On 2015-05-08 14:30:46 -0400, Tom Lane wrote:
> Maybe just hold an open transaction in another session while you do what
> the regression test does? I think this is probably not a matter of CCA
> per se but just timing. It's unfortunate that the test in question is
> run serially without other transactions occurring concurrently, as that
> would likely have exposed the problem far sooner.
I think Peter (on IM) just found a more likely explanation than mine.
index_close(idxRel, NoLock); heap_close(relation, NoLock); candidates = lappend_oid(candidates,
idxForm->indexrelid);
...
Yes. idxForm points into idxRel->rd_index.
He's working on a fix for both this and removal of the still unnecessary
indcheckxmin check.
Greetings,
Andres Freund