Line data Source code
1 : use std::cmp::min;
2 : use std::io::{self, ErrorKind};
3 : use std::ops::RangeInclusive;
4 : use std::sync::Arc;
5 :
6 : use anyhow::{Context, Result, anyhow, bail};
7 : use bytes::Bytes;
8 : use camino::Utf8PathBuf;
9 : use chrono::{DateTime, Utc};
10 : use futures::{SinkExt, StreamExt, TryStreamExt};
11 : use http_utils::error::ApiError;
12 : use postgres_ffi::{PG_TLI, XLogFileName, XLogSegNo};
13 : use remote_storage::GenericRemoteStorage;
14 : use reqwest::Certificate;
15 : use safekeeper_api::Term;
16 : use safekeeper_api::models::{PullTimelineRequest, PullTimelineResponse, TimelineStatus};
17 : use safekeeper_client::mgmt_api;
18 : use safekeeper_client::mgmt_api::Client;
19 : use serde::Deserialize;
20 : use tokio::fs::OpenOptions;
21 : use tokio::io::AsyncWrite;
22 : use tokio::sync::mpsc;
23 : use tokio::task;
24 : use tokio_tar::{Archive, Builder, Header};
25 : use tokio_util::io::{CopyToBytes, SinkWriter};
26 : use tokio_util::sync::PollSender;
27 : use tracing::{error, info, instrument};
28 : use utils::crashsafe::fsync_async_opt;
29 : use utils::id::{NodeId, TenantTimelineId};
30 : use utils::logging::SecretString;
31 : use utils::lsn::Lsn;
32 : use utils::pausable_failpoint;
33 :
34 : use crate::control_file::CONTROL_FILE_NAME;
35 : use crate::state::{EvictionState, TimelinePersistentState};
36 : use crate::timeline::{Timeline, TimelineError, WalResidentTimeline};
37 : use crate::timelines_global_map::{create_temp_timeline_dir, validate_temp_timeline};
38 : use crate::wal_storage::{open_wal_file, wal_file_paths};
39 : use crate::{GlobalTimelines, debug_dump, wal_backup};
40 :
41 : /// Stream tar archive of timeline to tx.
42 : #[instrument(name = "snapshot", skip_all, fields(ttid = %tli.ttid))]
43 : pub async fn stream_snapshot(
44 : tli: Arc<Timeline>,
45 : source: NodeId,
46 : destination: NodeId,
47 : tx: mpsc::Sender<Result<Bytes>>,
48 : storage: Option<Arc<GenericRemoteStorage>>,
49 : ) {
50 : match tli.try_wal_residence_guard().await {
51 : Err(e) => {
52 : tx.send(Err(anyhow!("Error checking residence: {:#}", e)))
53 : .await
54 : .ok();
55 : }
56 : Ok(maybe_resident_tli) => {
57 : if let Err(e) = match maybe_resident_tli {
58 : Some(resident_tli) => {
59 : stream_snapshot_resident_guts(
60 : resident_tli,
61 : source,
62 : destination,
63 : tx.clone(),
64 : storage,
65 : )
66 : .await
67 : }
68 : None => {
69 : if let Some(storage) = storage {
70 : stream_snapshot_offloaded_guts(
71 : tli,
72 : source,
73 : destination,
74 : tx.clone(),
75 : &storage,
76 : )
77 : .await
78 : } else {
79 : tx.send(Err(anyhow!("remote storage not configured")))
80 : .await
81 : .ok();
82 : return;
83 : }
84 : }
85 : } {
86 : // Error type/contents don't matter as they won't can't reach the client
87 : // (hyper likely doesn't do anything with it), but http stream will be
88 : // prematurely terminated. It would be nice to try to send the error in
89 : // trailers though.
90 : tx.send(Err(anyhow!("snapshot failed"))).await.ok();
91 : error!("snapshot failed: {:#}", e);
92 : }
93 : }
94 : }
95 : }
96 :
97 : /// State needed while streaming the snapshot.
98 : pub struct SnapshotContext {
99 : /// The interval of segment numbers. If None, the timeline hasn't had writes yet, so only send the control file
100 : pub from_to_segno: Option<RangeInclusive<XLogSegNo>>,
101 : pub term: Term,
102 : pub last_log_term: Term,
103 : pub flush_lsn: Lsn,
104 : pub wal_seg_size: usize,
105 : // used to remove WAL hold off in Drop.
106 : pub tli: WalResidentTimeline,
107 : }
108 :
109 : impl Drop for SnapshotContext {
110 0 : fn drop(&mut self) {
111 0 : let tli = self.tli.clone();
112 0 : task::spawn(async move {
113 0 : let mut shared_state = tli.write_shared_state().await;
114 0 : shared_state.wal_removal_on_hold = false;
115 0 : });
116 0 : }
117 : }
118 :
119 : /// Build a tokio_tar stream that sends encoded bytes into a Bytes channel.
120 0 : fn prepare_tar_stream(
121 0 : tx: mpsc::Sender<Result<Bytes>>,
122 0 : ) -> tokio_tar::Builder<impl AsyncWrite + Unpin + Send> {
123 : // tokio-tar wants Write implementor, but we have mpsc tx <Result<Bytes>>;
124 : // use SinkWriter as a Write impl. That is,
125 : // - create Sink from the tx. It returns PollSendError if chan is closed.
126 0 : let sink = PollSender::new(tx);
127 : // - SinkWriter needs sink error to be io one, map it.
128 0 : let sink_io_err = sink.sink_map_err(|_| io::Error::from(ErrorKind::BrokenPipe));
129 : // - SinkWriter wants sink type to be just Bytes, not Result<Bytes>, so map
130 : // it with with(). Note that with() accepts async function which we don't
131 : // need and allows the map to fail, which we don't need either, but hence
132 : // two Oks.
133 0 : let oksink = sink_io_err.with(|b: Bytes| async { io::Result::Ok(Result::Ok(b)) });
134 : // - SinkWriter (not surprisingly) wants sink of &[u8], not bytes, so wrap
135 : // into CopyToBytes. This is a data copy.
136 0 : let copy_to_bytes = CopyToBytes::new(oksink);
137 0 : let writer = SinkWriter::new(copy_to_bytes);
138 0 : let pinned_writer = Box::pin(writer);
139 :
140 : // Note that tokio_tar append_* funcs use tokio::io::copy with 8KB buffer
141 : // which is also likely suboptimal.
142 0 : Builder::new_non_terminated(pinned_writer)
143 0 : }
144 :
145 : /// Implementation of snapshot for an offloaded timeline, only reads control file
146 0 : pub(crate) async fn stream_snapshot_offloaded_guts(
147 0 : tli: Arc<Timeline>,
148 0 : source: NodeId,
149 0 : destination: NodeId,
150 0 : tx: mpsc::Sender<Result<Bytes>>,
151 0 : storage: &GenericRemoteStorage,
152 0 : ) -> Result<()> {
153 0 : let mut ar = prepare_tar_stream(tx);
154 :
155 0 : tli.snapshot_offloaded(&mut ar, source, destination, storage)
156 0 : .await?;
157 :
158 0 : ar.finish().await?;
159 :
160 0 : Ok(())
161 0 : }
162 :
163 : /// Implementation of snapshot for a timeline which is resident (includes some segment data)
164 0 : pub async fn stream_snapshot_resident_guts(
165 0 : tli: WalResidentTimeline,
166 0 : source: NodeId,
167 0 : destination: NodeId,
168 0 : tx: mpsc::Sender<Result<Bytes>>,
169 0 : storage: Option<Arc<GenericRemoteStorage>>,
170 0 : ) -> Result<()> {
171 0 : let mut ar = prepare_tar_stream(tx);
172 :
173 0 : let bctx = tli
174 0 : .start_snapshot(&mut ar, source, destination, storage)
175 0 : .await?;
176 0 : pausable_failpoint!("sk-snapshot-after-list-pausable");
177 :
178 0 : if let Some(from_to_segno) = &bctx.from_to_segno {
179 0 : let tli_dir = tli.get_timeline_dir();
180 0 : info!(
181 0 : "sending {} segments [{:#X}-{:#X}], term={}, last_log_term={}, flush_lsn={}",
182 0 : from_to_segno.end() - from_to_segno.start() + 1,
183 0 : from_to_segno.start(),
184 0 : from_to_segno.end(),
185 : bctx.term,
186 : bctx.last_log_term,
187 : bctx.flush_lsn,
188 : );
189 0 : for segno in from_to_segno.clone() {
190 0 : let Some((mut sf, is_partial)) =
191 0 : open_wal_file(&tli_dir, segno, bctx.wal_seg_size).await?
192 : else {
193 : // File is not found
194 0 : let (wal_file_path, _wal_file_partial_path) =
195 0 : wal_file_paths(&tli_dir, segno, bctx.wal_seg_size);
196 0 : tracing::warn!("couldn't find WAL segment file {wal_file_path}");
197 0 : bail!("couldn't find WAL segment file {wal_file_path}")
198 : };
199 0 : let mut wal_file_name = XLogFileName(PG_TLI, segno, bctx.wal_seg_size);
200 0 : if is_partial {
201 0 : wal_file_name.push_str(".partial");
202 0 : }
203 0 : ar.append_file(&wal_file_name, &mut sf).await?;
204 : }
205 : } else {
206 0 : info!("Not including any segments into the snapshot");
207 : }
208 :
209 : // Do the term check before ar.finish to make archive corrupted in case of
210 : // term change. Client shouldn't ignore abrupt stream end, but to be sure.
211 0 : tli.finish_snapshot(&bctx).await?;
212 :
213 0 : ar.finish().await?;
214 :
215 0 : Ok(())
216 0 : }
217 :
218 : impl Timeline {
219 : /// Simple snapshot for an offloaded timeline: we will only upload a renamed partial segment and
220 : /// pass a modified control file into the provided tar stream (nothing with data segments on disk, since
221 : /// we are offloaded and there aren't any)
222 0 : async fn snapshot_offloaded<W: AsyncWrite + Unpin + Send>(
223 0 : self: &Arc<Timeline>,
224 0 : ar: &mut tokio_tar::Builder<W>,
225 0 : source: NodeId,
226 0 : destination: NodeId,
227 0 : storage: &GenericRemoteStorage,
228 0 : ) -> Result<()> {
229 : // Take initial copy of control file, then release state lock
230 0 : let mut control_file = {
231 0 : let shared_state = self.write_shared_state().await;
232 :
233 0 : let control_file = TimelinePersistentState::clone(shared_state.sk.state());
234 :
235 : // Rare race: we got unevicted between entering function and reading control file.
236 : // We error out and let API caller retry.
237 0 : if !matches!(control_file.eviction_state, EvictionState::Offloaded(_)) {
238 0 : bail!("Timeline was un-evicted during snapshot, please retry");
239 0 : }
240 :
241 0 : control_file
242 : };
243 :
244 : // Modify the partial segment of the in-memory copy for the control file to
245 : // point to the destination safekeeper.
246 0 : let replace = control_file
247 0 : .partial_backup
248 0 : .replace_uploaded_segment(source, destination)?;
249 :
250 0 : let Some(replace) = replace else {
251 : // In Manager:: ready_for_eviction, we do not permit eviction unless the timeline
252 : // has a partial segment. It is unexpected that
253 0 : anyhow::bail!("Timeline has no partial segment, cannot generate snapshot");
254 : };
255 :
256 0 : tracing::info!("Replacing uploaded partial segment in in-mem control file: {replace:?}");
257 :
258 : // Optimistically try to copy the partial segment to the destination's path: this
259 : // can fail if the timeline was un-evicted and modified in the background.
260 0 : let remote_timeline_path = &self.remote_path;
261 0 : wal_backup::copy_partial_segment(
262 0 : storage,
263 0 : &replace.previous.remote_path(remote_timeline_path),
264 0 : &replace.current.remote_path(remote_timeline_path),
265 0 : )
266 0 : .await?;
267 :
268 : // Since the S3 copy succeeded with the path given in our control file snapshot, and
269 : // we are sending that snapshot in our response, we are giving the caller a consistent
270 : // snapshot even if our local Timeline was unevicted or otherwise modified in the meantime.
271 0 : let buf = control_file
272 0 : .write_to_buf()
273 0 : .with_context(|| "failed to serialize control store")?;
274 0 : let mut header = Header::new_gnu();
275 0 : header.set_size(buf.len().try_into().expect("never breaches u64"));
276 0 : ar.append_data(&mut header, CONTROL_FILE_NAME, buf.as_slice())
277 0 : .await
278 0 : .with_context(|| "failed to append to archive")?;
279 :
280 0 : Ok(())
281 0 : }
282 : }
283 :
284 : impl WalResidentTimeline {
285 : /// Start streaming tar archive with timeline:
286 : /// 1) stream control file under lock;
287 : /// 2) hold off WAL removal;
288 : /// 3) collect SnapshotContext to understand which WAL segments should be
289 : /// streamed.
290 : ///
291 : /// Snapshot streams data up to flush_lsn. To make this safe, we must check
292 : /// that term doesn't change during the procedure, or we risk sending mix of
293 : /// WAL from different histories. Term is remembered in the SnapshotContext
294 : /// and checked in finish_snapshot. Note that in the last segment some WAL
295 : /// higher than flush_lsn set here might be streamed; that's fine as long as
296 : /// terms doesn't change.
297 : ///
298 : /// Alternatively we could send only up to commit_lsn to get some valid
299 : /// state which later will be recovered by compute, in this case term check
300 : /// is not needed, but we likely don't want that as there might be no
301 : /// compute which could perform the recovery.
302 : ///
303 : /// When returned SnapshotContext is dropped WAL hold is removed.
304 0 : async fn start_snapshot<W: AsyncWrite + Unpin + Send>(
305 0 : &self,
306 0 : ar: &mut tokio_tar::Builder<W>,
307 0 : source: NodeId,
308 0 : destination: NodeId,
309 0 : storage: Option<Arc<GenericRemoteStorage>>,
310 0 : ) -> Result<SnapshotContext> {
311 0 : let mut shared_state = self.write_shared_state().await;
312 0 : let wal_seg_size = shared_state.get_wal_seg_size();
313 :
314 0 : let mut control_store = TimelinePersistentState::clone(shared_state.sk.state());
315 : // Modify the partial segment of the in-memory copy for the control file to
316 : // point to the destination safekeeper.
317 0 : let replace = control_store
318 0 : .partial_backup
319 0 : .replace_uploaded_segment(source, destination)?;
320 :
321 0 : if let Some(replace) = replace {
322 : // The deserialized control file has an uploaded partial. We upload a copy
323 : // of it to object storage for the destination safekeeper and send an updated
324 : // control file in the snapshot.
325 0 : tracing::info!(
326 0 : "Replacing uploaded partial segment in in-mem control file: {replace:?}"
327 : );
328 :
329 0 : let remote_timeline_path = &self.tli.remote_path;
330 0 : wal_backup::copy_partial_segment(
331 0 : &*storage.context("remote storage not configured")?,
332 0 : &replace.previous.remote_path(remote_timeline_path),
333 0 : &replace.current.remote_path(remote_timeline_path),
334 : )
335 0 : .await?;
336 0 : }
337 :
338 0 : let buf = control_store
339 0 : .write_to_buf()
340 0 : .with_context(|| "failed to serialize control store")?;
341 0 : let mut header = Header::new_gnu();
342 0 : header.set_size(buf.len().try_into().expect("never breaches u64"));
343 0 : ar.append_data(&mut header, CONTROL_FILE_NAME, buf.as_slice())
344 0 : .await
345 0 : .with_context(|| "failed to append to archive")?;
346 :
347 : // We need to stream since the oldest segment someone (s3 or pageserver)
348 : // still needs. This duplicates calc_horizon_lsn logic.
349 : //
350 : // We know that WAL wasn't removed up to this point because it cannot be
351 : // removed further than `backup_lsn`. Since we're holding shared_state
352 : // lock and setting `wal_removal_on_hold` later, it guarantees that WAL
353 : // won't be removed until we're done.
354 0 : let timeline_state = shared_state.sk.state();
355 0 : let from_lsn = min(
356 0 : timeline_state.remote_consistent_lsn,
357 0 : timeline_state.backup_lsn,
358 : );
359 0 : let flush_lsn = shared_state.sk.flush_lsn();
360 0 : let (send_segments, msg) = if from_lsn == Lsn::INVALID {
361 0 : (false, "snapshot is called on uninitialized timeline")
362 : } else {
363 0 : (true, "timeline is initialized")
364 : };
365 0 : tracing::info!(
366 0 : remote_consistent_lsn=%timeline_state.remote_consistent_lsn,
367 0 : backup_lsn=%timeline_state.backup_lsn,
368 : %flush_lsn,
369 0 : "{msg}"
370 : );
371 0 : let from_segno = from_lsn.segment_number(wal_seg_size);
372 0 : let term = shared_state.sk.state().acceptor_state.term;
373 0 : let last_log_term = shared_state.sk.last_log_term();
374 0 : let upto_segno = flush_lsn.segment_number(wal_seg_size);
375 : // have some limit on max number of segments as a sanity check
376 : const MAX_ALLOWED_SEGS: u64 = 1000;
377 0 : let num_segs = upto_segno - from_segno + 1;
378 0 : if num_segs > MAX_ALLOWED_SEGS {
379 0 : bail!(
380 0 : "snapshot is called on timeline with {} segments, but the limit is {}",
381 : num_segs,
382 : MAX_ALLOWED_SEGS
383 : );
384 0 : }
385 :
386 : // Prevent WAL removal while we're streaming data.
387 : //
388 : // Since this a flag, not a counter just bail out if already set; we
389 : // shouldn't need concurrent snapshotting.
390 0 : if shared_state.wal_removal_on_hold {
391 0 : bail!("wal_removal_on_hold is already true");
392 0 : }
393 0 : shared_state.wal_removal_on_hold = true;
394 :
395 : // Drop shared_state to release the lock, before calling wal_residence_guard().
396 0 : drop(shared_state);
397 :
398 0 : let tli_copy = self.wal_residence_guard().await?;
399 0 : let from_to_segno = send_segments.then_some(from_segno..=upto_segno);
400 0 : let bctx = SnapshotContext {
401 0 : from_to_segno,
402 0 : term,
403 0 : last_log_term,
404 0 : flush_lsn,
405 0 : wal_seg_size,
406 0 : tli: tli_copy,
407 0 : };
408 :
409 0 : Ok(bctx)
410 0 : }
411 :
412 : /// Finish snapshotting: check that term(s) hasn't changed.
413 : ///
414 : /// Note that WAL gc hold off is removed in Drop of SnapshotContext to not
415 : /// forget this if snapshotting fails mid the way.
416 0 : pub async fn finish_snapshot(&self, bctx: &SnapshotContext) -> Result<()> {
417 0 : let shared_state = self.read_shared_state().await;
418 0 : let term = shared_state.sk.state().acceptor_state.term;
419 0 : let last_log_term = shared_state.sk.last_log_term();
420 : // There are some cases to relax this check (e.g. last_log_term might
421 : // change, but as long as older history is strictly part of new that's
422 : // fine), but there is no need to do it.
423 0 : if bctx.term != term || bctx.last_log_term != last_log_term {
424 0 : bail!(
425 0 : "term(s) changed during snapshot: were term={}, last_log_term={}, now term={}, last_log_term={}",
426 : bctx.term,
427 : bctx.last_log_term,
428 : term,
429 : last_log_term
430 : );
431 0 : }
432 0 : Ok(())
433 0 : }
434 : }
435 :
436 : /// Response for debug dump request.
437 0 : #[derive(Debug, Deserialize)]
438 : pub struct DebugDumpResponse {
439 : pub start_time: DateTime<Utc>,
440 : pub finish_time: DateTime<Utc>,
441 : pub timelines: Vec<debug_dump::Timeline>,
442 : pub timelines_count: usize,
443 : pub config: debug_dump::Config,
444 : }
445 :
446 : /// Find the most advanced safekeeper and pull timeline from it.
447 0 : pub async fn handle_request(
448 0 : request: PullTimelineRequest,
449 0 : sk_auth_token: Option<SecretString>,
450 0 : ssl_ca_certs: Vec<Certificate>,
451 0 : global_timelines: Arc<GlobalTimelines>,
452 0 : ) -> Result<PullTimelineResponse, ApiError> {
453 0 : let existing_tli = global_timelines.get(TenantTimelineId::new(
454 0 : request.tenant_id,
455 0 : request.timeline_id,
456 0 : ));
457 0 : if existing_tli.is_ok() {
458 0 : info!("Timeline {} already exists", request.timeline_id);
459 0 : return Ok(PullTimelineResponse {
460 0 : safekeeper_host: None,
461 0 : });
462 0 : }
463 :
464 0 : let mut http_client = reqwest::Client::builder();
465 0 : for ssl_ca_cert in ssl_ca_certs {
466 0 : http_client = http_client.add_root_certificate(ssl_ca_cert);
467 0 : }
468 0 : let http_client = http_client
469 0 : .build()
470 0 : .map_err(|e| ApiError::InternalServerError(e.into()))?;
471 :
472 0 : let http_hosts = request.http_hosts.clone();
473 :
474 : // Figure out statuses of potential donors.
475 0 : let responses: Vec<Result<TimelineStatus, mgmt_api::Error>> =
476 0 : futures::future::join_all(http_hosts.iter().map(|url| async {
477 0 : let cclient = Client::new(http_client.clone(), url.clone(), sk_auth_token.clone());
478 0 : let info = cclient
479 0 : .timeline_status(request.tenant_id, request.timeline_id)
480 0 : .await?;
481 0 : Ok(info)
482 0 : }))
483 0 : .await;
484 :
485 0 : let mut statuses = Vec::new();
486 0 : for (i, response) in responses.into_iter().enumerate() {
487 0 : match response {
488 0 : Ok(status) => {
489 0 : statuses.push((status, i));
490 0 : }
491 0 : Err(e) => {
492 0 : info!("error fetching status from {}: {e}", http_hosts[i]);
493 : }
494 : }
495 : }
496 :
497 : // Allow missing responses from up to one safekeeper (say due to downtime)
498 : // e.g. if we created a timeline on PS A and B, with C being offline. Then B goes
499 : // offline and C comes online. Then we want a pull on C with A and B as hosts to work.
500 0 : let min_required_successful = (http_hosts.len() - 1).max(1);
501 0 : if statuses.len() < min_required_successful {
502 0 : return Err(ApiError::InternalServerError(anyhow::anyhow!(
503 0 : "only got {} successful status responses. required: {min_required_successful}",
504 0 : statuses.len()
505 0 : )));
506 0 : }
507 :
508 : // Find the most advanced safekeeper
509 0 : let (status, i) = statuses
510 0 : .into_iter()
511 0 : .max_by_key(|(status, _)| {
512 0 : (
513 0 : status.acceptor_state.epoch,
514 0 : status.flush_lsn,
515 0 : status.commit_lsn,
516 0 : )
517 0 : })
518 0 : .unwrap();
519 0 : let safekeeper_host = http_hosts[i].clone();
520 :
521 0 : assert!(status.tenant_id == request.tenant_id);
522 0 : assert!(status.timeline_id == request.timeline_id);
523 :
524 0 : let check_tombstone = !request.ignore_tombstone.unwrap_or_default();
525 :
526 0 : match pull_timeline(
527 0 : status,
528 0 : safekeeper_host,
529 0 : sk_auth_token,
530 0 : http_client,
531 0 : global_timelines,
532 0 : check_tombstone,
533 : )
534 0 : .await
535 : {
536 0 : Ok(resp) => Ok(resp),
537 0 : Err(e) => {
538 0 : match e.downcast_ref::<TimelineError>() {
539 0 : Some(TimelineError::AlreadyExists(_)) => Ok(PullTimelineResponse {
540 0 : safekeeper_host: None,
541 0 : }),
542 : Some(TimelineError::CreationInProgress(_)) => {
543 : // We don't return success here because creation might still fail.
544 0 : Err(ApiError::Conflict("Creation in progress".to_owned()))
545 : }
546 0 : _ => Err(ApiError::InternalServerError(e)),
547 : }
548 : }
549 : }
550 0 : }
551 :
552 0 : async fn pull_timeline(
553 0 : status: TimelineStatus,
554 0 : host: String,
555 0 : sk_auth_token: Option<SecretString>,
556 0 : http_client: reqwest::Client,
557 0 : global_timelines: Arc<GlobalTimelines>,
558 0 : check_tombstone: bool,
559 0 : ) -> Result<PullTimelineResponse> {
560 0 : let ttid = TenantTimelineId::new(status.tenant_id, status.timeline_id);
561 0 : info!(
562 0 : "pulling timeline {} from safekeeper {}, commit_lsn={}, flush_lsn={}, term={}, epoch={}",
563 : ttid,
564 : host,
565 : status.commit_lsn,
566 : status.flush_lsn,
567 : status.acceptor_state.term,
568 : status.acceptor_state.epoch
569 : );
570 :
571 0 : let conf = &global_timelines.get_global_config();
572 :
573 0 : let (_tmp_dir, tli_dir_path) = create_temp_timeline_dir(conf, ttid).await?;
574 0 : let client = Client::new(http_client, host.clone(), sk_auth_token.clone());
575 : // Request stream with basebackup archive.
576 0 : let bb_resp = client
577 0 : .snapshot(status.tenant_id, status.timeline_id, conf.my_id)
578 0 : .await?;
579 :
580 : // Make Stream of Bytes from it...
581 0 : let bb_stream = bb_resp.bytes_stream().map_err(std::io::Error::other);
582 : // and turn it into StreamReader implementing AsyncRead.
583 0 : let bb_reader = tokio_util::io::StreamReader::new(bb_stream);
584 :
585 : // Extract it on the fly to the disk. We don't use simple unpack() to fsync
586 : // files.
587 0 : let mut entries = Archive::new(bb_reader).entries()?;
588 0 : while let Some(base_tar_entry) = entries.next().await {
589 0 : let mut entry = base_tar_entry?;
590 0 : let header = entry.header();
591 0 : let file_path = header.path()?.into_owned();
592 0 : match header.entry_type() {
593 : tokio_tar::EntryType::Regular => {
594 0 : let utf8_file_path =
595 0 : Utf8PathBuf::from_path_buf(file_path).expect("non-Unicode path");
596 0 : let dst_path = tli_dir_path.join(utf8_file_path);
597 0 : let mut f = OpenOptions::new()
598 0 : .create(true)
599 0 : .truncate(true)
600 0 : .write(true)
601 0 : .open(&dst_path)
602 0 : .await?;
603 0 : tokio::io::copy(&mut entry, &mut f).await?;
604 : // fsync the file
605 0 : f.sync_all().await?;
606 : }
607 : _ => {
608 0 : bail!(
609 0 : "entry {} in backup tar archive is of unexpected type: {:?}",
610 0 : file_path.display(),
611 0 : header.entry_type()
612 : );
613 : }
614 : }
615 : }
616 :
617 : // fsync temp timeline directory to remember its contents.
618 0 : fsync_async_opt(&tli_dir_path, !conf.no_sync).await?;
619 :
620 : // Let's create timeline from temp directory and verify that it's correct
621 0 : let (commit_lsn, flush_lsn) = validate_temp_timeline(conf, ttid, &tli_dir_path).await?;
622 0 : info!(
623 0 : "finished downloading timeline {}, commit_lsn={}, flush_lsn={}",
624 : ttid, commit_lsn, flush_lsn
625 : );
626 0 : assert!(status.commit_lsn <= status.flush_lsn);
627 :
628 : // Finally, load the timeline.
629 0 : let _tli = global_timelines
630 0 : .load_temp_timeline(ttid, &tli_dir_path, check_tombstone)
631 0 : .await?;
632 :
633 0 : Ok(PullTimelineResponse {
634 0 : safekeeper_host: Some(host),
635 0 : })
636 0 : }
|
