Hi Rui, Thanks for the proposal. I think it makes a lot of sense to decouple the autoscaler from Kubernetes-related dependencies. A couple of notes when I read the proposal:
1. You propose AutoScalerEventHandler, AutoScalerStateStore, AutoScalerStateStoreFactory, and AutoScalerEventHandler. AutoscalerStateStore is a generic key/value database (methods: "get"/"put"/"delete"). I would propose to refine this interface and make it less general purpose, e.g. add a method for persisting scaling decisions as well as any metrics gathered for the current metric window. For simplicity, I'd even go so far to remove the state store entirely, but rather handle state in the AutoScalerEventHandler which will receive all related scaling and metric collection events, and can keep track of any state. 2. You propose to make the current autoscaler module Kubernetes-agnostic by moving the Kubernetes parts into the main operator module. I think that makes sense since the Kubernetes implementation will continue to be tightly coupled with Kubernetes. The goal of the separate module was to make the autoscaler logic pluggable, but this will continue to be possible with the new "flink-autoscaler" module which contains the autoscaling logic and interfaces. In the long run, the autoscaling logic can move to a separate repository, although this will complicate the release process, so I would defer this unless there is strong interest. 3. The proposal mentions some removal of tests. It is critical for us that all test coverage of the current implementation remains active. It is ok if some of the test coverage only covers the Kubernetes implementation. We can eventually move more tests without Kubernetes significance into the implementation-agnostic autoscaler tests. -Max On Tue, Aug 1, 2023 at 9:46 AM Rui Fan <fan...@apache.org> wrote: > > Hi all, > > I and Samrat(cc'ed) created the FLIP-334[1] to decoupling the autoscaler > and kubernetes. > > Currently, the flink-autoscaler is tightly integrated with Kubernetes. > There are compelling reasons to extend the use of flink-autoscaler to > more types of Flink jobs: > 1. With the recent merge of the Externalized Declarative Resource > Management (FLIP-291[2]), in-place scaling is now supported > across all types of Flink jobs. This development has made scaling Flink on > YARN a straightforward process. > 2. Several discussions[3] within the Flink user community, as observed in > the mail list , have emphasized the necessity of flink-autoscaler > supporting > Flink on YARN. > > Please refer to the FLIP[1] document for more details about the proposed > design and implementation. We welcome any feedback and opinions on > this proposal. > > [1] https://cwiki.apache.org/confluence/x/x4qzDw > [2] > https://cwiki.apache.org/confluence/display/FLINK/FLIP-291%3A+Externalized+Declarative+Resource+Management > [3] https://lists.apache.org/thread/pr0r8hq8kqpzk3q1zrzkl3rp1lz24v7v