Justin Pryzby <pryzby@telsasoft.com> writes:
> On Sat, Apr 15, 2023 at 11:33:58AM +1200, David Rowley wrote:
>> Any chance you could try and come up with a minimal reproducer?
> Try this
Thanks. I see the problem: finalize_aggregate is no longer forcing
a R/W expanded datum returned by the finalfn into R/O form. If
we re-use the aggregate result in multiple places, as this query
does, then the first use can clobber the value for later uses.
(The commit message specifically mentions this concern, so I wonder
how we failed to actually do it :-()
A minimal fix would be to force to R/O before returning from
finalize_aggregate, but I wonder if we should do it later.
By the by, I couldn't help noticing that ExecAggTransReparent
completely fails to do what its name promises it should do, ie
reparent a R/W datum into the proper context instead of physically
copying it. That looks suspiciously like something that got broken
during some other refactoring somewhere along the line. That'd be a
performance bug not a correctness bug, but it should be looked into.
regards, tom lane
