Re: [HACKERS] Statement-level rollback

Tsunakawa> The point raised in this thread was that that creates
Tsunakawa> too much network overhead, so a backend-based solution would be preferable.
Tsunakawa> We haven't seen any numbers or other evidence to quantify that claim, so
Tsunakawa> maybe it's worth looking into that some more

The performance overhead for "SELECT" statement (no columns, just select) statement over localhost is 36±4 us vs 38±3 us (savepoint is pipelined along with user-provided query). That is network overhead is close to negligible.

As far as I understand, the main problem with savepoints is they would consume memory even in case the same savepoint is reassigned again and again.

In other words, "savepoint; insert;savepoint; insert;savepoint; insert;savepoint; insert;savepoint; insert;" would allocate xids and might blow up backend's memory.

I see no way driver can workaround that, so it would be great if backend could release memory or provide a way to do so.

Adding protocol messages would blow pgbouncer, etc things, so it makes sense to refrain from new messages unless it is absolutely required.

Vladimir