Decoupling the PartitionData btw producer and fetch may not be a bad idea, especially the way PartitionData is sent to socket is kind of different (for fetch, the sendfile api is used).
Thanks, Jun On Wed, Sep 12, 2012 at 9:32 AM, Jay Kreps <jay.kr...@gmail.com> wrote: > Hey Guys, > > The request format in 0.8 has drifted a bit from the proposal ( > https://cwiki.apache.org/confluence/display/KAFKA/New+Wire+Format+Proposal > ). > A lot of this was new fields that were needed. But some oddities have > slipped in. > > For example we are reusing PartitionData.scala in both the produce request > and the fetch response. This seems like clever code reuse, but in reality > it changes the protocol to add a bunch of non-sensical fields into the > request format. For example in a request one now has to specify a error > code, and initial offset, and a high water mark! > > Is this intentional? I recommend we change this before the release. > Thoughts? > > -Jay >
