Allow just registerPredicate for the polling to be overridden/configured for PerResourcePollingDependentResources

[Inkromind]

Is your feature request related to a problem? Please describe.

When using the PerResourcePollingDependentResource and you do not want polling (yet) for certain resources, it is hard to discover and cumbersome that you need to override the createEventSource method (which happens to be the only implemented method of that class) just to add the single line withRegisterPredicate(mypredicate::test) to the PerResourcePollingConfigurationBuilder-builder (which also happens to be the only configuration property that is not already out-of-the-box overridable when using PerResourcePollingDependentResource). This has the additional downside that if the code for PerResourcePollingDependentResource.createEventSource is improved, we'd have to update our overridden method accordingly because we want the same behaviour, just only with the registerPredicate.

Describe the solution you'd like

A pollingRegisterPredicate method on PerResourcePollingDependentResource that could be overridden and by default returns null. Eg

@Override
protected Predicate<MyResource> pollingRegisterPredicate() {
   return resource -> "someType".equals(resource.getSpec().type());
}

PerResourcePollingDependentResource could then use this method to configure the registerPredicate when setting op the polling event source.

Additionally, documenting the fact that by default polling is started for every resource (even if the depending reconciler has informer filters) and how this could be adjusted would greatly help new consumers.

Alternatively, an annotation @PollingRegisterPredicate could be added to your class that would then get picked up by PerResourcePollingDependentResource.

Describe alternatives you've considered

For our case (see below), it would also work if the polling picks up on the genericFilter/onAddFilter in the @ControllerConfiguration/@Informer annotations on the depending reconciler, so that the polling does not start if no event is processed by the reconciler.

We recognize that his may be too specific to our use case which may not apply to other people and is also more complex than exposing the existing registerPredicate a bit more which would also be more generic.

Additional context

Our use case is that we have a reconciler with a PerResourcePollingDependentResource which polls an external system. Our reconciler should only reconcile specific resources within a kind (eg based on a static type field). We achieve this by adding a genericFilter to the @ControllerConfiguration/@Informer annotation:

@ControllerConfiguration(informer = @Informer(genericFilter = PredicateClass.class) ...)

We noticed that, even for resources not being handled by the reconciler, polling was being started and the external system was being queried. This puts load on the operator and the external system for resource we know will never need something to happen in the external system.

This behaviour is not (clearly) documented and required some digging around to discover that PerResourcePollingConfigurationBuilder has the method withRegisterPredicate which to control this behaviour, but this method is not used by the PerResourcePollingDependentResource. To use this register predicate, it's therefor required to override the createEventSource method of PerResourcePollingDependentResource, just to add the register predicate:

@Override
protected ExternalResourceCachingEventSource<MyInfo, MyResource> createEventSource(EventSourceContext<MyResource> context) {
		return new PerResourcePollingEventSource<>(
				resourceType(),
				context,
				new PerResourcePollingConfigurationBuilder<>(this, getPollingPeriod())
						.withCacheKeyMapper(this)
						.withName(name())
						.withRegisterPredicate(resource -> "someType".equals(resource.getSpec().type())) // <--- only difference with original method
						.build());
	}

[csviri]

Hi @Inkromind ,

Thank you for the issue, absolutely makes sense.

For our case (see below), it would also work if the polling picks up on the genericFilter/onAddFilter in the @ControllerConfiguration/@Informer annotations on the depending reconciler, so that the polling does not start if no event is processed by the reconciler. We recognize that his may be too specific to our use case which may not apply to other people and is also more complex than exposing the existing registerPredicate a bit more which would also be more generic.

This is an interesting idea might worth to further discuss it, maybe support it with a feature flag eventually (cc @xstefank @metacosm ), also note that you can override the ResourceEventAware related resources to do some additional filtering yourself:

https://github.com/operator-framework/java-operator-sdk/blob/e775349f9e23f46c1148245fe73b5e4b8d108117/operator-framework-core/src/main/java/io/javaoperatorsdk/operator/processing/event/source/polling/PerResourcePollingEventSource.java#L82-L102

Regarding supporting the preidicate with a method what you mentioned:

@Override
protected Predicate<MyResource> pollingRegisterPredicate() {
   return resource -> "someType".equals(resource.getSpec().type());
}

I'm fine with this approach, do you plan to create a PR for this?

[metacosm]

Note that we could probably use an annotation approach to configure things, using an approach similar to what's done for KubernetesDependentResource:

https://github.com/operator-framework/java-operator-sdk/blob/main/operator-framework-core/src/main/java/io/javaoperatorsdk/operator/processing/dependent/kubernetes/KubernetesDependentResource.java#L31-L35

[csviri]

That is an other option for sure. Note that this DR atm does not even implements ConfiguredDependentResource interface. So it might require more work. (Also we have to document it more)

(I personally have a bit issue passing such implementations via annotations, in this case a predicate, since an additional class needs to be created, but since we do it already, I'm fine with it)

[csviri]

@metacosm we don't have this annotation based configuration extension documented atm, will create an issue for that, since that is probably extremely hard to discover.

[Inkromind]

I'm fine with this approach, do you plan to create a PR for this?

Hi, I'm willing to make a PR with the method approach, but I'm currently time and budget constrained so I cannot say when I'd be able to do this. This would be at the earliest in a couple of weeks. If you prefer the annotation approach, it will take additional time (also for me to look into the how its done for KubernetesDependentResource), so that will be harder to fit in/will take longer before I can do it.

[csviri]

I'm personally fine with both approaches.