On Wed, Mar 14, 2018 at 12:29 PM, Tom Lane <t...@sss.pgh.pa.us> wrote: > Thomas Munro <thomas.mu...@enterprisedb.com> writes: >> There is a fundamental and complicated estimation problem lurking here >> of course and I'm not sure what to think about that yet. Maybe there >> is a very simple fix for this particular problem: > > Ah, I see you thought of the same hack I did. > > I think this may actually be a good fix, and here's the reason: this plan > is in fact being driven entirely off planner default estimates, because > we don't have any estimation code that knows what to do with > "wholerowvar *= wholerowvar". I'm suspicious that we could drop the > preceding ANALYZE as being a waste of cycles, except maybe it's finding > out the number of rows for us. In any case, LIMIT 1 is only a good idea > to the extent that the planner knows what it's doing, and this is an > example where it demonstrably doesn't and won't any time soon.
Hmm. I wonder if the ANALYZE might have been needed to avoid the
nested loop plan at some point in history.
Here's a patch to remove LIMIT 1, which fixes the plan for Jeff's test
scenario and some smaller and larger examples I tried. The query is
already executed with SPI_execute(..., 1) so it'll give up after one
row anyway. The regression test includes a case that causes a row to
be produced here and that's passing ('ERROR: new data for
materialized view "mvtest_mv" contains duplicate rows without any null
columns').
--
Thomas Munro
http://www.enterprisedb.com
0001-Fix-performance-regression-in-REFRESH-MATERIALIZED-V.patch
Description: Binary data
