> Michal, excellent work! You should get in touch with the MySQL maintainers;
> I met the principal maintainer about a year ago at Linux Expo in N.C. USA
> and he said he was interested in a PCP agent for MySQL, but I never had
> time to work on it [note: this reply is cc mysql@xxxxxxxxxxxxxxx].
>
I will cc this reply too. I am not personally on mysql maillist.
> > - Current version supports one server per running PMDA. I persobally think
> > this
> > is sufficient (see NOTES), but what do you think?
>
> This is fine.
OK, one less thing to worry about :)
> > - I cannot seem to be able to properly install help for metrics. When can
> > I find instructions?
>
> What is the nature of the problem? You should read the man pages for
> pmdaDaemon(3) and newhelp(1), and use the code for the trivial, sample
> and simple PMDAs as examples. There are also online help books at
> techpubs.sgi.com (search for the Performance Co-Pilot Programmer's Guide).
>
I have read the pages - I do pmdaDaemon with help file
/var/pcp/pmdas/mysql/help.
It is there and it is the plain text help file. However, 'pminfo -T' on any
metric
says that help text is not available. I have checked it with the other pmdas
and it seems I am doing it right... I have created the pmda from the
sendmail pmda, so it should be OK if the help works there...
> It is acceptable to cache results, but be careful with metrics that
> have counter semantics - client tools that do rate conversion will
> show rates in "bursts", because the counters will suddenly increase
> in value each time the cache period expires and you issue a new fetch
> to the sql server.
This should be OK - I didn't think of invalidate times higher than one
second... I ment it only for the tools like pminfo which get only 10 metrics
with each fetch (BTW, my pcpmon gets all metrics from one server with single
fetch ;)
> > - I would like to implement processlist ('show processlist' command). Do you
> > think it would provide useful information for PM?
>
> The per-client and per-process statistics have been historically of the
> most value in our use of PCP in tracking down oracle/sybase/informix
> performance problems, so I would say yes - the per process stuff is
> definitely worth doing. If you need help understanding instance domains,
> please ask!
>
I guess I understand them :) The instance philosophy is OK, I just have
to play with the way how the instances are got - seems via the inst parameter
to the fetch callback. But, er... how are they enumerated? I am currently
looking on the weblog pmda, it seems it doesn't use the callback but overwrites
whole fetch function. Anyway, I will look at it, I hope I am not too stupid
to not being able to figure it out ;)
Michal Kara
|