Line data Source code
1 : use std::cmp::min;
2 : use std::collections::HashSet;
3 : use std::num::NonZeroU32;
4 : use std::pin::Pin;
5 : use std::time::Duration;
6 :
7 : use anyhow::{Context, Result};
8 : use camino::{Utf8Path, Utf8PathBuf};
9 : use futures::StreamExt;
10 : use futures::stream::FuturesOrdered;
11 : use postgres_ffi::v14::xlog_utils::XLogSegNoOffsetToRecPtr;
12 : use postgres_ffi::{PG_TLI, XLogFileName, XLogSegNo};
13 : use remote_storage::{
14 : DownloadOpts, GenericRemoteStorage, ListingMode, RemotePath, StorageMetadata,
15 : };
16 : use safekeeper_api::models::PeerInfo;
17 : use tokio::fs::File;
18 : use tokio::select;
19 : use tokio::sync::mpsc::{self, Receiver, Sender};
20 : use tokio::sync::{OnceCell, watch};
21 : use tokio::task::JoinHandle;
22 : use tokio_util::sync::CancellationToken;
23 : use tracing::*;
24 : use utils::id::{NodeId, TenantTimelineId};
25 : use utils::lsn::Lsn;
26 : use utils::{backoff, pausable_failpoint};
27 :
28 : use crate::metrics::{BACKED_UP_SEGMENTS, BACKUP_ERRORS, WAL_BACKUP_TASKS};
29 : use crate::timeline::WalResidentTimeline;
30 : use crate::timeline_manager::{Manager, StateSnapshot};
31 : use crate::{SafeKeeperConf, WAL_BACKUP_RUNTIME};
32 :
33 : const UPLOAD_FAILURE_RETRY_MIN_MS: u64 = 10;
34 : const UPLOAD_FAILURE_RETRY_MAX_MS: u64 = 5000;
35 :
36 : /// Default buffer size when interfacing with [`tokio::fs::File`].
37 : const BUFFER_SIZE: usize = 32 * 1024;
38 :
39 : pub struct WalBackupTaskHandle {
40 : shutdown_tx: Sender<()>,
41 : handle: JoinHandle<()>,
42 : }
43 :
44 : impl WalBackupTaskHandle {
45 0 : pub(crate) async fn join(self) {
46 0 : if let Err(e) = self.handle.await {
47 0 : error!("WAL backup task panicked: {}", e);
48 0 : }
49 0 : }
50 : }
51 :
52 : /// Do we have anything to upload to S3, i.e. should safekeepers run backup activity?
53 35 : pub(crate) fn is_wal_backup_required(
54 35 : wal_seg_size: usize,
55 35 : num_computes: usize,
56 35 : state: &StateSnapshot,
57 35 : ) -> bool {
58 35 : num_computes > 0 ||
59 : // Currently only the whole segment is offloaded, so compare segment numbers.
60 24 : (state.commit_lsn.segment_number(wal_seg_size) > state.backup_lsn.segment_number(wal_seg_size))
61 35 : }
62 :
63 : /// Based on peer information determine which safekeeper should offload; if it
64 : /// is me, run (per timeline) task, if not yet. OTOH, if it is not me and task
65 : /// is running, kill it.
66 0 : pub(crate) async fn update_task(mgr: &mut Manager, need_backup: bool, state: &StateSnapshot) {
67 0 : let (offloader, election_dbg_str) =
68 0 : determine_offloader(&state.peers, state.backup_lsn, mgr.tli.ttid, &mgr.conf);
69 0 : let elected_me = Some(mgr.conf.my_id) == offloader;
70 :
71 0 : let should_task_run = need_backup && elected_me;
72 :
73 : // start or stop the task
74 0 : if should_task_run != (mgr.backup_task.is_some()) {
75 0 : if should_task_run {
76 0 : info!("elected for backup: {}", election_dbg_str);
77 :
78 0 : let (shutdown_tx, shutdown_rx) = mpsc::channel(1);
79 :
80 0 : let Ok(resident) = mgr.wal_resident_timeline() else {
81 0 : info!("Timeline shut down");
82 0 : return;
83 : };
84 :
85 0 : let async_task = backup_task_main(resident, mgr.conf.backup_parallel_jobs, shutdown_rx);
86 :
87 0 : let handle = if mgr.conf.current_thread_runtime {
88 0 : tokio::spawn(async_task)
89 : } else {
90 0 : WAL_BACKUP_RUNTIME.spawn(async_task)
91 : };
92 :
93 0 : mgr.backup_task = Some(WalBackupTaskHandle {
94 0 : shutdown_tx,
95 0 : handle,
96 0 : });
97 : } else {
98 0 : if !need_backup {
99 : // don't need backup at all
100 0 : info!("stepping down from backup, need_backup={}", need_backup);
101 : } else {
102 : // someone else has been elected
103 0 : info!("stepping down from backup: {}", election_dbg_str);
104 : }
105 0 : shut_down_task(&mut mgr.backup_task).await;
106 : }
107 0 : }
108 0 : }
109 :
110 0 : async fn shut_down_task(entry: &mut Option<WalBackupTaskHandle>) {
111 0 : if let Some(wb_handle) = entry.take() {
112 : // Tell the task to shutdown. Error means task exited earlier, that's ok.
113 0 : let _ = wb_handle.shutdown_tx.send(()).await;
114 : // Await the task itself. TODO: restart panicked tasks earlier.
115 0 : wb_handle.join().await;
116 0 : }
117 0 : }
118 :
119 : /// The goal is to ensure that normally only one safekeepers offloads. However,
120 : /// it is fine (and inevitable, as s3 doesn't provide CAS) that for some short
121 : /// time we have several ones as they PUT the same files. Also,
122 : /// - frequently changing the offloader would be bad;
123 : /// - electing seriously lagging safekeeper is undesirable;
124 : ///
125 : /// So we deterministically choose among the reasonably caught up candidates.
126 : /// TODO: take into account failed attempts to deal with hypothetical situation
127 : /// where s3 is unreachable only for some sks.
128 0 : fn determine_offloader(
129 0 : alive_peers: &[PeerInfo],
130 0 : wal_backup_lsn: Lsn,
131 0 : ttid: TenantTimelineId,
132 0 : conf: &SafeKeeperConf,
133 0 : ) -> (Option<NodeId>, String) {
134 0 : // TODO: remove this once we fill newly joined safekeepers since backup_lsn.
135 0 : let capable_peers = alive_peers
136 0 : .iter()
137 0 : .filter(|p| p.local_start_lsn <= wal_backup_lsn);
138 0 : match capable_peers.clone().map(|p| p.commit_lsn).max() {
139 0 : None => (None, "no connected peers to elect from".to_string()),
140 0 : Some(max_commit_lsn) => {
141 0 : let threshold = max_commit_lsn
142 0 : .checked_sub(conf.max_offloader_lag_bytes)
143 0 : .unwrap_or(Lsn(0));
144 0 : let mut caughtup_peers = capable_peers
145 0 : .clone()
146 0 : .filter(|p| p.commit_lsn >= threshold)
147 0 : .collect::<Vec<_>>();
148 0 : caughtup_peers.sort_by(|p1, p2| p1.sk_id.cmp(&p2.sk_id));
149 0 :
150 0 : // To distribute the load, shift by timeline_id.
151 0 : let offloader = caughtup_peers
152 0 : [(u128::from(ttid.timeline_id) % caughtup_peers.len() as u128) as usize]
153 0 : .sk_id;
154 0 :
155 0 : let mut capable_peers_dbg = capable_peers
156 0 : .map(|p| (p.sk_id, p.commit_lsn))
157 0 : .collect::<Vec<_>>();
158 0 : capable_peers_dbg.sort_by(|p1, p2| p1.0.cmp(&p2.0));
159 0 : (
160 0 : Some(offloader),
161 0 : format!(
162 0 : "elected {} among {:?} peers, with {} of them being caughtup",
163 0 : offloader,
164 0 : capable_peers_dbg,
165 0 : caughtup_peers.len()
166 0 : ),
167 0 : )
168 : }
169 : }
170 0 : }
171 :
172 : static REMOTE_STORAGE: OnceCell<Option<GenericRemoteStorage>> = OnceCell::const_new();
173 :
174 : // Storage must be configured and initialized when this is called.
175 0 : fn get_configured_remote_storage() -> &'static GenericRemoteStorage {
176 0 : REMOTE_STORAGE
177 0 : .get()
178 0 : .expect("failed to get remote storage")
179 0 : .as_ref()
180 0 : .unwrap()
181 0 : }
182 :
183 0 : pub async fn init_remote_storage(conf: &SafeKeeperConf) {
184 0 : // TODO: refactor REMOTE_STORAGE to avoid using global variables, and provide
185 0 : // dependencies to all tasks instead.
186 0 : REMOTE_STORAGE
187 0 : .get_or_init(|| async {
188 0 : if let Some(conf) = conf.remote_storage.as_ref() {
189 : Some(
190 0 : GenericRemoteStorage::from_config(conf)
191 0 : .await
192 0 : .expect("failed to create remote storage"),
193 : )
194 : } else {
195 0 : None
196 : }
197 0 : })
198 0 : .await;
199 0 : }
200 :
201 : struct WalBackupTask {
202 : timeline: WalResidentTimeline,
203 : timeline_dir: Utf8PathBuf,
204 : wal_seg_size: usize,
205 : parallel_jobs: usize,
206 : commit_lsn_watch_rx: watch::Receiver<Lsn>,
207 : }
208 :
209 : /// Offload single timeline.
210 : #[instrument(name = "wal_backup", skip_all, fields(ttid = %tli.ttid))]
211 : async fn backup_task_main(
212 : tli: WalResidentTimeline,
213 : parallel_jobs: usize,
214 : mut shutdown_rx: Receiver<()>,
215 : ) {
216 : let _guard = WAL_BACKUP_TASKS.guard();
217 : info!("started");
218 :
219 : let cancel = tli.tli.cancel.clone();
220 : let mut wb = WalBackupTask {
221 : wal_seg_size: tli.get_wal_seg_size().await,
222 : commit_lsn_watch_rx: tli.get_commit_lsn_watch_rx(),
223 : timeline_dir: tli.get_timeline_dir(),
224 : timeline: tli,
225 : parallel_jobs,
226 : };
227 :
228 : // task is spinned up only when wal_seg_size already initialized
229 : assert!(wb.wal_seg_size > 0);
230 :
231 : let mut canceled = false;
232 : select! {
233 : _ = wb.run() => {}
234 : _ = shutdown_rx.recv() => {
235 : canceled = true;
236 : },
237 : _ = cancel.cancelled() => {
238 : canceled = true;
239 : }
240 : }
241 : info!("task {}", if canceled { "canceled" } else { "terminated" });
242 : }
243 :
244 : impl WalBackupTask {
245 : /// This function must be called from a select! that also respects self.timeline's
246 : /// cancellation token. This is done in [`backup_task_main`].
247 : ///
248 : /// The future returned by this function is safe to drop at any time because it
249 : /// does not write to local disk.
250 0 : async fn run(&mut self) {
251 0 : let mut backup_lsn = Lsn(0);
252 0 :
253 0 : let mut retry_attempt = 0u32;
254 : // offload loop
255 0 : while !self.timeline.cancel.is_cancelled() {
256 0 : if retry_attempt == 0 {
257 : // wait for new WAL to arrive
258 0 : if let Err(e) = self.commit_lsn_watch_rx.changed().await {
259 : // should never happen, as we hold Arc to timeline and transmitter's lifetime
260 : // is within Timeline's
261 0 : error!("commit_lsn watch shut down: {:?}", e);
262 0 : return;
263 0 : };
264 : } else {
265 : // or just sleep if we errored previously
266 0 : let mut retry_delay = UPLOAD_FAILURE_RETRY_MAX_MS;
267 0 : if let Some(backoff_delay) = UPLOAD_FAILURE_RETRY_MIN_MS.checked_shl(retry_attempt)
268 0 : {
269 0 : retry_delay = min(retry_delay, backoff_delay);
270 0 : }
271 0 : tokio::time::sleep(Duration::from_millis(retry_delay)).await;
272 : }
273 :
274 0 : let commit_lsn = *self.commit_lsn_watch_rx.borrow();
275 0 :
276 0 : // Note that backup_lsn can be higher than commit_lsn if we
277 0 : // don't have much local WAL and others already uploaded
278 0 : // segments we don't even have.
279 0 : if backup_lsn.segment_number(self.wal_seg_size)
280 0 : >= commit_lsn.segment_number(self.wal_seg_size)
281 : {
282 0 : retry_attempt = 0;
283 0 : continue; /* nothing to do, common case as we wake up on every commit_lsn bump */
284 0 : }
285 0 : // Perhaps peers advanced the position, check shmem value.
286 0 : backup_lsn = self.timeline.get_wal_backup_lsn().await;
287 0 : if backup_lsn.segment_number(self.wal_seg_size)
288 0 : >= commit_lsn.segment_number(self.wal_seg_size)
289 : {
290 0 : retry_attempt = 0;
291 0 : continue;
292 0 : }
293 0 :
294 0 : match backup_lsn_range(
295 0 : &self.timeline,
296 0 : &mut backup_lsn,
297 0 : commit_lsn,
298 0 : self.wal_seg_size,
299 0 : &self.timeline_dir,
300 0 : self.parallel_jobs,
301 0 : )
302 0 : .await
303 : {
304 0 : Ok(()) => {
305 0 : retry_attempt = 0;
306 0 : }
307 0 : Err(e) => {
308 0 : // We might have managed to upload some segment even though
309 0 : // some later in the range failed, so log backup_lsn
310 0 : // separately.
311 0 : error!(
312 0 : "failed while offloading range {}-{}, backup_lsn {}: {:?}",
313 : backup_lsn, commit_lsn, backup_lsn, e
314 : );
315 :
316 0 : retry_attempt = retry_attempt.saturating_add(1);
317 : }
318 : }
319 : }
320 0 : }
321 : }
322 :
323 0 : async fn backup_lsn_range(
324 0 : timeline: &WalResidentTimeline,
325 0 : backup_lsn: &mut Lsn,
326 0 : end_lsn: Lsn,
327 0 : wal_seg_size: usize,
328 0 : timeline_dir: &Utf8Path,
329 0 : parallel_jobs: usize,
330 0 : ) -> Result<()> {
331 0 : if parallel_jobs < 1 {
332 0 : anyhow::bail!("parallel_jobs must be >= 1");
333 0 : }
334 0 :
335 0 : let remote_timeline_path = &timeline.remote_path;
336 0 : let start_lsn = *backup_lsn;
337 0 : let segments = get_segments(start_lsn, end_lsn, wal_seg_size);
338 0 :
339 0 : info!(
340 0 : "offloading segnos {:?} of range [{}-{})",
341 0 : segments.iter().map(|&s| s.seg_no).collect::<Vec<_>>(),
342 : start_lsn,
343 : end_lsn,
344 : );
345 :
346 : // Pool of concurrent upload tasks. We use `FuturesOrdered` to
347 : // preserve order of uploads, and update `backup_lsn` only after
348 : // all previous uploads are finished.
349 0 : let mut uploads = FuturesOrdered::new();
350 0 : let mut iter = segments.iter();
351 :
352 : loop {
353 0 : let added_task = match iter.next() {
354 0 : Some(s) => {
355 0 : uploads.push_back(backup_single_segment(s, timeline_dir, remote_timeline_path));
356 0 : true
357 : }
358 0 : None => false,
359 : };
360 :
361 : // Wait for the next segment to upload if we don't have any more segments,
362 : // or if we have too many concurrent uploads.
363 0 : if !added_task || uploads.len() >= parallel_jobs {
364 0 : let next = uploads.next().await;
365 0 : if let Some(res) = next {
366 : // next segment uploaded
367 0 : let segment = res?;
368 0 : let new_backup_lsn = segment.end_lsn;
369 0 : timeline
370 0 : .set_wal_backup_lsn(new_backup_lsn)
371 0 : .await
372 0 : .context("setting wal_backup_lsn")?;
373 0 : *backup_lsn = new_backup_lsn;
374 : } else {
375 : // no more segments to upload
376 0 : break;
377 : }
378 0 : }
379 : }
380 :
381 0 : info!(
382 0 : "offloaded segnos {:?} of range [{}-{})",
383 0 : segments.iter().map(|&s| s.seg_no).collect::<Vec<_>>(),
384 : start_lsn,
385 : end_lsn,
386 : );
387 0 : Ok(())
388 0 : }
389 :
390 0 : async fn backup_single_segment(
391 0 : seg: &Segment,
392 0 : timeline_dir: &Utf8Path,
393 0 : remote_timeline_path: &RemotePath,
394 0 : ) -> Result<Segment> {
395 0 : let segment_file_path = seg.file_path(timeline_dir)?;
396 0 : let remote_segment_path = seg.remote_path(remote_timeline_path);
397 :
398 0 : let res = backup_object(&segment_file_path, &remote_segment_path, seg.size()).await;
399 0 : if res.is_ok() {
400 0 : BACKED_UP_SEGMENTS.inc();
401 0 : } else {
402 0 : BACKUP_ERRORS.inc();
403 0 : }
404 0 : res?;
405 0 : debug!("Backup of {} done", segment_file_path);
406 :
407 0 : Ok(*seg)
408 0 : }
409 :
410 : #[derive(Debug, Copy, Clone)]
411 : pub struct Segment {
412 : seg_no: XLogSegNo,
413 : start_lsn: Lsn,
414 : end_lsn: Lsn,
415 : }
416 :
417 : impl Segment {
418 0 : pub fn new(seg_no: u64, start_lsn: Lsn, end_lsn: Lsn) -> Self {
419 0 : Self {
420 0 : seg_no,
421 0 : start_lsn,
422 0 : end_lsn,
423 0 : }
424 0 : }
425 :
426 0 : pub fn object_name(self) -> String {
427 0 : XLogFileName(PG_TLI, self.seg_no, self.size())
428 0 : }
429 :
430 0 : pub fn file_path(self, timeline_dir: &Utf8Path) -> Result<Utf8PathBuf> {
431 0 : Ok(timeline_dir.join(self.object_name()))
432 0 : }
433 :
434 0 : pub fn remote_path(self, remote_timeline_path: &RemotePath) -> RemotePath {
435 0 : remote_timeline_path.join(self.object_name())
436 0 : }
437 :
438 0 : pub fn size(self) -> usize {
439 0 : (u64::from(self.end_lsn) - u64::from(self.start_lsn)) as usize
440 0 : }
441 : }
442 :
443 0 : fn get_segments(start: Lsn, end: Lsn, seg_size: usize) -> Vec<Segment> {
444 0 : let first_seg = start.segment_number(seg_size);
445 0 : let last_seg = end.segment_number(seg_size);
446 0 :
447 0 : let res: Vec<Segment> = (first_seg..last_seg)
448 0 : .map(|s| {
449 0 : let start_lsn = XLogSegNoOffsetToRecPtr(s, 0, seg_size);
450 0 : let end_lsn = XLogSegNoOffsetToRecPtr(s + 1, 0, seg_size);
451 0 : Segment::new(s, Lsn::from(start_lsn), Lsn::from(end_lsn))
452 0 : })
453 0 : .collect();
454 0 : res
455 0 : }
456 :
457 0 : async fn backup_object(
458 0 : source_file: &Utf8Path,
459 0 : target_file: &RemotePath,
460 0 : size: usize,
461 0 : ) -> Result<()> {
462 0 : let storage = get_configured_remote_storage();
463 :
464 0 : let file = File::open(&source_file)
465 0 : .await
466 0 : .with_context(|| format!("Failed to open file {source_file:?} for wal backup"))?;
467 :
468 0 : let file = tokio_util::io::ReaderStream::with_capacity(file, BUFFER_SIZE);
469 0 :
470 0 : let cancel = CancellationToken::new();
471 0 :
472 0 : storage
473 0 : .upload_storage_object(file, size, target_file, &cancel)
474 0 : .await
475 0 : }
476 :
477 0 : pub(crate) async fn backup_partial_segment(
478 0 : source_file: &Utf8Path,
479 0 : target_file: &RemotePath,
480 0 : size: usize,
481 0 : ) -> Result<()> {
482 0 : let storage = get_configured_remote_storage();
483 :
484 0 : let file = File::open(&source_file)
485 0 : .await
486 0 : .with_context(|| format!("Failed to open file {source_file:?} for wal backup"))?;
487 :
488 : // limiting the file to read only the first `size` bytes
489 0 : let limited_file = tokio::io::AsyncReadExt::take(file, size as u64);
490 0 :
491 0 : let file = tokio_util::io::ReaderStream::with_capacity(limited_file, BUFFER_SIZE);
492 0 :
493 0 : let cancel = CancellationToken::new();
494 0 :
495 0 : storage
496 0 : .upload(
497 0 : file,
498 0 : size,
499 0 : target_file,
500 0 : Some(StorageMetadata::from([("sk_type", "partial_segment")])),
501 0 : &cancel,
502 0 : )
503 0 : .await
504 0 : }
505 :
506 0 : pub(crate) async fn copy_partial_segment(
507 0 : source: &RemotePath,
508 0 : destination: &RemotePath,
509 0 : ) -> Result<()> {
510 0 : let storage = get_configured_remote_storage();
511 0 : let cancel = CancellationToken::new();
512 0 :
513 0 : storage.copy_object(source, destination, &cancel).await
514 0 : }
515 :
516 0 : pub async fn read_object(
517 0 : file_path: &RemotePath,
518 0 : offset: u64,
519 0 : ) -> anyhow::Result<Pin<Box<dyn tokio::io::AsyncRead + Send + Sync>>> {
520 0 : let storage = REMOTE_STORAGE
521 0 : .get()
522 0 : .context("Failed to get remote storage")?
523 0 : .as_ref()
524 0 : .context("No remote storage configured")?;
525 :
526 0 : info!("segment download about to start from remote path {file_path:?} at offset {offset}");
527 :
528 0 : let cancel = CancellationToken::new();
529 0 :
530 0 : let opts = DownloadOpts {
531 0 : byte_start: std::ops::Bound::Included(offset),
532 0 : ..Default::default()
533 0 : };
534 0 : let download = storage
535 0 : .download(file_path, &opts, &cancel)
536 0 : .await
537 0 : .with_context(|| {
538 0 : format!("Failed to open WAL segment download stream for remote path {file_path:?}")
539 0 : })?;
540 :
541 0 : let reader = tokio_util::io::StreamReader::new(download.download_stream);
542 0 :
543 0 : let reader = tokio::io::BufReader::with_capacity(BUFFER_SIZE, reader);
544 0 :
545 0 : Ok(Box::pin(reader))
546 0 : }
547 :
548 : /// Delete WAL files for the given timeline. Remote storage must be configured
549 : /// when called.
550 0 : pub async fn delete_timeline(ttid: &TenantTimelineId) -> Result<()> {
551 0 : let storage = get_configured_remote_storage();
552 0 : let remote_path = remote_timeline_path(ttid)?;
553 :
554 : // see DEFAULT_MAX_KEYS_PER_LIST_RESPONSE
555 : // const Option unwrap is not stable, otherwise it would be const.
556 0 : let batch_size: NonZeroU32 = NonZeroU32::new(1000).unwrap();
557 0 :
558 0 : // A backoff::retry is used here for two reasons:
559 0 : // - To provide a backoff rather than busy-polling the API on errors
560 0 : // - To absorb transient 429/503 conditions without hitting our error
561 0 : // logging path for issues deleting objects.
562 0 : //
563 0 : // Note: listing segments might take a long time if there are many of them.
564 0 : // We don't currently have http requests timeout cancellation, but if/once
565 0 : // we have listing should get streaming interface to make progress.
566 0 :
567 0 : pausable_failpoint!("sk-delete-timeline-remote-pause");
568 :
569 0 : fail::fail_point!("sk-delete-timeline-remote", |_| {
570 0 : Err(anyhow::anyhow!("failpoint: sk-delete-timeline-remote"))
571 0 : });
572 :
573 0 : let cancel = CancellationToken::new(); // not really used
574 0 : backoff::retry(
575 0 : || async {
576 : // Do list-delete in batch_size batches to make progress even if there a lot of files.
577 : // Alternatively we could make remote storage list return iterator, but it is more complicated and
578 : // I'm not sure deleting while iterating is expected in s3.
579 : loop {
580 0 : let files = storage
581 0 : .list(
582 0 : Some(&remote_path),
583 0 : ListingMode::NoDelimiter,
584 0 : Some(batch_size),
585 0 : &cancel,
586 0 : )
587 0 : .await?
588 : .keys
589 0 : .into_iter()
590 0 : .map(|o| o.key)
591 0 : .collect::<Vec<_>>();
592 0 : if files.is_empty() {
593 0 : return Ok(()); // done
594 0 : }
595 0 : // (at least) s3 results are sorted, so can log min/max:
596 0 : // "List results are always returned in UTF-8 binary order."
597 0 : info!(
598 0 : "deleting batch of {} WAL segments [{}-{}]",
599 0 : files.len(),
600 0 : files.first().unwrap().object_name().unwrap_or(""),
601 0 : files.last().unwrap().object_name().unwrap_or("")
602 : );
603 0 : storage.delete_objects(&files, &cancel).await?;
604 : }
605 0 : },
606 0 : // consider TimeoutOrCancel::caused_by_cancel when using cancellation
607 0 : |_| false,
608 0 : 3,
609 0 : 10,
610 0 : "executing WAL segments deletion batch",
611 0 : &cancel,
612 0 : )
613 0 : .await
614 0 : .ok_or_else(|| anyhow::anyhow!("canceled"))
615 0 : .and_then(|x| x)?;
616 :
617 0 : Ok(())
618 0 : }
619 :
620 : /// Used by wal_backup_partial.
621 0 : pub async fn delete_objects(paths: &[RemotePath]) -> Result<()> {
622 0 : let cancel = CancellationToken::new(); // not really used
623 0 : let storage = get_configured_remote_storage();
624 0 : storage.delete_objects(paths, &cancel).await
625 0 : }
626 :
627 : /// Copy segments from one timeline to another. Used in copy_timeline.
628 0 : pub async fn copy_s3_segments(
629 0 : wal_seg_size: usize,
630 0 : src_ttid: &TenantTimelineId,
631 0 : dst_ttid: &TenantTimelineId,
632 0 : from_segment: XLogSegNo,
633 0 : to_segment: XLogSegNo,
634 0 : ) -> Result<()> {
635 : const SEGMENTS_PROGRESS_REPORT_INTERVAL: u64 = 1024;
636 :
637 0 : let storage = REMOTE_STORAGE
638 0 : .get()
639 0 : .expect("failed to get remote storage")
640 0 : .as_ref()
641 0 : .unwrap();
642 :
643 0 : let remote_dst_path = remote_timeline_path(dst_ttid)?;
644 :
645 0 : let cancel = CancellationToken::new();
646 :
647 0 : let files = storage
648 0 : .list(
649 0 : Some(&remote_dst_path),
650 0 : ListingMode::NoDelimiter,
651 0 : None,
652 0 : &cancel,
653 0 : )
654 0 : .await?
655 : .keys;
656 :
657 0 : let uploaded_segments = &files
658 0 : .iter()
659 0 : .filter_map(|o| o.key.object_name().map(ToOwned::to_owned))
660 0 : .collect::<HashSet<_>>();
661 0 :
662 0 : debug!(
663 0 : "these segments have already been uploaded: {:?}",
664 : uploaded_segments
665 : );
666 :
667 0 : for segno in from_segment..to_segment {
668 0 : if segno % SEGMENTS_PROGRESS_REPORT_INTERVAL == 0 {
669 0 : info!("copied all segments from {} until {}", from_segment, segno);
670 0 : }
671 :
672 0 : let segment_name = XLogFileName(PG_TLI, segno, wal_seg_size);
673 0 : if uploaded_segments.contains(&segment_name) {
674 0 : continue;
675 0 : }
676 0 : debug!("copying segment {}", segment_name);
677 :
678 0 : let from = remote_timeline_path(src_ttid)?.join(&segment_name);
679 0 : let to = remote_dst_path.join(&segment_name);
680 0 :
681 0 : storage.copy_object(&from, &to, &cancel).await?;
682 : }
683 :
684 0 : info!(
685 0 : "finished copying segments from {} until {}",
686 : from_segment, to_segment
687 : );
688 0 : Ok(())
689 0 : }
690 :
691 : /// Get S3 (remote_storage) prefix path used for timeline files.
692 13 : pub fn remote_timeline_path(ttid: &TenantTimelineId) -> Result<RemotePath> {
693 13 : RemotePath::new(&Utf8Path::new(&ttid.tenant_id.to_string()).join(ttid.timeline_id.to_string()))
694 13 : }
|
