On 10/10/14 12:38, Jim Nasby wrote:
> On 10/8/14, 5:51 PM, Peter Geoghegan wrote:
>> On Wed, Oct 8, 2014 at 2:01 PM, Kevin Grittner<kgrittn@ymail.com>
>> wrote:
>>> >Although the last go-around does suggest that there is at least one
>>> >point of difference on the semantics. You seem to want to fire the
>>> >BEFORE INSERT triggers before determining whether this will be an
>>> >INSERT or an UPDATE. That seems like a bad idea to me, but if the
>>> >consensus is that we want to do that, it does argue for your plan
>>> >of making UPSERT a variant of the INSERT command.
>> Well, it isn't that I'm doing it because I think that it is a great
>> idea, with everything to recommend it. It's more like I don't see any
>> practical alternative. We need the before row insert triggers to fire
>> to figure out what to insert (or if we should update instead). No way
>> around that. At the same time, those triggers are at liberty to do
>> almost anything, and so in general we have no way of totally
>> nullifying their effects (or side effects). Surely you see the
>> dilemma.
>
> FWIW, if each row was handled in a subtransaction then an insert that
> turned out to be an update could be rolled back... but the performance
> impact of going that route might be pretty horrid. :( There's also the
> potential to get stuck in a loop where a BEFORE INSERT trigger turns
> the tuple into an UPDATE and a BEFORE UPDATE trigger turns it into an
> INSERT.
Perhaps you need an UPSERT trigger?