- The first issue is that in certain cases, a broker can use significantly more CPU compared with the 0.8.1 release. This happens when (1) replication is used, (2) the number of partitions per broker is high (say a few thousand) and (3) the cluster is small (say less than 5 brokers).
- The second issue is that when starting a broker with a conflicting broker id, the process of shutting down this broker accidentally removes critical broker registration in ZooKeeper. This may cause some of the topic partitions to be unavailable unexpectedly later on.
- The last two issues are in the new Java producer. We fixed an issue that a message may not be delivered since the producer doesn’t correctly pick an available partition. This tends to happen when the replication factor is only one and one or more brokers fail. Another issue that we fixed is that on an empty Kafka cluster, the CPU usage of the new producer is high after initialization.
Since those issues are relatively critical, we recommend everyone upgrade to the 0.8.2.1 release. Note that all those changes are already included in the Kafka jar shipped with Confluent Platform 1.0.