Add response histograms to apache_metrics example. #46
Conversation
| counter apache_http_received_bytes_total by server_port, handler, request_method, request_status, request_protocol | ||
| counter apache_http_request_time_microseconds by quantile, server_port, handler, request_method, request_status, request_protocol |
There was a problem hiding this comment.
"quantile" doesn't seem like the right term here. It's the boundary of the latency bucket we are talking about here. So "bucket" or "upper_bound" or something?
There was a problem hiding this comment.
I was just using the labeling used by other similar metrics. This should result in normal Prometheus-style cumulative histogram labeling.
|
10x buckets are really coarse. Check out our default buckets for a "network service": https://github.com/prometheus/client_golang/blob/master/prometheus/histogram.go#L54-L59 Even those are pretty coarse, and only meant as a starting point before you manually tailor the buckets and zoom into the area you are interested in. |
|
This is an example file, I didn't want to get overly granular, but just provided a starting point example for "normal" apache servers. But I can adjust them. Part of the problem is apache servers sometimes also handle very long-lived requests. |
|
@beorn7 |
|
OK, so the labels should be fine now, but it would be good to see an example output. WRT bucket scheme: Especially because it is an example and because Apache may server very long running queries, I'd use the same bucketing as in https://github.com/prometheus/client_golang/blob/master/prometheus/histogram.go#L54-L59 . The buckets there were laid out under the exact same premises. Some monitoring systems use 4x buckets, and that's already mostly useless. 10x is really not a good role model, so it shouldn't be used in an example. |
1 similar comment
Use the standard recommended network service buckets[0]. [0]: https://github.com/prometheus/client_golang/blob/master/prometheus/histogram.go#L54-L59
|
@beorn7 Ok, I've updated the buckets to follow the recommended defaults. |
1 similar comment
|
Thanks. Could you attach a sample output to this PR so that we can check the final outcome? |
|
Thanks! |
Sorry, microseconds is the only reasonably compatible time until mtail supports floats metrics.
CC @beorn7