pgsql: Attempt to stabilize flapping regression test - Mailing list pgsql-committers

From David Rowley
Subject pgsql: Attempt to stabilize flapping regression test
Date
Msg-id E1ranY0-006RsA-Tk@gemulon.postgresql.org
Whole thread Raw
List pgsql-committers
Attempt to stabilize flapping regression test

Per buildfarm animal mylodon, the plan for this test was sometimes
swapping the join order for tenk1 and tenk2.  Given that add_path() has
no code that would cause this fluctuation when given paths with consistent
costs, this indicates that the costs must be fluctuating in some runs.
The only proven reason I've seen where that could happen was slight
variations in pg_class.relpages for some tables.  This was demonstrated to
be true by f03a9ca43 and related discussion.  Manually adjusting tenk2's
pg_class.relpages by subtracting just 1 page does cause the plan to change
for this test.

Here we've not gone to the same lengths to prove that's what's going on
in this case.  Proving that does not seem worth the time.  Let's just
shrink one side of the join so the additional cost of the swapped join
order is sufficiently different that if the relpages estimate is off a few
pages that the planner still shouldn't swap the join order.

Reported-by: Thomas Munro
Author: Andy Fan, David Rowley
Discussion: https://postgr.es/m/CA+hUKGLqC-NobKYfjxNM3Gexv9OJ-Fhvy9bugUcXsZjTqH7W=Q@mail.gmail.com

Branch
------
master

Details
-------
https://git.postgresql.org/pg/commitdiff/1fe66680c09b6cc1ed20236c84f0913a7b786bbc

Modified Files
--------------
src/test/regress/expected/union.out | 5 +++--
src/test/regress/sql/union.sql      | 2 +-
2 files changed, 4 insertions(+), 3 deletions(-)


pgsql-committers by date:

Previous
From: Alexander Korotkov
Date:
Subject: pgsql: Followup fixes for transaction_timeout
Next
From: Amit Kapila
Date:
Subject: Re: pgsql: Allow upgrades to preserve the full subscription's state.