Re: Getting "cache lookup failed for aggregate" error

On Tue, Jun 24, 2014 at 4:25 PM, Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> wrote:

> Patrick Krecker <patrick(at)judicata(dot)com> writes:
> > Hello everyone -- We received a strange error today on our production
> write
> > master. During a routine maintenance script, we got the following error:
>
> > "ERROR: cache lookup failed for aggregate 5953992"
>
> > I think I found the culprit. We have a cron script that (among other
> > things) recreates the aggregate function array_cat_aggregate()once every
> > minute.
>
> Um. Why's it do that?
>

Well, it's admittedly a hack. The cron script uses the function just after
creating it. It was an easy way of guaranteeing that the database will have
the function when it's necessary. However, I suppose that assumption is
wrong because you have confirmed that aggregate functions are not part of
the snapshot provided by beginning a transaction.

>
> > My guess is that a) transactions do not guarantee a snapshot of
> > custom functions and b) we got unlucky and the aggregate was deleted
> during
> > the execution of the UPDATE.
>
> Yeah, there's no guarantee that dropping the aggregate wouldn't leave a
> window for this type of failure.
>
> 9.4 might be a little better about this because it doesn't use SnapshotNow
> for catalog fetches anymore, but I think you'd still be at some risk.
>
> > Here is the query that produced the error:
>
> It'd be more interesting to see what the cron script was doing to the
> aggregate definition.
>

FWIW the SQL is

DROP AGGREGATE IF EXISTS array_cat_aggregate(anyarray);
CREATE AGGREGATE array_cat_aggregate(anyarray) (
SFUNC = array_cat,
STYPE = anyarray,
INITCOND = '{}'
);

Followed by the other statement given in my previous email. But, I think
you've thoroughly answered by question. Thanks!

>
> regards, tom lane
>