On Fri, Sep 30, 2016 at 5:33 PM, Konstantin Knizhnik
<k.knizhnik@postgrespro.ru> wrote:
> So the question is whether it is correct that ExecOnConflictUpdate tries to
> access and update tuple without holding lock on the buffer?
You're right -- this is a bug in Postgres.
I'm travelling from Ireland to the USA this weekend, but will work on
this early next week. I don't think it's a particularly tricky fix --
as you say, it is necessary to have at least a shared buffer lock to
call HeapTupleSatisfiesVisibility(), and we quite simply fail to
ensure that. We must have a shared buffer lock in the visibility-check
path for ON CONFLICT DO UPDATE where isolation level > READ COMMITTED
-- a buffer pin is not enough.
It also looks like the DO NOTHING variant is similarly affected, even
when the isolation level is READ COMMITTED. :-(
Thanks for the detailed report.
--
Peter Geoghegan