Re: Vacuum ERRORs out considering freezing dead tuples from before OldestXmin

On Wed, Jul 24, 2024 at 5:40 AM Masahiko Sawada <sawada(dot)mshk(at)gmail(dot)com> wrote:

> Without MEMORY_CONTEXT_CHECK, if size is 16 bytes, required_size is
> also 16 bytes as it's already 8-byte aligned and Bump_CHUNKHDRSZ is 0.
> On the other hand with MEMORY_CONTEXT_CHECK, the requied_size is
> bumped to 40 bytes as chunk_size is 24 bytes and Bump_CHUNKHDRSZ is 16
> bytes. Therefore, with MEMORY_CONTEXT_CHECK, we allocate more memory
> and use more Bump memory blocks, resulting in filling up TidStore in
> the test cases. We can easily reproduce this test failure with
> PostgreSQL server built without --enable-cassert. It seems that
> copperhead is the sole BF animal that doesn't use --enable-cassert but
> runs recovery-check.

It seems we could force the bitmaps to be larger, and also reduce the
number of updated tuples by updating only the last few tuples (say
5-10) by looking at the ctid's offset. This requires some trickery,
but I believe I've done it in the past by casting to text and
extracting with a regex. (I'm assuming the number of tuples updated is
more important than the number of tuples inserted on a newly created
table.)

As for lowering the limit, we've experimented with 256kB here:

https://www.postgresql.org/message-id/CANWCAZZUTvZ3LsYpauYQVzcEZXZ7Qe+9ntnHgYZDTWxPuL++zA@mail.gmail.com

As I mention there, going lower than that would need a small amount of
reorganization in the radix tree. Not difficult -- the thing I'm
concerned about is that we'd likely need to document a separate
minimum for DSA, since that behaves strangely with 256kB and might not
work at all lower than that.