Incorrect CHUNKHDRSZ in nodeAgg.c

David Rowley <dgrowleyml@gmail.com>

From: David Rowley <dgrowleyml@gmail.com>
To: PostgreSQL Developers <pgsql-hackers@lists.postgresql.org>
Date: 2025-01-01T22:55:32Z
Lists: pgsql-hackers

Attachments

While reading nodeAgg.c, I noticed code that uses CHUNKHDRSZ to help
figure out how much memory a tuple uses so the code knows when to
spill to disk.  CHUNKHDRSZ is currently set to 16 bytes, which was
fine when that code was added, but it's a few years out-of-date since
c6e0fe1f2 in 2022.

The attached adjusts the 16 to sizeof(MemoryChunk), which is normally
8, but 16 in assert builds.

The memory accounting should be more accurate with the patch, however,
that accounting doesn't account for blocks that the chunks are on. For
that reason, I'm thinking of not backpatching this as it will make
hash aggregates use a bit more memory than unpatched before spilling,
albeit, for most cases, closer to the hash_mem limit, which is when
the spilling should be happening.

Sound ok?

David

Commits

  1. HashAgg: use Bump allocator for hash TupleHashTable entries.

  2. Fix outdated CHUNKHDRSZ value in nodeAgg.c