option1 would take a throughput hit as you are trying to commit one message at a time. Option 2 is pretty widely used at LinkedIn and am pretty sure at several other places as well. Option 3 is essentially what the high level consumer does under the covers already. It prefetches data in batches from the server to provide high throughput.
On Wed, Aug 13, 2014 at 2:20 AM, Anand Nalya <anand.na...@gmail.com> wrote: > Hi Jim, > > In one of the applications, we implemented option #1: > > messageList = getNext(1000) > process(messageList) > commit() > > In case of failure, this resulted in duplicate processing for at most 1000 > records per partition. > > Regards, > Anand > > > On 1 August 2014 20:35, Jim <jimi...@gmail.com> wrote: > > > Thanks Guozhang, > > > > I was looking for actual real world workflows. I realize you can commit > > after each message but if you’re using ZK for offsets for instance you’ll > > put too much write load on the nodes and crush your throughput. So I was > > interested in batching strategies people have used that balance high/full > > throughput and fully committed events. > > > > > > On Thu, Jul 31, 2014 at 8:16 AM, Guozhang Wang <wangg...@gmail.com> > wrote: > > > > > Hi Jim, > > > > > > Whether to use high level or simple consumer depends on your use case. > If > > > you need to manually manage partition assignments among your consumers, > > or > > > you need to commit your offsets elsewhere than ZK, or you do not want > > auto > > > rebalancing of consumers upon failures etc, you will use simple > > consumers; > > > otherwise you use high level consumer. > > > > > > From your description of pulling a batch of messages it seems you are > > > currently using the simple consumer. Suppose you are using the high > level > > > consumer, to achieve at-lease-once basically you can do sth like: > > > > > > message = consumer.iter.next() > > > process(message) > > > consumer.commit() > > > > > > which is effectively the same as option 2 for using a simple consumer. > Of > > > course, doing so has a heavy overhead of one-commit-per-message, you > can > > > also do option 1, by the cost of duplicates, which is tolerable for > > > at-least-once. > > > > > > Guozhang > > > > > > > > > On Wed, Jul 30, 2014 at 8:25 PM, Jim <jimi...@gmail.com> wrote: > > > > > > > Curious on a couple questions... > > > > > > > > Are most people(are you?) using the simple consumer vs the high level > > > > consumer in production? > > > > > > > > > > > > What is the common processing paradigm for maintaining a full > pipeline > > > for > > > > kafka consumers for at-least-once messaging? E.g. you pull a batch of > > > 1000 > > > > messages and: > > > > > > > > option 1. > > > > you wait for the slowest worker to finish working on that message, > when > > > you > > > > get back 1000 acks internally you commit your offset and pull another > > > batch > > > > > > > > option 2. > > > > you feed your workers n msgs at a time in sequence and move your > offset > > > up > > > > as you work through your batch > > > > > > > > option 3. > > > > you maintain a full stream of 1000 messages ideally and as you get > acks > > > > back from your workers you see if you can move your offset up in the > > > stream > > > > to pull n more messages to fill up your pipeline so you're not > blocked > > by > > > > the slowest consumer (probability wise) > > > > > > > > > > > > any good docs or articles on the subject would be great, thanks! > > > > > > > > > > > > > > > > -- > > > -- Guozhang > > > > > >
