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