On Thu, Dec 12, 2013 at 11:00 AM, Kyotaro HORIGUCHI
<horiguchi.kyotaro@lab.ntt.co.jp> wrote:
> Hello, we happened to see server crash on archive recovery under
> some condition.
>
> After TLI was incremented, there should be the case that the WAL
> file for older timeline is archived but not for that of the same
> segment id but for newer timeline. Archive recovery should fail
> for the case with PANIC error like follows,
>
> | PANIC: record with zero length at 0/1820D40
>
> Replay script is attached. This issue occured for 9.4dev, 9.3.2,
> and not for 9.2.6 and 9.1.11. The latter search pg_xlog for the
> TLI before trying archive for older TLIs.
>
> This occurrs during fetching checkpoint redo record in archive
> recovery.
>
>> if (checkPoint.redo < RecPtr)
>> {
>> /* back up to find the record */
>> record = ReadRecord(xlogreader, checkPoint.redo, PANIC, false);
>
> And this is caused by that the segment file for older timeline in
> archive directory is preferred to that for newer timeline in
> pg_xlog.
>
> Looking into pg_xlog before trying the older TLIs in archive like
> 9.2- fixes this issue. The attached patch is one possible
> solution for 9.4dev.
>
> Attached files are,
>
> - recvtest.sh: Replay script. Step 1 and 2 makes the condition
> and step 3 causes the issue.
>
> - archrecvfix_20131212.patch: The patch fixes the issue. Archive
> recovery reads pg_xlog before trying older TLI in archive
> similarly to 9.1- by this patch.
>
> regards,
>
Horiguchi-san,
Wonder if the following commit in 9.2 branch was to address a
similar/same problem?
http://git.postgresql.org/gitweb/?p=postgresql.git;a=commit;h=4aed94f1660fb55bc825bf7f3135379dab28eb55
--
Amit Langote