Обсуждение: Autovac cancellation is broken in v14

Hi,

On 2020-08-27 16:20:30 -0400, Jeff Janes wrote:
> On Thu, Aug 27, 2020 at 3:10 PM Jeff Janes <jeff.janes@gmail.com> wrote:
> 
> > If I create a large table with "CREATE TABLE ... AS SELECT ... from
> > generate_series(1,3e7)" with no explicit transactions, then once it is done
> > I wait for autovac to kick in, then when I try to build an index on that
> > table (or drop the table) the autovac doesn't go away on its own.
> >
> 
> After a bit more poking at this, I think we are checking if we ourselves
> are an autovac process, not doing the intended check of whether the other
> guy is one.

Ugh, good catch.


> Where would be a good spot to add a regression test for this?
> "isolation_regression" ?

I'm not immediately sure how we could write a good test for this,
particularly not in the isolation tests. We'd basically have to make
sure that a table needs autovacuuming, then sleep for long enough for
autovacuum to have come around, and block autovacuum from making
progress. That latter is doable by holding a pin on a page it needs to
freeze, e.g. using a cursor.  I suspect all of that would at least
require a TAP test, and might still be too fragile.

Other ideas?

Regards,

Andres

Hi,

On 2020-08-27 14:35:06 -0700, Andres Freund wrote:
> On 2020-08-27 16:20:30 -0400, Jeff Janes wrote:
> > On Thu, Aug 27, 2020 at 3:10 PM Jeff Janes <jeff.janes@gmail.com> wrote:
> > 
> > > If I create a large table with "CREATE TABLE ... AS SELECT ... from
> > > generate_series(1,3e7)" with no explicit transactions, then once it is done
> > > I wait for autovac to kick in, then when I try to build an index on that
> > > table (or drop the table) the autovac doesn't go away on its own.
> > >
> > 
> > After a bit more poking at this, I think we are checking if we ourselves
> > are an autovac process, not doing the intended check of whether the other
> > guy is one.
> 
> Ugh, good catch.

Pushed the fix.


> > Where would be a good spot to add a regression test for this?
> > "isolation_regression" ?
> 
> I'm not immediately sure how we could write a good test for this,
> particularly not in the isolation tests. We'd basically have to make
> sure that a table needs autovacuuming, then sleep for long enough for
> autovacuum to have come around, and block autovacuum from making
> progress. That latter is doable by holding a pin on a page it needs to
> freeze, e.g. using a cursor.  I suspect all of that would at least
> require a TAP test, and might still be too fragile.

Perhaps the easiest way for this would be to have an option to have
manual VACUUMs be interruptible by other backends. That seems like a
useful option anyway? I'll start a new thread.

Greetings,

Andres Freund