On Aug 12, 2013, at 12:47 PM, Tom Lane <tgl@xxxxxxxxxxxxx> wrote: > Matt Solnit <msolnit@xxxxxxxxxx> writes: >> 2. The function seems to work consistently when I do a SELECT >> SUM(mycol) without any GROUP BY. It's only when I add grouping that >> the failures happen. I'm not sure if this is a real clue or a red >> herring. > > That isn't enormously surprising, since the memory management for > the transition values is different in the two cases. > >> Finally, can you tell me what precisely happens when you call >> datumCopy() with ArrayType? If it's only returning a copy of >> the TOAST reference, then how is it safe for the transition function >> to modify the content? I'm probably *completely* misunderstanding >> how this works, so I would love to be enlightened :-). > > You're right, datumCopy() won't expand a TOAST reference. What does > expand it is PG_GETARG_ARRAYTYPE_P(). So if you have a case where the > system picks up a copy of an array input that happens to be toasted, > it's the GETARG step in the next invocation of the aggregate transition > function that expands the TOAST reference, and then after that you have an > in-memory copy that's safe to modify. Maybe you're missing that somehow? > The code fragment you showed looked okay but ... > > regards, tom lane I think I figured it out. The problem is this line: Datum *arrayData1, *arrayData2; Datum* was correct when I first started this journey, using deconstruct_array(), but is incorrect when accessing the array's content directly using ARR_DATA_PTR(). Changing these to int* fixes the problem, at least on all the systems I've tried so far. I've been wondering why the broken code worked without a GROUP BY, and I think it was just dumb luck. With no GROUP BY, I was only overrunning a single buffer, and maybe the effects were not immediately apparent. With GROUP BY, however, there's a buffer overrun for each group, and each one increases the chance of doing something catastrophic. Sincerely, Matt Solnit -- Sent via pgsql-general mailing list (pgsql-general@xxxxxxxxxxxxxx) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-general
