graham sanderson 2012-07-07, 17:18
Niek Sanders 2012-07-07, 20:30
Graham Sanderson 2012-07-07, 23:37
Jay Kreps 2012-07-09, 16:13
Yes, w.r.t. 3, yes I later saw the client redesign doc which does mention not tracking/balancing consumers for a topic, along with a possible non-blocking consumer
+1 for all of those
One particular use case of interest to me (along with the standard data bus for loading into say HDFS), involve many parts of our system watching the flow of traffic over certain topics and partitions (that they can figure out based on what they are looking for). Particularly in cases where we may be wanting to learn something statistical (estimates) from the data in real time, the playback from the past feature is one of the handiest features of kafka - if we spin up a new box, it can simply playback the last N hours of data . We are also (will be) highly into filtering, with small int values (GCed/reused over time) providing a lookup into zookeeper for metadata including, source machine, application, message encoding (e.g. AVRO) and schema and other metadata. This allows us to do simple bit set tests on the message header (and indeed with messages full of other messages) to trivially reject messages that don't interest us (over and above us having rejected topics we don't care about). Sometimes we will need to listen on all partitions of a topic, and potentially multiple topics, so the non-blocking consumer feature would be nice, but lower priority since we can easily enough listen on multiple threads.
We also plan to feed low level and potentially high volume instrumentation/log events, timed periods etc into kafka (though not with the intention of it actually being persisted after the fact), (Again playback is wonderful for seeing what happened this morning). Here filtering can be very efficient, because if you're looking for sessionId="foo", then you can skip any messages that don't have sessionId in their schema, also based on the pluggable encoding, do 1) a byte level grep for "foo", 2) if that passes a more semantic grep. Note with AVRO being our choice, and having the schema defined by the first word, we avoid repetitious encoding of field names. Also knowing the schema we can do more interesting filters like price < 10000. Anyway the only reason I mention it (and we're at the very early stages) is that we may want to find a way to push this filtering away from our clients (to save them downloading then immediately discarding) whether that being into kafka in some way, or simply via a set of services which share a fast network with the kafka cluster.
On Jul 9, 2012, at 11:13 AM, Jay Kreps wrote:
> 1,2. We don't really provide guarantees outside the message boundary. You
> are right that in the current implementation the nested message sets we use
> for compression do actually fullfill this purpose. Our intention with these
> was really to allow batch compression, and they don't really have any
> advantage over putting multiple records in a single kafka message.
> Likewise, our delivery guarantees are at the message boundary.
> 3. We would be interested in decoupling the offset tracking and also in
> adding some facility to garbage collect unused groups (e.g. if no update to
> a group for > 1 week, delete it). This feature doesn't exist though. If
> anyone wants to add it, it would probably be pretty easy--just a background
> task in the broker that periodically scanned the groups in zk.
> On Sat, Jul 7, 2012 at 10:18 AM, graham sanderson <[EMAIL PROTECTED]> wrote:
>> 1) I would like to guarantee that a group of messages are always delivered
>> in their entirety together (because there is contextual information in
>> messages which precede other messages). I'm a little confused by the use of
>> the term "nested message sets" since I don't really see much in the code
>> (though II don't really know Scala) - perhaps this refers to the fact that
>> you can have a set of messages within a message set file on disk. Anyway, I
>> was curious (and I'm using the Java api now, but may move to the Scala