Comment # 1
on bug 1133
from Ken McDonell
Frank,
We _could_ register these earlier ... it probably needs to be in the common
lock initialization code to make sure all uses are captured. But this would
mean we'd always have derived metrics for every single context, even those that
don't use derived metrics and don't use event records (which would be the
majority of contexts I'd expect).
The problem here is that this adds some non-trivial overhead on the code path
for many libpcp calls ... and I'm loathed to do that if it can be avoided.
Is a fetchgroup client likely to be asking for event.* metrics?
If yes, is this likely to be explicit or are they expected to be found in a
PMNS traversal from the root of the PMNS?
If you could give me some use cases, I'll see if there is (a) a different place
to pre-emptively register the metrics that would work for you, but not burden
everyone else, or (b) suggest a new libpcp routine to explicitly trigger the
registration.