Обсуждение: [PATCH] Add TOAST support for several system tables
Hi, hackers
ACL lists in tables may potentially be large in size. I suggest adding TOAST support for system tables, namely pg_class, pg_attribute and pg_largeobject_metadata, for they include ACL columns.
In commit 96cdeae0 these tables were missed because of possible conflicts with VACUUM FULL and problem with seeing a non-empty new cluster by pg_upgrade. Given patch fixes both expected bugs. Also patch adds a new regress test for checking TOAST properly working with ACL attributes. Suggested feature has been used in Postgres Pro Enterprise for a long time, and it helps with large ACL's.
The VACUUM FULL bug is detailed here:
Looking forward to your comments
Вложения
=?UTF-8?B?U29maWEgS29waWtvdmE=?= <s.kopikova@postgrespro.ru> writes:
> ACL lists in tables may potentially be large in size. I suggest adding TOAST support for system tables, namely
pg_class,pg_attribute and pg_largeobject_metadata, for they include ACL columns.
TBH, the idea of adding a toast table to pg_class scares the daylights
out of me. I do not for one minute believe that you've fixed every
problem that will cause, nor do I think "allow wider ACLs" is a
compelling enough reason to take the risk.
I wonder if it'd be practical to move the ACLs for relations
and attributes into some new table, indexed like pg_depend or
pg_description, so that we could dodge that whole problem.
pg_attrdef is a prototype for how this could work.
(Obviously, once we had such a table the ACLs for other things
could be put in it, but I'm not sure that the ensuing breakage
would be justified for other sorts of objects.)
regards, tom lane
On 2022-02-28 18:08:48 -0500, Tom Lane wrote: > =?UTF-8?B?U29maWEgS29waWtvdmE=?= <s.kopikova@postgrespro.ru> writes: > > ACL lists in tables may potentially be large in size. I suggest adding TOAST support for system tables, namely pg_class,pg_attribute and pg_largeobject_metadata, for they include ACL columns. > > TBH, the idea of adding a toast table to pg_class scares the daylights > out of me. I do not for one minute believe that you've fixed every > problem that will cause, nor do I think "allow wider ACLs" is a > compelling enough reason to take the risk. > > I wonder if it'd be practical to move the ACLs for relations > and attributes into some new table, indexed like pg_depend or > pg_description, so that we could dodge that whole problem. > pg_attrdef is a prototype for how this could work. > > (Obviously, once we had such a table the ACLs for other things > could be put in it, but I'm not sure that the ensuing breakage > would be justified for other sorts of objects.) As there has been no progress since this email, I'm marking this CF entry as returned with feedback for now. - Andres