On Mon, Mar 27, 2017 at 12:36 PM, Rushabh Lathia
<rushabh.lathia@gmail.com> wrote:
> Hmm I agree that it's good idea, and I will work on that as separate patch.
Maybe you want to start with what David already posted?
>> Possibly we
>> should fix the crash bug first, though, and then do that afterwards.
>> What bugs me a little about Rushabh's fix is that it looks like magic.
>> You have to know that we're looping over two things and freeing them
>> up, but there's one more of one thing than the other thing. I think
>> that at least needs some comments or something.
>>
> So in my second version of patch I change gather_merge_clear_slots() to
> just clear the slot for the worker and some other clean up. Also throwing
> NULL from gather_merge_getnext() when all the queues and heap are
> exhausted - which earlier gather_merge_clear_slots() was returning clear
> slot. This way we make sure that we don't run over freeing the slot for
> the leader and gather_merge_getnext() don't need to depend on that
> clear slot.
Ah, I missed that. That does seem cleaner. Anybody see a problem
with that approach?
--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company