Bruce Momjian <pgman@candle.pha.pa.us> writes:
> Right. What if the postmaster signals the backend once a second to do
> their reporting. I am not sure what the final solution will be, but we
> _need_ one based on the performance numbers I and others have seen.
> Could we have PGPROC have a reporting boolean that is set every second
> and somehow checked by each backend?
I don't see any point in involving the postmaster in it. What might be
interesting is to replace the current backend timer-interrupt handling
by a free-running cyclic interrupt every N milliseconds (the resolution
of the statement_timeout and deadlock_check delays would then be no
better than N milliseconds, since those actions would occur at the next
cyclic interrupt after the desired time expires). We could possibly
drive stats reports and ps_status updates from this, as well as sampling
EXPLAIN ANALYZE if anyone cares to pursue that.
Complaints I can foresee:
* lots of cycles wasted in idle backends. Possibly a backend that's not
received any command for a second or two could shut down its interrupt
until it next gets a command.
* not clear whether the interrupt happens when waiting for I/O. I
already mentioned that this would be a problem for EXPLAIN ANALYZE,
but it might be no big deal for the other uses.
regards, tom lane
---------------------------(end of broadcast)---------------------------
TIP 2: Don't 'kill -9' the postmaster
