tbeijen / klutch Goto Github PK

Thoughts:

• Readme
• Consider moving dev & vscode things to separate files under /docs
• Use case (blog post?)

Scenario:

• Trigger fires, hpas scale up
• Traffic comes in
• Sequence ends, hpas reverted. Traffic still high.
• 2nd trigger fires
• This is not expected to once again cause an increase in traffic, now based on still ongoing peak traffic, klutch would again ramp up deployments.

It might be good to have the option to set a cooldown: Never scale up a 2nd time within that period.
However, this might also have averse effect, not ramping up where it would have been needed.

Note that this will likely lead to renaming the cooldown option in the Config class to something else.

Removing latency between trigger appearing and being picked up by klutch.

Thoughts:

• Auth? Lo-fi starting point would be a secret holding bearer token that needs to be present in request.
• How to run control loop and webserver in parallel.

First thoughts: Generic webhook, suitable for slack but also other platforms.

Currently publishing docker image and chart on tag push.

Good enough initially, later real releases could bring benefits:

• Release notes
• Helm chart .tar.gz files being true release artifacts instead of being committed to gh-pages branch.

Resources:

• https://github.com/python-poetry/poetry/blob/master/.github/workflows/release.yml
• https://github.com/orbit/orbit/blob/master/.github/workflows/release_orbit.yml

Example: If RBAC is not set up correctly exceptions will be logged. But the process doesn't crash.

Probably best to crash the process, resulting in CrashLoop which will be more visible and trigger any alerts already set up.

Using a real or simulated cluster.

• Kind?
• K3d? https://github.com/rancher/k3d

Resources:

• https://github.com/rancher/k3d
• https://radu-matei.com/blog/kubernetes-e2e-github-actions/

Prometheus metrics. Something to put in dashboards.

Possibile metrics include:

• Active y/n
• Pods added
• HPAs elegible
• HPAs examined
• Control loop duration
• Errors
• ...

Relates to #3, also requiring a webserver to run in parellel with control loop.

First of all: Needed, or sticking to add your own trigger service, posting to API endpoint? (#3)