I wrote:
> Looking back at your original discussion of the bug,
> http://archives.postgresql.org/pgsql-hackers/2007-06/msg00234.php
> I'm wondering why you chose option #3 rather than option #4?
> I still find the proposed patch a bit crufty.
In particular, it seems like a patch per #4 would be a one-liner:
*** src/backend/access/transam/xlog.c.orig Wed Sep 26 18:36:30 2007
--- src/backend/access/transam/xlog.c Thu Sep 27 12:20:56 2007
***************
*** 5092,5101 **** * * If we stopped short of the end of WAL during recovery, then we are * generating a
newtimeline and must assign it a unique new ID. * Otherwise, we can just extend the timeline we were in when we
ranout * of WAL. */
! if (needNewTimeLine) { ThisTimeLineID = findNewestTimeLine(recoveryTargetTLI) + 1;
ereport(LOG,
--- 5092,5103 ---- * * If we stopped short of the end of WAL during recovery, then we are * generating a
newtimeline and must assign it a unique new ID.
+ * We also force a new timeline when recovering from an archive, to avoid
+ * problems with trying to overwrite existing archived segments. * Otherwise, we can just extend the
timelinewe were in when we ran out * of WAL. */
! if (needNewTimeLine || (InArchiveRecovery && XLogArchivingActive())) { ThisTimeLineID =
findNewestTimeLine(recoveryTargetTLI)+ 1; ereport(LOG,
though I admit I've not tested this. The comments in
exitArchiveRecovery probably need adjustment too --- re-reading them,
it seems obvious that the current approach is broken by design, because
it *must* lead to an attempt to overwrite a previously archived version
of the last segment.
regards, tom lane