Re: reindex concurrently and two toast indexes

On Thu, Feb 27, 2020 at 09:07:35AM +0100, Julien Rouhaud wrote:
> While looking at it, I see that invalid indexes seem to leaked when the table
> is dropped, with no way to get rid of them:
>
> Shouldn't DROP TABLE be fixed to also drop invalid indexes?

Hmm. The problem here is that I think that we don't have the correct
correct interface to handle the dependency switching between the old
and new indexes from the start, and 68ac9cf made things better in some
aspects (like non-cancellation and old index drop) but not in others
(like yours, or even a column drop). changeDependenciesOf/On() have
been added especially for REINDEX CONCURRENTLY, but they are not
actually able to handle the case we want them to handle: do a switch
for both relations within the same scan. It is possible to use three
times the existing routines with a couple of CCIs in-between and what
I would call a fake placeholder OID to switch all the records cleanly,
but it would be actually cleaner to do a single scan of pg_depend and
switch the dependencies of both objects at once.

Attached is a draft patch to take care of that problem for HEAD. It
still needs a lot of polishing (variable names are not actually old
or new anymore, etc.) but that's enough to show the idea. If a version
reaches PG12, we would need to keep around the past routines to avoid
an ABI breakage, even if I doubt there are callers of it, but who
knows..
--
Michael