You are viewing an old version of this page. View the current version.

Compare with Current View Page History

« Previous Version 6 Current »

Status

Current state: Under Discussion

Discussion thread: here

JIRA: KAFKA-13285

Please keep the discussion on the mailing list rather than commenting on the wiki (wiki discussions get unwieldy fast).

Motivation

The *Result classes returned by the Admin client originally had package access constructors. This allowed these constructors to be changed without needing to reason about compatibility, since client code could not directly instantiate these classes.

Over time a minority of the Result classes have ended up with public constructors. So there is an inconsistency in these classes. This caused a problem in the Kafka 3.0 development cycle because a couple of constructor signatures were changed, but one of those constructors was public, which was missed in the original code review.

So to have a consistent API and avoid similar problems in the future the inconsistency ought to be resolved, either by making all the constructors public or or all package access.

KIP-692 proposed to make all these constructors public, using the rationale that it simplifies the creation of mock Admin clients for testing. Note that when the vote for KIP-692 was called it was done on the [DISCUSS] thread, so the lack of votes might not be an sure indication of how the community feels about it.

This KIP proposes taking a step in the opposite direction: To deprecate the public use of these constructors and eventually to make them package access.

Trade off is:

  • While the creation of Admin mocks with package constructors is not _ergonomic_, it is _possible_. The example code in KIP-692 requires two line of codes for each result instance.
  • Having public constructors can hinder the evolution of the Admin client. A number of these constructors have changed their signature during the life of the Admin client. Further change can never be discounted. Sometimes these changes have been non-trivial. For example, while it's easy to overload a constructor (e.g. to add an extra future in a compatible way), overloading can't be used for making compatible changes when a generic type argument needs to be changed. Since we can't anticipate how the Admin Result classes might need to evolve in the future it's simpler and safer to avoid exposing these constructors than it is to need to reason about making source and binary compatible changes.

Public Interfaces

Deprecate the public use of the constructors of the following classes:

  • AlterClientQuotasResult
  • AlterUserScramCredentialsResult
  • DeleteRecordsResult
  • DescribeClientQuotasResult
  • DescribeConsumerGroupsResult
  • ListOffsetsResult

Proposed Changes

Per public interfaces section.

Compatibility, Deprecation, and Migration Plan

The constructors will become package access in a future major release. Existing out-of-package callers will need to use mock libraries or reflection to instantiate these classes. The only anticipated existing out-of-package callers are in other project's test code.

Rejected Alternatives

  • Do nothing.
  • Achieve consistency by making the constructors public, as proposed by KIP-692, and deal with compatibility problems on a KIP-by-KIP- basis.
  • Provide some other public API for the creation of instances of these classes to address the verbosity of mock creation. Providing a supportable API in the face of changing underlying constructors seems like a lot of work for only marginal benefit.

  • Change the Result classes to interfaces. This would be quite a bit more code, for similar protection as a package-private constructor and not improve the situation for people mocking the Admin client.
  • No labels