On Sun, Oct 24, 2010 at 5:50 AM, Martijn van Oosterhout
<kleptog@svana.org> wrote:
> On Sat, Oct 23, 2010 at 03:59:13PM -0700, Greg Stark wrote:
>> I think the blocker with MERGE has always been that the standard is
>> *so* general that it could apply to conditions that *aren't* covered
>> by a unique constraint or exclusion constriant. Personally, I'm fine
>> saying that those cases will perform poorly, locking the table, or
>> aren't allowed and print a warning explaining exactly what exclusion
>> constraint or btree unique index would be necessary to allow it. I
>> think virtually every case where people will want to use it will be on
>> a primary key anyways.
>
> Can we please not get MERGE hung up on trying to make it atomic. The
> standard requires no such thing and there are plenty of uses of MERGE
> that don't involve high concurrency updates of the same row by
> different processes. If we want an atomic UPSERT, make that a seperate
> command, MERGE without atomicity is useful enough already.
>
> Simply, if the row is visible in the current snapshot you do an update
> otherwise an insert. If you get an error, raise it. Icing can be added
> later.
Long term, I'm wondering if the permanent fix for this problem might
involve something similar to what EXCLUDE does - use a dirty snapshot
for the scan of the target table, and block on in=d
Yeah, I couldn't have said it better. It's an annoying implementation
restriction but only that. I think it's a perfectly acceptable
restriction for an initial commit. It's no secret that our process
has trouble absorbing large patches.
I am also wondering exactly what the semantics are supposed to be
under concurrency. We can't assume that it's OK to treat WHEN NOT
MATCHED as WHEN MATCHED if a unique-key violation is encountered. The
join condition could be something else entirely. I guess we could
scan the target table with a dirty snapshot and block on any in-doubt
tuples, similar to what EXCLUDE constraints do internally, but
throwing MVCC out the window doesn't seem right either.
--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company