Closed victor-sudakov closed 2 years ago
Hey @victor-sudakov. Yep, there are differences between Gauge
and currently used Summary
. The summary
was chosen because Gauge
panics on negative numbers which I've seen sometimes. So, I wanted to use the safest approach from the stability POV. I'll retest Gauge
one more time.
The
summary
was chosen becausegauge
panics on negative numbers which I've seen sometimes.
It is very strange because a gauge is in 64-Bit floating point format which officially supports negative numbers. If it's a bug in Prometheus or in the library, it should be reported.
It is very strange because a gauge is in 64-Bit floating point format which officially supports negative numbers.
Yep, I know. I need to re-check this, might be that was my issue somewhere. Will be available in 2.7
.
Yeah, I have tested Gauge
on high load, no errors were reported. Let's switch this type as semantically more suitable.
@victor-sudakov Enjoy: https://github.com/spiral/roadrunner-binary/releases/tag/v2.6.5 ))
Are you sure it's fixed?
/app $ /usr/bin/rr --version
rr version 2.6.5 (build time: 2021-12-14T13:31:00+0000, go1.17.5)
/app $ curl -s http://localhost:9090/metrics | grep queue
# HELP rr_http_requests_queue Total number of queued requests.
# TYPE rr_http_requests_queue summary
rr_http_requests_queue_sum 0
rr_http_requests_queue_count 0
/app $
Related: spiral/roadrunner-binary#151
It's fixed in the master branch, but I guess by some reason cherry-pick
into stable was failed. Ok, sorry, have to re-release.
And I'm afraid the rr_http_requests_queue_sum is also broken now, it's always zero. So 2.6.5 is kind of between 2 worlds.
Yep, cherry-pick
failed. The new version will be in a moment.
@victor-sudakov https://github.com/spiral/roadrunner-binary/releases/tag/v2.6.6
Works as expected in v2.6.6, thank you!
No duplicates 🥲.
What happened?
I'm afraid the rr_http_requests_queue metrics has incorrect type. It has the type of "summary" while in fact it is a gauge.
Quoting the Prometheus docs
So basically rr_http_requests_queue_sum should never go down, but of course it does, and it has to be used as a gauge in Prometheus/Grafana which is counter-intuitive.
Version
2.6.3
Relevant log output
No response