Re: postgresql's MVCC implementation

Kent Tong <kent(at)cpttm(dot)org(dot)mo> writes:
> 1: T1 sets isolation to serializable & begins a transaction
> 2: T2 sets isolation to serializable & begins a transaction
> 3: T1 reads X into v1
> 4: T2 reads Y into v2
> 5: T1 writes v1 into Y
> 6: T2 writes v2 into X
> 7: T1 commits
> 8: T2 commits

> Obviously, this sequence is also not a serializable execution. However, it
> is allowed by
> PostgreSQL. Moreover, according to the MVCC reference above, step 5 should
> really
> fail because the read timestamp of Y is that of T2, which is greater than
> that of T1.

If you want that to fail, use a SELECT FOR UPDATE at steps 3/4.

My interpretation of MVCC is that the above example isn't even
meaningful, because it assumes that "writing into Y" is an overwrite,
which it is not in Postgres --- that is, if T2 reads Y again, it'll
get the same value as before.

regards, tom lane