On Mon, Jul 08, 2019 at 06:49:44PM +1200, Thomas Munro wrote:
> "simple" has 20k rows and "extremely_skewed" has 20k rows but the
> planner thinks it only has 2. So this going to take O(n^2) time and n
> is 20k. Not sure what to do about that. Maybe "join_hash" should be
> skipped for the -h (= no hash joins please) case?
Ah, thanks. Yes that's going to take a while :)
Well, another thing I'd like to think about is if there is any point
to keep regressplans.sh and the relevant options in postgres at this
stage. From the top of the file one can read that:
# This script runs the Postgres regression tests with all useful combinations
# of the backend options that disable various query plan types. If the
# results are not all the same, it may indicate a bug in a particular
# plan type, or perhaps just a regression test whose results aren't fully
# determinate (eg, due to lack of an ORDER BY keyword).
However if you run any option with make check, then in all runs there
are tests failing. We can improve the situation for some of them with
ORDER BY queries by looking at the query outputs, but some EXPLAIN
queries are sensitive to that, and the history around regressplans.sh
does not play in favor of it (some people really use it?). If you
look at the latest commits for it, it has not been really touched in
19 years.
So I would be rather in favor in nuking it.
--
Michael
