Page History

...

Deprecating a feature version is an incompatible change, which requires a major release of Kafka (see Non-goals section). This requirement translates to increasing the cluster-wide finalized minimum version level of one or more features in the ZK '/features' node. It is important to note that the minimum version level can not be mutated via the Controller API. This is because, the minimum version level is usually increased only to indicate the intent to stop support for a certain feature version. We would usually deprecate features during broker releases, after prior announcements. Therefore, this is not a dynamic operation, and such a mutation is not supported through the ApiKeys.UPDATE_FEATURES controller API.

Instead, it is best if the version deprecation is activated through the controller. This is the proposed approach to achieve itthe same:
Feature advertisements from each broker in the cluster are reactively picked up by the controller via ZK watches. Today these watches are already set up by the controller on every broker’s ephemeral ZK node (at '/brokers/ids/<id>'). The contents of the broker nodes are cached in memory in the Controller. Using these contents, the controller should be able to track the highest supported feature minimum version exported by any broker in the cluster. Then, whenever a Kafka release advances the supported feature minimum version (to deprecate feature version), the controller picks up this change and persists the highest supported minimum version to ZK as the finalized minimum version level for the feature. This change by the controller will deprecate the finalized feature versions, unto 1 below the highest supported minimum version.

It is important that an admin should make sure that no clients are using a deprecated feature version (e.g. using the client version metric) before deploying a release that deprecates it.

Client discovery of finalized feature versions

...