On Thu, Oct 17, 2024 at 9:59 AM Shayon Mukherjee <shayonj@gmail.com> wrote:
> My take away from whether or not an in-place update is needed on pg_index [1]
>
> - It’s unclear to me why it’s needed.
> - I understand the xmin would get incremented when using CatalogTupleUpdate to update indisenabled.
> - However, I haven’t been able to replicate any odd behavior locally or CI.
> - FWIW - REINDEX CONCURRENTLY (via index_swap), index_constraint_create and few other places perform
CatalogTupleUpdateto update the relevant attributes as well.
>
> Based on the above summary and after my testing I would like to propose a v3 of the patch. The only difference
betweenthis and v1 [2] is that the update of pg_index row happens via CatalogTupleUpdate.
>
> [1] https://www.postgresql.org/message-id/20180618215635.m5vrnxdxhxytvmcm@alap3.anarazel.de
> [2] https://www.postgresql.org/message-id/EF2313B8-A017-4869-9B7F-A24EDD8795DE%40gmail.com
In-place updates are generally bad news, so I think this patch
shouldn't use them. However, someone will need to investigate whether
that breaks the indcheckxmin thing that Andres mentions in your [1],
and if it does, figure out what to do about it. Creating a test case
to show the breakage would probably be a good first step, to frame the
discussion.
--
Robert Haas
EDB: http://www.enterprisedb.com
