On Mon, Nov 5, 2012 at 12:49 PM, Bruce Momjian <bruce@momjian.us> wrote:
> On Mon, Nov 5, 2012 at 03:30:32PM -0500, Tom Lane wrote:
>> Magnus Hagander <magnus@hagander.net> writes:
>> > On Mon, Nov 5, 2012 at 9:14 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
>> >> BTW, does pg_upgrade run pg_restore in --single-transaction mode?
>> >> That would probably make synchronous_commit moot, at least for that
>> >> step.
>>
>> > It doesn't use pg_restore at all - it uses the dump from pg_dumpall, which
>> > you can't reload with pg_restore.
>>
>> Sorry, I should've said psql --single-transaction. Although that isn't
>> going to work either given the presence of \connect commands in the
>> script. I wonder whether pg_dumpall ought to have some sort of "one
>> transaction per database please" option.
>
> pg_dumpall is already doing lots of gymnastics with SQL, and pg_upgrade
> splits the output file into db/user creation and object creation, so I
> am hesitant to add anything more in there.
>
> I was surprised by the scale of the performance improvement, but a
> simple table creation test confirmed that improvement, irregardless of
> pg_upgrade. Perhaps we should suggest synchronous_commit=off for
> pg_dumpall restores, particularly when using --schema-only.
Or have options for pg_dump and pg_restore to insert "set
synchronous_commit=off" into the SQL stream?
Cheers,
Jeff
