On 2013-08-30 17:23:51 -0400, Tom Lane wrote:
> Andres Freund <andres@2ndquadrant.com> writes:
> > No need, found the bug. And I think can build a testcase myself.
>
> > ExecReScanMergeAppend resets ms_initialized, but doesn't clear the
> > binaryheap. Thus no new elements fit.
>
> Sounds like a bug all right, but I'm not convinced it explains Terje's
> report. The thing this theory doesn't explain is why would Terje be
> having trouble reproducing the failure? Seems like re-running the same
> query ought to produce the same failure.
Even better explanation: The merge append is some steps below a LockRows
node, so the number of rescans might depend on the concurrency because
we'll do the EvalPlanQual dance on a concurrent row update.
Terje, do you use read committed or repeatable read/serializable?
Greetings,
Andres Freund
--
Andres Freund http://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Training & Services