I think I've discovered the answer to my second question: according to the code in ZookeeperConsumerConnector.scala, a rebalance derives its offsets from what's already in Zookeeper. Therefore, uncommitted but consumed messages from a given partition will be replayed when the partition is reassigned.
On Fri, Jun 13, 2014 at 3:01 PM, Kyle Banker <kyleban...@gmail.com> wrote: > I'm using Kafka 0.8.1.1. > > I have a simple goal: use the high-level consumer to consume a message > from Kafka, publish the message to a different system, and then commit the > message in Kafka. Based on my reading of the docs and the mailing list, it > seems like this isn't so easy to achieve. Here is my current understanding: > > First, I have to disable auto-commit. If the consumer automatically > commits, then I may lose messages if, for example, my process dies after > consuming but before publishing my message. > > Next, if my app is multi-threaded, I need to either > > a) use a separate consumer per thread (memory-intensive, hard on > Zookeeper) or > b) use a single consumer and assign a KafkaStream to each thread. Then, > when I want to commit, first synchronize all threads using a barrier. > > First question: is this correct so far? > > > Still, it appears that rebalancing may be a problem. In particular, this > sequence of events: > > 1. I'm consuming from a stream tied to two partitions, A and B. > 2. I consume a message, M, from partition A. > 3. Partition A gets assigned to a different consumer. > 4. I choose not to commit M or my process fails. > > Second question: When the partition is reassigned, will the message that I > consumed be automatically committed? If so, then there's no way to get the > reliability I want. > > > Third question: How do the folks at LinkedIn handle this overall use case? > What about other users? > > It seems to me that a lot of the complexity here could be easily addressed > by changing the way in which a partition's message pointer is advanced. > That is, when I consume message M, advance the pointer to message (M - 1) > rather than to M. In other words, calling iterator.next() would imply that > the previously consumed message may be safely committed. If this were the > case, I could simply enable auto-commit and be happy. >
