On Mon, Jun 13, 2016 at 1:45 PM, Peter Geoghegan <pg@heroku.com> wrote:
> Ideally, you'd be able to build Postgres yourself, with the attached
> patch applied (the 9.5 release branch should be used, of course). This
> will provoke a hard crash in the event of hitting this problem. You
> can set things up such that this produces a core dump for further
> debugging here (for the directly affected backend):
>
> https://wiki.postgresql.org/wiki/Getting_a_stack_trace_of_a_running_PostgreSQL_backend_on_Linux/BSD
>
> I'm particularly interested in seeing a backtrace from the core dump.
> That should make the general nature of the problem clear.
For the record, it's not clear that this is an independent bug to the
old bug that Andres described on the "HeapTupleSatisfiesToast()
busted?" thread:
https://www.postgresql.org/message-id/20160510070540.lyvintwdiw4ogmnw@alap3.anarazel.de
It does not seem at all unlikely that this is just another way of
hitting the problem. It might be that UPSERT makes it more likely that
the issue is hit in practice, since I've now seen two of these
reports.
--
Peter Geoghegan