azure-pipelines scaler: exact match demands (don't match jobs without capability)

[patst]

Proposal

We are using the azure-pipelines scaler and configuring the capability support introduced with #2328 .

Example trigger config:

  triggers:
    - type: azure-pipelines
      metadata:
        targetPipelinesQueueLength: "{{ $.Values.targetPipelinesQueueLength }}"
        poolName: "{{ $.Values.poolName }}"
        organizationURLFromEnv: "AZP_URL"
        demands: "cap-4gb"
      authenticationRef:
        name: pipeline-trigger-auth

I would like to have an attribute like requireDemands to configure that a scaled object for above trigger is only scaled if the Azure DevOps job is requesting a capability named "4gb".

Use-Case

We have different types of build agents running inside a kubernetes cluster.
Our internal developer use these buildagents and they are autoscaled by KEDAs azure-pipelines scaler.

All buildagents are in the same Nodepool (e.g. DEV) but have different configurations regarding CPU and RAM limits.

At the moment everything works fine, when a demand is configured an the Azure DevOPs pipeline:

# from an pipeline.yaml
demands: [ "cap-4gb" ]

If the pipelines to not specify a demand all buildagent types are scaled and one auf dem randomly picks the waiting job. The others are scaled even if there is no work left for them.

This behaviour is documented in the docs (Note: If more than one scaling definition is able to fulfill the demands of the job then they will both spin up an agent.).

Wouldn't it be great to add the configuration to only scale up if the demands configured in an trigger are all requested by and job?

I think the logic could be implemented here: https://github.com/kedacore/keda/blob/main/pkg/scalers/azure_pipelines_scaler.go#L367

Just check if every demand from the trigger is requested by the job.

Is this a feature you are interested in implementing yourself?

Yes

Anything else?

An alternative to this would be a configuration flag like matchEmptyDemands: false (with a default value true) which only checks if the Job requests anything as demand. GitLab runners allow this with a flag called --run-untagged=false

This way we would be able to provide a "default" scaledObject without any demands.

Our usecase would be fulfilled by this behaviour and it is probably even easier to implement.

WDYT?

[JorTurFer]

Hi @patst,
recaping: You would like to have another parameter like requireDemands and use it to scale ONLY if the pipeline has that requirement, to avoid cases where this agent is scaled when a "regular" job is scheduled, right?
This makes sense totally to me, but IDK if I'd do it with another parameter or just with a boolean like forDemandsOnly or similar, to reuse the same values given for demands (to reduce the possibilities of misconfiguration).
WDYT?
I see you are open to contribute with this improvement, I can assign this issue to you if you are still open to it

[patst]

@JorTurFer thanks for your reply.
Your summary is correct. To keep it simple

• I would go for a another parameter with type boolean (named: requireAllDemands )
• if true: check if not only the trigger fulfils a demands but the job requests all demands configured for the trigger
• change to docs to reflect that behaviour

I will try to come up with a pull request the next days

[JorTurFer]

nice!
I assign this issue to you. Thanks a lot! ❤️

[EugeneLugovtsov]

@patst @JorTurFer thank you. This feature is very important and useful for us as well.

[Eldarrin]

HI @patst ,

Could you look how this intersects with: #4195

Cheers, Andy

[patst]

HI @patst ,

Could you look how this intersects with: #4195

Cheers, Andy

@Eldarrin tried to add my understanding to #4195 . Hope that helps. If there any additions to the proposed solution here feel free to add.