Re: [HACKERS] GSoC 2017: Foreign Key Arrays

Dear All,

I have retested the patch on a windows build and it passes the regression
tests thanks to Justin's recommendations. Hopefully, it will pass CI too.

Changelog:
- v7 (compatible with current master 2021-3-12,
commit 02b5940dbea17d07a1dbcba3cbe113cc8b70f228)
* re-add failing regression test with fixes
* rebase patch

/Mark

On Thu, Mar 4, 2021 at 12:22 AM Mark Rofail <markm(dot)rofail(at)gmail(dot)com> wrote:

> Hello Justin,
>
> It doesn't just rebase: it also removes the test which was failing on
>> windows
>> CI:
>>
> My apologies, I didn’t include it in the changelog since this is not a
> code change, just wanted to see if any other test would fail on the windows
> CI
>
> I think the SELECT, when it works, is actually doing a seq scan and not
>> using
>> the index. On my PC, the index scan is used until an autovacuum/analyze
>> run,
>> after which it uses seqscan. I'm not sure how the other CIs all managed
>> to run
>> autovacuum between creating a table and running a query on it, though.
>
> This is genius! That explains it. I have been racking my brain for two
> weeks now and you figured it out.
>
> I guess you should first run the query with "explain (costs off)" to show
>> what
>> plan it's using, and add things like "SET enable_seqscan=off" as needed to
>> guarantee that everyone will use the same plan, regardless of minor cost
>> differences and vacuum timing.
>
> I think that will solve the test discrepancy.
>
> Honestly Justin, hats off!
>
> /Mark
>
>>