I have just checked in a little test script that I've been using for a
while (since before 6.5) to pound on lesser-used paths in the optimizer.
It's called src/test/regress/regressplans.sh, and it just runs the
regular regression tests with different PGOPTIONS settings to force
varying plan type selections.
The reason I bring it up is that recently the thing has been failing
with backend messages "ERROR: out of free buffers: time to abort" (often
followed by a core dump) at what seem to be random places. Running the
regression test standalone with the same PGOPTIONS settings does not
reproduce the error, and in fact it happens to different tests if you
run the script over and over.
I have also sometimes seen failures out of mdblindwrt, apparently trying
to dump a dirty buffer for a no-longer-existing database.
Anyone have any idea how to debug this, or what might be triggering it?
The best theory I've come up with is that it's got something to do with
the repeated destruction and re-creation of the "regression" database.
But usually the failure occurs during the later tests within a
particular regression set, so you'd think any effects of destroying
the previous incarnation of the DB would be long gone.
regards, tom lane
