MIME-Version: 1.0
References: <CANqtF-oXKe0M=0QOih6H+sZRjE2BWAbkW_1+9nMEAMLxUJg5jA@mail.gmail.com>
In-Reply-To: <CANqtF-oXKe0M=0QOih6H+sZRjE2BWAbkW_1+9nMEAMLxUJg5jA@mail.gmail.com>
From: David Rowley <dgrowleyml@gmail.com>
Date: Tue, 10 Sep 2024 10:16:34 +1200
Message-ID: <CAApHDvoVL57C-pWtBO_1bvzTHRvSGN3BY_5us5F9mDbKdqG=YA@mail.gmail.com>
Subject: Re: Proposal to Enable/Disable Index using ALTER INDEX
To: Shayon Mukherjee <shayonj@gmail.com>
Cc: pgsql-hackers@lists.postgresql.org
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable
Archived-At: <https://www.postgresql.org/message-id/CAApHDvoVL57C-pWtBO_1bvzTHRvSGN3BY_5us5F9mDbKdqG%3DYA%40mail.gmail.com>
Precedence: bulk

On Tue, 10 Sept 2024 at 09:39, Shayon Mukherjee <shayonj@gmail.com> wrote:
> Adding and removing indexes is a common operation in PostgreSQL. On large=
r databases, however, these operations can be resource-intensive. When eval=
uating the performance impact of one or more indexes, dropping them might n=
ot be ideal since as a user you may want a quicker way to test their effect=
s without fully committing to removing & adding them back again. Which can =
be a time taking operation on larger tables.
>
> Proposal:
> I propose adding an ALTER INDEX command that allows for enabling or disab=
ling an index globally. This could look something like:
>
> ALTER INDEX index_name ENABLE;
> ALTER INDEX index_name DISABLE;
>
> A disabled index would still receive updates and enforce constraints as u=
sual but would not be used for queries. This allows users to assess whether=
 an index impacts query performance before deciding to drop it entirely.

I personally think having some way to alter an index to stop it from
being used in query plans would be very useful for the reasons you
mentioned.  I don't have any arguments against the syntax you've
proposed.  We'd certainly have to clearly document that constraints
are still enforced. Perhaps there is some other syntax which would
self-document slightly better. I just can't think of it right now.

> Implementation:
> To keep this simple, I suggest toggling the indisvalid flag in pg_index d=
uring the enable/disable operation.

That's not a good idea as it would allow ALTER INDEX ... ENABLE; to be
used to make valid a failed concurrently created index.  I think this
would need a new flag and everywhere in the planner would need to be
adjusted to ignore indexes when that flag is false.

> Additional Considerations:
> - Keeping the index up-to-date while it=E2=80=99s disabled seems preferab=
le, as it avoids the need to rebuild the index if it=E2=80=99s re-enabled l=
ater. The alternative would be dropping and rebuilding the index upon re-en=
abling, which I believe would introduce additional overhead in terms of app=
lication logic & complexity.

I think the primary use case here is to assist in dropping useless
indexes in a way that can very quickly be undone if the index is more
useful than thought. If you didn't keep the index up-to-date then that
would make the feature useless for that purpose.

If we get the skip scan feature for PG18, then there's likely going to
be lots of people with indexes that they might want to consider
removing after upgrading. Maybe this is a good time to consider this
feature as it possibly won't ever be more useful than it will be after
we get skip scans.

David