Re: BTScanOpaqueData size slows down tests

Hi,

On 2025-04-03 11:26:08 +1300, David Rowley wrote:
> On Thu, 3 Apr 2025 at 04:21, Andres Freund <andres(at)anarazel(dot)de> wrote:
> > I was mildly
> > surprised to see how expensive the new compact attribute checks are.
>
> Is this a fairly deform-heavy workload?

It was our regression tests, although I was playing around with combining
various groups in the schedule to get the tests to pass faster overall (on
average we just use 2.8 cores, which seems depressingly low).

I do see a about a 6-7% increase in test time with just the default schedule.

> I feel having something to check these are in sync is worthwhile. It did
> find bugs during the development of the patch.

I agree. But I wonder if we couldn't make it a bit smarter than right
now. Afaict we'll recheck it over and over, which does seem a bit extreme.

When running queries that deform a lot, the overhead is rather substantial.

With pgbench of f <(echo 'SELECT sum(pronargs) FROM pg_proc;') I see 2155 TPS
with the call to verify_compact_attribute() commented out and 827 TPS without.

Greetings,

Andres Freund