Re: [pcp] pmie support for string-typed metrics

[Mark Goodwin <mgoodwin@xxxxxxxxxx>]

On 05/02/2016 11:31 AM, Nathan Scott wrote:
> Hi Mark,
>
> ----- Original Message -----
> > [...]
> >
> > I've been thinking about this topic - not just for string valued metrics
> > (btw: hinv.map.scsi is another one in this class), but also for discrete
> > integer typed metrics, such as hinv.ndisk etc.
> Its also a bit like the proc.psinfo.environ case that Martins came across,
> I think.
>
> > These metrics need to be logged more than "once"; rather more like "when
> > changed".
> > Once logged, they retain their discrete value (string or integer) until
> > logged
> > again with a new value. As an alternative to pmie monitoring these metrics
> > and telling
> > pmlogger to log them again via pmlc, maybe pmlogger could just maintain a
> > cache
> > (suitably indexed for efficiency) - when a discrete metric/instance/value
> > with logging
> > frequency of "when changed" changes, write it to the log again.
> I guess that would involve fetching these metrics like normal, but modifying
> the pmResult to drop them, before writing to disk.  A bit like the inverse of
> derived metrics, where the pmresult is added to with new metrics.
>
> Might need a new pmlogger-language grouping keyword to indicate this new kind
> of on-change-only behaviour for certain metrics?  (Or, automatically done for
> all string-type metrics?)
How about "async" logging frequency

e.g. :

log advisory on async {
        hinv
        kernel.uname
        filesys.mountdir
        filesys.blocksize
        filesys.capacity
}