Line data Source code
1 : use std::collections::HashSet;
2 : use std::sync::Arc;
3 :
4 : use anyhow::Context;
5 : use http_utils::error::ApiError;
6 : use pageserver_api::models::DetachBehavior;
7 : use pageserver_api::models::detach_ancestor::AncestorDetached;
8 : use pageserver_api::shard::ShardIdentity;
9 : use tokio::sync::Semaphore;
10 : use tokio_util::sync::CancellationToken;
11 : use tracing::Instrument;
12 : use utils::completion;
13 : use utils::generation::Generation;
14 : use utils::id::TimelineId;
15 : use utils::lsn::Lsn;
16 : use utils::sync::gate::GateError;
17 :
18 : use super::layer_manager::LayerManager;
19 : use super::{FlushLayerError, Timeline};
20 : use crate::context::{DownloadBehavior, RequestContext};
21 : use crate::task_mgr::TaskKind;
22 : use crate::tenant::Tenant;
23 : use crate::tenant::remote_timeline_client::index::GcBlockingReason::DetachAncestor;
24 : use crate::tenant::storage_layer::layer::local_layer_path;
25 : use crate::tenant::storage_layer::{AsLayerDesc as _, DeltaLayerWriter, Layer, ResidentLayer};
26 : use crate::virtual_file::{MaybeFatalIo, VirtualFile};
27 :
28 : #[derive(Debug, thiserror::Error)]
29 : pub(crate) enum Error {
30 : #[error("no ancestors")]
31 : NoAncestor,
32 :
33 : #[error("too many ancestors")]
34 : TooManyAncestors,
35 :
36 : #[error("ancestor is not empty")]
37 : AncestorNotEmpty,
38 :
39 : #[error("shutting down, please retry later")]
40 : ShuttingDown,
41 :
42 : #[error("archived: {}", .0)]
43 : Archived(TimelineId),
44 :
45 : #[error(transparent)]
46 : NotFound(crate::tenant::GetTimelineError),
47 :
48 : #[error("failed to reparent all candidate timelines, please retry")]
49 : FailedToReparentAll,
50 :
51 : #[error("ancestor is already being detached by: {}", .0)]
52 : OtherTimelineDetachOngoing(TimelineId),
53 :
54 : #[error("preparing to timeline ancestor detach failed")]
55 : Prepare(#[source] anyhow::Error),
56 :
57 : #[error("detaching and reparenting failed")]
58 : DetachReparent(#[source] anyhow::Error),
59 :
60 : #[error("completing ancestor detach failed")]
61 : Complete(#[source] anyhow::Error),
62 :
63 : #[error("failpoint: {}", .0)]
64 : Failpoint(&'static str),
65 : }
66 :
67 : impl Error {
68 : /// Try to catch cancellation from within the `anyhow::Error`, or wrap the anyhow as the given
69 : /// variant or fancier `or_else`.
70 0 : fn launder<F>(e: anyhow::Error, or_else: F) -> Error
71 0 : where
72 0 : F: Fn(anyhow::Error) -> Error,
73 0 : {
74 : use remote_storage::TimeoutOrCancel;
75 :
76 : use crate::tenant::remote_timeline_client::WaitCompletionError;
77 : use crate::tenant::upload_queue::NotInitialized;
78 :
79 0 : if e.is::<NotInitialized>()
80 0 : || TimeoutOrCancel::caused_by_cancel(&e)
81 0 : || e.downcast_ref::<remote_storage::DownloadError>()
82 0 : .is_some_and(|e| e.is_cancelled())
83 0 : || e.is::<WaitCompletionError>()
84 : {
85 0 : Error::ShuttingDown
86 : } else {
87 0 : or_else(e)
88 : }
89 0 : }
90 : }
91 :
92 : impl From<Error> for ApiError {
93 0 : fn from(value: Error) -> Self {
94 0 : match value {
95 0 : Error::NoAncestor => ApiError::Conflict(value.to_string()),
96 : Error::TooManyAncestors | Error::AncestorNotEmpty => {
97 0 : ApiError::BadRequest(anyhow::anyhow!("{value}"))
98 : }
99 0 : Error::ShuttingDown => ApiError::ShuttingDown,
100 0 : Error::Archived(_) => ApiError::BadRequest(anyhow::anyhow!("{value}")),
101 : Error::OtherTimelineDetachOngoing(_) | Error::FailedToReparentAll => {
102 0 : ApiError::ResourceUnavailable(value.to_string().into())
103 : }
104 0 : Error::NotFound(e) => ApiError::from(e),
105 : // these variants should have no cancellation errors because of Error::launder
106 : Error::Prepare(_)
107 : | Error::DetachReparent(_)
108 : | Error::Complete(_)
109 0 : | Error::Failpoint(_) => ApiError::InternalServerError(value.into()),
110 : }
111 0 : }
112 : }
113 :
114 : impl From<crate::tenant::upload_queue::NotInitialized> for Error {
115 0 : fn from(_: crate::tenant::upload_queue::NotInitialized) -> Self {
116 0 : // treat all as shutting down signals, even though that is not entirely correct
117 0 : // (uninitialized state)
118 0 : Error::ShuttingDown
119 0 : }
120 : }
121 : impl From<super::layer_manager::Shutdown> for Error {
122 0 : fn from(_: super::layer_manager::Shutdown) -> Self {
123 0 : Error::ShuttingDown
124 0 : }
125 : }
126 :
127 : pub(crate) enum Progress {
128 : Prepared(Attempt, PreparedTimelineDetach),
129 : Done(AncestorDetached),
130 : }
131 :
132 : pub(crate) struct PreparedTimelineDetach {
133 : layers: Vec<Layer>,
134 : }
135 :
136 : // TODO: this should be part of PageserverConf because we cannot easily modify cplane arguments.
137 : #[derive(Debug)]
138 : pub(crate) struct Options {
139 : pub(crate) rewrite_concurrency: std::num::NonZeroUsize,
140 : pub(crate) copy_concurrency: std::num::NonZeroUsize,
141 : }
142 :
143 : impl Default for Options {
144 0 : fn default() -> Self {
145 0 : Self {
146 0 : rewrite_concurrency: std::num::NonZeroUsize::new(2).unwrap(),
147 0 : copy_concurrency: std::num::NonZeroUsize::new(100).unwrap(),
148 0 : }
149 0 : }
150 : }
151 :
152 : /// Represents an across tenant reset exclusive single attempt to detach ancestor.
153 : #[derive(Debug)]
154 : pub(crate) struct Attempt {
155 : pub(crate) timeline_id: TimelineId,
156 : pub(crate) ancestor_timeline_id: TimelineId,
157 : pub(crate) ancestor_lsn: Lsn,
158 : _guard: completion::Completion,
159 : gate_entered: Option<utils::sync::gate::GateGuard>,
160 : }
161 :
162 : impl Attempt {
163 0 : pub(crate) fn before_reset_tenant(&mut self) {
164 0 : let taken = self.gate_entered.take();
165 0 : assert!(taken.is_some());
166 0 : }
167 :
168 0 : pub(crate) fn new_barrier(&self) -> completion::Barrier {
169 0 : self._guard.barrier()
170 0 : }
171 : }
172 :
173 : /// See [`Timeline::prepare_to_detach_from_ancestor`]
174 0 : pub(super) async fn prepare(
175 0 : detached: &Arc<Timeline>,
176 0 : tenant: &Tenant,
177 0 : behavior: DetachBehavior,
178 0 : options: Options,
179 0 : ctx: &RequestContext,
180 0 : ) -> Result<Progress, Error> {
181 : use Error::*;
182 :
183 0 : let Some((mut ancestor, mut ancestor_lsn)) = detached
184 0 : .ancestor_timeline
185 0 : .as_ref()
186 0 : .map(|tl| (tl.clone(), detached.ancestor_lsn))
187 : else {
188 : let ancestor_id;
189 : let ancestor_lsn;
190 0 : let still_in_progress = {
191 0 : let accessor = detached.remote_client.initialized_upload_queue()?;
192 :
193 : // we are safe to inspect the latest uploaded, because we can only witness this after
194 : // restart is complete and ancestor is no more.
195 0 : let latest = accessor.latest_uploaded_index_part();
196 0 : let Some((id, lsn)) = latest.lineage.detached_previous_ancestor() else {
197 0 : return Err(NoAncestor);
198 : };
199 0 : ancestor_id = id;
200 0 : ancestor_lsn = lsn;
201 0 :
202 0 : latest
203 0 : .gc_blocking
204 0 : .as_ref()
205 0 : .is_some_and(|b| b.blocked_by(DetachAncestor))
206 0 : };
207 0 :
208 0 : if still_in_progress {
209 : // gc is still blocked, we can still reparent and complete.
210 : // we are safe to reparent remaining, because they were locked in in the beginning.
211 0 : let attempt =
212 0 : continue_with_blocked_gc(detached, tenant, ancestor_id, ancestor_lsn).await?;
213 :
214 : // because the ancestor of detached is already set to none, we have published all
215 : // of the layers, so we are still "prepared."
216 0 : return Ok(Progress::Prepared(
217 0 : attempt,
218 0 : PreparedTimelineDetach { layers: Vec::new() },
219 0 : ));
220 0 : }
221 :
222 0 : let reparented_timelines = reparented_direct_children(detached, tenant)?;
223 0 : return Ok(Progress::Done(AncestorDetached {
224 0 : reparented_timelines,
225 0 : }));
226 : };
227 :
228 0 : if detached.is_archived() != Some(false) {
229 0 : return Err(Archived(detached.timeline_id));
230 0 : }
231 0 :
232 0 : if !ancestor_lsn.is_valid() {
233 : // rare case, probably wouldn't even load
234 0 : tracing::error!("ancestor is set, but ancestor_lsn is invalid, this timeline needs fixing");
235 0 : return Err(NoAncestor);
236 0 : }
237 0 :
238 0 : check_no_archived_children_of_ancestor(tenant, detached, &ancestor, ancestor_lsn)?;
239 :
240 0 : if let DetachBehavior::MultiLevelAndNoReparent = behavior {
241 : // If the ancestor has an ancestor, we might be able to fast-path detach it if the current ancestor does not have any data written/used by the detaching timeline.
242 0 : while let Some(ancestor_of_ancestor) = ancestor.ancestor_timeline.clone() {
243 0 : if ancestor_lsn != ancestor.ancestor_lsn {
244 : // non-technical requirement; we could flatten still if ancestor LSN does not match but that needs
245 : // us to copy and cut more layers.
246 0 : return Err(AncestorNotEmpty);
247 0 : }
248 0 : // Use the ancestor of the ancestor as the new ancestor (only when the ancestor LSNs are the same)
249 0 : ancestor_lsn = ancestor.ancestor_lsn; // Get the LSN first before resetting the `ancestor` variable
250 0 : ancestor = ancestor_of_ancestor;
251 0 : // TODO: do we still need to check if we don't want to reparent?
252 0 : check_no_archived_children_of_ancestor(tenant, detached, &ancestor, ancestor_lsn)?;
253 : }
254 0 : } else if ancestor.ancestor_timeline.is_some() {
255 : // non-technical requirement; we could flatten N ancestors just as easily but we chose
256 : // not to, at least initially
257 0 : return Err(TooManyAncestors);
258 0 : }
259 :
260 0 : tracing::info!(
261 0 : "attempt to detach the timeline from the ancestor: {}@{}, behavior={:?}",
262 0 : ancestor.timeline_id,
263 : ancestor_lsn,
264 : behavior
265 : );
266 :
267 0 : let attempt = start_new_attempt(detached, tenant, ancestor.timeline_id, ancestor_lsn).await?;
268 :
269 0 : utils::pausable_failpoint!("timeline-detach-ancestor::before_starting_after_locking-pausable");
270 :
271 0 : fail::fail_point!(
272 0 : "timeline-detach-ancestor::before_starting_after_locking",
273 0 : |_| Err(Error::Failpoint(
274 0 : "timeline-detach-ancestor::before_starting_after_locking"
275 0 : ))
276 0 : );
277 :
278 0 : if ancestor_lsn >= ancestor.get_disk_consistent_lsn() {
279 0 : let span =
280 0 : tracing::info_span!("freeze_and_flush", ancestor_timeline_id=%ancestor.timeline_id);
281 0 : async {
282 0 : let started_at = std::time::Instant::now();
283 0 : let freeze_and_flush = ancestor.freeze_and_flush0();
284 0 : let mut freeze_and_flush = std::pin::pin!(freeze_and_flush);
285 :
286 0 : let res =
287 0 : tokio::time::timeout(std::time::Duration::from_secs(1), &mut freeze_and_flush)
288 0 : .await;
289 :
290 0 : let res = match res {
291 0 : Ok(res) => res,
292 0 : Err(_elapsed) => {
293 0 : tracing::info!("freezing and flushing ancestor is still ongoing");
294 0 : freeze_and_flush.await
295 : }
296 : };
297 :
298 0 : res.map_err(|e| {
299 : use FlushLayerError::*;
300 0 : match e {
301 : Cancelled | NotRunning(_) => {
302 : // FIXME(#6424): technically statically unreachable right now, given how we never
303 : // drop the sender
304 0 : Error::ShuttingDown
305 : }
306 0 : CreateImageLayersError(_) | Other(_) => Error::Prepare(e.into()),
307 : }
308 0 : })?;
309 :
310 : // we do not need to wait for uploads to complete but we do need `struct Layer`,
311 : // copying delta prefix is unsupported currently for `InMemoryLayer`.
312 0 : tracing::info!(
313 0 : elapsed_ms = started_at.elapsed().as_millis(),
314 0 : "froze and flushed the ancestor"
315 : );
316 0 : Ok::<_, Error>(())
317 0 : }
318 0 : .instrument(span)
319 0 : .await?;
320 0 : }
321 :
322 0 : let end_lsn = ancestor_lsn + 1;
323 :
324 0 : let (filtered_layers, straddling_branchpoint, rest_of_historic) = {
325 : // we do not need to start from our layers, because they can only be layers that come
326 : // *after* ancestor_lsn
327 0 : let layers = tokio::select! {
328 0 : guard = ancestor.layers.read() => guard,
329 0 : _ = detached.cancel.cancelled() => {
330 0 : return Err(ShuttingDown);
331 : }
332 0 : _ = ancestor.cancel.cancelled() => {
333 0 : return Err(ShuttingDown);
334 : }
335 : };
336 :
337 : // between retries, these can change if compaction or gc ran in between. this will mean
338 : // we have to redo work.
339 0 : partition_work(ancestor_lsn, &layers)?
340 : };
341 :
342 : // TODO: layers are already sorted by something: use that to determine how much of remote
343 : // copies are already done -- gc is blocked, but a compaction could had happened on ancestor,
344 : // which is something to keep in mind if copy skipping is implemented.
345 0 : tracing::info!(filtered=%filtered_layers, to_rewrite = straddling_branchpoint.len(), historic=%rest_of_historic.len(), "collected layers");
346 :
347 : // TODO: copying and lsn prefix copying could be done at the same time with a single fsync after
348 0 : let mut new_layers: Vec<Layer> =
349 0 : Vec::with_capacity(straddling_branchpoint.len() + rest_of_historic.len());
350 0 :
351 0 : {
352 0 : tracing::debug!(to_rewrite = %straddling_branchpoint.len(), "copying prefix of delta layers");
353 :
354 0 : let mut tasks = tokio::task::JoinSet::new();
355 0 :
356 0 : let mut wrote_any = false;
357 0 :
358 0 : let limiter = Arc::new(Semaphore::new(options.rewrite_concurrency.get()));
359 :
360 0 : for layer in straddling_branchpoint {
361 0 : let limiter = limiter.clone();
362 0 : let timeline = detached.clone();
363 0 : let ctx = ctx.detached_child(TaskKind::DetachAncestor, DownloadBehavior::Download);
364 :
365 0 : let span = tracing::info_span!("upload_rewritten_layer", %layer);
366 0 : tasks.spawn(
367 0 : async move {
368 0 : let _permit = limiter.acquire().await;
369 0 : let copied =
370 0 : upload_rewritten_layer(end_lsn, &layer, &timeline, &timeline.cancel, &ctx)
371 0 : .await?;
372 0 : if let Some(copied) = copied.as_ref() {
373 0 : tracing::info!(%copied, "rewrote and uploaded");
374 0 : }
375 0 : Ok(copied)
376 0 : }
377 0 : .instrument(span),
378 0 : );
379 0 : }
380 :
381 0 : while let Some(res) = tasks.join_next().await {
382 0 : match res {
383 0 : Ok(Ok(Some(copied))) => {
384 0 : wrote_any = true;
385 0 : new_layers.push(copied);
386 0 : }
387 0 : Ok(Ok(None)) => {}
388 0 : Ok(Err(e)) => return Err(e),
389 0 : Err(je) => return Err(Error::Prepare(je.into())),
390 : }
391 : }
392 :
393 : // FIXME: the fsync should be mandatory, after both rewrites and copies
394 0 : if wrote_any {
395 0 : fsync_timeline_dir(detached, ctx).await;
396 0 : }
397 : }
398 :
399 0 : let mut tasks = tokio::task::JoinSet::new();
400 0 : let limiter = Arc::new(Semaphore::new(options.copy_concurrency.get()));
401 0 : let cancel_eval = CancellationToken::new();
402 :
403 0 : for adopted in rest_of_historic {
404 0 : let limiter = limiter.clone();
405 0 : let timeline = detached.clone();
406 0 : let cancel_eval = cancel_eval.clone();
407 0 :
408 0 : tasks.spawn(
409 0 : async move {
410 0 : let _permit = tokio::select! {
411 0 : permit = limiter.acquire() => {
412 0 : permit
413 : }
414 : // Wait for the cancellation here instead of letting the entire task be cancelled.
415 : // Cancellations are racy in that they might leave layers on disk.
416 0 : _ = cancel_eval.cancelled() => {
417 0 : Err(Error::ShuttingDown)?
418 : }
419 : };
420 0 : let (owned, did_hardlink) = remote_copy(
421 0 : &adopted,
422 0 : &timeline,
423 0 : timeline.generation,
424 0 : timeline.shard_identity,
425 0 : &timeline.cancel,
426 0 : )
427 0 : .await?;
428 0 : tracing::info!(layer=%owned, did_hard_link=%did_hardlink, "remote copied");
429 0 : Ok((owned, did_hardlink))
430 0 : }
431 0 : .in_current_span(),
432 0 : );
433 0 : }
434 :
435 0 : fn delete_layers(timeline: &Timeline, layers: Vec<Layer>) -> Result<(), Error> {
436 : // We are deleting layers, so we must hold the gate
437 0 : let _gate = timeline.gate.enter().map_err(|e| match e {
438 0 : GateError::GateClosed => Error::ShuttingDown,
439 0 : })?;
440 0 : {
441 0 : layers.into_iter().for_each(|l: Layer| {
442 0 : l.delete_on_drop();
443 0 : std::mem::drop(l);
444 0 : });
445 0 : }
446 0 : Ok(())
447 0 : }
448 :
449 0 : let mut should_fsync = false;
450 0 : let mut first_err = None;
451 0 : while let Some(res) = tasks.join_next().await {
452 0 : match res {
453 0 : Ok(Ok((owned, did_hardlink))) => {
454 0 : if did_hardlink {
455 0 : should_fsync = true;
456 0 : }
457 0 : new_layers.push(owned);
458 : }
459 :
460 : // Don't stop the evaluation on errors, so that we get the full set of hardlinked layers to delete.
461 0 : Ok(Err(failed)) => {
462 0 : cancel_eval.cancel();
463 0 : first_err.get_or_insert(failed);
464 0 : }
465 0 : Err(je) => {
466 0 : cancel_eval.cancel();
467 0 : first_err.get_or_insert(Error::Prepare(je.into()));
468 0 : }
469 : }
470 : }
471 :
472 0 : if let Some(failed) = first_err {
473 0 : delete_layers(detached, new_layers)?;
474 0 : return Err(failed);
475 0 : }
476 0 :
477 0 : // fsync directory again if we hardlinked something
478 0 : if should_fsync {
479 0 : fsync_timeline_dir(detached, ctx).await;
480 0 : }
481 :
482 0 : let prepared = PreparedTimelineDetach { layers: new_layers };
483 0 :
484 0 : Ok(Progress::Prepared(attempt, prepared))
485 0 : }
486 :
487 0 : async fn start_new_attempt(
488 0 : detached: &Timeline,
489 0 : tenant: &Tenant,
490 0 : ancestor_timeline_id: TimelineId,
491 0 : ancestor_lsn: Lsn,
492 0 : ) -> Result<Attempt, Error> {
493 0 : let attempt = obtain_exclusive_attempt(detached, tenant, ancestor_timeline_id, ancestor_lsn)?;
494 :
495 : // insert the block in the index_part.json, if not already there.
496 0 : let _dont_care = tenant
497 0 : .gc_block
498 0 : .insert(
499 0 : detached,
500 0 : crate::tenant::remote_timeline_client::index::GcBlockingReason::DetachAncestor,
501 0 : )
502 0 : .await
503 0 : .map_err(|e| Error::launder(e, Error::Prepare))?;
504 :
505 0 : Ok(attempt)
506 0 : }
507 :
508 0 : async fn continue_with_blocked_gc(
509 0 : detached: &Timeline,
510 0 : tenant: &Tenant,
511 0 : ancestor_timeline_id: TimelineId,
512 0 : ancestor_lsn: Lsn,
513 0 : ) -> Result<Attempt, Error> {
514 0 : // FIXME: it would be nice to confirm that there is an in-memory version, since we've just
515 0 : // verified there is a persistent one?
516 0 : obtain_exclusive_attempt(detached, tenant, ancestor_timeline_id, ancestor_lsn)
517 0 : }
518 :
519 0 : fn obtain_exclusive_attempt(
520 0 : detached: &Timeline,
521 0 : tenant: &Tenant,
522 0 : ancestor_timeline_id: TimelineId,
523 0 : ancestor_lsn: Lsn,
524 0 : ) -> Result<Attempt, Error> {
525 : use Error::{OtherTimelineDetachOngoing, ShuttingDown};
526 :
527 : // ensure we are the only active attempt for this tenant
528 0 : let (guard, barrier) = completion::channel();
529 0 : {
530 0 : let mut guard = tenant.ongoing_timeline_detach.lock().unwrap();
531 0 : if let Some((tl, other)) = guard.as_ref() {
532 0 : if !other.is_ready() {
533 0 : return Err(OtherTimelineDetachOngoing(*tl));
534 0 : }
535 : // FIXME: no test enters here
536 0 : }
537 0 : *guard = Some((detached.timeline_id, barrier));
538 : }
539 :
540 : // ensure the gate is still open
541 0 : let _gate_entered = detached.gate.enter().map_err(|_| ShuttingDown)?;
542 :
543 0 : Ok(Attempt {
544 0 : timeline_id: detached.timeline_id,
545 0 : ancestor_timeline_id,
546 0 : ancestor_lsn,
547 0 : _guard: guard,
548 0 : gate_entered: Some(_gate_entered),
549 0 : })
550 0 : }
551 :
552 0 : fn reparented_direct_children(
553 0 : detached: &Arc<Timeline>,
554 0 : tenant: &Tenant,
555 0 : ) -> Result<HashSet<TimelineId>, Error> {
556 0 : let mut all_direct_children = tenant
557 0 : .timelines
558 0 : .lock()
559 0 : .unwrap()
560 0 : .values()
561 0 : .filter_map(|tl| {
562 0 : let is_direct_child = matches!(tl.ancestor_timeline.as_ref(), Some(ancestor) if Arc::ptr_eq(ancestor, detached));
563 :
564 0 : if is_direct_child {
565 0 : Some(tl.clone())
566 : } else {
567 0 : if let Some(timeline) = tl.ancestor_timeline.as_ref() {
568 0 : assert_ne!(timeline.timeline_id, detached.timeline_id, "we cannot have two timelines with the same timeline_id live");
569 0 : }
570 0 : None
571 : }
572 0 : })
573 0 : // Collect to avoid lock taking order problem with Tenant::timelines and
574 0 : // Timeline::remote_client
575 0 : .collect::<Vec<_>>();
576 0 :
577 0 : let mut any_shutdown = false;
578 0 :
579 0 : all_direct_children.retain(|tl| match tl.remote_client.initialized_upload_queue() {
580 0 : Ok(accessor) => accessor
581 0 : .latest_uploaded_index_part()
582 0 : .lineage
583 0 : .is_reparented(),
584 0 : Err(_shutdownalike) => {
585 0 : // not 100% a shutdown, but let's bail early not to give inconsistent results in
586 0 : // sharded enviroment.
587 0 : any_shutdown = true;
588 0 : true
589 : }
590 0 : });
591 0 :
592 0 : if any_shutdown {
593 : // it could be one or many being deleted; have client retry
594 0 : return Err(Error::ShuttingDown);
595 0 : }
596 0 :
597 0 : Ok(all_direct_children
598 0 : .into_iter()
599 0 : .map(|tl| tl.timeline_id)
600 0 : .collect())
601 0 : }
602 :
603 0 : fn partition_work(
604 0 : ancestor_lsn: Lsn,
605 0 : source: &LayerManager,
606 0 : ) -> Result<(usize, Vec<Layer>, Vec<Layer>), Error> {
607 0 : let mut straddling_branchpoint = vec![];
608 0 : let mut rest_of_historic = vec![];
609 0 :
610 0 : let mut later_by_lsn = 0;
611 :
612 0 : for desc in source.layer_map()?.iter_historic_layers() {
613 : // off by one chances here:
614 : // - start is inclusive
615 : // - end is exclusive
616 0 : if desc.lsn_range.start > ancestor_lsn {
617 0 : later_by_lsn += 1;
618 0 : continue;
619 0 : }
620 :
621 0 : let target = if desc.lsn_range.start <= ancestor_lsn
622 0 : && desc.lsn_range.end > ancestor_lsn
623 0 : && desc.is_delta
624 : {
625 : // TODO: image layer at Lsn optimization
626 0 : &mut straddling_branchpoint
627 : } else {
628 0 : &mut rest_of_historic
629 : };
630 :
631 0 : target.push(source.get_from_desc(&desc));
632 : }
633 :
634 0 : Ok((later_by_lsn, straddling_branchpoint, rest_of_historic))
635 0 : }
636 :
637 0 : async fn upload_rewritten_layer(
638 0 : end_lsn: Lsn,
639 0 : layer: &Layer,
640 0 : target: &Arc<Timeline>,
641 0 : cancel: &CancellationToken,
642 0 : ctx: &RequestContext,
643 0 : ) -> Result<Option<Layer>, Error> {
644 0 : let copied = copy_lsn_prefix(end_lsn, layer, target, ctx).await?;
645 :
646 0 : let Some(copied) = copied else {
647 0 : return Ok(None);
648 : };
649 :
650 0 : target
651 0 : .remote_client
652 0 : .upload_layer_file(&copied, cancel)
653 0 : .await
654 0 : .map_err(|e| Error::launder(e, Error::Prepare))?;
655 :
656 0 : Ok(Some(copied.into()))
657 0 : }
658 :
659 0 : async fn copy_lsn_prefix(
660 0 : end_lsn: Lsn,
661 0 : layer: &Layer,
662 0 : target_timeline: &Arc<Timeline>,
663 0 : ctx: &RequestContext,
664 0 : ) -> Result<Option<ResidentLayer>, Error> {
665 0 : if target_timeline.cancel.is_cancelled() {
666 0 : return Err(Error::ShuttingDown);
667 0 : }
668 0 :
669 0 : tracing::debug!(%layer, %end_lsn, "copying lsn prefix");
670 :
671 0 : let mut writer = DeltaLayerWriter::new(
672 0 : target_timeline.conf,
673 0 : target_timeline.timeline_id,
674 0 : target_timeline.tenant_shard_id,
675 0 : layer.layer_desc().key_range.start,
676 0 : layer.layer_desc().lsn_range.start..end_lsn,
677 0 : ctx,
678 0 : )
679 0 : .await
680 0 : .with_context(|| format!("prepare to copy lsn prefix of ancestors {layer}"))
681 0 : .map_err(Error::Prepare)?;
682 :
683 0 : let resident = layer.download_and_keep_resident(ctx).await.map_err(|e| {
684 0 : if e.is_cancelled() {
685 0 : Error::ShuttingDown
686 : } else {
687 0 : Error::Prepare(e.into())
688 : }
689 0 : })?;
690 :
691 0 : let records = resident
692 0 : .copy_delta_prefix(&mut writer, end_lsn, ctx)
693 0 : .await
694 0 : .with_context(|| format!("copy lsn prefix of ancestors {layer}"))
695 0 : .map_err(Error::Prepare)?;
696 :
697 0 : drop(resident);
698 0 :
699 0 : tracing::debug!(%layer, records, "copied records");
700 :
701 0 : if records == 0 {
702 0 : drop(writer);
703 0 : // TODO: we might want to store an empty marker in remote storage for this
704 0 : // layer so that we will not needlessly walk `layer` on repeated attempts.
705 0 : Ok(None)
706 : } else {
707 : // reuse the key instead of adding more holes between layers by using the real
708 : // highest key in the layer.
709 0 : let reused_highest_key = layer.layer_desc().key_range.end;
710 0 : let (desc, path) = writer
711 0 : .finish(reused_highest_key, ctx)
712 0 : .await
713 0 : .map_err(Error::Prepare)?;
714 0 : let copied = Layer::finish_creating(target_timeline.conf, target_timeline, desc, &path)
715 0 : .map_err(Error::Prepare)?;
716 :
717 0 : tracing::debug!(%layer, %copied, "new layer produced");
718 :
719 0 : Ok(Some(copied))
720 : }
721 0 : }
722 :
723 : /// Creates a new Layer instance for the adopted layer, and ensures it is found in the remote
724 : /// storage on successful return. without the adopted layer being added to `index_part.json`.
725 : /// Returns (Layer, did hardlink)
726 0 : async fn remote_copy(
727 0 : adopted: &Layer,
728 0 : adoptee: &Arc<Timeline>,
729 0 : generation: Generation,
730 0 : shard_identity: ShardIdentity,
731 0 : cancel: &CancellationToken,
732 0 : ) -> Result<(Layer, bool), Error> {
733 0 : let mut metadata = adopted.metadata();
734 0 : debug_assert!(metadata.generation <= generation);
735 0 : metadata.generation = generation;
736 0 : metadata.shard = shard_identity.shard_index();
737 0 :
738 0 : let conf = adoptee.conf;
739 0 : let file_name = adopted.layer_desc().layer_name();
740 :
741 : // We don't want to shut the timeline down during this operation because we do `delete_on_drop` below
742 0 : let _gate = adoptee.gate.enter().map_err(|e| match e {
743 0 : GateError::GateClosed => Error::ShuttingDown,
744 0 : })?;
745 :
746 : // depending if Layer::keep_resident, do a hardlink
747 : let did_hardlink;
748 0 : let owned = if let Some(adopted_resident) = adopted.keep_resident().await {
749 0 : let adopted_path = adopted_resident.local_path();
750 0 : let adoptee_path = local_layer_path(
751 0 : conf,
752 0 : &adoptee.tenant_shard_id,
753 0 : &adoptee.timeline_id,
754 0 : &file_name,
755 0 : &metadata.generation,
756 0 : );
757 0 :
758 0 : match std::fs::hard_link(adopted_path, &adoptee_path) {
759 0 : Ok(()) => {}
760 0 : Err(e) if e.kind() == std::io::ErrorKind::AlreadyExists => {
761 0 : // In theory we should not get into this situation as we are doing cleanups of the layer file after errors.
762 0 : // However, we don't do cleanups for errors past `prepare`, so there is the slight chance to get to this branch.
763 0 :
764 0 : // Double check that the file is orphan (probably from an earlier attempt), then delete it
765 0 : let key = file_name.clone().into();
766 0 : if adoptee.layers.read().await.contains_key(&key) {
767 : // We are supposed to filter out such cases before coming to this function
768 0 : return Err(Error::Prepare(anyhow::anyhow!(
769 0 : "layer file {file_name} already present and inside layer map"
770 0 : )));
771 0 : }
772 0 : tracing::info!("Deleting orphan layer file to make way for hard linking");
773 : // Delete orphan layer file and try again, to ensure this layer has a well understood source
774 0 : std::fs::remove_file(adopted_path)
775 0 : .map_err(|e| Error::launder(e.into(), Error::Prepare))?;
776 0 : std::fs::hard_link(adopted_path, &adoptee_path)
777 0 : .map_err(|e| Error::launder(e.into(), Error::Prepare))?;
778 : }
779 0 : Err(e) => {
780 0 : return Err(Error::launder(e.into(), Error::Prepare));
781 : }
782 : };
783 0 : did_hardlink = true;
784 0 : Layer::for_resident(conf, adoptee, adoptee_path, file_name, metadata).drop_eviction_guard()
785 : } else {
786 0 : did_hardlink = false;
787 0 : Layer::for_evicted(conf, adoptee, file_name, metadata)
788 : };
789 :
790 0 : let layer = match adoptee
791 0 : .remote_client
792 0 : .copy_timeline_layer(adopted, &owned, cancel)
793 0 : .await
794 : {
795 0 : Ok(()) => owned,
796 0 : Err(e) => {
797 0 : {
798 0 : // Clean up the layer so that on a retry we don't get errors that the file already exists
799 0 : owned.delete_on_drop();
800 0 : std::mem::drop(owned);
801 0 : }
802 0 : return Err(Error::launder(e, Error::Prepare));
803 : }
804 : };
805 :
806 0 : Ok((layer, did_hardlink))
807 0 : }
808 :
809 : pub(crate) enum DetachingAndReparenting {
810 : /// All of the following timeline ids were reparented and the timeline ancestor detach must be
811 : /// marked as completed.
812 : Reparented(HashSet<TimelineId>),
813 :
814 : /// Some of the reparentings failed. The timeline ancestor detach must **not** be marked as
815 : /// completed.
816 : ///
817 : /// Nested `must_reset_tenant` is set to true when any restart requiring changes were made.
818 : SomeReparentingFailed { must_reset_tenant: bool },
819 :
820 : /// Detaching and reparentings were completed in a previous attempt. Timeline ancestor detach
821 : /// must be marked as completed.
822 : AlreadyDone(HashSet<TimelineId>),
823 : }
824 :
825 : impl DetachingAndReparenting {
826 0 : pub(crate) fn reset_tenant_required(&self) -> bool {
827 : use DetachingAndReparenting::*;
828 0 : match self {
829 0 : Reparented(_) => true,
830 0 : SomeReparentingFailed { must_reset_tenant } => *must_reset_tenant,
831 0 : AlreadyDone(_) => false,
832 : }
833 0 : }
834 :
835 0 : pub(crate) fn completed(self) -> Option<HashSet<TimelineId>> {
836 : use DetachingAndReparenting::*;
837 0 : match self {
838 0 : Reparented(x) | AlreadyDone(x) => Some(x),
839 0 : SomeReparentingFailed { .. } => None,
840 : }
841 0 : }
842 : }
843 :
844 : /// See [`Timeline::detach_from_ancestor_and_reparent`].
845 0 : pub(super) async fn detach_and_reparent(
846 0 : detached: &Arc<Timeline>,
847 0 : tenant: &Tenant,
848 0 : prepared: PreparedTimelineDetach,
849 0 : ancestor_timeline_id: TimelineId,
850 0 : ancestor_lsn: Lsn,
851 0 : behavior: DetachBehavior,
852 0 : _ctx: &RequestContext,
853 0 : ) -> Result<DetachingAndReparenting, Error> {
854 0 : let PreparedTimelineDetach { layers } = prepared;
855 :
856 : #[derive(Debug)]
857 : enum Ancestor {
858 : NotDetached(Arc<Timeline>, Lsn),
859 : Detached(Arc<Timeline>, Lsn),
860 : }
861 :
862 0 : let (recorded_branchpoint, still_ongoing) = {
863 0 : let access = detached.remote_client.initialized_upload_queue()?;
864 0 : let latest = access.latest_uploaded_index_part();
865 0 :
866 0 : (
867 0 : latest.lineage.detached_previous_ancestor(),
868 0 : latest
869 0 : .gc_blocking
870 0 : .as_ref()
871 0 : .is_some_and(|b| b.blocked_by(DetachAncestor)),
872 0 : )
873 0 : };
874 0 : assert!(
875 0 : still_ongoing,
876 0 : "cannot (detach? reparent)? complete if the operation is not still ongoing"
877 : );
878 :
879 0 : let ancestor_to_detach = match detached.ancestor_timeline.as_ref() {
880 0 : Some(mut ancestor) => {
881 0 : while ancestor.timeline_id != ancestor_timeline_id {
882 0 : match ancestor.ancestor_timeline.as_ref() {
883 0 : Some(found) => {
884 0 : if ancestor_lsn != ancestor.ancestor_lsn {
885 0 : return Err(Error::DetachReparent(anyhow::anyhow!(
886 0 : "cannot find the ancestor timeline to detach from: wrong ancestor lsn"
887 0 : )));
888 0 : }
889 0 : ancestor = found;
890 : }
891 : None => {
892 0 : return Err(Error::DetachReparent(anyhow::anyhow!(
893 0 : "cannot find the ancestor timeline to detach from"
894 0 : )));
895 : }
896 : }
897 : }
898 0 : Some(ancestor)
899 : }
900 0 : None => None,
901 : };
902 0 : let ancestor = match (ancestor_to_detach, recorded_branchpoint) {
903 0 : (Some(ancestor), None) => {
904 0 : assert!(
905 0 : !layers.is_empty(),
906 0 : "there should always be at least one layer to inherit"
907 : );
908 0 : Ancestor::NotDetached(ancestor.clone(), detached.ancestor_lsn)
909 : }
910 : (Some(_), Some(_)) => {
911 0 : panic!(
912 0 : "it should be impossible to get to here without having gone through the tenant reset; if the tenant was reset, then the ancestor_timeline would be None"
913 0 : );
914 : }
915 0 : (None, Some((ancestor_id, ancestor_lsn))) => {
916 0 : // it has been either:
917 0 : // - detached but still exists => we can try reparenting
918 0 : // - detached and deleted
919 0 : //
920 0 : // either way, we must complete
921 0 : assert!(
922 0 : layers.is_empty(),
923 0 : "no layers should had been copied as detach is done"
924 : );
925 :
926 0 : let existing = tenant.timelines.lock().unwrap().get(&ancestor_id).cloned();
927 :
928 0 : if let Some(ancestor) = existing {
929 0 : Ancestor::Detached(ancestor, ancestor_lsn)
930 : } else {
931 0 : let direct_children = reparented_direct_children(detached, tenant)?;
932 0 : return Ok(DetachingAndReparenting::AlreadyDone(direct_children));
933 : }
934 : }
935 : (None, None) => {
936 : // TODO: make sure there are no `?` before tenant_reset from after a questionmark from
937 : // here.
938 0 : panic!(
939 0 : "bug: detach_and_reparent called on a timeline which has not been detached or which has no live ancestor"
940 0 : );
941 : }
942 : };
943 :
944 : // publish the prepared layers before we reparent any of the timelines, so that on restart
945 : // reparented timelines find layers. also do the actual detaching.
946 : //
947 : // if we crash after this operation, a retry will allow reparenting the remaining timelines as
948 : // gc is blocked.
949 :
950 0 : let (ancestor, ancestor_lsn, was_detached) = match ancestor {
951 0 : Ancestor::NotDetached(ancestor, ancestor_lsn) => {
952 0 : // this has to complete before any reparentings because otherwise they would not have
953 0 : // layers on the new parent.
954 0 : detached
955 0 : .remote_client
956 0 : .schedule_adding_existing_layers_to_index_detach_and_wait(
957 0 : &layers,
958 0 : (ancestor.timeline_id, ancestor_lsn),
959 0 : )
960 0 : .await
961 0 : .context("publish layers and detach ancestor")
962 0 : .map_err(|e| Error::launder(e, Error::DetachReparent))?;
963 :
964 0 : tracing::info!(
965 0 : ancestor=%ancestor.timeline_id,
966 0 : %ancestor_lsn,
967 0 : inherited_layers=%layers.len(),
968 0 : "detached from ancestor"
969 : );
970 0 : (ancestor, ancestor_lsn, true)
971 : }
972 0 : Ancestor::Detached(ancestor, ancestor_lsn) => (ancestor, ancestor_lsn, false),
973 : };
974 :
975 0 : if let DetachBehavior::MultiLevelAndNoReparent = behavior {
976 : // Do not reparent if the user requests to behave so.
977 0 : return Ok(DetachingAndReparenting::Reparented(HashSet::new()));
978 0 : }
979 0 :
980 0 : let mut tasks = tokio::task::JoinSet::new();
981 0 :
982 0 : // Returns a single permit semaphore which will be used to make one reparenting succeed,
983 0 : // others will fail as if those timelines had been stopped for whatever reason.
984 0 : #[cfg(feature = "testing")]
985 0 : let failpoint_sem = || -> Option<Arc<Semaphore>> {
986 0 : fail::fail_point!("timeline-detach-ancestor::allow_one_reparented", |_| Some(
987 0 : Arc::new(Semaphore::new(1))
988 0 : ));
989 0 : None
990 0 : }();
991 0 :
992 0 : // because we are now keeping the slot in progress, it is unlikely that there will be any
993 0 : // timeline deletions during this time. if we raced one, then we'll just ignore it.
994 0 : {
995 0 : let g = tenant.timelines.lock().unwrap();
996 0 : reparentable_timelines(g.values(), detached, &ancestor, ancestor_lsn)
997 0 : .cloned()
998 0 : .for_each(|timeline| {
999 : // important in this scope: we are holding the Tenant::timelines lock
1000 0 : let span = tracing::info_span!("reparent", reparented=%timeline.timeline_id);
1001 0 : let new_parent = detached.timeline_id;
1002 0 : #[cfg(feature = "testing")]
1003 0 : let failpoint_sem = failpoint_sem.clone();
1004 0 :
1005 0 : tasks.spawn(
1006 0 : async move {
1007 0 : let res = async {
1008 : #[cfg(feature = "testing")]
1009 0 : if let Some(failpoint_sem) = failpoint_sem {
1010 0 : let _permit = failpoint_sem.acquire().await.map_err(|_| {
1011 0 : anyhow::anyhow!(
1012 0 : "failpoint: timeline-detach-ancestor::allow_one_reparented",
1013 0 : )
1014 0 : })?;
1015 0 : failpoint_sem.close();
1016 0 : }
1017 :
1018 0 : timeline
1019 0 : .remote_client
1020 0 : .schedule_reparenting_and_wait(&new_parent)
1021 0 : .await
1022 0 : }
1023 0 : .await;
1024 :
1025 0 : match res {
1026 : Ok(()) => {
1027 0 : tracing::info!("reparented");
1028 0 : Some(timeline)
1029 : }
1030 0 : Err(e) => {
1031 0 : // with the use of tenant slot, raced timeline deletion is the most
1032 0 : // likely reason.
1033 0 : tracing::warn!("reparenting failed: {e:#}");
1034 0 : None
1035 : }
1036 : }
1037 0 : }
1038 0 : .instrument(span),
1039 0 : );
1040 0 : });
1041 0 : }
1042 0 :
1043 0 : let reparenting_candidates = tasks.len();
1044 0 : let mut reparented = HashSet::with_capacity(tasks.len());
1045 :
1046 0 : while let Some(res) = tasks.join_next().await {
1047 0 : match res {
1048 0 : Ok(Some(timeline)) => {
1049 0 : assert!(
1050 0 : reparented.insert(timeline.timeline_id),
1051 0 : "duplicate reparenting? timeline_id={}",
1052 0 : timeline.timeline_id
1053 : );
1054 : }
1055 0 : Err(je) if je.is_cancelled() => unreachable!("not used"),
1056 : // just ignore failures now, we can retry
1057 0 : Ok(None) => {}
1058 0 : Err(je) if je.is_panic() => {}
1059 0 : Err(je) => tracing::error!("unexpected join error: {je:?}"),
1060 : }
1061 : }
1062 :
1063 0 : let reparented_all = reparenting_candidates == reparented.len();
1064 0 :
1065 0 : if reparented_all {
1066 0 : Ok(DetachingAndReparenting::Reparented(reparented))
1067 : } else {
1068 0 : tracing::info!(
1069 0 : reparented = reparented.len(),
1070 0 : candidates = reparenting_candidates,
1071 0 : "failed to reparent all candidates; they can be retried after the tenant_reset",
1072 : );
1073 :
1074 0 : let must_reset_tenant = !reparented.is_empty() || was_detached;
1075 0 : Ok(DetachingAndReparenting::SomeReparentingFailed { must_reset_tenant })
1076 : }
1077 0 : }
1078 :
1079 0 : pub(super) async fn complete(
1080 0 : detached: &Arc<Timeline>,
1081 0 : tenant: &Tenant,
1082 0 : mut attempt: Attempt,
1083 0 : _ctx: &RequestContext,
1084 0 : ) -> Result<(), Error> {
1085 0 : assert_eq!(detached.timeline_id, attempt.timeline_id);
1086 :
1087 0 : if attempt.gate_entered.is_none() {
1088 0 : let entered = detached.gate.enter().map_err(|_| Error::ShuttingDown)?;
1089 0 : attempt.gate_entered = Some(entered);
1090 0 : } else {
1091 0 : // Some(gate_entered) means the tenant was not restarted, as is not required
1092 0 : }
1093 :
1094 0 : assert!(detached.ancestor_timeline.is_none());
1095 :
1096 : // this should be an 503 at least...?
1097 0 : fail::fail_point!(
1098 0 : "timeline-detach-ancestor::complete_before_uploading",
1099 0 : |_| Err(Error::Failpoint(
1100 0 : "timeline-detach-ancestor::complete_before_uploading"
1101 0 : ))
1102 0 : );
1103 :
1104 0 : tenant
1105 0 : .gc_block
1106 0 : .remove(
1107 0 : detached,
1108 0 : crate::tenant::remote_timeline_client::index::GcBlockingReason::DetachAncestor,
1109 0 : )
1110 0 : .await
1111 0 : .map_err(|e| Error::launder(e, Error::Complete))?;
1112 :
1113 0 : Ok(())
1114 0 : }
1115 :
1116 : /// Query against a locked `Tenant::timelines`.
1117 : ///
1118 : /// A timeline is reparentable if:
1119 : ///
1120 : /// - It is not the timeline being detached.
1121 : /// - It has the same ancestor as the timeline being detached. Note that the ancestor might not be the direct ancestor.
1122 0 : fn reparentable_timelines<'a, I>(
1123 0 : timelines: I,
1124 0 : detached: &'a Arc<Timeline>,
1125 0 : ancestor: &'a Arc<Timeline>,
1126 0 : ancestor_lsn: Lsn,
1127 0 : ) -> impl Iterator<Item = &'a Arc<Timeline>> + 'a
1128 0 : where
1129 0 : I: Iterator<Item = &'a Arc<Timeline>> + 'a,
1130 0 : {
1131 0 : timelines.filter_map(move |tl| {
1132 0 : if Arc::ptr_eq(tl, detached) {
1133 0 : return None;
1134 0 : }
1135 :
1136 0 : let tl_ancestor = tl.ancestor_timeline.as_ref()?;
1137 0 : let is_same = Arc::ptr_eq(ancestor, tl_ancestor);
1138 0 : let is_earlier = tl.get_ancestor_lsn() <= ancestor_lsn;
1139 0 :
1140 0 : let is_deleting = tl
1141 0 : .delete_progress
1142 0 : .try_lock()
1143 0 : .map(|flow| !flow.is_not_started())
1144 0 : .unwrap_or(true);
1145 0 :
1146 0 : if is_same && is_earlier && !is_deleting {
1147 0 : Some(tl)
1148 : } else {
1149 0 : None
1150 : }
1151 0 : })
1152 0 : }
1153 :
1154 0 : fn check_no_archived_children_of_ancestor(
1155 0 : tenant: &Tenant,
1156 0 : detached: &Arc<Timeline>,
1157 0 : ancestor: &Arc<Timeline>,
1158 0 : ancestor_lsn: Lsn,
1159 0 : ) -> Result<(), Error> {
1160 0 : let timelines = tenant.timelines.lock().unwrap();
1161 0 : let timelines_offloaded = tenant.timelines_offloaded.lock().unwrap();
1162 0 : for timeline in reparentable_timelines(timelines.values(), detached, ancestor, ancestor_lsn) {
1163 0 : if timeline.is_archived() == Some(true) {
1164 0 : return Err(Error::Archived(timeline.timeline_id));
1165 0 : }
1166 : }
1167 0 : for timeline_offloaded in timelines_offloaded.values() {
1168 0 : if timeline_offloaded.ancestor_timeline_id != Some(ancestor.timeline_id) {
1169 0 : continue;
1170 0 : }
1171 : // This forbids the detach ancestor feature if flattened timelines are present,
1172 : // even if the ancestor_lsn is from after the branchpoint of the detached timeline.
1173 : // But as per current design, we don't record the ancestor_lsn of flattened timelines.
1174 : // This is a bit unfortunate, but as of writing this we don't support flattening
1175 : // anyway. Maybe we can evolve the data model in the future.
1176 0 : if let Some(retain_lsn) = timeline_offloaded.ancestor_retain_lsn {
1177 0 : let is_earlier = retain_lsn <= ancestor_lsn;
1178 0 : if !is_earlier {
1179 0 : continue;
1180 0 : }
1181 0 : }
1182 0 : return Err(Error::Archived(timeline_offloaded.timeline_id));
1183 : }
1184 0 : Ok(())
1185 0 : }
1186 :
1187 0 : async fn fsync_timeline_dir(timeline: &Timeline, ctx: &RequestContext) {
1188 0 : let path = &timeline
1189 0 : .conf
1190 0 : .timeline_path(&timeline.tenant_shard_id, &timeline.timeline_id);
1191 0 : let timeline_dir = VirtualFile::open(&path, ctx)
1192 0 : .await
1193 0 : .fatal_err("VirtualFile::open for timeline dir fsync");
1194 0 : timeline_dir
1195 0 : .sync_all()
1196 0 : .await
1197 0 : .fatal_err("VirtualFile::sync_all timeline dir");
1198 0 : }
|
