[KAFKA-1011] Decompression and re-compression on MirrorMaker could result in messages being dropped in the pipeline

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Assignee: Guozhang Wang
• Reporter: Guozhang Wang
• Created: 15/Aug/13 21:10
• Updated: 24/May/16 20:57
• Resolved: 24/May/16 20:57

Description

The way MirrorMaker works today is that its consumers could use deep iterator to decompress messages received from the source brokers and its producers could re-compress the messages while sending them to the target brokers. Since MirrorMakers use a centralized data channel for its consumers to pipe messages to its producers, and since producers would compress messages with the same topic within a batch as a single produce request, this could result in messages accepted at the front end of the pipeline being dropped at the target brokers of the MirrorMaker due to MesageSizeTooLargeException if it happens that one batch of messages contain too many messages of the same topic in MirrorMaker's producer. If we can use shallow iterator at the MirrorMaker's consumer side to directly pipe compressed messages this issue can be fixed.

Also as Swapnil pointed out, currently if the MirrorMaker lags and there are large messages in the MirrorMaker queue (large after decompression), it can run into an OutOfMemoryException. Shallow iteration will be very helpful in avoiding this exception.

The proposed solution of this issue is also related to KAFKA-527.