On 2016/11/02 2:44, Robert Haas wrote:
> On Fri, Oct 28, 2016 at 3:53 AM, Amit Langote
> <Langote_Amit_f8@lab.ntt.co.jp> wrote:
>>> Insisting that you can't drop a child without detaching it first seems
>>> wrong to me. If I already made this comment and you responded to it,
>>> please point me back to whatever you said. However, my feeling is
>>> this is flat wrong and absolutely must be changed.
>>
>> I said the following [1]:
>>
>> | Hmm, I don't think I like this. Why should it be necessary to detach
>> | a partition before dropping it? That seems like an unnecessary step.
>>
>> I thought we had better lock the parent table when removing one of its
>> partitions and it seemed a bit odd to lock the parent table when dropping
>> a partition using DROP TABLE? OTOH, with ALTER TABLE parent DETACH
>> PARTITION, the parent table is locked anyway.
>
> That "OTOH" part seems like a pretty relevant point.
>
> Basically, I think people expect to be able to say "DROP THINGTYPE
> thingname" or at most "DROP THINGTYPE thingname CASCADE" and have that
> thing go away. I'm opposed to anything which requires some other
> series of steps without a very good reason, and possible surprise
> about the precise locks that the command requires isn't a good enough
> reason from my point of view.
OK, I changed things so that DROP TABLE a_partition no longer complains
about requiring to detach first. Much like how index_drop() locks the
parent table ('parent' in a different sense, of course) and later
invalidates its relcache, heap_drop_with_catalog(), if the passed in relid
is a partition, locks the parent table using AccessExclusiveLock, performs
its usual business, and finally invalidates the parent's relcache before
closing it without relinquishing the lock. Does that sounds sane? One
downside is that if the specified command is DROP TABLE parent CASCADE,
the above described invalidation is a waste of cycles because the parent
will be dropped anyway after all the partitions are dropped.
Thanks,
Amit