Informer Pools #2415

How it works

• An informer's identity is captured by an InformerClassifier composed of:
  • target cluster (API server URL, including the remote client used for multi-cluster event sources)
  • resource type (or group/version/kind for generic resources)
  • watched namespace
  • label, field, and shard selectors
  • configured item store
• Event sources whose classifiers are equal share one SharedIndexInformer.
• The pool is reference-counted: the informer is created on first use and stopped only when the last user releases it.
• informerListLimit is intentionally not part of the identity — if two otherwise-equivalent sources request different limits, the existing informer is reused (a warning is logged, first-configured limit wins).
• Indexers are not part of the identity, since they can be added to a shared informer independently (avoid indexer name collisions).
• Works with dynamic event source registration: registering a source for a resource already backed by a running informer reuses it, and the informer replays its current cache to the newly added handler so initial state is
  not missed.

API / configuration

New strategy interface InformerPool with two implementations:

• DefaultInformerPool (default) — shares informers as described above.
• AlwaysNewInformerPool — never shares; one informer per event source (previous behavior). Use to opt out.

Selectable via the configuration service:

Operator operator = new Operator(overrider ->
    overrider.withInformerPool(new AlwaysNewInformerPool()));

Status / notes

• Marked @Experimental: the runtime behavior is production-ready, but the configuration API may still change in a non-backwards-compatible way.
• ConfigurationService.informerPool() must return the same instance per config service (a per-ConfigurationService singleton); AbstractConfigurationService caches it.

Tasks

• InformerPool abstraction + DefaultInformerPool (sharing, reference-counted) and AlwaysNewInformerPool
• ConfigurationServiceOverrider.withInformerPool(...)
• Integration tests: basic sharing, dynamic registration, de-registration/release — run against both pool strategies
• Unit tests: classifier equality, reference counting
• Documentation