On Thu, Nov 30, 2017 at 4:13 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
> Andres Freund <andres@anarazel.de> writes:
>> Add some regression tests that exercise hash join code.
>
> At least one buildfarm member doesn't like this ...
$$); initially_multibatch | increased_batches ----------------------+-------------------
! t | f (1 row)
rollback to settings;
--- 6002,6008 ---- $$); initially_multibatch | increased_batches ----------------------+-------------------
! | (1 row)
Hmm. aholehole didn't give me the hash join EXPLAIN ANALYZE output.
All other animals are fine so far. Gah. My guess is that this is the
following unlikely sequence:
1. Worker launched.
2. Leader descheduled.
3. Worker runs whole join.
4. Leader awakes from slumber, begins running plan and tries to scan
outer relation, sees EOF, takes empty-outer optimisation and skips
building hash table.
5. Explain has no data.
It's arguably a bug that EXPLAIN ANALYZE can fail to give you what you
asked for because of details of timing and I have a patch in
development to fix that but it's not yet baked. Hmm. Wondering if
there is a quick way to avoid that case in the meantime...
--
Thomas Munro
http://www.enterprisedb.com