[KAFKA-3766] Unhandled "not enough replicas" errors in SyncGroup

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Duplicate
• Assignee: Jason Gustafson
• Reporter: Jason Gustafson
• Created: 28/May/16 01:01
• Updated: 29/Jul/16 18:35
• Resolved: 29/Jul/16 18:35

Description

Caught by ijuma. We seem to be missing at least a couple error codes when handling the append log response when writing group metadata to the offsets topic in the SyncGroup handler. In particular, we are missing checks for NOT_ENOUGH_REPLICAS and NOT_ENOUGH_REPLICAS_AFTER_APPEND. Currently these errors are returned directly in the SyncGroup response and cause an exception to be raised to the user.

There are two options to fix this problem:
1. We can continue to return these error codes in the sync group response and add a handler on the client side to retry.
2. We can convert the errors on the server to something like COORDINATOR_NOT_AVAILABLE, which will cause the client to retry with the existing logic.

The second option seems a little nicer to avoid exposing the internal implementation of the SyncGroup request (i.e. that we write group metadata to a partition). It also has the nice side effect of fixing old clients automatically when the server is upgraded.

Issue Links

• duplicates: KAFKA-3590 KafkaConsumer fails with "Messages are rejected since there are fewer in-sync replicas than required." when polling