From: Tom Lane <tgl@sss.pgh.pa.us>
To: David Rowley <dgrowleyml@gmail.com>
cc: PostgreSQL Developers <pgsql-hackers@lists.postgresql.org>
Subject: Re: Incorrect CHUNKHDRSZ in nodeAgg.c
In-reply-to: <CAApHDvozg3jDN4oVYvtpA91P1DA50gdRZ5071SnPg_24oL5PbA@mail.gmail.com>
References: <CAApHDvpMpRQvsTqZo3FinXkgytwxwF8sCyZm83xDj-1s_hLe+w@mail.gmail.com> <151884.1735773536@sss.pgh.pa.us> <CAApHDvozg3jDN4oVYvtpA91P1DA50gdRZ5071SnPg_24oL5PbA@mail.gmail.com>
Comments: In-reply-to David Rowley <dgrowleyml@gmail.com>
	message dated "Thu, 02 Jan 2025 13:23:54 +1300"
MIME-Version: 1.0
Content-Type: text/plain; charset="us-ascii"
Content-ID: <160824.1735778003.1@sss.pgh.pa.us>
Date: Wed, 01 Jan 2025 19:33:23 -0500
Message-ID: <160825.1735778003@sss.pgh.pa.us>
Archived-At: <https://www.postgresql.org/message-id/160825.1735778003%40sss.pgh.pa.us>
Precedence: bulk

David Rowley <dgrowleyml@gmail.com> writes:
> On Thu, 2 Jan 2025 at 12:18, Tom Lane <tgl@sss.pgh.pa.us> wrote:
>> I thought for a bit about whether we shouldn't try to account for
>> palloc power-of-2-block-size overhead here.  That omission would
>> typically be a far larger error than the one you are fixing.  However,
>> given that the inputs to hash_agg_entry_size are only estimates,
>> I'm not sure that we can hope to do better than the current behavior.

> Likely the most correct way would be to use GetMemoryChunkSpace(), but
> there might be some additional overhead to consider there.

Nah, you've got the wrong mental model.  hash_agg_entry_size is
trying to predict the average hash entry size in advance of seeing
any actual data, so that we can estimate how many entries will fit
in work_mem.  By the time we can use GetMemoryChunkSpace on an
actual entry, it's too late for that.

>> Could we use a generation or even bump context?

> Bump wouldn't work due to the SH_FREE() in SH_GROW() when resizing the
> table.

Meh.  I guess we'd have to keep that structure in a context separate
from the tuples.  Might not be worth the trouble.

> I think what would be more interesting is seeing if we can store the
> TupleHashEntryData.firstTuple in a bump context.

Are you saying the same as above, or something different?

			regards, tom lane