Horton

Re: improve performance of pg_dump with many sequences

Nathan Bossart <nathandbossart@gmail.com>

View thread View on postgr.es
From: Nathan Bossart <nathandbossart@gmail.com>
To: Tom Lane <tgl@sss.pgh.pa.us>
Cc: Michael Paquier <michael@paquier.xyz>, Euler Taveira <euler@eulerto.com>, pgsql-hackers <pgsql-hackers@postgresql.org>
Date: 2026-01-08T19:34:28Z
Lists: pgsql-hackers

Attachments

On Thu, Jan 08, 2026 at 01:19:39PM -0500, Tom Lane wrote:
> One nitpicky point is that try_sequence_open() will still error out
> if it is given an OID that is a non-sequence relation.  I think it'd
> be more desirable for it to close the relation again and return NULL.
> That's probably insignificant for pg_dump's usage, because we could
> only hit the case with very improbable OID wraparound timing.  But
> I think our experience with catalog-inspection functions similar to
> pg_get_sequence_data is that it's usually better to return NULL than
> throw an error.

Hm.  That makes sense, but both try_table_open and try_index_open error for
wrong relkinds.  I could change all of the try_*_open functions to return
NULL in that case, or I could just open-code the relkind check in
pg_get_sequence_data after try_relation_open (and have it return NULL for
non-sequences).  I'm leaning towards the latter, if for no other reason
than it might be slightly nicer for back-patching (e.g., smaller, no new
extern functions).

-- 
nathan

Commits

  1. pg_dump: Fix gathering of sequence information.

  2. Improve performance of dumpSequence().

  3. Improve performance of dumpSequenceData().

  4. Introduce pg_sequence_read_tuple().

  5. Parse sequence type and integer metadata in dumpSequence().