Re: Tracking wait event for latches

On Thu, Sep 22, 2016 at 6:49 AM, Thomas Munro
<thomas(dot)munro(at)enterprisedb(dot)com> wrote:
> On Thu, Sep 22, 2016 at 1:23 AM, Robert Haas <robertmhaas(at)gmail(dot)com> wrote:
>> Moreover, it's pretty confusing that we have this general concept of
>> wait events in pg_stat_activity, and then here the specific type of
>> wait event we're waiting for is the ... wait event kind. Uh, what?
>
> Yeah, that is confusing. It comes about because of the coincidence
> that pg_stat_activity finished up with a wait_event column, and our IO
> multiplexing abstraction finished up with the name WaitEventSet.
> <stuck-record-mode>We could instead call these new things "wait
> points", because, well, erm, they name points in the code at which we
> wait. They don't name events (they just happen to use the
> WaitEventSet mechanism, which itself does involve events: but those
> events are things like "hey, this socket is now ready for
> writing").</>

What about trying first to come up with a class name generic enough
that would cover the set of events we are trying to cover. One idea is
to simply call the class "Process" ("Point", "Poll"), and mention in
the docs that this can wait for a socket, a timeout, postmaster death
or another process to tell it to move on. The idea is to come with a
name generic enough. In the first patch we don't detail much what a
process is waiting for at a given point, like what has been submitted
does. This would still be helpful for the user, because it would be
possible to look back at the code and guess where this is waiting.

Then comes the interesting bits: I don't think that it is a good idea
to associate only one event for a waiting point in the system views.
Say that at a waiting point WaitLatch is called with
WL_POSTMASTER_DEATH and WL_TIMEOUT, I'd rather let the user know that
both of them have been set up and not choose just one of them using
some hierarchy. But I rather think that we should list all of them
depending on the WL_* flags set up by the caller. Here comes a second
patch: let's use the last byte of the wait events and add this new
text[] column in pg_stat_activity, say wait_details. So for a waiting
point it would be possible to tell to the user that it is waiting on
'{socket,timeout,postmaster_death}' for example.

Then comes a third patch: addition of a system view to list all the
activity happening for processes that are not dependent on
max_connections. pg_stat_activity has the advantage, as Tom mentioned
a couple of days ago, that one can just run count(*) on it to estimate
the number of connections left. I think this is quite helpful. Or we
could just add a column in pg_stat_activity to mark a process type: is
it a system process or not? This deserves its own discussion for sure.

Patches 2 and 3 are independent things. Patch 1 is a requirement for 2 and 3.

>> Another thing to think about is that there's no way to actually see
>> wait event information for a number of the processes which this patch
>> instruments, because they don't appear in pg_stat_activity.
>
> Good point. Perhaps there could be another extended view, or system
> process view, or some other mechanism. That could be material for a
> separate patch?

Definitely a separate patch..
--
Michael