All,

Just a side note, separated from arguing about implementations.

If I'd start this discussion from beginning, I'd leave the read
consistency aside and offer just the locking/modification behavior Ann
was suggesting: wait for concurrent commit and then lock/overwrite. It
can be bound to read-commmitted (PGSQL, Oracle) or repeatable-read
(InnoDB) isolation levels. We're free to discuss both. Let's just treat
this as an option (alternative or default) of the existing isolation
levels, not a brand new one.


Dmitry