Open Stack

Hi Synaps Folks,

IIUC the code, it seems Synaps does a pretty strict interpretation of
INSUFFICIENT_DATA, i.e. one of the evaluation periods without samples
is enough for this state transition to occur.

Whereas a little experimentation shows that CW is much more forgiving,
in the sense that it requires no samples whatsoever across the entire
evaluation periods, plus some leeway, in order for an alarm to flip
into this state.

So before investing time in preparing a patch, I was wondering if you
folks would be amenable to the idea of loosening this constraint?

The motivation would be to avoid causing lots of spurious state
transitions (and the associated actions firing) when there's a short
gap in the metric stream (due, say, to back-pressure in the AMQP
queues feeding the storm cluster). This I suspect might be a problem
in large production deployments.

The situation is complicated a bit by the Synaps concept of evaluation
period as overlapping windows offset by 60s (as discussed here[1]), so
a subsidiary question would be how wedded to that concept are you guys?

(BTW I can see how the moving-average style of interpretation might
useful, by smoothening out transient spikes and tending to reveal the
underlying trend. So I'm not questioning its utility, say as an
optional/additional set of statistics. However it seems to me that the
default behavior should follow the CW semantics.)

Cheers,
Eoghan

[1] https://answers.launchpad.net/synaps/+question/211970