Visualization of performance tests results

[nader-ziada]

/area test-and-release

Describe the feature

Background:

• The old Knative performance testing framework was based on Mako, which is not open to accept data from non-Google developers, making it not possible for community members to run benchmark tests.
• This got replaced with some perf tests running on kperf and the old suite using Mako running as well but using a side-car stub to collect the data instead of sending it to Mako backend.
• These two methods collect data in files, but do not provide a graphical representation of the data.

The feature request now is to have a graphical representation on the continuously collected data

• Have a central location the prow test jobs can send the data to.
• Run nightly jobs to collect the results.
• Build a dashboard to show performance results over time.
• Document the performance tests and what measurements are captured in each test

[nader-ziada]

/assign

[nader-ziada]

The approach I'm planning currently is add the data points in an influxdb instance and then have a grafana dashboard showing the results

This is an example of how the dashboard could look like, still working on a POC locally

If folks have comments or feedback about the approach or the selected tools, please comment on the issue here

[psschwei]

Where's the db / dashboard going to live?

[nader-ziada]

Where's the db / dashboard going to live?

somewhere in knative gcp project I guess, I need to figure this out with productivity wg

[daisy-ycguo]

@nader-ziada how is it going? I think we could request a server at CNCF cluster and host the performance dashboard there.

And I think the dashboard could be a part of kperf project. Maybe we can create an issue at kperf and track the progress there.

[nader-ziada]

There is a community cluster in the knative project that is meant to host the influxdb and dashboard, currently working on it here #13192

[nader-ziada]

http://34.170.87.98:443/d/HFw0qn74k/serving-performance-testing?from=1663269760987.5151&to=1663323727899.558&orgId=1

[psschwei]

asks for a username/password

[nader-ziada]

asks for a username/password

knative/knative
not sure how to make it public

[dprotaso]

You can setup a domain name ie. perf.knative.dev and point it to your dashboard

https://github.com/knative/test-infra/blob/main/infra/gcp/dns/dns.tf

[nader-ziada]

You can setup a domain name ie. perf.knative.dev and point it to your dashboard

https://github.com/knative/test-infra/blob/main/infra/gcp/dns/dns.tf

knative/test-infra#3584

[dprotaso]

Playing with the dashboard - it's a bit cumbersome.

Questions/thoughts

1. If I run the performance test on a PR how do I visualize the metrics for just that single run?
2. We need to label the axes and have better names for the lines in the graph
   • see ie. https://mako.dev/benchmark?benchmark_key=5143375149793280
   • I'm assuming there a way to just configure the dashboard to re-label the series
3. Is there a way to toggle on/off a specific series. For example in the above mako link just deployment latency.
4. Graphs load slow - not sure what to do here - maybe split up the dashboard so it's not loading data for four graphs
5. Is there a way to filter data based on tags in the grafana UI - ie. if we start doing perf tests on release branches and different k8s versions how would I do that.

[nader-ziada]

thanks for the feedback, will investigate how to do these changes on the dashboard

[dprotaso]

Also wondering if there's a way to drop the interpolation between the successive runs - it seems a bit noisy

[github-actions]

This issue is stale because it has been open for 90 days with no
activity. It will automatically close after 30 more days of
inactivity. Reopen the issue with /reopen. Mark the issue as
fresh by adding the comment /remove-lifecycle stale.