Line data Source code
1 : //! Acceptor part of proposer-acceptor consensus algorithm.
2 :
3 : use anyhow::{bail, Context, Result};
4 : use byteorder::{LittleEndian, ReadBytesExt};
5 : use bytes::{Buf, BufMut, Bytes, BytesMut};
6 :
7 : use postgres_ffi::{TimeLineID, MAX_SEND_SIZE};
8 : use serde::{Deserialize, Serialize};
9 : use std::cmp::max;
10 : use std::cmp::min;
11 : use std::fmt;
12 : use std::io::Read;
13 : use storage_broker::proto::SafekeeperTimelineInfo;
14 :
15 : use tracing::*;
16 :
17 : use crate::control_file;
18 : use crate::metrics::MISC_OPERATION_SECONDS;
19 : use crate::send_wal::HotStandbyFeedback;
20 :
21 : use crate::state::TimelineState;
22 : use crate::wal_storage;
23 : use pq_proto::SystemId;
24 : use utils::pageserver_feedback::PageserverFeedback;
25 : use utils::{
26 : bin_ser::LeSer,
27 : id::{NodeId, TenantId, TimelineId},
28 : lsn::Lsn,
29 : };
30 :
31 : const SK_PROTOCOL_VERSION: u32 = 2;
32 : pub const UNKNOWN_SERVER_VERSION: u32 = 0;
33 :
34 : /// Consensus logical timestamp.
35 : pub type Term = u64;
36 : pub const INVALID_TERM: Term = 0;
37 :
38 8 : #[derive(Debug, Clone, Copy, Serialize, Deserialize, PartialEq, Eq, PartialOrd, Ord)]
39 : pub struct TermLsn {
40 : pub term: Term,
41 : pub lsn: Lsn,
42 : }
43 :
44 : // Creation from tuple provides less typing (e.g. for unit tests).
45 : impl From<(Term, Lsn)> for TermLsn {
46 36 : fn from(pair: (Term, Lsn)) -> TermLsn {
47 36 : TermLsn {
48 36 : term: pair.0,
49 36 : lsn: pair.1,
50 36 : }
51 36 : }
52 : }
53 :
54 12 : #[derive(Clone, Serialize, Deserialize, PartialEq)]
55 : pub struct TermHistory(pub Vec<TermLsn>);
56 :
57 : impl TermHistory {
58 11455 : pub fn empty() -> TermHistory {
59 11455 : TermHistory(Vec::new())
60 11455 : }
61 :
62 : // Parse TermHistory as n_entries followed by TermLsn pairs
63 7017 : pub fn from_bytes(bytes: &mut Bytes) -> Result<TermHistory> {
64 7017 : if bytes.remaining() < 4 {
65 0 : bail!("TermHistory misses len");
66 7017 : }
67 7017 : let n_entries = bytes.get_u32_le();
68 7017 : let mut res = Vec::with_capacity(n_entries as usize);
69 7017 : for _ in 0..n_entries {
70 62011 : if bytes.remaining() < 16 {
71 0 : bail!("TermHistory is incomplete");
72 62011 : }
73 62011 : res.push(TermLsn {
74 62011 : term: bytes.get_u64_le(),
75 62011 : lsn: bytes.get_u64_le().into(),
76 62011 : })
77 : }
78 7017 : Ok(TermHistory(res))
79 7017 : }
80 :
81 : /// Return copy of self with switches happening strictly after up_to
82 : /// truncated.
83 32308 : pub fn up_to(&self, up_to: Lsn) -> TermHistory {
84 32308 : let mut res = Vec::with_capacity(self.0.len());
85 185052 : for e in &self.0 {
86 152808 : if e.lsn > up_to {
87 64 : break;
88 152744 : }
89 152744 : res.push(*e);
90 : }
91 32308 : TermHistory(res)
92 32308 : }
93 :
94 : /// Find point of divergence between leader (walproposer) term history and
95 : /// safekeeper. Arguments are not symmetrics as proposer history ends at
96 : /// +infinity while safekeeper at flush_lsn.
97 : /// C version is at walproposer SendProposerElected.
98 8 : pub fn find_highest_common_point(
99 8 : prop_th: &TermHistory,
100 8 : sk_th: &TermHistory,
101 8 : sk_wal_end: Lsn,
102 8 : ) -> Option<TermLsn> {
103 8 : let (prop_th, sk_th) = (&prop_th.0, &sk_th.0); // avoid .0 below
104 :
105 8 : if let Some(sk_th_last) = sk_th.last() {
106 8 : assert!(
107 8 : sk_th_last.lsn <= sk_wal_end,
108 0 : "safekeeper term history end {:?} LSN is higher than WAL end {:?}",
109 : sk_th_last,
110 : sk_wal_end
111 : );
112 0 : }
113 :
114 : // find last common term, if any...
115 8 : let mut last_common_idx = None;
116 16 : for i in 0..min(sk_th.len(), prop_th.len()) {
117 16 : if prop_th[i].term != sk_th[i].term {
118 4 : break;
119 12 : }
120 12 : // If term is the same, LSN must be equal as well.
121 12 : assert!(
122 12 : prop_th[i].lsn == sk_th[i].lsn,
123 0 : "same term {} has different start LSNs: prop {}, sk {}",
124 0 : prop_th[i].term,
125 0 : prop_th[i].lsn,
126 0 : sk_th[i].lsn
127 : );
128 12 : last_common_idx = Some(i);
129 : }
130 8 : let last_common_idx = match last_common_idx {
131 2 : None => return None, // no common point
132 6 : Some(lci) => lci,
133 6 : };
134 6 : // Now find where it ends at both prop and sk and take min. End of
135 6 : // (common) term is the start of the next except it is the last one;
136 6 : // there it is flush_lsn in case of safekeeper or, in case of proposer
137 6 : // +infinity, so we just take flush_lsn then.
138 6 : if last_common_idx == prop_th.len() - 1 {
139 2 : Some(TermLsn {
140 2 : term: prop_th[last_common_idx].term,
141 2 : lsn: sk_wal_end,
142 2 : })
143 : } else {
144 4 : let prop_common_term_end = prop_th[last_common_idx + 1].lsn;
145 4 : let sk_common_term_end = if last_common_idx + 1 < sk_th.len() {
146 2 : sk_th[last_common_idx + 1].lsn
147 : } else {
148 2 : sk_wal_end
149 : };
150 4 : Some(TermLsn {
151 4 : term: prop_th[last_common_idx].term,
152 4 : lsn: min(prop_common_term_end, sk_common_term_end),
153 4 : })
154 : }
155 8 : }
156 : }
157 :
158 : /// Display only latest entries for Debug.
159 : impl fmt::Debug for TermHistory {
160 400 : fn fmt(&self, fmt: &mut fmt::Formatter) -> fmt::Result {
161 400 : let n_printed = 20;
162 400 : write!(
163 400 : fmt,
164 400 : "{}{:?}",
165 400 : if self.0.len() > n_printed { "... " } else { "" },
166 400 : self.0
167 400 : .iter()
168 400 : .rev()
169 400 : .take(n_printed)
170 2216 : .map(|&e| (e.term, e.lsn)) // omit TermSwitchEntry
171 400 : .collect::<Vec<_>>()
172 400 : )
173 400 : }
174 : }
175 :
176 : /// Unique id of proposer. Not needed for correctness, used for monitoring.
177 : pub type PgUuid = [u8; 16];
178 :
179 : /// Persistent consensus state of the acceptor.
180 12 : #[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
181 : pub struct AcceptorState {
182 : /// acceptor's last term it voted for (advanced in 1 phase)
183 : pub term: Term,
184 : /// History of term switches for safekeeper's WAL.
185 : /// Actually it often goes *beyond* WAL contents as we adopt term history
186 : /// from the proposer before recovery.
187 : pub term_history: TermHistory,
188 : }
189 :
190 : impl AcceptorState {
191 : /// acceptor's last_log_term is the term of the highest entry in the log
192 7023 : pub fn get_last_log_term(&self, flush_lsn: Lsn) -> Term {
193 7023 : let th = self.term_history.up_to(flush_lsn);
194 7023 : match th.0.last() {
195 5894 : Some(e) => e.term,
196 1129 : None => 0,
197 : }
198 7023 : }
199 : }
200 :
201 : /// Information about Postgres. Safekeeper gets it once and then verifies
202 : /// all further connections from computes match.
203 8 : #[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
204 : pub struct ServerInfo {
205 : /// Postgres server version
206 : pub pg_version: u32,
207 : pub system_id: SystemId,
208 : pub wal_seg_size: u32,
209 : }
210 :
211 4 : #[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
212 : pub struct PersistedPeerInfo {
213 : /// LSN up to which safekeeper offloaded WAL to s3.
214 : pub backup_lsn: Lsn,
215 : /// Term of the last entry.
216 : pub term: Term,
217 : /// LSN of the last record.
218 : pub flush_lsn: Lsn,
219 : /// Up to which LSN safekeeper regards its WAL as committed.
220 : pub commit_lsn: Lsn,
221 : }
222 :
223 : impl PersistedPeerInfo {
224 0 : pub fn new() -> Self {
225 0 : Self {
226 0 : backup_lsn: Lsn::INVALID,
227 0 : term: INVALID_TERM,
228 0 : flush_lsn: Lsn(0),
229 0 : commit_lsn: Lsn(0),
230 0 : }
231 0 : }
232 : }
233 :
234 : // make clippy happy
235 : impl Default for PersistedPeerInfo {
236 0 : fn default() -> Self {
237 0 : Self::new()
238 0 : }
239 : }
240 :
241 : // protocol messages
242 :
243 : /// Initial Proposer -> Acceptor message
244 157617 : #[derive(Debug, Deserialize)]
245 : pub struct ProposerGreeting {
246 : /// proposer-acceptor protocol version
247 : pub protocol_version: u32,
248 : /// Postgres server version
249 : pub pg_version: u32,
250 : pub proposer_id: PgUuid,
251 : pub system_id: SystemId,
252 : pub timeline_id: TimelineId,
253 : pub tenant_id: TenantId,
254 : pub tli: TimeLineID,
255 : pub wal_seg_size: u32,
256 : }
257 :
258 : /// Acceptor -> Proposer initial response: the highest term known to me
259 : /// (acceptor voted for).
260 : #[derive(Debug, Serialize)]
261 : pub struct AcceptorGreeting {
262 : term: u64,
263 : node_id: NodeId,
264 : }
265 :
266 : /// Vote request sent from proposer to safekeepers
267 25281 : #[derive(Debug, Deserialize)]
268 : pub struct VoteRequest {
269 : pub term: Term,
270 : }
271 :
272 : /// Vote itself, sent from safekeeper to proposer
273 : #[derive(Debug, Serialize)]
274 : pub struct VoteResponse {
275 : pub term: Term, // safekeeper's current term; if it is higher than proposer's, the compute is out of date.
276 : vote_given: u64, // fixme u64 due to padding
277 : // Safekeeper flush_lsn (end of WAL) + history of term switches allow
278 : // proposer to choose the most advanced one.
279 : pub flush_lsn: Lsn,
280 : truncate_lsn: Lsn,
281 : pub term_history: TermHistory,
282 : timeline_start_lsn: Lsn,
283 : }
284 :
285 : /*
286 : * Proposer -> Acceptor message announcing proposer is elected and communicating
287 : * term history to it.
288 : */
289 : #[derive(Debug)]
290 : pub struct ProposerElected {
291 : pub term: Term,
292 : pub start_streaming_at: Lsn,
293 : pub term_history: TermHistory,
294 : pub timeline_start_lsn: Lsn,
295 : }
296 :
297 : /// Request with WAL message sent from proposer to safekeeper. Along the way it
298 : /// communicates commit_lsn.
299 : #[derive(Debug)]
300 : pub struct AppendRequest {
301 : pub h: AppendRequestHeader,
302 : pub wal_data: Bytes,
303 : }
304 23215 : #[derive(Debug, Clone, Deserialize)]
305 : pub struct AppendRequestHeader {
306 : // safekeeper's current term; if it is higher than proposer's, the compute is out of date.
307 : pub term: Term,
308 : // TODO: remove this field from the protocol, it in unused -- LSN of term
309 : // switch can be taken from ProposerElected (as well as from term history).
310 : pub term_start_lsn: Lsn,
311 : /// start position of message in WAL
312 : pub begin_lsn: Lsn,
313 : /// end position of message in WAL
314 : pub end_lsn: Lsn,
315 : /// LSN committed by quorum of safekeepers
316 : pub commit_lsn: Lsn,
317 : /// minimal LSN which may be needed by proposer to perform recovery of some safekeeper
318 : pub truncate_lsn: Lsn,
319 : // only for logging/debugging
320 : pub proposer_uuid: PgUuid,
321 : }
322 :
323 : /// Report safekeeper state to proposer
324 : #[derive(Debug, Serialize, Clone)]
325 : pub struct AppendResponse {
326 : // Current term of the safekeeper; if it is higher than proposer's, the
327 : // compute is out of date.
328 : pub term: Term,
329 : // Flushed end of wal on safekeeper; one should be always mindful from what
330 : // term history this value comes, either checking history directly or
331 : // observing term being set to one for which WAL truncation is known to have
332 : // happened.
333 : pub flush_lsn: Lsn,
334 : // We report back our awareness about which WAL is committed, as this is
335 : // a criterion for walproposer --sync mode exit
336 : pub commit_lsn: Lsn,
337 : pub hs_feedback: HotStandbyFeedback,
338 : pub pageserver_feedback: Option<PageserverFeedback>,
339 : }
340 :
341 : impl AppendResponse {
342 0 : fn term_only(term: Term) -> AppendResponse {
343 0 : AppendResponse {
344 0 : term,
345 0 : flush_lsn: Lsn(0),
346 0 : commit_lsn: Lsn(0),
347 0 : hs_feedback: HotStandbyFeedback::empty(),
348 0 : pageserver_feedback: None,
349 0 : }
350 0 : }
351 : }
352 :
353 : /// Proposer -> Acceptor messages
354 : #[derive(Debug)]
355 : pub enum ProposerAcceptorMessage {
356 : Greeting(ProposerGreeting),
357 : VoteRequest(VoteRequest),
358 : Elected(ProposerElected),
359 : AppendRequest(AppendRequest),
360 : NoFlushAppendRequest(AppendRequest),
361 : FlushWAL,
362 : }
363 :
364 : impl ProposerAcceptorMessage {
365 : /// Parse proposer message.
366 213130 : pub fn parse(msg_bytes: Bytes) -> Result<ProposerAcceptorMessage> {
367 213130 : // xxx using Reader is inefficient but easy to work with bincode
368 213130 : let mut stream = msg_bytes.reader();
369 : // u64 is here to avoid padding; it will be removed once we stop packing C structs into the wire as is
370 213130 : let tag = stream.read_u64::<LittleEndian>()? as u8 as char;
371 213130 : match tag {
372 : 'g' => {
373 157617 : let msg = ProposerGreeting::des_from(&mut stream)?;
374 157617 : Ok(ProposerAcceptorMessage::Greeting(msg))
375 : }
376 : 'v' => {
377 25281 : let msg = VoteRequest::des_from(&mut stream)?;
378 25281 : Ok(ProposerAcceptorMessage::VoteRequest(msg))
379 : }
380 : 'e' => {
381 7017 : let mut msg_bytes = stream.into_inner();
382 7017 : if msg_bytes.remaining() < 16 {
383 0 : bail!("ProposerElected message is not complete");
384 7017 : }
385 7017 : let term = msg_bytes.get_u64_le();
386 7017 : let start_streaming_at = msg_bytes.get_u64_le().into();
387 7017 : let term_history = TermHistory::from_bytes(&mut msg_bytes)?;
388 7017 : if msg_bytes.remaining() < 8 {
389 0 : bail!("ProposerElected message is not complete");
390 7017 : }
391 7017 : let timeline_start_lsn = msg_bytes.get_u64_le().into();
392 7017 : let msg = ProposerElected {
393 7017 : term,
394 7017 : start_streaming_at,
395 7017 : timeline_start_lsn,
396 7017 : term_history,
397 7017 : };
398 7017 : Ok(ProposerAcceptorMessage::Elected(msg))
399 : }
400 : 'a' => {
401 : // read header followed by wal data
402 23215 : let hdr = AppendRequestHeader::des_from(&mut stream)?;
403 23215 : let rec_size = hdr
404 23215 : .end_lsn
405 23215 : .checked_sub(hdr.begin_lsn)
406 23215 : .context("begin_lsn > end_lsn in AppendRequest")?
407 : .0 as usize;
408 23215 : if rec_size > MAX_SEND_SIZE {
409 0 : bail!(
410 0 : "AppendRequest is longer than MAX_SEND_SIZE ({})",
411 0 : MAX_SEND_SIZE
412 0 : );
413 23215 : }
414 23215 :
415 23215 : let mut wal_data_vec: Vec<u8> = vec![0; rec_size];
416 23215 : stream.read_exact(&mut wal_data_vec)?;
417 23215 : let wal_data = Bytes::from(wal_data_vec);
418 23215 : let msg = AppendRequest { h: hdr, wal_data };
419 23215 :
420 23215 : Ok(ProposerAcceptorMessage::AppendRequest(msg))
421 : }
422 0 : _ => bail!("unknown proposer-acceptor message tag: {}", tag),
423 : }
424 213130 : }
425 : }
426 :
427 : /// Acceptor -> Proposer messages
428 : #[derive(Debug)]
429 : pub enum AcceptorProposerMessage {
430 : Greeting(AcceptorGreeting),
431 : VoteResponse(VoteResponse),
432 : AppendResponse(AppendResponse),
433 : }
434 :
435 : impl AcceptorProposerMessage {
436 : /// Serialize acceptor -> proposer message.
437 202041 : pub fn serialize(&self, buf: &mut BytesMut) -> Result<()> {
438 202041 : match self {
439 157617 : AcceptorProposerMessage::Greeting(msg) => {
440 157617 : buf.put_u64_le('g' as u64);
441 157617 : buf.put_u64_le(msg.term);
442 157617 : buf.put_u64_le(msg.node_id.0);
443 157617 : }
444 25281 : AcceptorProposerMessage::VoteResponse(msg) => {
445 25281 : buf.put_u64_le('v' as u64);
446 25281 : buf.put_u64_le(msg.term);
447 25281 : buf.put_u64_le(msg.vote_given);
448 25281 : buf.put_u64_le(msg.flush_lsn.into());
449 25281 : buf.put_u64_le(msg.truncate_lsn.into());
450 25281 : buf.put_u32_le(msg.term_history.0.len() as u32);
451 123766 : for e in &msg.term_history.0 {
452 98485 : buf.put_u64_le(e.term);
453 98485 : buf.put_u64_le(e.lsn.into());
454 98485 : }
455 25281 : buf.put_u64_le(msg.timeline_start_lsn.into());
456 : }
457 19143 : AcceptorProposerMessage::AppendResponse(msg) => {
458 19143 : buf.put_u64_le('a' as u64);
459 19143 : buf.put_u64_le(msg.term);
460 19143 : buf.put_u64_le(msg.flush_lsn.into());
461 19143 : buf.put_u64_le(msg.commit_lsn.into());
462 19143 : buf.put_i64_le(msg.hs_feedback.ts);
463 19143 : buf.put_u64_le(msg.hs_feedback.xmin);
464 19143 : buf.put_u64_le(msg.hs_feedback.catalog_xmin);
465 :
466 : // AsyncReadMessage in walproposer.c will not try to decode pageserver_feedback
467 : // if it is not present.
468 19143 : if let Some(ref msg) = msg.pageserver_feedback {
469 0 : msg.serialize(buf);
470 19143 : }
471 : }
472 : }
473 :
474 202041 : Ok(())
475 202041 : }
476 : }
477 :
478 : /// Safekeeper implements consensus to reliably persist WAL across nodes.
479 : /// It controls all WAL disk writes and updates of control file.
480 : ///
481 : /// Currently safekeeper processes:
482 : /// - messages from compute (proposers) and provides replies
483 : /// - messages from broker peers
484 : pub struct SafeKeeper<CTRL: control_file::Storage, WAL: wal_storage::Storage> {
485 : /// LSN since the proposer safekeeper currently talking to appends WAL;
486 : /// determines last_log_term switch point.
487 : pub term_start_lsn: Lsn,
488 :
489 : pub state: TimelineState<CTRL>, // persistent state storage
490 : pub wal_store: WAL,
491 :
492 : node_id: NodeId, // safekeeper's node id
493 : }
494 :
495 : impl<CTRL, WAL> SafeKeeper<CTRL, WAL>
496 : where
497 : CTRL: control_file::Storage,
498 : WAL: wal_storage::Storage,
499 : {
500 : /// Accepts a control file storage containing the safekeeper state.
501 : /// State must be initialized, i.e. contain filled `tenant_id`, `timeline_id`
502 : /// and `server` (`wal_seg_size` inside it) fields.
503 70501 : pub fn new(
504 70501 : state: TimelineState<CTRL>,
505 70501 : wal_store: WAL,
506 70501 : node_id: NodeId,
507 70501 : ) -> Result<SafeKeeper<CTRL, WAL>> {
508 70501 : if state.tenant_id == TenantId::from([0u8; 16])
509 70501 : || state.timeline_id == TimelineId::from([0u8; 16])
510 : {
511 0 : bail!(
512 0 : "Calling SafeKeeper::new with empty tenant_id ({}) or timeline_id ({})",
513 0 : state.tenant_id,
514 0 : state.timeline_id
515 0 : );
516 70501 : }
517 70501 :
518 70501 : Ok(SafeKeeper {
519 70501 : term_start_lsn: Lsn(0),
520 70501 : state,
521 70501 : wal_store,
522 70501 : node_id,
523 70501 : })
524 70501 : }
525 :
526 : /// Get history of term switches for the available WAL
527 25285 : fn get_term_history(&self) -> TermHistory {
528 25285 : self.state
529 25285 : .acceptor_state
530 25285 : .term_history
531 25285 : .up_to(self.flush_lsn())
532 25285 : }
533 :
534 7023 : pub fn get_last_log_term(&self) -> Term {
535 7023 : self.state
536 7023 : .acceptor_state
537 7023 : .get_last_log_term(self.flush_lsn())
538 7023 : }
539 :
540 : /// wal_store wrapper avoiding commit_lsn <= flush_lsn violation when we don't have WAL yet.
541 88616 : pub fn flush_lsn(&self) -> Lsn {
542 88616 : max(self.wal_store.flush_lsn(), self.state.timeline_start_lsn)
543 88616 : }
544 :
545 : /// Process message from proposer and possibly form reply. Concurrent
546 : /// callers must exclude each other.
547 232283 : pub async fn process_msg(
548 232283 : &mut self,
549 232283 : msg: &ProposerAcceptorMessage,
550 232283 : ) -> Result<Option<AcceptorProposerMessage>> {
551 232283 : match msg {
552 157617 : ProposerAcceptorMessage::Greeting(msg) => self.handle_greeting(msg).await,
553 25285 : ProposerAcceptorMessage::VoteRequest(msg) => self.handle_vote_request(msg).await,
554 7019 : ProposerAcceptorMessage::Elected(msg) => self.handle_elected(msg).await,
555 4 : ProposerAcceptorMessage::AppendRequest(msg) => {
556 4 : self.handle_append_request(msg, true).await
557 : }
558 23215 : ProposerAcceptorMessage::NoFlushAppendRequest(msg) => {
559 23215 : self.handle_append_request(msg, false).await
560 : }
561 19143 : ProposerAcceptorMessage::FlushWAL => self.handle_flush().await,
562 : }
563 232283 : }
564 :
565 : /// Handle initial message from proposer: check its sanity and send my
566 : /// current term.
567 157617 : async fn handle_greeting(
568 157617 : &mut self,
569 157617 : msg: &ProposerGreeting,
570 157617 : ) -> Result<Option<AcceptorProposerMessage>> {
571 157617 : // Check protocol compatibility
572 157617 : if msg.protocol_version != SK_PROTOCOL_VERSION {
573 0 : bail!(
574 0 : "incompatible protocol version {}, expected {}",
575 0 : msg.protocol_version,
576 0 : SK_PROTOCOL_VERSION
577 0 : );
578 157617 : }
579 157617 : /* Postgres major version mismatch is treated as fatal error
580 157617 : * because safekeepers parse WAL headers and the format
581 157617 : * may change between versions.
582 157617 : */
583 157617 : if msg.pg_version / 10000 != self.state.server.pg_version / 10000
584 0 : && self.state.server.pg_version != UNKNOWN_SERVER_VERSION
585 : {
586 0 : bail!(
587 0 : "incompatible server version {}, expected {}",
588 0 : msg.pg_version,
589 0 : self.state.server.pg_version
590 0 : );
591 157617 : }
592 157617 :
593 157617 : if msg.tenant_id != self.state.tenant_id {
594 0 : bail!(
595 0 : "invalid tenant ID, got {}, expected {}",
596 0 : msg.tenant_id,
597 0 : self.state.tenant_id
598 0 : );
599 157617 : }
600 157617 : if msg.timeline_id != self.state.timeline_id {
601 0 : bail!(
602 0 : "invalid timeline ID, got {}, expected {}",
603 0 : msg.timeline_id,
604 0 : self.state.timeline_id
605 0 : );
606 157617 : }
607 157617 : if self.state.server.wal_seg_size != msg.wal_seg_size {
608 0 : bail!(
609 0 : "invalid wal_seg_size, got {}, expected {}",
610 0 : msg.wal_seg_size,
611 0 : self.state.server.wal_seg_size
612 0 : );
613 157617 : }
614 157617 :
615 157617 : // system_id will be updated on mismatch
616 157617 : // sync-safekeepers doesn't know sysid and sends 0, ignore it
617 157617 : if self.state.server.system_id != msg.system_id && msg.system_id != 0 {
618 0 : if self.state.server.system_id != 0 {
619 0 : warn!(
620 0 : "unexpected system ID arrived, got {}, expected {}",
621 0 : msg.system_id, self.state.server.system_id
622 : );
623 0 : }
624 :
625 0 : let mut state = self.state.start_change();
626 0 : state.server.system_id = msg.system_id;
627 0 : if msg.pg_version != UNKNOWN_SERVER_VERSION {
628 0 : state.server.pg_version = msg.pg_version;
629 0 : }
630 0 : self.state.finish_change(&state).await?;
631 157617 : }
632 :
633 157617 : info!(
634 0 : "processed greeting from walproposer {}, sending term {:?}",
635 4832 : msg.proposer_id.map(|b| format!("{:X}", b)).join(""),
636 0 : self.state.acceptor_state.term
637 : );
638 157617 : Ok(Some(AcceptorProposerMessage::Greeting(AcceptorGreeting {
639 157617 : term: self.state.acceptor_state.term,
640 157617 : node_id: self.node_id,
641 157617 : })))
642 157617 : }
643 :
644 : /// Give vote for the given term, if we haven't done that previously.
645 25285 : async fn handle_vote_request(
646 25285 : &mut self,
647 25285 : msg: &VoteRequest,
648 25285 : ) -> Result<Option<AcceptorProposerMessage>> {
649 25285 : // Once voted, we won't accept data from older proposers; flush
650 25285 : // everything we've already received so that new proposer starts
651 25285 : // streaming at end of our WAL, without overlap. Currently we truncate
652 25285 : // WAL at streaming point, so this avoids truncating already committed
653 25285 : // WAL.
654 25285 : //
655 25285 : // TODO: it would be smoother to not truncate committed piece at
656 25285 : // handle_elected instead. Currently not a big deal, as proposer is the
657 25285 : // only source of WAL; with peer2peer recovery it would be more
658 25285 : // important.
659 25285 : self.wal_store.flush_wal().await?;
660 : // initialize with refusal
661 25285 : let mut resp = VoteResponse {
662 25285 : term: self.state.acceptor_state.term,
663 25285 : vote_given: false as u64,
664 25285 : flush_lsn: self.flush_lsn(),
665 25285 : truncate_lsn: self.state.inmem.peer_horizon_lsn,
666 25285 : term_history: self.get_term_history(),
667 25285 : timeline_start_lsn: self.state.timeline_start_lsn,
668 25285 : };
669 25285 : if self.state.acceptor_state.term < msg.term {
670 24107 : let mut state = self.state.start_change();
671 24107 : state.acceptor_state.term = msg.term;
672 24107 : // persist vote before sending it out
673 24107 : self.state.finish_change(&state).await?;
674 :
675 24107 : resp.term = self.state.acceptor_state.term;
676 24107 : resp.vote_given = true as u64;
677 1178 : }
678 25285 : info!("processed VoteRequest for term {}: {:?}", msg.term, &resp);
679 25285 : Ok(Some(AcceptorProposerMessage::VoteResponse(resp)))
680 25285 : }
681 :
682 : /// Form AppendResponse from current state.
683 19147 : fn append_response(&self) -> AppendResponse {
684 19147 : let ar = AppendResponse {
685 19147 : term: self.state.acceptor_state.term,
686 19147 : flush_lsn: self.flush_lsn(),
687 19147 : commit_lsn: self.state.commit_lsn,
688 19147 : // will be filled by the upper code to avoid bothering safekeeper
689 19147 : hs_feedback: HotStandbyFeedback::empty(),
690 19147 : pageserver_feedback: None,
691 19147 : };
692 19147 : trace!("formed AppendResponse {:?}", ar);
693 19147 : ar
694 19147 : }
695 :
696 7019 : async fn handle_elected(
697 7019 : &mut self,
698 7019 : msg: &ProposerElected,
699 7019 : ) -> Result<Option<AcceptorProposerMessage>> {
700 7019 : let _timer = MISC_OPERATION_SECONDS
701 7019 : .with_label_values(&["handle_elected"])
702 7019 : .start_timer();
703 7019 :
704 7019 : info!("received ProposerElected {:?}", msg);
705 7019 : if self.state.acceptor_state.term < msg.term {
706 2 : let mut state = self.state.start_change();
707 2 : state.acceptor_state.term = msg.term;
708 2 : self.state.finish_change(&state).await?;
709 7017 : }
710 :
711 : // If our term is higher, ignore the message (next feedback will inform the compute)
712 7019 : if self.state.acceptor_state.term > msg.term {
713 0 : return Ok(None);
714 7019 : }
715 7019 :
716 7019 : // This might happen in a rare race when another (old) connection from
717 7019 : // the same walproposer writes + flushes WAL after this connection
718 7019 : // already sent flush_lsn in VoteRequest. It is generally safe to
719 7019 : // proceed, but to prevent commit_lsn surprisingly going down we should
720 7019 : // either refuse the session (simpler) or skip the part we already have
721 7019 : // from the stream (can be implemented).
722 7019 : if msg.term == self.get_last_log_term() && self.flush_lsn() > msg.start_streaming_at {
723 0 : bail!("refusing ProposerElected which is going to overwrite correct WAL: term={}, flush_lsn={}, start_streaming_at={}; restarting the handshake should help",
724 0 : msg.term, self.flush_lsn(), msg.start_streaming_at)
725 7019 : }
726 7019 : // Otherwise we must never attempt to truncate committed data.
727 7019 : assert!(
728 7019 : msg.start_streaming_at >= self.state.inmem.commit_lsn,
729 0 : "attempt to truncate committed data: start_streaming_at={}, commit_lsn={}",
730 : msg.start_streaming_at,
731 : self.state.inmem.commit_lsn
732 : );
733 :
734 : // Before first WAL write initialize its segment. It makes first segment
735 : // pg_waldump'able because stream from compute doesn't include its
736 : // segment and page headers.
737 : //
738 : // If we fail before first WAL write flush this action would be
739 : // repeated, that's ok because it is idempotent.
740 7019 : if self.wal_store.flush_lsn() == Lsn::INVALID {
741 1127 : self.wal_store
742 1127 : .initialize_first_segment(msg.start_streaming_at)
743 0 : .await?;
744 5892 : }
745 :
746 : // TODO: cross check divergence point, check if msg.start_streaming_at corresponds to
747 : // intersection of our history and history from msg
748 :
749 : // truncate wal, update the LSNs
750 7019 : self.wal_store.truncate_wal(msg.start_streaming_at).await?;
751 :
752 : // and now adopt term history from proposer
753 : {
754 7019 : let mut state = self.state.start_change();
755 7019 :
756 7019 : // Here we learn initial LSN for the first time, set fields
757 7019 : // interested in that.
758 7019 :
759 7019 : if state.timeline_start_lsn == Lsn(0) {
760 : // Remember point where WAL begins globally.
761 1127 : state.timeline_start_lsn = msg.timeline_start_lsn;
762 1127 : info!(
763 0 : "setting timeline_start_lsn to {:?}",
764 : state.timeline_start_lsn
765 : );
766 5892 : }
767 7019 : if state.peer_horizon_lsn == Lsn(0) {
768 1127 : // Update peer_horizon_lsn as soon as we know where timeline starts.
769 1127 : // It means that peer_horizon_lsn cannot be zero after we know timeline_start_lsn.
770 1127 : state.peer_horizon_lsn = msg.timeline_start_lsn;
771 5892 : }
772 7019 : if state.local_start_lsn == Lsn(0) {
773 1127 : state.local_start_lsn = msg.start_streaming_at;
774 1127 : info!("setting local_start_lsn to {:?}", state.local_start_lsn);
775 5892 : }
776 : // Initializing commit_lsn before acking first flushed record is
777 : // important to let find_end_of_wal skip the hole in the beginning
778 : // of the first segment.
779 : //
780 : // NB: on new clusters, this happens at the same time as
781 : // timeline_start_lsn initialization, it is taken outside to provide
782 : // upgrade.
783 7019 : state.commit_lsn = max(state.commit_lsn, state.timeline_start_lsn);
784 7019 :
785 7019 : // Initializing backup_lsn is useful to avoid making backup think it should upload 0 segment.
786 7019 : state.backup_lsn = max(state.backup_lsn, state.timeline_start_lsn);
787 7019 : // similar for remote_consistent_lsn
788 7019 : state.remote_consistent_lsn =
789 7019 : max(state.remote_consistent_lsn, state.timeline_start_lsn);
790 7019 :
791 7019 : state.acceptor_state.term_history = msg.term_history.clone();
792 7019 : self.state.finish_change(&state).await?;
793 : }
794 :
795 7019 : info!("start receiving WAL since {:?}", msg.start_streaming_at);
796 :
797 : // Cache LSN where term starts to immediately fsync control file with
798 : // commit_lsn once we reach it -- sync-safekeepers finishes when
799 : // persisted commit_lsn on majority of safekeepers aligns.
800 7019 : self.term_start_lsn = match msg.term_history.0.last() {
801 0 : None => bail!("proposer elected with empty term history"),
802 7019 : Some(term_lsn_start) => term_lsn_start.lsn,
803 7019 : };
804 7019 :
805 7019 : Ok(None)
806 7019 : }
807 :
808 : /// Advance commit_lsn taking into account what we have locally.
809 : ///
810 : /// Note: it is assumed that 'WAL we have is from the right term' check has
811 : /// already been done outside.
812 11495 : async fn update_commit_lsn(&mut self, mut candidate: Lsn) -> Result<()> {
813 11495 : // Both peers and walproposer communicate this value, we might already
814 11495 : // have a fresher (higher) version.
815 11495 : candidate = max(candidate, self.state.inmem.commit_lsn);
816 11495 : let commit_lsn = min(candidate, self.flush_lsn());
817 11495 : assert!(
818 11495 : commit_lsn >= self.state.inmem.commit_lsn,
819 0 : "commit_lsn monotonicity violated: old={} new={}",
820 : self.state.inmem.commit_lsn,
821 : commit_lsn
822 : );
823 :
824 11495 : self.state.inmem.commit_lsn = commit_lsn;
825 11495 :
826 11495 : // If new commit_lsn reached term switch, force sync of control
827 11495 : // file: walproposer in sync mode is very interested when this
828 11495 : // happens. Note: this is for sync-safekeepers mode only, as
829 11495 : // otherwise commit_lsn might jump over term_start_lsn.
830 11495 : if commit_lsn >= self.term_start_lsn && self.state.commit_lsn < self.term_start_lsn {
831 907 : self.state.flush().await?;
832 10588 : }
833 :
834 11495 : Ok(())
835 11495 : }
836 :
837 : /// Handle request to append WAL.
838 : #[allow(clippy::comparison_chain)]
839 23219 : async fn handle_append_request(
840 23219 : &mut self,
841 23219 : msg: &AppendRequest,
842 23219 : require_flush: bool,
843 23219 : ) -> Result<Option<AcceptorProposerMessage>> {
844 23219 : if self.state.acceptor_state.term < msg.h.term {
845 0 : bail!("got AppendRequest before ProposerElected");
846 23219 : }
847 23219 :
848 23219 : // If our term is higher, immediately refuse the message.
849 23219 : if self.state.acceptor_state.term > msg.h.term {
850 0 : let resp = AppendResponse::term_only(self.state.acceptor_state.term);
851 0 : return Ok(Some(AcceptorProposerMessage::AppendResponse(resp)));
852 23219 : }
853 23219 :
854 23219 : // Now we know that we are in the same term as the proposer,
855 23219 : // processing the message.
856 23219 :
857 23219 : self.state.inmem.proposer_uuid = msg.h.proposer_uuid;
858 23219 :
859 23219 : // do the job
860 23219 : if !msg.wal_data.is_empty() {
861 4267 : self.wal_store
862 4267 : .write_wal(msg.h.begin_lsn, &msg.wal_data)
863 0 : .await?;
864 18952 : }
865 :
866 : // flush wal to the disk, if required
867 23219 : if require_flush {
868 4 : self.wal_store.flush_wal().await?;
869 23215 : }
870 :
871 : // Update commit_lsn.
872 23219 : if msg.h.commit_lsn != Lsn(0) {
873 11495 : self.update_commit_lsn(msg.h.commit_lsn).await?;
874 11724 : }
875 : // Value calculated by walproposer can always lag:
876 : // - safekeepers can forget inmem value and send to proposer lower
877 : // persisted one on restart;
878 : // - if we make safekeepers always send persistent value,
879 : // any compute restart would pull it down.
880 : // Thus, take max before adopting.
881 23219 : self.state.inmem.peer_horizon_lsn =
882 23219 : max(self.state.inmem.peer_horizon_lsn, msg.h.truncate_lsn);
883 23219 :
884 23219 : // Update truncate and commit LSN in control file.
885 23219 : // To avoid negative impact on performance of extra fsync, do it only
886 23219 : // when commit_lsn delta exceeds WAL segment size.
887 23219 : if self.state.commit_lsn + (self.state.server.wal_seg_size as u64)
888 23219 : < self.state.inmem.commit_lsn
889 : {
890 0 : self.state.flush().await?;
891 23219 : }
892 :
893 23219 : trace!(
894 0 : "processed AppendRequest of len {}, end_lsn={:?}, commit_lsn={:?}, truncate_lsn={:?}, flushed={:?}",
895 0 : msg.wal_data.len(),
896 : msg.h.end_lsn,
897 : msg.h.commit_lsn,
898 : msg.h.truncate_lsn,
899 : require_flush,
900 : );
901 :
902 : // If flush_lsn hasn't updated, AppendResponse is not very useful.
903 23219 : if !require_flush {
904 23215 : return Ok(None);
905 4 : }
906 4 :
907 4 : let resp = self.append_response();
908 4 : Ok(Some(AcceptorProposerMessage::AppendResponse(resp)))
909 23219 : }
910 :
911 : /// Flush WAL to disk. Return AppendResponse with latest LSNs.
912 19143 : async fn handle_flush(&mut self) -> Result<Option<AcceptorProposerMessage>> {
913 19143 : self.wal_store.flush_wal().await?;
914 19143 : Ok(Some(AcceptorProposerMessage::AppendResponse(
915 19143 : self.append_response(),
916 19143 : )))
917 19143 : }
918 :
919 : /// Update commit_lsn from peer safekeeper data.
920 0 : pub async fn record_safekeeper_info(&mut self, sk_info: &SafekeeperTimelineInfo) -> Result<()> {
921 0 : if (Lsn(sk_info.commit_lsn) != Lsn::INVALID) && (sk_info.last_log_term != INVALID_TERM) {
922 : // Note: the check is too restrictive, generally we can update local
923 : // commit_lsn if our history matches (is part of) history of advanced
924 : // commit_lsn provider.
925 0 : if sk_info.last_log_term == self.get_last_log_term() {
926 0 : self.update_commit_lsn(Lsn(sk_info.commit_lsn)).await?;
927 0 : }
928 0 : }
929 0 : Ok(())
930 0 : }
931 : }
932 :
933 : #[cfg(test)]
934 : mod tests {
935 : use futures::future::BoxFuture;
936 : use postgres_ffi::{XLogSegNo, WAL_SEGMENT_SIZE};
937 :
938 : use super::*;
939 : use crate::{
940 : state::{EvictionState, PersistedPeers, TimelinePersistentState},
941 : wal_storage::Storage,
942 : };
943 : use std::{ops::Deref, str::FromStr, time::Instant};
944 :
945 : // fake storage for tests
946 : struct InMemoryState {
947 : persisted_state: TimelinePersistentState,
948 : }
949 :
950 : #[async_trait::async_trait]
951 : impl control_file::Storage for InMemoryState {
952 6 : async fn persist(&mut self, s: &TimelinePersistentState) -> Result<()> {
953 6 : self.persisted_state = s.clone();
954 6 : Ok(())
955 6 : }
956 :
957 0 : fn last_persist_at(&self) -> Instant {
958 0 : Instant::now()
959 0 : }
960 : }
961 :
962 : impl Deref for InMemoryState {
963 : type Target = TimelinePersistentState;
964 :
965 126 : fn deref(&self) -> &Self::Target {
966 126 : &self.persisted_state
967 126 : }
968 : }
969 :
970 4 : fn test_sk_state() -> TimelinePersistentState {
971 4 : let mut state = TimelinePersistentState::empty();
972 4 : state.server.wal_seg_size = WAL_SEGMENT_SIZE as u32;
973 4 : state.tenant_id = TenantId::from([1u8; 16]);
974 4 : state.timeline_id = TimelineId::from([1u8; 16]);
975 4 : state
976 4 : }
977 :
978 : struct DummyWalStore {
979 : lsn: Lsn,
980 : }
981 :
982 : #[async_trait::async_trait]
983 : impl wal_storage::Storage for DummyWalStore {
984 20 : fn flush_lsn(&self) -> Lsn {
985 20 : self.lsn
986 20 : }
987 :
988 2 : async fn initialize_first_segment(&mut self, _init_lsn: Lsn) -> Result<()> {
989 2 : Ok(())
990 2 : }
991 :
992 4 : async fn write_wal(&mut self, startpos: Lsn, buf: &[u8]) -> Result<()> {
993 4 : self.lsn = startpos + buf.len() as u64;
994 4 : Ok(())
995 4 : }
996 :
997 4 : async fn truncate_wal(&mut self, end_pos: Lsn) -> Result<()> {
998 4 : self.lsn = end_pos;
999 4 : Ok(())
1000 4 : }
1001 :
1002 8 : async fn flush_wal(&mut self) -> Result<()> {
1003 8 : Ok(())
1004 8 : }
1005 :
1006 0 : fn remove_up_to(&self, _segno_up_to: XLogSegNo) -> BoxFuture<'static, anyhow::Result<()>> {
1007 0 : Box::pin(async { Ok(()) })
1008 0 : }
1009 :
1010 0 : fn get_metrics(&self) -> crate::metrics::WalStorageMetrics {
1011 0 : crate::metrics::WalStorageMetrics::default()
1012 0 : }
1013 : }
1014 :
1015 : #[tokio::test]
1016 2 : async fn test_voting() {
1017 2 : let storage = InMemoryState {
1018 2 : persisted_state: test_sk_state(),
1019 2 : };
1020 2 : let wal_store = DummyWalStore { lsn: Lsn(0) };
1021 2 : let mut sk = SafeKeeper::new(TimelineState::new(storage), wal_store, NodeId(0)).unwrap();
1022 2 :
1023 2 : // check voting for 1 is ok
1024 2 : let vote_request = ProposerAcceptorMessage::VoteRequest(VoteRequest { term: 1 });
1025 2 : let mut vote_resp = sk.process_msg(&vote_request).await;
1026 2 : match vote_resp.unwrap() {
1027 2 : Some(AcceptorProposerMessage::VoteResponse(resp)) => assert!(resp.vote_given != 0),
1028 2 : r => panic!("unexpected response: {:?}", r),
1029 2 : }
1030 2 :
1031 2 : // reboot...
1032 2 : let state = sk.state.deref().clone();
1033 2 : let storage = InMemoryState {
1034 2 : persisted_state: state,
1035 2 : };
1036 2 :
1037 2 : sk = SafeKeeper::new(TimelineState::new(storage), sk.wal_store, NodeId(0)).unwrap();
1038 2 :
1039 2 : // and ensure voting second time for 1 is not ok
1040 2 : vote_resp = sk.process_msg(&vote_request).await;
1041 2 : match vote_resp.unwrap() {
1042 2 : Some(AcceptorProposerMessage::VoteResponse(resp)) => assert!(resp.vote_given == 0),
1043 2 : r => panic!("unexpected response: {:?}", r),
1044 2 : }
1045 2 : }
1046 :
1047 : #[tokio::test]
1048 2 : async fn test_last_log_term_switch() {
1049 2 : let storage = InMemoryState {
1050 2 : persisted_state: test_sk_state(),
1051 2 : };
1052 2 : let wal_store = DummyWalStore { lsn: Lsn(0) };
1053 2 :
1054 2 : let mut sk = SafeKeeper::new(TimelineState::new(storage), wal_store, NodeId(0)).unwrap();
1055 2 :
1056 2 : let mut ar_hdr = AppendRequestHeader {
1057 2 : term: 1,
1058 2 : term_start_lsn: Lsn(3),
1059 2 : begin_lsn: Lsn(1),
1060 2 : end_lsn: Lsn(2),
1061 2 : commit_lsn: Lsn(0),
1062 2 : truncate_lsn: Lsn(0),
1063 2 : proposer_uuid: [0; 16],
1064 2 : };
1065 2 : let mut append_request = AppendRequest {
1066 2 : h: ar_hdr.clone(),
1067 2 : wal_data: Bytes::from_static(b"b"),
1068 2 : };
1069 2 :
1070 2 : let pem = ProposerElected {
1071 2 : term: 1,
1072 2 : start_streaming_at: Lsn(1),
1073 2 : term_history: TermHistory(vec![TermLsn {
1074 2 : term: 1,
1075 2 : lsn: Lsn(3),
1076 2 : }]),
1077 2 : timeline_start_lsn: Lsn(0),
1078 2 : };
1079 2 : sk.process_msg(&ProposerAcceptorMessage::Elected(pem))
1080 2 : .await
1081 2 : .unwrap();
1082 2 :
1083 2 : // check that AppendRequest before term_start_lsn doesn't switch last_log_term.
1084 2 : let resp = sk
1085 2 : .process_msg(&ProposerAcceptorMessage::AppendRequest(append_request))
1086 2 : .await;
1087 2 : assert!(resp.is_ok());
1088 2 : assert_eq!(sk.get_last_log_term(), 0);
1089 2 :
1090 2 : // but record at term_start_lsn does the switch
1091 2 : ar_hdr.begin_lsn = Lsn(2);
1092 2 : ar_hdr.end_lsn = Lsn(3);
1093 2 : append_request = AppendRequest {
1094 2 : h: ar_hdr,
1095 2 : wal_data: Bytes::from_static(b"b"),
1096 2 : };
1097 2 : let resp = sk
1098 2 : .process_msg(&ProposerAcceptorMessage::AppendRequest(append_request))
1099 2 : .await;
1100 2 : assert!(resp.is_ok());
1101 2 : sk.wal_store.truncate_wal(Lsn(3)).await.unwrap(); // imitate the complete record at 3 %)
1102 2 : assert_eq!(sk.get_last_log_term(), 1);
1103 2 : }
1104 :
1105 : #[test]
1106 2 : fn test_find_highest_common_point_none() {
1107 2 : let prop_th = TermHistory(vec![(0, Lsn(1)).into()]);
1108 2 : let sk_th = TermHistory(vec![(1, Lsn(1)).into(), (2, Lsn(2)).into()]);
1109 2 : assert_eq!(
1110 2 : TermHistory::find_highest_common_point(&prop_th, &sk_th, Lsn(3),),
1111 2 : None
1112 2 : );
1113 2 : }
1114 :
1115 : #[test]
1116 2 : fn test_find_highest_common_point_middle() {
1117 2 : let prop_th = TermHistory(vec![
1118 2 : (1, Lsn(10)).into(),
1119 2 : (2, Lsn(20)).into(),
1120 2 : (4, Lsn(40)).into(),
1121 2 : ]);
1122 2 : let sk_th = TermHistory(vec![
1123 2 : (1, Lsn(10)).into(),
1124 2 : (2, Lsn(20)).into(),
1125 2 : (3, Lsn(30)).into(), // sk ends last common term 2 at 30
1126 2 : ]);
1127 2 : assert_eq!(
1128 2 : TermHistory::find_highest_common_point(&prop_th, &sk_th, Lsn(40),),
1129 2 : Some(TermLsn {
1130 2 : term: 2,
1131 2 : lsn: Lsn(30),
1132 2 : })
1133 2 : );
1134 2 : }
1135 :
1136 : #[test]
1137 2 : fn test_find_highest_common_point_sk_end() {
1138 2 : let prop_th = TermHistory(vec![
1139 2 : (1, Lsn(10)).into(),
1140 2 : (2, Lsn(20)).into(), // last common term 2, sk will end it at 32 sk_end_lsn
1141 2 : (4, Lsn(40)).into(),
1142 2 : ]);
1143 2 : let sk_th = TermHistory(vec![(1, Lsn(10)).into(), (2, Lsn(20)).into()]);
1144 2 : assert_eq!(
1145 2 : TermHistory::find_highest_common_point(&prop_th, &sk_th, Lsn(32),),
1146 2 : Some(TermLsn {
1147 2 : term: 2,
1148 2 : lsn: Lsn(32),
1149 2 : })
1150 2 : );
1151 2 : }
1152 :
1153 : #[test]
1154 2 : fn test_find_highest_common_point_walprop() {
1155 2 : let prop_th = TermHistory(vec![(1, Lsn(10)).into(), (2, Lsn(20)).into()]);
1156 2 : let sk_th = TermHistory(vec![(1, Lsn(10)).into(), (2, Lsn(20)).into()]);
1157 2 : assert_eq!(
1158 2 : TermHistory::find_highest_common_point(&prop_th, &sk_th, Lsn(32),),
1159 2 : Some(TermLsn {
1160 2 : term: 2,
1161 2 : lsn: Lsn(32),
1162 2 : })
1163 2 : );
1164 2 : }
1165 :
1166 : #[test]
1167 2 : fn test_sk_state_bincode_serde_roundtrip() {
1168 2 : use utils::Hex;
1169 2 : let tenant_id = TenantId::from_str("cf0480929707ee75372337efaa5ecf96").unwrap();
1170 2 : let timeline_id = TimelineId::from_str("112ded66422aa5e953e5440fa5427ac4").unwrap();
1171 2 : let state = TimelinePersistentState {
1172 2 : tenant_id,
1173 2 : timeline_id,
1174 2 : acceptor_state: AcceptorState {
1175 2 : term: 42,
1176 2 : term_history: TermHistory(vec![TermLsn {
1177 2 : lsn: Lsn(0x1),
1178 2 : term: 41,
1179 2 : }]),
1180 2 : },
1181 2 : server: ServerInfo {
1182 2 : pg_version: 14,
1183 2 : system_id: 0x1234567887654321,
1184 2 : wal_seg_size: 0x12345678,
1185 2 : },
1186 2 : proposer_uuid: {
1187 2 : let mut arr = timeline_id.as_arr();
1188 2 : arr.reverse();
1189 2 : arr
1190 2 : },
1191 2 : timeline_start_lsn: Lsn(0x12345600),
1192 2 : local_start_lsn: Lsn(0x12),
1193 2 : commit_lsn: Lsn(1234567800),
1194 2 : backup_lsn: Lsn(1234567300),
1195 2 : peer_horizon_lsn: Lsn(9999999),
1196 2 : remote_consistent_lsn: Lsn(1234560000),
1197 2 : peers: PersistedPeers(vec![(
1198 2 : NodeId(1),
1199 2 : PersistedPeerInfo {
1200 2 : backup_lsn: Lsn(1234567000),
1201 2 : term: 42,
1202 2 : flush_lsn: Lsn(1234567800 - 8),
1203 2 : commit_lsn: Lsn(1234567600),
1204 2 : },
1205 2 : )]),
1206 2 : partial_backup: crate::wal_backup_partial::State::default(),
1207 2 : eviction_state: EvictionState::Present,
1208 2 : };
1209 2 :
1210 2 : let ser = state.ser().unwrap();
1211 2 :
1212 2 : #[rustfmt::skip]
1213 2 : let expected = [
1214 2 : // tenant_id as length prefixed hex
1215 2 : 0x20, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
1216 2 : 0x63, 0x66, 0x30, 0x34, 0x38, 0x30, 0x39, 0x32, 0x39, 0x37, 0x30, 0x37, 0x65, 0x65, 0x37, 0x35, 0x33, 0x37, 0x32, 0x33, 0x33, 0x37, 0x65, 0x66, 0x61, 0x61, 0x35, 0x65, 0x63, 0x66, 0x39, 0x36,
1217 2 : // timeline_id as length prefixed hex
1218 2 : 0x20, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
1219 2 : 0x31, 0x31, 0x32, 0x64, 0x65, 0x64, 0x36, 0x36, 0x34, 0x32, 0x32, 0x61, 0x61, 0x35, 0x65, 0x39, 0x35, 0x33, 0x65, 0x35, 0x34, 0x34, 0x30, 0x66, 0x61, 0x35, 0x34, 0x32, 0x37, 0x61, 0x63, 0x34,
1220 2 : // term
1221 2 : 0x2a, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
1222 2 : // length prefix
1223 2 : 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
1224 2 : // unsure why this order is swapped
1225 2 : 0x29, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
1226 2 : 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
1227 2 : // pg_version
1228 2 : 0x0e, 0x00, 0x00, 0x00,
1229 2 : // systemid
1230 2 : 0x21, 0x43, 0x65, 0x87, 0x78, 0x56, 0x34, 0x12,
1231 2 : // wal_seg_size
1232 2 : 0x78, 0x56, 0x34, 0x12,
1233 2 : // pguuid as length prefixed hex
1234 2 : 0x20, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
1235 2 : 0x63, 0x34, 0x37, 0x61, 0x34, 0x32, 0x61, 0x35, 0x30, 0x66, 0x34, 0x34, 0x65, 0x35, 0x35, 0x33, 0x65, 0x39, 0x61, 0x35, 0x32, 0x61, 0x34, 0x32, 0x36, 0x36, 0x65, 0x64, 0x32, 0x64, 0x31, 0x31,
1236 2 :
1237 2 : // timeline_start_lsn
1238 2 : 0x00, 0x56, 0x34, 0x12, 0x00, 0x00, 0x00, 0x00,
1239 2 : 0x12, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
1240 2 : 0x78, 0x02, 0x96, 0x49, 0x00, 0x00, 0x00, 0x00,
1241 2 : 0x84, 0x00, 0x96, 0x49, 0x00, 0x00, 0x00, 0x00,
1242 2 : 0x7f, 0x96, 0x98, 0x00, 0x00, 0x00, 0x00, 0x00,
1243 2 : 0x00, 0xe4, 0x95, 0x49, 0x00, 0x00, 0x00, 0x00,
1244 2 : // length prefix for persistentpeers
1245 2 : 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
1246 2 : // nodeid
1247 2 : 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
1248 2 : // backuplsn
1249 2 : 0x58, 0xff, 0x95, 0x49, 0x00, 0x00, 0x00, 0x00,
1250 2 : 0x2a, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
1251 2 : 0x70, 0x02, 0x96, 0x49, 0x00, 0x00, 0x00, 0x00,
1252 2 : 0xb0, 0x01, 0x96, 0x49, 0x00, 0x00, 0x00, 0x00,
1253 2 : // partial_backup
1254 2 : 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
1255 2 : // eviction_state
1256 2 : 0x00, 0x00, 0x00, 0x00,
1257 2 : ];
1258 2 :
1259 2 : assert_eq!(Hex(&ser), Hex(&expected));
1260 :
1261 2 : let deser = TimelinePersistentState::des(&ser).unwrap();
1262 2 :
1263 2 : assert_eq!(deser, state);
1264 2 : }
1265 : }
|