Re: pgsql: Fix deadlock hazard in CREATE INDEX CONCURRENTLY

Hello,

Andres Freund wrote:

> On 2018-01-03 02:20:14 +0000, Alvaro Herrera wrote:

> > Include an isolationtester spec that 8 out of 10 times reproduces the
> > deadlock with the unpatched code for me (Álvaro).
>
> The isolation test doesn't appear to be fully stable:
>
> https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=baiji&dt=2018-01-03%2003%3A00%3A01

Yeah, thanks for reporting that. It's annoying, but obvious in
hindsight -- the second process occasionally gets stuck waiting for the
first one to release its snapshot. I can reproduce this about half the
time, by adding a 100ms sleep in ProcessUtilitySlow right after the
DefineIndex call.

I can turn the isolation test to produce the output where s2 waits
(without the sleep) if I introduce a third session s3 to lock s2; then
releasing the lock in s3 causes both s1 and s2 to complete. I think the
completions would be reported always in that order because of the way
isolationtester verifies suspension.

However this means that the test will get removed in 9.4 and 9.5 because
isolationtester is not smart enough there.

I suppose the other option would be to add an alternate expected file
for the test.

--
Álvaro Herrera https://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services