Re: pgsql: Fix deadlock hazard in CREATE INDEX CONCURRENTLY

Tom Lane wrote:
> Alvaro Herrera <alvherre@alvh.no-ip.org> writes:
> > Tom Lane wrote:
> >> Meh.  I'd rather have the more stable test going forward; I think
> >> alternate expected-files too easily hide unexpected behavior.  We could
> >> try leaving 9.4/9.5 alone and see if it's true that it doesn't fail
> >> there.  If not, I wouldn't mind losing the test in those branches
> >> --- it's mainly intended to catch future breakage, after all.
> 
> > Makes sense.  Pushed to 9.6 and up.
> 
> Some of the buildfarm machines still don't like this.  It looks
> like the buildfarm script is only capturing the postmaster log
> and not regression.diffs, making it hard to diagnose :-(

Actually three of these recent failures are showing the diff:
https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=protosciurus&dt=2018-01-03%2018%3A39%3A46
https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=castoroides&dt=2018-01-03%2017%3A31%3A24
https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=woodlouse&dt=2018-01-03%2015%3A42%3A03

It proves me wrong about the ordering in which the steps return
completion being consistent:

***************
*** 20,24 ****
  unlck          
  
  t              
- step s1i: <... completed>
  step s2i: <... completed>
--- 20,24 ----
  unlck          
  
  t              
  step s2i: <... completed>
+ step s1i: <... completed>

Again this could be solved by just including an alternate file, or we
could go a bit further and report all completed steps in a single line
rather than each in its own line.  This would require patching
isolationtester back to 9.6, but it should be a small fix ...  Will look
into it after pushing another patch.

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