Line data Source code
1 : //! Safekeeper timeline has a background task which is subscribed to `commit_lsn`
2 : //! and `flush_lsn` updates. After the partial segment was updated (`flush_lsn`
3 : //! was changed), the segment will be uploaded to S3 in about 15 minutes.
4 : //!
5 : //! The filename format for partial segments is
6 : //! `Segment_Term_Flush_Commit_skNN.partial`, where:
7 : //! - `Segment` – the segment name, like `000000010000000000000001`
8 : //! - `Term` – current term
9 : //! - `Flush` – flush_lsn in hex format `{:016X}`, e.g. `00000000346BC568`
10 : //! - `Commit` – commit_lsn in the same hex format
11 : //! - `NN` – safekeeper_id, like `1`
12 : //!
13 : //! The full object name example:
14 : //! `000000010000000000000002_2_0000000002534868_0000000002534410_sk1.partial`
15 : //!
16 : //! Each safekeeper will keep info about remote partial segments in its control
17 : //! file. Code updates state in the control file before doing any S3 operations.
18 : //! This way control file stores information about all potentially existing
19 : //! remote partial segments and can clean them up after uploading a newer version.
20 :
21 : use camino::Utf8PathBuf;
22 : use postgres_ffi::{XLogFileName, XLogSegNo, PG_TLI};
23 : use remote_storage::RemotePath;
24 : use serde::{Deserialize, Serialize};
25 :
26 : use tracing::{debug, error, info, instrument, warn};
27 : use utils::lsn::Lsn;
28 :
29 : use crate::{
30 : metrics::{MISC_OPERATION_SECONDS, PARTIAL_BACKUP_UPLOADED_BYTES, PARTIAL_BACKUP_UPLOADS},
31 : rate_limit::{rand_duration, RateLimiter},
32 : safekeeper::Term,
33 : timeline::WalResidentTimeline,
34 : timeline_manager::StateSnapshot,
35 : wal_backup::{self, remote_timeline_path},
36 : SafeKeeperConf,
37 : };
38 :
39 0 : #[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
40 : pub enum UploadStatus {
41 : /// Upload is in progress. This status should be used only for garbage collection,
42 : /// don't read data from the remote storage with this status.
43 : InProgress,
44 : /// Upload is finished. There is always at most one segment with this status.
45 : /// It means that the segment is actual and can be used.
46 : Uploaded,
47 : /// Deletion is in progress. This status should be used only for garbage collection,
48 : /// don't read data from the remote storage with this status.
49 : Deleting,
50 : }
51 :
52 0 : #[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
53 : pub struct PartialRemoteSegment {
54 : pub status: UploadStatus,
55 : pub name: String,
56 : pub commit_lsn: Lsn,
57 : pub flush_lsn: Lsn,
58 : // We should use last_log_term here, otherwise it's possible to have inconsistent data in the
59 : // remote storage.
60 : //
61 : // More info here: https://github.com/neondatabase/neon/pull/8022#discussion_r1654738405
62 : pub term: Term,
63 : }
64 :
65 : impl PartialRemoteSegment {
66 0 : fn eq_without_status(&self, other: &Self) -> bool {
67 0 : self.name == other.name
68 0 : && self.commit_lsn == other.commit_lsn
69 0 : && self.flush_lsn == other.flush_lsn
70 0 : && self.term == other.term
71 0 : }
72 :
73 0 : pub(crate) fn remote_path(&self, remote_timeline_path: &RemotePath) -> RemotePath {
74 0 : remote_timeline_path.join(&self.name)
75 0 : }
76 : }
77 :
78 : // NB: these structures are a part of a control_file, you can't change them without
79 : // changing the control file format version.
80 6 : #[derive(Debug, Clone, Serialize, Deserialize, PartialEq, Default)]
81 : pub struct State {
82 : pub segments: Vec<PartialRemoteSegment>,
83 : }
84 :
85 : impl State {
86 : /// Find an Uploaded segment. There should be only one Uploaded segment at a time.
87 0 : pub(crate) fn uploaded_segment(&self) -> Option<PartialRemoteSegment> {
88 0 : self.segments
89 0 : .iter()
90 0 : .find(|seg| seg.status == UploadStatus::Uploaded)
91 0 : .cloned()
92 0 : }
93 : }
94 :
95 : struct PartialBackup {
96 : wal_seg_size: usize,
97 : tli: WalResidentTimeline,
98 : conf: SafeKeeperConf,
99 : local_prefix: Utf8PathBuf,
100 : remote_timeline_path: RemotePath,
101 :
102 : state: State,
103 : }
104 :
105 : // Read-only methods for getting segment names
106 : impl PartialBackup {
107 0 : fn segno(&self, lsn: Lsn) -> XLogSegNo {
108 0 : lsn.segment_number(self.wal_seg_size)
109 0 : }
110 :
111 0 : fn segment_name(&self, segno: u64) -> String {
112 0 : XLogFileName(PG_TLI, segno, self.wal_seg_size)
113 0 : }
114 :
115 0 : fn remote_segment_name(
116 0 : &self,
117 0 : segno: u64,
118 0 : term: u64,
119 0 : commit_lsn: Lsn,
120 0 : flush_lsn: Lsn,
121 0 : ) -> String {
122 0 : format!(
123 0 : "{}_{}_{:016X}_{:016X}_sk{}.partial",
124 0 : self.segment_name(segno),
125 0 : term,
126 0 : flush_lsn.0,
127 0 : commit_lsn.0,
128 0 : self.conf.my_id.0,
129 0 : )
130 0 : }
131 :
132 0 : fn local_segment_name(&self, segno: u64) -> String {
133 0 : format!("{}.partial", self.segment_name(segno))
134 0 : }
135 : }
136 :
137 : impl PartialBackup {
138 : /// Takes a lock to read actual safekeeper state and returns a segment that should be uploaded.
139 0 : async fn prepare_upload(&self) -> PartialRemoteSegment {
140 : // this operation takes a lock to get the actual state
141 0 : let sk_info = self.tli.get_safekeeper_info(&self.conf).await;
142 0 : let flush_lsn = Lsn(sk_info.flush_lsn);
143 0 : let commit_lsn = Lsn(sk_info.commit_lsn);
144 0 : let last_log_term = sk_info.last_log_term;
145 0 : let segno = self.segno(flush_lsn);
146 0 :
147 0 : let name = self.remote_segment_name(segno, last_log_term, commit_lsn, flush_lsn);
148 0 :
149 0 : PartialRemoteSegment {
150 0 : status: UploadStatus::InProgress,
151 0 : name,
152 0 : commit_lsn,
153 0 : flush_lsn,
154 0 : term: last_log_term,
155 0 : }
156 0 : }
157 :
158 : /// Reads segment from disk and uploads it to the remote storage.
159 0 : async fn upload_segment(&mut self, prepared: PartialRemoteSegment) -> anyhow::Result<()> {
160 0 : let flush_lsn = prepared.flush_lsn;
161 0 : let segno = self.segno(flush_lsn);
162 0 :
163 0 : // We're going to backup bytes from the start of the segment up to flush_lsn.
164 0 : let backup_bytes = flush_lsn.segment_offset(self.wal_seg_size);
165 0 :
166 0 : let local_path = self.local_prefix.join(self.local_segment_name(segno));
167 0 : let remote_path = prepared.remote_path(&self.remote_timeline_path);
168 0 :
169 0 : // Upload first `backup_bytes` bytes of the segment to the remote storage.
170 0 : wal_backup::backup_partial_segment(&local_path, &remote_path, backup_bytes).await?;
171 0 : PARTIAL_BACKUP_UPLOADED_BYTES.inc_by(backup_bytes as u64);
172 :
173 : // We uploaded the segment, now let's verify that the data is still actual.
174 : // If the term changed, we cannot guarantee the validity of the uploaded data.
175 : // If the term is the same, we know the data is not corrupted.
176 0 : let sk_info = self.tli.get_safekeeper_info(&self.conf).await;
177 0 : if sk_info.last_log_term != prepared.term {
178 0 : anyhow::bail!("term changed during upload");
179 0 : }
180 0 : assert!(prepared.commit_lsn <= Lsn(sk_info.commit_lsn));
181 0 : assert!(prepared.flush_lsn <= Lsn(sk_info.flush_lsn));
182 :
183 0 : Ok(())
184 0 : }
185 :
186 : /// Write new state to disk. If in-memory and on-disk states diverged, returns an error.
187 0 : async fn commit_state(&mut self, new_state: State) -> anyhow::Result<()> {
188 0 : self.tli
189 0 : .map_control_file(|cf| {
190 0 : if cf.partial_backup != self.state {
191 0 : let memory = self.state.clone();
192 0 : self.state = cf.partial_backup.clone();
193 0 : anyhow::bail!(
194 0 : "partial backup state diverged, memory={:?}, disk={:?}",
195 0 : memory,
196 0 : cf.partial_backup
197 0 : );
198 0 : }
199 0 :
200 0 : cf.partial_backup = new_state.clone();
201 0 : Ok(())
202 0 : })
203 0 : .await?;
204 : // update in-memory state
205 0 : self.state = new_state;
206 0 : Ok(())
207 0 : }
208 :
209 : /// Upload the latest version of the partial segment and garbage collect older versions.
210 0 : #[instrument(name = "upload", skip_all, fields(name = %prepared.name))]
211 : async fn do_upload(&mut self, prepared: &PartialRemoteSegment) -> anyhow::Result<()> {
212 : let _timer = MISC_OPERATION_SECONDS
213 : .with_label_values(&["partial_do_upload"])
214 : .start_timer();
215 : info!("starting upload {:?}", prepared);
216 :
217 : let state_0 = self.state.clone();
218 : let state_1 = {
219 : let mut state = state_0.clone();
220 : state.segments.push(prepared.clone());
221 : state
222 : };
223 :
224 : // we're going to upload a new segment, let's write it to disk to make GC later
225 : self.commit_state(state_1).await?;
226 :
227 : self.upload_segment(prepared.clone()).await?;
228 :
229 : let state_2 = {
230 : let mut state = state_0.clone();
231 : for seg in state.segments.iter_mut() {
232 : seg.status = UploadStatus::Deleting;
233 : }
234 : let mut actual_remote_segment = prepared.clone();
235 : actual_remote_segment.status = UploadStatus::Uploaded;
236 : state.segments.push(actual_remote_segment);
237 : state
238 : };
239 :
240 : // we've uploaded new segment, it's actual, all other segments should be GCed
241 : self.commit_state(state_2).await?;
242 : self.gc().await?;
243 :
244 : Ok(())
245 : }
246 :
247 : /// Delete all non-Uploaded segments from the remote storage. There should be only one
248 : /// Uploaded segment at a time.
249 0 : #[instrument(name = "gc", skip_all)]
250 : async fn gc(&mut self) -> anyhow::Result<()> {
251 : let mut segments_to_delete = vec![];
252 :
253 : let new_segments: Vec<PartialRemoteSegment> = self
254 : .state
255 : .segments
256 : .iter()
257 0 : .filter_map(|seg| {
258 0 : if seg.status == UploadStatus::Uploaded {
259 0 : Some(seg.clone())
260 : } else {
261 0 : segments_to_delete.push(seg.name.clone());
262 0 : None
263 : }
264 0 : })
265 : .collect();
266 :
267 : if new_segments.len() == 1 {
268 : // we have an uploaded segment, it must not be deleted from remote storage
269 0 : segments_to_delete.retain(|name| name != &new_segments[0].name);
270 : } else {
271 : // there should always be zero or one uploaded segment
272 : assert!(
273 : new_segments.is_empty(),
274 : "too many uploaded segments: {:?}",
275 : new_segments
276 : );
277 : }
278 :
279 : info!("deleting objects: {:?}", segments_to_delete);
280 : let mut objects_to_delete = vec![];
281 : for seg in segments_to_delete.iter() {
282 : let remote_path = self.remote_timeline_path.join(seg);
283 : objects_to_delete.push(remote_path);
284 : }
285 :
286 : // removing segments from remote storage
287 : wal_backup::delete_objects(&objects_to_delete).await?;
288 :
289 : // now we can update the state on disk
290 : let new_state = {
291 : let mut state = self.state.clone();
292 : state.segments = new_segments;
293 : state
294 : };
295 : self.commit_state(new_state).await?;
296 :
297 : Ok(())
298 : }
299 : }
300 :
301 : /// Check if everything is uploaded and partial backup task doesn't need to run.
302 0 : pub(crate) fn needs_uploading(
303 0 : state: &StateSnapshot,
304 0 : uploaded: &Option<PartialRemoteSegment>,
305 0 : ) -> bool {
306 0 : match uploaded {
307 0 : Some(uploaded) => {
308 0 : uploaded.status != UploadStatus::Uploaded
309 0 : || uploaded.flush_lsn != state.flush_lsn
310 0 : || uploaded.commit_lsn != state.commit_lsn
311 0 : || uploaded.term != state.last_log_term
312 : }
313 0 : None => true,
314 : }
315 0 : }
316 :
317 : /// Main task for partial backup. It waits for the flush_lsn to change and then uploads the
318 : /// partial segment to the remote storage. It also does garbage collection of old segments.
319 : ///
320 : /// When there is nothing more to do and the last segment was successfully uploaded, the task
321 : /// returns PartialRemoteSegment, to signal readiness for offloading the timeline.
322 0 : #[instrument(name = "Partial backup", skip_all, fields(ttid = %tli.ttid))]
323 : pub async fn main_task(
324 : tli: WalResidentTimeline,
325 : conf: SafeKeeperConf,
326 : limiter: RateLimiter,
327 : ) -> Option<PartialRemoteSegment> {
328 : debug!("started");
329 : let await_duration = conf.partial_backup_timeout;
330 : let mut first_iteration = true;
331 :
332 : let (_, persistent_state) = tli.get_state().await;
333 : let mut commit_lsn_rx = tli.get_commit_lsn_watch_rx();
334 : let mut flush_lsn_rx = tli.get_term_flush_lsn_watch_rx();
335 : let wal_seg_size = tli.get_wal_seg_size().await;
336 :
337 : let local_prefix = tli.get_timeline_dir();
338 : let remote_timeline_path = match remote_timeline_path(&tli.ttid) {
339 : Ok(path) => path,
340 : Err(e) => {
341 : error!("failed to create remote path: {:?}", e);
342 : return None;
343 : }
344 : };
345 :
346 : let mut backup = PartialBackup {
347 : wal_seg_size,
348 : tli,
349 : state: persistent_state.partial_backup,
350 : conf,
351 : local_prefix,
352 : remote_timeline_path,
353 : };
354 :
355 : debug!("state: {:?}", backup.state);
356 :
357 : // The general idea is that each safekeeper keeps only one partial segment
358 : // both in remote storage and in local state. If this is not true, something
359 : // went wrong.
360 : const MAX_SIMULTANEOUS_SEGMENTS: usize = 10;
361 :
362 : 'outer: loop {
363 : if backup.state.segments.len() > MAX_SIMULTANEOUS_SEGMENTS {
364 : warn!(
365 : "too many segments in control_file state, running gc: {}",
366 : backup.state.segments.len()
367 : );
368 :
369 0 : backup.gc().await.unwrap_or_else(|e| {
370 0 : error!("failed to run gc: {:#}", e);
371 0 : });
372 : }
373 :
374 : // wait until we have something to upload
375 : let uploaded_segment = backup.state.uploaded_segment();
376 : if let Some(seg) = &uploaded_segment {
377 : // check if uploaded segment matches the current state
378 : if flush_lsn_rx.borrow().lsn == seg.flush_lsn
379 : && *commit_lsn_rx.borrow() == seg.commit_lsn
380 : && flush_lsn_rx.borrow().term == seg.term
381 : {
382 : // we have nothing to do, the last segment is already uploaded
383 : return Some(seg.clone());
384 : }
385 : }
386 :
387 : // if we don't have any data and zero LSNs, wait for something
388 : while flush_lsn_rx.borrow().lsn == Lsn(0) {
389 : tokio::select! {
390 : _ = backup.tli.cancel.cancelled() => {
391 : info!("timeline canceled");
392 : return None;
393 : }
394 : _ = flush_lsn_rx.changed() => {}
395 : }
396 : }
397 :
398 : // smoothing the load after restart, by sleeping for a random time.
399 : // if this is not the first iteration, we will wait for the full await_duration
400 : let await_duration = if first_iteration {
401 : first_iteration = false;
402 : rand_duration(&await_duration)
403 : } else {
404 : await_duration
405 : };
406 :
407 : // fixing the segno and waiting some time to prevent reuploading the same segment too often
408 : let pending_segno = backup.segno(flush_lsn_rx.borrow().lsn);
409 : let timeout = tokio::time::sleep(await_duration);
410 : tokio::pin!(timeout);
411 : let mut timeout_expired = false;
412 :
413 : // waiting until timeout expires OR segno changes
414 : 'inner: loop {
415 : tokio::select! {
416 : _ = backup.tli.cancel.cancelled() => {
417 : info!("timeline canceled");
418 : return None;
419 : }
420 : _ = commit_lsn_rx.changed() => {}
421 : _ = flush_lsn_rx.changed() => {
422 : let segno = backup.segno(flush_lsn_rx.borrow().lsn);
423 : if segno != pending_segno {
424 : // previous segment is no longer partial, aborting the wait
425 : break 'inner;
426 : }
427 : }
428 : _ = &mut timeout => {
429 : // timeout expired, now we are ready for upload
430 : timeout_expired = true;
431 : break 'inner;
432 : }
433 : }
434 : }
435 :
436 : if !timeout_expired {
437 : // likely segno has changed, let's try again in the next iteration
438 : continue 'outer;
439 : }
440 :
441 : // limit concurrent uploads
442 : let _upload_permit = limiter.acquire_partial_backup().await;
443 :
444 : let prepared = backup.prepare_upload().await;
445 : if let Some(seg) = &uploaded_segment {
446 : if seg.eq_without_status(&prepared) {
447 : // we already uploaded this segment, nothing to do
448 : continue 'outer;
449 : }
450 : }
451 :
452 : match backup.do_upload(&prepared).await {
453 : Ok(()) => {
454 : debug!(
455 : "uploaded {} up to flush_lsn {}",
456 : prepared.name, prepared.flush_lsn
457 : );
458 : PARTIAL_BACKUP_UPLOADS.with_label_values(&["ok"]).inc();
459 : }
460 : Err(e) => {
461 : info!("failed to upload {}: {:#}", prepared.name, e);
462 : PARTIAL_BACKUP_UPLOADS.with_label_values(&["error"]).inc();
463 : }
464 : }
465 : }
466 : }
|
