Hello Michael, Kyotaro, Please find attached the patch fixing the problem and the updated TAP test that addresses Nit.
-- Regards, -- Alexander Kukushkin
042_no_contrecord_switch.pl
Description: Perl program
diff --git a/src/backend/access/transam/xlogrecovery.c b/src/backend/access/transam/xlogrecovery.c index 853b540945..f98ad350a9 100644 --- a/src/backend/access/transam/xlogrecovery.c +++ b/src/backend/access/transam/xlogrecovery.c @@ -3301,6 +3301,7 @@ XLogPageRead(XLogReaderState *xlogreader, XLogRecPtr targetPagePtr, int reqLen, uint32 targetPageOff; XLogSegNo targetSegNo PG_USED_FOR_ASSERTS_ONLY; int r; + bool timelineChanged = false; XLByteToSeg(targetPagePtr, targetSegNo, wal_segment_size); targetPageOff = XLogSegmentOffset(targetPagePtr, wal_segment_size); @@ -3425,6 +3426,16 @@ retry: Assert(targetPageOff == readOff); Assert(reqLen <= readLen); + /* + * Check if timeline changed while we are trying to read the page. If a + * read page is still invalid after timeline change we want to return from + * the function with the XLREAD_FAIL error instead or retrying, because we + * might be trying to decode record that spans multiple pages. If under + * this condition other replica was promoted and switched to the new WAL + * segment, than the page we are trying to read might never be written. + */ + timelineChanged = xlogreader->seg.ws_tli != curFileTLI; + xlogreader->seg.ws_tli = curFileTLI; /* @@ -3494,8 +3505,12 @@ next_record_is_invalid: readLen = 0; readSource = XLOG_FROM_ANY; - /* In standby-mode, keep trying */ - if (StandbyMode) + /* + * In standby-mode, keep trying, but only if we know that timeline didn't + * change, because after timeline change the new primary might have + * overwrote previous parts of the record that spans multiple pages. + */ + if (StandbyMode && !timelineChanged) goto retry; else return XLREAD_FAIL;
