feat: implement filter and score

[aramase]

Signed-off-by: Anish Ramasekar [email protected]

• Adds Dockerfile
• Adds build targets for the container
• Implements the filter and score plugins
  • Use PreFilter and PreScore to perform the required computation one time and annotate the pod with the node preference and placement policy used.
  • The node preference annotation is used by Filter and Score to make a decision for the node

[helayoty]

We can use cache instead of adding annotation and create a new addEventHandler from podInformer and update the cache with the unassigned pods. An example can be found here capacity scheduling.

[helayoty]

What if we introduce an internal counter let's call it currentTargetSize and update it everytime we assign a pod to a node?

Then filter the pods with the pp.PodSelector from (nodeList).Get(i).NodeInfo.Requested to know how many pods have already been scheduled and compare the number with the currentTargetSize, wdyt?

I'm trying to get rid of annotating every pod we will schedule.

[aramase]

IMO, having the annotation on every single is more deterministic because we are annotating the required action on it. The annotation indicates where the plugin wanted the pod to end up and we can use to validate if that did get honored. Even if we maintain the internal counter, we're still making the same set of API calls to construct the state?

[helayoty]

I get the value of adding annotations for observability reasons but I am opposed to depending on them in the scheduler core logic.
Also, trying to optimize the query we are asking apiServer to perform has +ve weight. Especially if we have a large cluster with many nodes and thousand of pods.

[helayoty]

As we agreed, we will have this PR merged to be able to continue the integration test and @aramase with open an issue with the opened comments to discuss it later with @khenidak

[aramase]

Merging this PR with #25 so we have core logic to use with integration and e2e tests.