Jan Wieck <JanWieck@yahoo.com> writes:
>> You can't count heapscan info via the relcache entry --- what if there's
>> more than one heapscan in progress on the same rel? What if the
>> relcache entry gets rebuilt by SI invalidation? I don't think the stats
>> stuff should touch the relcache at all.
> The counting end's up in one and the same global slot of the
> tabstat messages. These slots are collected per frontend
> command and are identified by the tables oid. Does it matter
> how often the pointer to that slot is updated to the same
> value?
In that case, forget about the fields in the heapscan and indexscan
structs. You can easily keep both the heap and index access counts
in (or referenced by) the relcache entry, no? It just bothers me
that the treatment of heap and index counts is so inconsistent.
One or the other is wrong, or at least inefficient.
BTW, does it really matter whether we are counting an index-driven
or heap-driven fetch? Seems to me that if you count indextuple and
heaptuple fetches, you can figure out what's going on without having
to have heap_fetch know the difference.
>> Also, I do not like the pgStatPmPipe mechanism for detecting postmaster
>> shutdown. That eats two file descriptors in every backend, which is a
>> pretty substantial waste of resources. Why not just have the postmaster
>> send a signal to the stats collector when it's time to shut down?
> Don't kill -9 the postmaster :-)
What's your point? If the postmaster crashes, whether the stats
collector goes away without help seems like the least of your worries.
You'll end up killing backends manually anyhow, so why not the stats
process too? (For that matter, do you care if the stats process doesn't
die? Once there is no one left to send messages to it, it'll never do
anything, no?)
regards, tom lane
