Details
-
Sub-task
-
Status: Resolved
-
Blocker
-
Resolution: Fixed
-
3.6.0
-
None
-
None
Description
KIP-890 Part 1 tries to address hanging transactions on old clients. Thus, the produce version can not be bumped and no new errors can be added. Currently we use the java client's notion of retriable and abortable errors – retriable errors are defined as such by extending the retriable error class, fatal errors are defined explicitly, and abortable errors are the remaining. However, many other clients treat non specified errors as fatal and that means many retriable errors kill the application. This is not ideal.
While reviewing https://github.com/apache/kafka/pull/14370 I added some of the documentation for the returned errors in the produce response as well.
There were concerns about the new errors:
- {@link Errors#COORDINATOR_LOAD_IN_PROGRESS}
- {@link Errors#COORDINATOR_NOT_AVAILABLE}
- {@link Errors#INVALID_TXN_STATE}
- {@link Errors#INVALID_PRODUCER_ID_MAPPING}
- {@link Errors#CONCURRENT_TRANSACTIONS}
The coordinator load, not available, and concurrent transactions errors should be retriable.
The invalid txn state and pid mapping errors should be abortable.
This is how older java clients handle the errors, but it is unclear how other clients handle them. It seems that rdkafka (for example) treats the abortable errors as fatal instead. The coordinator errors are retriable but not the concurrent transactions error. Generally anything not specified otherwise is fatal.
It seems acceptable for the abortable errors to be fatal on some clients since the error is likely on a zombie producer or in a state that may be harder to recover from. However, for the retriable errors, we can return NOT_ENOUGH_REPLICAS which is a known retriable response. We can use the produce api's response string to specify the real cause of the error for debugging.
There were trade-offs between making the older clients work and for clarity in errors. This seems to be the best compromise.
Attachments
Issue Links
- links to