Keda Support

[willplatnick-okta]

Keda is a useful tool for doing event-driven autoscaling in Kubernetes. To do its job, you create ScaledObject manifests, and Keda then creates and manages the HPAs itself. It would be nice if Flagger supported ScaledObject cloning and management during deploys like it does now with HPA's.

[aryan9600]

Looking into this 👍

[keegancwinchester]

@aryan9600 hey! any update on this?

[aryan9600]

These two issues currently block KEDA support for Flagger:

• Provide support for explicitly pausing autoscaling of workloads. kedacore/keda#944
• Add support for templating in spec.triggers.metadata.query kedacore/keda#2757

[pragmaticivan]

It appears kedacore/keda#2757 is not going to fly.

What are your thoughts on allowing the triggers metadata block to be fully overwritten for canaries by using inputs from Flagger yaml instead?

[aryan9600]

IMO, we'd be better off going with a find and replace approach. Having a field in the Canary spec just for this doesn't sit right with me and I can't think of which section of the spec it'd make sense to have it, semantically speaking.

[pragmaticivan]

What are your thoughts on having Keda autoscale the Canary CRD instead? (Like it does with Argo Rollout). Ref: https://keda.sh/blog/2020-11-04-keda-2.0-release/#scaledobject-improvements

[aryan9600]

That's not possible. An Argo Rollout is basically a thin wrapper around a Deployment. Thus it exposes a Scale subresource, which KEDA can use to scale the Rollout object. Canary does not have that subresource, because it uses Deployments natively. Flagger does not edit your actual/stable deployment it just makes a copy of it and then routes traffic to them according to calculated rules.