Proposal
HPA-class PodAutoscalers should be able to scale on concurrency as well as CPU. This requires exposing the Knative calculated concurrency metric as a K8s custom metric. And creating a v2beta2 HPA with a pointer to the custom metric. (v1 HPA does not support custom metrics).
The custom metrics server implementation can vary from provider-to-provider. But we will probably want to configure the Prometheus instance (which comes out-of-the-box) to collect the custom metrics and export it with the Promethus adapter.
josephburnett
All 2 comments
I think we should implement a very lightweight version of the custom-metrics-server (https://github.com/kubernetes-incubator/custom-metrics-apiserver). After that's done and it works we can write up guides on how to connect the Knative installation to prometheus + configuring the prometheus custom-metrics adapter to work just as well.
That way we'd prevent people to need to buy into Prometheus (there have been concerns about its "heaviness").
markusthoemmes
on 27 Feb 2019
I'm going to knick this off as part of the custom-metrics work I'm doing
/assign
markusthoemmes
on 17 May 2019
Related issues
evankanderson
路
3Comments
tcnghia
路
3Comments
greghaynes
路
6Comments
greghaynes
路
4Comments
ysjjovo
路
5Comments
Most helpful comment
I think we should implement a very lightweight version of the custom-metrics-server (https://github.com/kubernetes-incubator/custom-metrics-apiserver). After that's done and it works we can write up guides on how to connect the Knative installation to prometheus + configuring the prometheus custom-metrics adapter to work just as well.
That way we'd prevent people to need to buy into Prometheus (there have been concerns about its "heaviness").