Vadim, I don't know under exactly what conditions it happens but the behavior in this test seems to reliably reproduce it: https://github.com/v-a/poseidon/commit/d0ac928e0967e1eaf5b92b403103c4f0dc8fd7f7
-Bob On 27 June 2013 14:29, Vadim Keylis <vkeylis2...@gmail.com> wrote: > Jay. I assume this is problem exists in the consumer. How this can this > problem be triggered so I could test my high level consumer. > > Thanks > > On Jun 26, 2013, at 9:21 AM, Jay Kreps <jay.kr...@gmail.com> wrote: > > > Yeah, that is true. I thought I documented that, but looking at the > > protocol docs, it looks like I didn't. > > > > I agree this is kind of a pain in the ass. It was an important > > optimization in 0.7 because we didn't know where the message > > boundaries were but in 0.8 we have a fast way to compute message > > boundaries and in fact we normally don't give out partial messages, I > > think this happens when you hit the size threshold of your fetch > > request (e.g. 1MB) instead of searching for the nearest message > > boundary we give you that chunk of log. I think we should consider > > just fixing it entirely in the next release--the perf hit is pretty > > minor and it is an annoyance and source of bugs for clients. > > > > For now you have to handle it, so I added documentation to the protocol > wiki. > > > > Sorry! > > > > -Jay > > > > On Wed, Jun 26, 2013 at 8:59 AM, Bob Potter <bobby.pot...@gmail.com> > wrote: > >> Howdy, > >> > >> I'm developing a client for kafka 0.8. It looks like a fetch response > will > >> sometimes end with a partial message. I understand why this might be the > >> case but it was unexpected and as far as I can tell undocumented. > >> > >> Is my understanding correct or am i missing something? > >> > >> -Bob > -- Bob Potter