Line data Source code
1 : //! This module implements pulling WAL from peer safekeepers if compute can't
2 : //! provide it, i.e. safekeeper lags too much.
3 :
4 : use std::time::SystemTime;
5 : use std::{fmt, pin::pin};
6 :
7 : use anyhow::{bail, Context};
8 : use futures::StreamExt;
9 : use postgres_protocol::message::backend::ReplicationMessage;
10 : use tokio::sync::mpsc::{channel, Receiver, Sender};
11 : use tokio::time::timeout;
12 : use tokio::{
13 : select,
14 : time::sleep,
15 : time::{self, Duration},
16 : };
17 : use tokio_postgres::replication::ReplicationStream;
18 : use tokio_postgres::types::PgLsn;
19 : use tracing::*;
20 : use utils::{id::NodeId, lsn::Lsn, postgres_client::wal_stream_connection_config};
21 :
22 : use crate::receive_wal::{WalAcceptor, REPLY_QUEUE_SIZE};
23 : use crate::safekeeper::{AppendRequest, AppendRequestHeader};
24 : use crate::timeline::FullAccessTimeline;
25 : use crate::{
26 : http::routes::TimelineStatus,
27 : receive_wal::MSG_QUEUE_SIZE,
28 : safekeeper::{
29 : AcceptorProposerMessage, ProposerAcceptorMessage, ProposerElected, Term, TermHistory,
30 : TermLsn, VoteRequest,
31 : },
32 : timeline::PeerInfo,
33 : SafeKeeperConf,
34 : };
35 :
36 : /// Entrypoint for per timeline task which always runs, checking whether
37 : /// recovery for this safekeeper is needed and starting it if so.
38 0 : #[instrument(name = "recovery task", skip_all, fields(ttid = %tli.ttid))]
39 : pub async fn recovery_main(tli: FullAccessTimeline, conf: SafeKeeperConf) {
40 : info!("started");
41 :
42 : let cancel = tli.cancel.clone();
43 : select! {
44 : _ = recovery_main_loop(tli, conf) => { unreachable!() }
45 : _ = cancel.cancelled() => {
46 : info!("stopped");
47 : }
48 : }
49 : }
50 :
51 : /// Should we start fetching WAL from a peer safekeeper, and if yes, from
52 : /// which? Answer is yes, i.e. .donors is not empty if 1) there is something
53 : /// to fetch, and we can do that without running elections; 2) there is no
54 : /// actively streaming compute, as we don't want to compete with it.
55 : ///
56 : /// If donor(s) are choosen, theirs last_log_term is guaranteed to be equal
57 : /// to its last_log_term so we are sure such a leader ever had been elected.
58 : ///
59 : /// All possible donors are returned so that we could keep connection to the
60 : /// current one if it is good even if it slightly lags behind.
61 : ///
62 : /// Note that term conditions above might be not met, but safekeepers are
63 : /// still not aligned on last flush_lsn. Generally in this case until
64 : /// elections are run it is not possible to say which safekeeper should
65 : /// recover from which one -- history which would be committed is different
66 : /// depending on assembled quorum (e.g. classic picture 8 from Raft paper).
67 : /// Thus we don't try to predict it here.
68 0 : async fn recovery_needed(
69 0 : tli: &FullAccessTimeline,
70 0 : heartbeat_timeout: Duration,
71 0 : ) -> RecoveryNeededInfo {
72 0 : let ss = tli.read_shared_state().await;
73 0 : let term = ss.sk.state.acceptor_state.term;
74 0 : let last_log_term = ss.sk.get_last_log_term();
75 0 : let flush_lsn = ss.sk.flush_lsn();
76 0 : // note that peers contain myself, but that's ok -- we are interested only in peers which are strictly ahead of us.
77 0 : let mut peers = ss.get_peers(heartbeat_timeout);
78 0 : // Sort by <last log term, lsn> pairs.
79 0 : peers.sort_by(|p1, p2| {
80 0 : let tl1 = TermLsn {
81 0 : term: p1.last_log_term,
82 0 : lsn: p1.flush_lsn,
83 0 : };
84 0 : let tl2 = TermLsn {
85 0 : term: p2.last_log_term,
86 0 : lsn: p2.flush_lsn,
87 0 : };
88 0 : tl2.cmp(&tl1) // desc
89 0 : });
90 0 : let num_streaming_computes = tli.get_walreceivers().get_num_streaming();
91 0 : let donors = if num_streaming_computes > 0 {
92 0 : vec![] // If there is a streaming compute, don't try to recover to not intervene.
93 : } else {
94 0 : peers
95 0 : .iter()
96 0 : .filter_map(|candidate| {
97 0 : // Are we interested in this candidate?
98 0 : let candidate_tl = TermLsn {
99 0 : term: candidate.last_log_term,
100 0 : lsn: candidate.flush_lsn,
101 0 : };
102 0 : let my_tl = TermLsn {
103 0 : term: last_log_term,
104 0 : lsn: flush_lsn,
105 0 : };
106 0 : if my_tl < candidate_tl {
107 : // Yes, we are interested. Can we pull from it without
108 : // (re)running elections? It is possible if 1) his term
109 : // is equal to his last_log_term so we could act on
110 : // behalf of leader of this term (we must be sure he was
111 : // ever elected) and 2) our term is not higher, or we'll refuse data.
112 0 : if candidate.term == candidate.last_log_term && candidate.term >= term {
113 0 : Some(Donor::from(candidate))
114 : } else {
115 0 : None
116 : }
117 : } else {
118 0 : None
119 : }
120 0 : })
121 0 : .collect()
122 : };
123 0 : RecoveryNeededInfo {
124 0 : term,
125 0 : last_log_term,
126 0 : flush_lsn,
127 0 : peers,
128 0 : num_streaming_computes,
129 0 : donors,
130 0 : }
131 0 : }
132 : /// Result of Timeline::recovery_needed, contains donor(s) if recovery needed and
133 : /// fields to explain the choice.
134 : #[derive(Debug)]
135 : pub struct RecoveryNeededInfo {
136 : /// my term
137 : pub term: Term,
138 : /// my last_log_term
139 : pub last_log_term: Term,
140 : /// my flush_lsn
141 : pub flush_lsn: Lsn,
142 : /// peers from which we can fetch WAL, for observability.
143 : pub peers: Vec<PeerInfo>,
144 : /// for observability
145 : pub num_streaming_computes: usize,
146 : pub donors: Vec<Donor>,
147 : }
148 :
149 : // Custom to omit not important fields from PeerInfo.
150 : impl fmt::Display for RecoveryNeededInfo {
151 0 : fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
152 0 : write!(f, "{{")?;
153 0 : write!(
154 0 : f,
155 0 : "term: {}, last_log_term: {}, flush_lsn: {}, peers: {{",
156 0 : self.term, self.last_log_term, self.flush_lsn
157 0 : )?;
158 0 : for p in self.peers.iter() {
159 0 : write!(
160 0 : f,
161 0 : "PeerInfo {{ sk_id: {}, term: {}, last_log_term: {}, flush_lsn: {} }}, ",
162 0 : p.sk_id, p.term, p.last_log_term, p.flush_lsn
163 0 : )?;
164 : }
165 0 : write!(
166 0 : f,
167 0 : "}} num_streaming_computes: {}, donors: {:?}",
168 0 : self.num_streaming_computes, self.donors
169 0 : )
170 0 : }
171 : }
172 :
173 : #[derive(Clone, Debug)]
174 : pub struct Donor {
175 : pub sk_id: NodeId,
176 : /// equals to last_log_term
177 : pub term: Term,
178 : pub flush_lsn: Lsn,
179 : pub pg_connstr: String,
180 : pub http_connstr: String,
181 : }
182 :
183 : impl From<&PeerInfo> for Donor {
184 0 : fn from(p: &PeerInfo) -> Self {
185 0 : Donor {
186 0 : sk_id: p.sk_id,
187 0 : term: p.term,
188 0 : flush_lsn: p.flush_lsn,
189 0 : pg_connstr: p.pg_connstr.clone(),
190 0 : http_connstr: p.http_connstr.clone(),
191 0 : }
192 0 : }
193 : }
194 :
195 : const CHECK_INTERVAL_MS: u64 = 2000;
196 :
197 : /// Check regularly whether we need to start recovery.
198 0 : async fn recovery_main_loop(tli: FullAccessTimeline, conf: SafeKeeperConf) {
199 0 : let check_duration = Duration::from_millis(CHECK_INTERVAL_MS);
200 : loop {
201 0 : let recovery_needed_info = recovery_needed(&tli, conf.heartbeat_timeout).await;
202 0 : match recovery_needed_info.donors.first() {
203 0 : Some(donor) => {
204 0 : info!(
205 0 : "starting recovery from donor {}: {}",
206 : donor.sk_id, recovery_needed_info
207 : );
208 0 : match recover(tli.clone(), donor, &conf).await {
209 : // Note: 'write_wal rewrites WAL written before' error is
210 : // expected here and might happen if compute and recovery
211 : // concurrently write the same data. Eventually compute
212 : // should win.
213 0 : Err(e) => warn!("recovery failed: {:#}", e),
214 0 : Ok(msg) => info!("recovery finished: {}", msg),
215 : }
216 : }
217 : None => {
218 0 : trace!(
219 0 : "recovery not needed or not possible: {}",
220 : recovery_needed_info
221 : );
222 : }
223 : }
224 0 : sleep(check_duration).await;
225 : }
226 : }
227 :
228 : /// Recover from the specified donor. Returns message explaining normal finish
229 : /// reason or error.
230 0 : async fn recover(
231 0 : tli: FullAccessTimeline,
232 0 : donor: &Donor,
233 0 : conf: &SafeKeeperConf,
234 0 : ) -> anyhow::Result<String> {
235 0 : // Learn donor term switch history to figure out starting point.
236 0 : let client = reqwest::Client::new();
237 0 : let timeline_info: TimelineStatus = client
238 0 : .get(format!(
239 0 : "http://{}/v1/tenant/{}/timeline/{}",
240 0 : donor.http_connstr, tli.ttid.tenant_id, tli.ttid.timeline_id
241 0 : ))
242 0 : .send()
243 0 : .await?
244 0 : .json()
245 0 : .await?;
246 0 : if timeline_info.acceptor_state.term != donor.term {
247 0 : bail!(
248 0 : "donor term changed from {} to {}",
249 0 : donor.term,
250 0 : timeline_info.acceptor_state.term
251 0 : );
252 0 : }
253 0 : // convert from API TermSwitchApiEntry into TermLsn.
254 0 : let donor_th = TermHistory(
255 0 : timeline_info
256 0 : .acceptor_state
257 0 : .term_history
258 0 : .iter()
259 0 : .map(|tl| Into::<TermLsn>::into(*tl))
260 0 : .collect(),
261 0 : );
262 0 :
263 0 : // Now understand our term history.
264 0 : let vote_request = ProposerAcceptorMessage::VoteRequest(VoteRequest { term: donor.term });
265 0 : let vote_response = match tli
266 0 : .process_msg(&vote_request)
267 0 : .await
268 0 : .context("VoteRequest handling")?
269 : {
270 0 : Some(AcceptorProposerMessage::VoteResponse(vr)) => vr,
271 : _ => {
272 0 : bail!("unexpected VoteRequest response"); // unreachable
273 : }
274 : };
275 0 : if vote_response.term != donor.term {
276 0 : bail!(
277 0 : "our term changed from {} to {}",
278 0 : donor.term,
279 0 : vote_response.term
280 0 : );
281 0 : }
282 :
283 0 : let last_common_point = match TermHistory::find_highest_common_point(
284 0 : &donor_th,
285 0 : &vote_response.term_history,
286 0 : vote_response.flush_lsn,
287 0 : ) {
288 0 : None => bail!(
289 0 : "couldn't find common point in histories, donor {:?}, sk {:?}",
290 0 : donor_th,
291 0 : vote_response.term_history,
292 0 : ),
293 0 : Some(lcp) => lcp,
294 0 : };
295 0 : info!("found last common point at {:?}", last_common_point);
296 :
297 : // truncate WAL locally
298 0 : let pe = ProposerAcceptorMessage::Elected(ProposerElected {
299 0 : term: donor.term,
300 0 : start_streaming_at: last_common_point.lsn,
301 0 : term_history: donor_th,
302 0 : timeline_start_lsn: Lsn::INVALID,
303 0 : });
304 0 : // Successful ProposerElected handling always returns None. If term changed,
305 0 : // we'll find out that during the streaming. Note: it is expected to get
306 0 : // 'refusing to overwrite correct WAL' here if walproposer reconnected
307 0 : // concurrently, restart helps here.
308 0 : tli.process_msg(&pe)
309 0 : .await
310 0 : .context("ProposerElected handling")?;
311 :
312 0 : recovery_stream(tli, donor, last_common_point.lsn, conf).await
313 0 : }
314 :
315 : // Pull WAL from donor, assuming handshake is already done.
316 0 : async fn recovery_stream(
317 0 : tli: FullAccessTimeline,
318 0 : donor: &Donor,
319 0 : start_streaming_at: Lsn,
320 0 : conf: &SafeKeeperConf,
321 0 : ) -> anyhow::Result<String> {
322 : // TODO: pass auth token
323 0 : let cfg = wal_stream_connection_config(tli.ttid, &donor.pg_connstr, None, None)?;
324 0 : let mut cfg = cfg.to_tokio_postgres_config();
325 0 : // It will make safekeeper give out not committed WAL (up to flush_lsn).
326 0 : cfg.application_name(&format!("safekeeper_{}", conf.my_id));
327 0 : cfg.replication_mode(tokio_postgres::config::ReplicationMode::Physical);
328 0 :
329 0 : let connect_timeout = Duration::from_millis(10000);
330 0 : let (client, connection) = match time::timeout(connect_timeout, cfg.connect(postgres::NoTls))
331 0 : .await
332 : {
333 0 : Ok(client_and_conn) => client_and_conn?,
334 0 : Err(_elapsed) => {
335 0 : bail!("timed out while waiting {connect_timeout:?} for connection to peer safekeeper to open");
336 : }
337 : };
338 0 : trace!("connected to {:?}", donor);
339 :
340 : // The connection object performs the actual communication with the
341 : // server, spawn it off to run on its own.
342 0 : let ttid = tli.ttid;
343 0 : tokio::spawn(async move {
344 0 : if let Err(e) = connection
345 0 : .instrument(info_span!("recovery task connection poll", ttid = %ttid))
346 0 : .await
347 : {
348 : // This logging isn't very useful as error is anyway forwarded to client.
349 0 : trace!(
350 0 : "tokio_postgres connection object finished with error: {}",
351 : e
352 : );
353 0 : }
354 0 : });
355 0 :
356 0 : let query = format!(
357 0 : "START_REPLICATION PHYSICAL {} (term='{}')",
358 0 : start_streaming_at, donor.term
359 0 : );
360 :
361 0 : let copy_stream = client.copy_both_simple(&query).await?;
362 0 : let physical_stream = ReplicationStream::new(copy_stream);
363 0 :
364 0 : // As in normal walreceiver, do networking and writing to disk in parallel.
365 0 : let (msg_tx, msg_rx) = channel(MSG_QUEUE_SIZE);
366 0 : let (reply_tx, reply_rx) = channel(REPLY_QUEUE_SIZE);
367 0 : let wa = WalAcceptor::spawn(tli.clone(), msg_rx, reply_tx, None);
368 :
369 0 : let res = tokio::select! {
370 : r = network_io(physical_stream, msg_tx, donor.clone(), tli.clone(), conf.clone()) => r,
371 0 : r = read_replies(reply_rx, donor.term) => r.map(|()| None),
372 : };
373 :
374 : // Join the spawned WalAcceptor. At this point chans to/from it passed to
375 : // network routines are dropped, so it will exit as soon as it touches them.
376 0 : match wa.await {
377 : Ok(Ok(())) => {
378 : // WalAcceptor finished normally, termination reason is different
379 0 : match res {
380 0 : Ok(Some(success_desc)) => Ok(success_desc),
381 0 : Ok(None) => bail!("unexpected recovery end without error/success"), // can't happen
382 0 : Err(e) => Err(e), // network error or term change
383 : }
384 : }
385 0 : Ok(Err(e)) => Err(e), // error while processing message
386 0 : Err(e) => bail!("WalAcceptor panicked: {}", e),
387 : }
388 0 : }
389 :
390 : // Perform network part of streaming: read data and push it to msg_tx, send KA
391 : // to make sender hear from us. If there is nothing coming for a while, check
392 : // for termination.
393 : // Returns
394 : // - Ok(None) if channel to WalAcceptor closed -- its task should return error.
395 : // - Ok(Some(String)) if recovery successfully completed.
396 : // - Err if error happened while reading/writing to socket.
397 0 : async fn network_io(
398 0 : physical_stream: ReplicationStream,
399 0 : msg_tx: Sender<ProposerAcceptorMessage>,
400 0 : donor: Donor,
401 0 : tli: FullAccessTimeline,
402 0 : conf: SafeKeeperConf,
403 0 : ) -> anyhow::Result<Option<String>> {
404 0 : let mut physical_stream = pin!(physical_stream);
405 0 : let mut last_received_lsn = Lsn::INVALID;
406 0 : // tear down connection if no data arrives withing this period
407 0 : let no_data_timeout = Duration::from_millis(30000);
408 :
409 : loop {
410 0 : let msg = match timeout(no_data_timeout, physical_stream.next()).await {
411 0 : Ok(next) => match next {
412 0 : None => bail!("unexpected end of replication stream"),
413 0 : Some(msg) => msg.context("get replication message")?,
414 : },
415 0 : Err(_) => bail!("no message received within {:?}", no_data_timeout),
416 : };
417 :
418 0 : match msg {
419 0 : ReplicationMessage::XLogData(xlog_data) => {
420 0 : let ar_hdr = AppendRequestHeader {
421 0 : term: donor.term,
422 0 : term_start_lsn: Lsn::INVALID, // unused
423 0 : begin_lsn: Lsn(xlog_data.wal_start()),
424 0 : end_lsn: Lsn(xlog_data.wal_start()) + xlog_data.data().len() as u64,
425 0 : commit_lsn: Lsn::INVALID, // do not attempt to advance, peer communication anyway does it
426 0 : truncate_lsn: Lsn::INVALID, // do not attempt to advance
427 0 : proposer_uuid: [0; 16],
428 0 : };
429 0 : let ar = AppendRequest {
430 0 : h: ar_hdr,
431 0 : wal_data: xlog_data.into_data(),
432 0 : };
433 0 : trace!(
434 0 : "processing AppendRequest {}-{}, len {}",
435 0 : ar.h.begin_lsn,
436 0 : ar.h.end_lsn,
437 0 : ar.wal_data.len()
438 : );
439 0 : last_received_lsn = ar.h.end_lsn;
440 0 : if msg_tx
441 0 : .send(ProposerAcceptorMessage::AppendRequest(ar))
442 0 : .await
443 0 : .is_err()
444 : {
445 0 : return Ok(None); // chan closed, WalAcceptor terminated
446 0 : }
447 : }
448 : ReplicationMessage::PrimaryKeepAlive(_) => {
449 : // keepalive means nothing is being streamed for a while. Check whether we need to stop.
450 0 : let recovery_needed_info = recovery_needed(&tli, conf.heartbeat_timeout).await;
451 : // do current donors still contain one we currently connected to?
452 0 : if !recovery_needed_info
453 0 : .donors
454 0 : .iter()
455 0 : .any(|d| d.sk_id == donor.sk_id)
456 : {
457 : // Most likely it means we are caughtup.
458 : // note: just exiting makes tokio_postgres send CopyFail to the far end.
459 0 : return Ok(Some(format!(
460 0 : "terminating at {} as connected safekeeper {} with term {} is not a donor anymore: {}",
461 0 : last_received_lsn, donor.sk_id, donor.term, recovery_needed_info
462 0 : )));
463 0 : }
464 : }
465 0 : _ => {}
466 : }
467 : // Send reply to each message to keep connection alive. Ideally we
468 : // should do that once in a while instead, but this again requires
469 : // stream split or similar workaround, and recovery is anyway not that
470 : // performance critical.
471 : //
472 : // We do not know here real write/flush LSNs (need to take mutex again
473 : // or check replies which are read in different future), but neither
474 : // sender much cares about them, so just send last received.
475 0 : physical_stream
476 0 : .as_mut()
477 0 : .standby_status_update(
478 0 : PgLsn::from(last_received_lsn.0),
479 0 : PgLsn::from(last_received_lsn.0),
480 0 : PgLsn::from(last_received_lsn.0),
481 0 : SystemTime::now(),
482 0 : 0,
483 0 : )
484 0 : .await?;
485 : }
486 0 : }
487 :
488 : // Read replies from WalAcceptor. We are not interested much in sending them to
489 : // donor safekeeper, so don't route them anywhere. However, we should check if
490 : // term changes and exit if it does.
491 : // Returns Ok(()) if channel closed, Err in case of term change.
492 0 : async fn read_replies(
493 0 : mut reply_rx: Receiver<AcceptorProposerMessage>,
494 0 : donor_term: Term,
495 0 : ) -> anyhow::Result<()> {
496 : loop {
497 0 : match reply_rx.recv().await {
498 0 : Some(msg) => {
499 0 : if let AcceptorProposerMessage::AppendResponse(ar) = msg {
500 0 : if ar.term != donor_term {
501 0 : bail!("donor term changed from {} to {}", donor_term, ar.term);
502 0 : }
503 0 : }
504 : }
505 0 : None => return Ok(()), // chan closed, WalAcceptor terminated
506 : }
507 : }
508 0 : }
|
