"Simon@2ndquadrant.com" <simon@2ndquadrant.com> writes:
> This is not a provably correct state machine
I think the discussion ends right there. You are assuming that the
commit is guaranteed to finish in X amount of time, when it is not
possible to make any such guarantee. We are not putting in an
unreliable commit mechanism in order to save a small amount of lock
contention. (As I tried to point out already, but it doesn't seem
to have sunk in: this newly-added lock is not likely to be that much
more contention added to the commit path, seeing that the path of
control it protects already involves taking at least two exclusive
LWLocks. Those locks will likely each cause as much or more SMP
cache thrashing as this one.)
What we could use is a better way to build LWLocks in general. I do not
know how to do that, though, in the face of SMP machines that seem to
fundamentally not have any cheap locking mechanisms...
regards, tom lane