Re: Fail Fast In CTAS/CMV If Relation Already Exists To Avoid Unnecessary Rewrite, Planning Costs

On Thu, Dec 17, 2020 at 03:06:59PM +0530, Bharath Rupireddy wrote:
> The behavior of the ctas/cmv, in case the relation already exists is as
> shown in [1]. The things that have been changed with the patch are: 1) In
> any case we do not rewrite or plan the select part if the relation already
> exists 2) For explain ctas/cmv (without analyze), now the relation
> existence is checked early and the error is thrown as highlighted in [1].
>
> With patch, there is no behavioral change(from that of master branch) in
> explain analyze ctas/cmv with if-not-exists i.e. error is thrown not the
> notice.
>
> Thoughts?

HEAD is already a mixed bad of behaviors, and the set of results you
are presenting here is giving a similar impression. It brings in some
sanity by just ignoring the effects of the IF NOT EXISTS clause all
the time still that's not consistent with the queries not using
EXPLAIN. Hmm. Looking for similar behaviors, I can see one case in
select_into.sql where we just never execute the plan when using WITH
NO DATA but still show the plan, meaning that the query gets planned
but it just gets marked as "(never executed)" if attempting to use
ANALYZE. There may be use cases for that as the user directly asked
directly for an EXPLAIN.

Note: the patch needs tests for all the patterns you would like to
stress. This way it is easier to follow the patterns that are
changing with your patch and compare them with the HEAD behavior (like
looking at the diffs with the tests of the patch, but without the
diffs in src/backend/).
--
Michael