> That doesn't follow. A transaction only has to see data from
> committed transactions that were visible when it started. A
> transaction that has committed on another node but whose
> notification hasn't arrive is no different from a transaction that
> committed after our transaction has started. Same logic, same rule.

Agree with Jim here.

Note that on a distributed system one has to take a broader view of
snapshot isolation. From Mrs. Kemme's 2005 paper:

"Using this scheme the following can happen. If Ti and Tj do not
conflict, one replica R might commit Ti, start a transaction Tk reading
both X and Y , and then commit Tj , and another replica R' might commit
Tj , start a transaction Tl reading both X and Y , and then commit Ti.
Although snapshot isolation is provided individually at each replica, a
centralized system executing the same set of transactions could not
have provided these different snapshots to Tk and Tl."