Greg Stark <gsstark@mit.edu> wrote:
> Robert Haas <robertmhaas@gmail.com> wrote:
>> But let's back up and talk about MVCC for a minute. Suppose we
>> have three source tuples, (1), (2), and (3); and the target table
>> contains tuples (1) and (2), of which only (1) is visible to our
>> MVCC snapshot; suppose also an equijoin. Clearly, source tuple
>> (1) should fire the MATCHED rule and source tuple (3) should fire
>> the NOT MATCHED rule, but what in the world should source tuple
>> (2) do? AFAICS, the only sensible behavior is to throw a
>> serialization error, because no matter what you do the results
>> won't be equivalent to a serial execution of the transaction that
>> committed target tuple (2) and the transaction that contains the
>> MERGE.
>
> So the behaviour we get with UPDATE in this situation is that we
> update (2) so I would expect this to execute the MATCHED rule.
Certainly that would be consistent with the behavior of READ
COMMITTED -- wait for commit or rollback of the concurrent
transaction, and then proceed with whatever data is there after
completion of the other transaction. With REPEATABLE READ or
SERIALIZABLE you would block until commit of the other transaction
and terminate with a write conflict -- a form of serialization
failure. If the other transaction rolls back you INSERT.
At least, that would be the least surprising behavior to me.
-Kevin
