Horton

Re: Faster "SET search_path"

Jeff Davis <pgsql@j-davis.com>

View thread View on postgr.es
From: Jeff Davis <pgsql@j-davis.com>
To: Noah Misch <noah@leadboat.com>
Cc: Nathan Bossart <nathandbossart@gmail.com>, Isaac Morland <isaac.morland@gmail.com>, pgsql-hackers@postgresql.org
Date: 2024-07-08T23:39:21Z
Lists: pgsql-hackers

Attachments

On Sun, 2024-06-30 at 15:30 -0700, Noah Misch wrote:
> You're caching the result of object_aclcheck(NamespaceRelationId,
> ...), so
> pg_auth_members changes

Thank you for the report.

Question: to check for changes to pg_auth_members, it seems like either
AUTHMEMROLEMEM or AUTHMEMMEMROLE work, and to use both would be
redundant. Am I missing something, or should I just pick one
arbitrarily (or by some convention)?

>  and pg_database changes also need to invalidate this
> cache.  (pg_database affects the ACL_CREATE_TEMP case in
> pg_namespace_aclmask_ext()

I am having trouble finding an actual problem with ACL_CREATE_TEMP.
search_path ACL checks are normally bypassed for the temp namespace, so
it can only happen when the actual temp namespace name (e.g.
pg_temp_NNN) is explicitly included. In that case, the mask is
ACL_USAGE, so the two branches in pg_namespace_aclmask_ext() are
equivalent, right?

This question is not terribly important for the fix, because
invalidating for each pg_database change is still necessary as you
point out in here:

>  and affects ROLE_PG_DATABASE_OWNER membership.)

Another good catch, thank you.

Patch attached. Contains a bit of cleanup and is intended for 17+18.

Regards,
	Jeff Davis

Commits

  1. Fix missing invalidations for search_path cache.

  2. Optimize SearchPathCache by saving the last entry.

  3. Optimize check_search_path() by using SearchPathCache.

  4. Be more paranoid about OOM in search_path cache.

  5. Add cache for recomputeNamespacePath().

  6. Transform proconfig for faster execution.