database/postgres
database/postgres
1
0
Fork 0
mirror of https://github.com/postgres/postgres.git synced 2025-05-06 19:59:18 +03:00
Code

Allow WAL summarization to back up when timeline changes.

Browse Source 
The old code believed that it was not possible to switch timelines
without first replaying all of the WAL from the old timeline, but
that turns out to be false, as demonstrated by an example from Fujii
Masao. As a result, it assumed that summarization would always
continue from the LSN where summarization previously ended. But in
fact, when a timeline switch occurs without replaying all the WAL
from the previous timeline, we can need to back up to an earlier
LSN. Adjust accordingly.

Discussion: https://postgr.es/m/CA+TgmoZGEsZodXC4f=XZNkAeyuDmWTSkpkjCEOcF19Am0mt_OA@mail.gmail.com
This commit is contained in:
Robert Haas 2024-07-26 09:50:31 -04:00
parent c0c0050708
commit c7cfbc5157
1 changed files with 19 additions and 13 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

32
src/backend/postmaster/walsummarizer.c
View File

@ -388,13 +388,27 @@ WalSummarizerMain(char *startup_data, size_t startup_data_len)
/* /*
* If we've reached the switch LSN, we can't summarize anything else * If we've reached the switch LSN, we can't summarize anything else
* on this timeline. Switch to the next timeline and go around again. * on this timeline. Switch to the next timeline and go around again,
* backing up to the exact switch point if we passed it.
*/ */
if (!XLogRecPtrIsInvalid(switch_lsn) && current_lsn >= switch_lsn) if (!XLogRecPtrIsInvalid(switch_lsn) && current_lsn >= switch_lsn)
{ {
/* Restart summarization from switch point. */
current_tli = switch_tli; current_tli = switch_tli;
current_lsn = switch_lsn;
/* Next timeline and switch point, if any, not yet known. */
switch_lsn = InvalidXLogRecPtr; switch_lsn = InvalidXLogRecPtr;
switch_tli = 0; switch_tli = 0;
/* Update (really, rewind, if needed) state in shared memory. */
LWLockAcquire(WALSummarizerLock, LW_EXCLUSIVE);
WalSummarizerCtl->summarized_lsn = current_lsn;
WalSummarizerCtl->summarized_tli = current_tli;
WalSummarizerCtl->lsn_is_exact = true;
WalSummarizerCtl->pending_lsn = current_lsn;
LWLockRelease(WALSummarizerLock);
continue; continue;
} }
@ -415,7 +429,6 @@ WalSummarizerMain(char *startup_data, size_t startup_data_len)
/* Update state in shared memory. */ /* Update state in shared memory. */
LWLockAcquire(WALSummarizerLock, LW_EXCLUSIVE); LWLockAcquire(WALSummarizerLock, LW_EXCLUSIVE);
Assert(WalSummarizerCtl->pending_lsn <= end_of_summary_lsn);
WalSummarizerCtl->summarized_lsn = end_of_summary_lsn; WalSummarizerCtl->summarized_lsn = end_of_summary_lsn;
WalSummarizerCtl->summarized_tli = current_tli; WalSummarizerCtl->summarized_tli = current_tli;
WalSummarizerCtl->lsn_is_exact = true; WalSummarizerCtl->lsn_is_exact = true;
@ -1060,7 +1073,6 @@ SummarizeWAL(TimeLineID tli, XLogRecPtr start_lsn, bool exact,
/* Also update shared memory. */ /* Also update shared memory. */
LWLockAcquire(WALSummarizerLock, LW_EXCLUSIVE); LWLockAcquire(WALSummarizerLock, LW_EXCLUSIVE);
Assert(summary_end_lsn >= WalSummarizerCtl->pending_lsn);
Assert(summary_end_lsn >= WalSummarizerCtl->summarized_lsn); Assert(summary_end_lsn >= WalSummarizerCtl->summarized_lsn);
WalSummarizerCtl->pending_lsn = summary_end_lsn; WalSummarizerCtl->pending_lsn = summary_end_lsn;
LWLockRelease(WALSummarizerLock); LWLockRelease(WALSummarizerLock);
@ -1460,17 +1472,11 @@ summarizer_read_local_xlog_page(XLogReaderState *state,
* Allow reads up to exactly the switch point. * Allow reads up to exactly the switch point.
* *
* It's possible that this will cause read_upto to move * It's possible that this will cause read_upto to move
* backwards, because walreceiver might have read a * backwards, because we might have been promoted before
* partial record and flushed it to disk, and we'd view * reaching the end of the previous timeline. In that case,
* that data as safe to read. However, the * the next loop iteration will likely conclude that we've
* XLOG_END_OF_RECOVERY record will be written at the end * reached end of WAL.
* of the last complete WAL record, not at the end of the
* WAL that we've flushed to disk.
*
* So switchpoint < private->read_upto is possible here,
* but switchpoint < state->EndRecPtr should not be.
*/ */
Assert(switchpoint >= state->EndRecPtr);
private_data->read_upto = switchpoint; private_data->read_upto = switchpoint;
/* Debugging output. */ /* Debugging output. */