On Thu, Nov 19, 2015 at 7:19 AM, Thom Brown <thom@linux.com> wrote:
> On 27 March 2015 at 04:54, Kyotaro HORIGUCHI
> <horiguchi.kyotaro@lab.ntt.co.jp> wrote:
>> Hello,
>>
>> At Thu, 26 Mar 2015 18:50:24 +0100, Andres Freund <andres@2ndquadrant.com> wrote in
<20150326175024.GJ451@alap3.anarazel.de>
>>> I think the problem here is that the *primary* makes no such
>>> assumptions. Init forks are logged via stuff like
>>> smgrwrite(index->rd_smgr, INIT_FORKNUM, BTREE_METAPAGE,
>> ..
>>> i.e. the data is written out directly to disk, circumventing
>>> shared_buffers. It's pretty bad that we don't do the same on the
>>> standby. For master I think we should just add a bit to the XLOG_FPI
>>> record saying the data should be forced out to disk. I'm less sure
>>> what's to be done in the back branches. Flushing every HEAP_NEWPAGE
>>> record isn't really an option.
>>
>> The problem exists only for INIT_FORKNUM. So I suppose it is
>> enough to check forknum to decide whether to sync immediately.
>>
>> Specifically for this instance, syncing buffers of INIT_FORKNUM
>> at the end of XLOG_FPI block in xlog_redo fixed the problem.
>>
>> The another (ugly!) solution sould be syncing only buffers for
>> INIT_FORKNUM and is BM_DIRTY in ResetUnlogggedRelations(op =
>> UNLOGGED_RELATION_INIT). This is catching-all-at-once solution
>> though it is a kind of reversion of fast promotion. But buffers
>> to be synced here should be pretty few.
>
> This bug still exists.
Hmm. This probably should have been on the open items list. I didn't
pay too much attention this to this before because it seemed like
Andres and Michael were all over it.
--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company