On Sat, Jan 3, 2015 at 12:42 PM, Peter Geoghegan <pg@heroku.com> wrote:
> There are probably additional factors that make the situation more
> complicated than it is for the ON CONFLICT patch, but it's clear to me
> that the mutual dependency that can be involved with two ordinary
> exclusion constraint inserters is reason enough for those to deadlock.
I looked at the code in more detail, and realized that there were old
bugs in the exclusion constraint related modifications. I attach a
delta patch that fixes them. This is a combined patch that is all that
is needed to apply on top of v1.8.vallock2.tar.gz [1] to have all
available bugfixes.
This unbreaks my previous exclusion constraint test case, as far as it
goes. I am still concerned about the risk of lock starvation/livelocks
here. But I must admit, having stressed the patch with this latest
fix, that it's possible that the real risks have been overestimated.
Maybe this is in the same category as the way L&Y's algorithm could
theoretically starve a session with a pagesplit that needs to move
right indefinitely. It's a theoretical risk that turns out to not
matter in practice, for reasons that lack a real theoretical
justification beyond the fact that sustained very bad luck - a
sustained "perfect storm" - is required for starvation to occur. OTOH,
maybe my OS scheduler doesn't have the characteristic that provoke a
the suspected bug. I really don't know, but I suppose that one or the
other concurrent inserters will tend to win the race often enough - a
draw may be a very rare thing.
[1] http://www.postgresql.org/message-id/CAM3SWZRg_hTrOL-6_wfe6_d_UcUYc28JfaPsFh_tra76GkkdNw@mail.gmail.com
--
Peter Geoghegan