Re: VAC FULL/CLUSTER on system catalogs is prone to deadlock

Tom Lane wrote:
> 2. In VAC FULL and CLUSTER, tell index rebuild not to wait around for
> INSERT_IN_PROGRESS or DELETE_IN_PROGRESS tuples. I believe that the
> price of this would be not re-verifying the integrity of unique indexes.
> Which is kind of annoying, but then again rechecking data consistency
> is not the purpose of these commands. It would not be too unreasonable
> to tell people to use REINDEX if they suspect the validity of an index
> --- that takes a significantly weaker lock and wouldn't be nearly as
> big a problem to use on system catalogs.
>
> 3. Try to ameliorate the problem by making sure it's the VACUUM FULL
> that fails, and not the other process, when there's a deadlock.
>
> BTW, the reason we don't see this type of problem with the existing
> VAC FULL INPLACE logic is that it essentially silently adopts #2
> --- it suppresses uniqueness checks even when it's making new index
> insertions.
>
> At the moment #2 is seeming like the most attractive alternative;
> both of the others look like research projects. Comments?

Yea, #2. No one has complained about this behavior in the past, as I
remember.

--
Bruce Momjian <bruce(at)momjian(dot)us> http://momjian.us
EnterpriseDB http://enterprisedb.com

+ If your life is a hard drive, Christ can be your backup. +