This adds support for histogram-style metrics instead of using
summaries. It means we can sum on a cluster level and present the user's
experienced latency instead of looking at it on a per-node level.
The current version limits the range of histogram buckets between 0.1ms
and 60s, to avoid exporting huge amounts of buckets that are likely
empty. Further patches could limit this further, for example by going
for a 1.44x increment instead of the 1.2x increment, or by specifying
the ranges in the configuration.
Even with the limits in place, this exports 76 metrics to 3 metric
families per histogram. The original summaries-based code only exports 8
metrics (in 3 families), though in theory those are no longer needed and
could be disabled with a flag.
This adds support for histogram-style metrics instead of using summaries. It means we can sum on a cluster level and present the user's experienced latency instead of looking at it on a per-node level.
The current version limits the range of histogram buckets between 0.1ms and 60s, to avoid exporting huge amounts of buckets that are likely empty. Further patches could limit this further, for example by going for a 1.44x increment instead of the 1.2x increment, or by specifying the ranges in the configuration.
Even with the limits in place, this exports 76 metrics to 3 metric families per histogram. The original summaries-based code only exports 8 metrics (in 3 families), though in theory those are no longer needed and could be disabled with a flag.