Re: [COMMITTERS] pgsql: Sync pg_dump and pg_dumpall output

On Thu, Mar 23, 2017 at 1:10 AM, Stephen Frost <sfrost(at)snowman(dot)net> wrote:
> Andrew,
>
> * Andrew Dunstan (andrew(dot)dunstan(at)2ndquadrant(dot)com) wrote:
>> On 03/22/2017 11:39 AM, Stephen Frost wrote:
>> > * Andrew Dunstan (andrew(at)dunslane(dot)net) wrote:
>> >> Sync pg_dump and pg_dumpall output
>> > This probably should have adjusted all callers of pg_dump in the
>> > regression tests to use the --no-sync option, otherwise we'll end up
>> > spending possibly a good bit of time calling fsync() during the
>> > regression tests unnecessairly.
>>
>> All of them? The imnpact is not likely to be huge in most cases
>> (possibly different on Windows). On crake, the bin-check stage actually
>> took less time after the change than before, so I suspect that the
>> impact will be pretty small.
>
> Well, perhaps not all, but I'd think --no-sync would be better to have
> in most cases. We turn off fsync for all of the TAP tests and all
> initdb calls, so it seems like we should here too. Perhaps it won't be
> much overhead on an unloaded system, but not all of the buildfarm
> animals seem to be on unloaded systems, nor are they particularly fast
> in general or have fast i/o..

Agreed.

>> Still I agree that we should have tests for both cases.
>
> Perhaps, though if I recall correctly, we've basically had zero calls
> for fsync() until this. If we don't feel like we need to test that in
> the backend then it seems a bit silly to feel like we need it for
> pg_dump's regression coverage.
>
> That said, perhaps the right answer is to have a couple tests for both
> the backend and for pg_dump which do exercise the fsync-enabled paths.

And agreed.

The patch attached uses --no-sync in most places for pg_dump, except
in 4 places of 002_pg_dump.pl to stress as well the sync code path.
--
Michael