Open ryandotsmith opened 11 years ago
Please correct me if my understanding librato's aggregation is off, but I think it might be best to sum both the sum
and count
metrics, as in both cases you want the sum of all values for a time period, not the number of times librato received the metric. e.g.
1...10
, at 12 second intervals over the space of 2 minutesmetric.sum=15, metric.count=5
metric.sum=40, metric.count=5
When it comes to aggregating the two periods into one, we should see metric.sum=55, metric.count=10
. If the aggregation of metric.count
were count then I think we'd see "2" rather than "10".
WRT to percentiles, I think max
is best, as you want to know the worst case for each percentile, however I'm not sure about mean, median and last.
Thoughts?
.count
-> sum.sum
-> sum.max
-> max.min
-> min.p99
-> max.p95
-> max.mean
-> average/max.median
-> average/max.last
-> average/maxI am actually not sure if l2met should be responsible for this. I am thinking that librato-dash will take care of all this type of data.
You need to specify how librato down-samples data over time, currently there are options for: sum, average, max. L2met metrics like sum and count should be count aggregated over time while things like mean should be averaged. When it comes to percentiles, I am not sure what to do here. Perhaps the max function would be the best.