Investigate CI performance

[ulope]

trafficstars

Problem Definition

In the last couple of months our CI performance has noticeably decreased. The current average workflow runtime seems to be 20+ minutes.

Ideally we want to return to the the 10 - 15 min range.

Things to investigate:

• Where is time being spent?
  • Workspace up-/download
    • This seems relatively slow esp. compared to cache
    • Maybe move some files (e.g. python venv) to cache and only keep config / setting in workspace
  • Caching (is it working as intended?)
• Are we being throttled by concurrency limits? (Conversation with Circle @karlb @czepluch)
• Can we improve test run-time by further parallelization? (diminishing returns due to spin up overhead)
• Can the integration tests be optimized? E.g.:
  • Possibility to not restart eth node / synapse / etc. for every test
  • Do the contracts have to be deployed for every test?

[czepluch]

I can setup a meeting with CircleCI if we believe that would be helpful.

[karlb]

Are we being throttled by concurrency limits? (Conversation with Circle @karlb @czepluch)

Yes, we are. I have been contacted by a CircleCI employee who would like to talk to us about it. I'll forward the contact details to both of you, again.

[istankovic]

Today we mentioned a possibility of bumping up the required Python version to 3.8 (@palango, @ezdac). If we decide to go that way, we can simply remove all the python-3.7 stuff from our CI, which I guess would help somewhat.