On Mon, Jan 2, 2012 at 5:17 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
> Simon Riggs <simon@2ndQuadrant.com> writes:
>> Assuming that is the right approach, some timings
>
> Um ... timings of what?
Apologies for being terse, no problem to give a full explanation.
You suggested earlier that putting calls to GetSnapshotData() in place
of using SnapshotNow might increase the time taken to do catalog
scans. That matters most in places where stuff isn't cached, one such
place you suggested might be the dependency check code, so a test of
dropping thousands of functions might show up any bad regressions.
From what I can see, there are no bad performance regressions from
making SnapshotNow use MVCC.
We're benefiting in 9.2 from good reductions in GetSnapshotData()
contention and improved performance anyway, so it looks like a
reasonable balance.
I'm proposing that we apply the above patch (v4) to allow SnapshotNow
scans to return consistent, trustable results. That fixes some of the
corner case bugs we've seen and paves the way for me to re-enable the
lock reduction code.
--
Simon Riggs http://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Training & Services