 10100 : for _ in 0..NUM_KEYS {
4859 10000 : lsn = Lsn(lsn.0 + 0x10);
4860 10000 : let blknum = thread_rng().gen_range(0..NUM_KEYS);
4861 10000 : test_key.field6 = blknum as u32;
4862 10000 : let writer = tline.writer().await;
4863 10000 : writer
4864 10000 : .put(
4865 10000 : test_key,
4866 10000 : lsn,
4867 10000 : &Value::Image(test_img(&format!("{} {} at {}", idx, blknum, lsn))),
4868 10000 : &ctx,
4869 10000 : )
4870 178 : .await?;
4871 10000 : println!("updating [{}][{}] at {}", idx, blknum, lsn);
4872 10000 : writer.finish_write(lsn);
4873 10000 : drop(writer);
4874 10000 : updated[idx][blknum] = lsn;
4875 2 : }
4876 2 : }
4877 2 :
4878 2 : // Read pages from leaf timeline across all ancestors.
4879 100 : for (idx, lsns) in updated.iter().enumerate() {
4880 10000 : for (blknum, lsn) in lsns.iter().enumerate() {
4881 2 : // Skip empty mutations.
4882 10000 : if lsn.0 == 0 {
4883 3646 : continue;
4884 6354 : }
4885 6354 : println!("checking [{idx}][{blknum}] at {lsn}");
4886 6354 : test_key.field6 = blknum as u32;
4887 6354 : assert_eq!(
4888 6354 : tline.get(test_key, *lsn, &ctx).await?,
4889 6354 : test_img(&format!("{idx} {blknum} at {lsn}"))
4890 2 : );
4891 2 : }
4892 2 : }
4893 2 : Ok(())
4894 2 : }
4895 :
4896 2 : #[tokio::test]
4897 2 : async fn test_write_at_initdb_lsn_takes_optimization_code_path() -> anyhow::Result<()> {
4898 2 : let (tenant, ctx) = TenantHarness::create("test_empty_test_timeline_is_usable")?
4899 2 : .load()
4900 2 : .await;
4901 2 :
4902 2 : let initdb_lsn = Lsn(0x20);
4903 2 : let utline = tenant
4904 2 : .create_empty_timeline(TIMELINE_ID, initdb_lsn, DEFAULT_PG_VERSION, &ctx)
4905 2 : .await?;
4906 2 : let tline = utline.raw_timeline().unwrap();
4907 2 :
4908 2 : // Spawn flush loop now so that we can set the `expect_initdb_optimization`
4909 2 : tline.maybe_spawn_flush_loop();
4910 2 :
4911 2 : // Make sure the timeline has the minimum set of required keys for operation.
4912 2 : // The only operation you can always do on an empty timeline is to `put` new data.
4913 2 : // Except if you `put` at `initdb_lsn`.
4914 2 : // In that case, there's an optimization to directly create image layers instead of delta layers.
4915 2 : // It uses `repartition()`, which assumes some keys to be present.
4916 2 : // Let's make sure the test timeline can handle that case.
4917 2 : {
4918 2 : let mut state = tline.flush_loop_state.lock().unwrap();
4919 2 : assert_eq!(
4920 2 : timeline::FlushLoopState::Running {
4921 2 : expect_initdb_optimization: false,
4922 2 : initdb_optimization_count: 0,
4923 2 : },
4924 2 : *state
4925 2 : );
4926 2 : *state = timeline::FlushLoopState::Running {
4927 2 : expect_initdb_optimization: true,
4928 2 : initdb_optimization_count: 0,
4929 2 : };
4930 2 : }
4931 2 :
4932 2 : // Make writes at the initdb_lsn. When we flush it below, it should be handled by the optimization.
4933 2 : // As explained above, the optimization requires some keys to be present.
4934 2 : // As per `create_empty_timeline` documentation, use init_empty to set them.
4935 2 : // This is what `create_test_timeline` does, by the way.
4936 2 : let mut modification = tline.begin_modification(initdb_lsn);
4937 2 : modification
4938 2 : .init_empty_test_timeline()
4939 2 : .context("init_empty_test_timeline")?;
4940 2 : modification
4941 2 : .commit(&ctx)
4942 2 : .await
4943 2 : .context("commit init_empty_test_timeline modification")?;
4944 2 :
4945 2 : // Do the flush. The flush code will check the expectations that we set above.
4946 2 : tline.freeze_and_flush().await?;
4947 2 :
4948 2 : // assert freeze_and_flush exercised the initdb optimization
4949 2 : {
4950 2 : let state = tline.flush_loop_state.lock().unwrap();
4951 2 : let timeline::FlushLoopState::Running {
4952 2 : expect_initdb_optimization,
4953 2 : initdb_optimization_count,
4954 2 : } = *state
4955 2 : else {
4956 2 : panic!("unexpected state: {:?}", *state);
4957 2 : };
4958 2 : assert!(expect_initdb_optimization);
4959 2 : assert!(initdb_optimization_count > 0);
4960 2 : }
4961 2 : Ok(())
4962 2 : }
4963 :
4964 2 : #[tokio::test]
4965 2 : async fn test_uninit_mark_crash() -> anyhow::Result<()> {
4966 2 : let name = "test_uninit_mark_crash";
4967 2 : let harness = TenantHarness::create(name)?;
4968 2 : {
4969 2 : let (tenant, ctx) = harness.load().await;
4970 2 : let tline = tenant
4971 2 : .create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION, &ctx)
4972 2 : .await?;
4973 2 : // Keeps uninit mark in place
4974 2 : let raw_tline = tline.raw_timeline().unwrap();
4975 2 : raw_tline
4976 2 : .shutdown()
4977 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))
4978 2 : .await;
4979 2 : std::mem::forget(tline);
4980 2 : }
4981 2 :
4982 2 : let (tenant, _) = harness.load().await;
4983 2 : match tenant.get_timeline(TIMELINE_ID, false) {
4984 2 : Ok(_) => panic!("timeline should've been removed during load"),
4985 2 : Err(e) => {
4986 2 : assert_eq!(
4987 2 : e,
4988 2 : GetTimelineError::NotFound {
4989 2 : tenant_id: tenant.tenant_shard_id,
4990 2 : timeline_id: TIMELINE_ID,
4991 2 : }
4992 2 : )
4993 2 : }
4994 2 : }
4995 2 :
4996 2 : assert!(!harness
4997 2 : .conf
4998 2 : .timeline_path(&tenant.tenant_shard_id, &TIMELINE_ID)
4999 2 : .exists());
5000 2 :
5001 2 : assert!(!harness
5002 2 : .conf
5003 2 : .timeline_uninit_mark_file_path(tenant.tenant_shard_id, TIMELINE_ID)
5004 2 : .exists());
5005 2 :
5006 2 : Ok(())
5007 2 : }
5008 : }
|
