I looked at the newly added batch API to Kinesis for inspiration. The response on the batch put is a list of message-ids and their status (offset if success else a failure code).
Ideally, I think the server should fail the entire batch or succeed the entire batch (i.e. no duplicates), but this is pretty hard to implement. Given that, what Kinesis did is probably good compromise (perhaps while we wait for exactly once semantics :)) In addition, perhaps adding a flush() method to the producer to allow for control over when flushes happen might be another good starting point. With the addition of a flush, it's easier to implement a "SyncProducer" by doing something like, flush() -> n x send() -> flush(). This doesn't guarantee that a particular batch isn't broken into two, but with sane batch sizes and sane record sizes, we can assume the guarantee. On Mon, Feb 2, 2015 at 1:48 PM, Gwen Shapira <gshap...@cloudera.com> wrote: > I've been thinking about that too, since both Flume and Sqoop rely on > send(List) API of the old API. > > I'd like to see this API come back, but I'm debating how we'd handle > errors. IIRC, the old API would fail an entire batch on a single > error, which can lead to duplicates. Having N callbacks lets me retry > / save / whatever just the messages that had issues. > > If messages had identifiers from the producer side, we could have the > API call the callback with a list of message-ids and their status. But > they don't :) > > Any thoughts on how you'd like it to work? > > Gwen > > > On Mon, Feb 2, 2015 at 1:38 PM, Pradeep Gollakota <pradeep...@gmail.com> > wrote: > > This is a great question Otis. Like Gwen said, you can accomplish Sync > mode > > by setting the batch size to 1. But this does highlight a shortcoming of > > the new producer API. > > > > I really like the design of the new API and it has really great > properties > > and I'm enjoying working with it. However, once API that I think we're > > lacking is a "batch" API. Currently, I have to iterate over a batch and > > call .send() on each record, which returns n callbacks instead of 1 > > callback for the whole batch. This significantly complicates recovery > logic > > where we need to commit a batch as opposed 1 record at a time. > > > > Do you guys have any plans to add better semantics around batches? > > > > On Mon, Feb 2, 2015 at 1:34 PM, Gwen Shapira <gshap...@cloudera.com> > wrote: > > > >> If I understood the code and Jay correctly - if you wait for the > >> future it will be a similar delay to that of the old sync producer. > >> > >> Put another way, if you test it out and see longer delays than the > >> sync producer had, we need to find out why and fix it. > >> > >> Gwen > >> > >> On Mon, Feb 2, 2015 at 1:27 PM, Otis Gospodnetic > >> <otis.gospodne...@gmail.com> wrote: > >> > Hi, > >> > > >> > Nope, unfortunately it can't do that. X is a remote app, doesn't > listen > >> to > >> > anything external, calls Y via HTTPS. So X has to decide what to do > with > >> > its data based on Y's synchronous response. It has to block until Y > >> > responds. And it wouldn't be pretty, I think, because nobody wants to > >> run > >> > apps that talk to remove servers and hang on to connections more than > >> they > >> > have to. But perhaps that is the only way? Or maybe the answer to > "I'm > >> > guessing the delay would be more or less the same as if the Producer > was > >> > using SYNC mode?" is YES, in which case the connection from X to Y > would > >> be > >> > open for just as long as with a SYNC producer running in Y? > >> > > >> > Thanks, > >> > Otis > >> > -- > >> > Monitoring * Alerting * Anomaly Detection * Centralized Log Management > >> > Solr & Elasticsearch Support * http://sematext.com/ > >> > > >> > > >> > On Mon, Feb 2, 2015 at 4:03 PM, Gwen Shapira <gshap...@cloudera.com> > >> wrote: > >> > > >> >> Can Y have a callback that will handle the notification to X? > >> >> In this case, perhaps Y can be async and X can buffer the data until > >> >> the callback triggers and says "all good" (or resend if the callback > >> >> indicates an error) > >> >> > >> >> On Mon, Feb 2, 2015 at 12:56 PM, Otis Gospodnetic > >> >> <otis.gospodne...@gmail.com> wrote: > >> >> > Hi, > >> >> > > >> >> > Thanks for the info. Here's the use case. We have something up > >> stream > >> >> > sending data, say a log shipper called X. It sends it to some > remote > >> >> > component Y. Y is the Kafka Producer and it puts data into Kafka. > >> But Y > >> >> > needs to send a reply to X and tell it whether it successfully put > all > >> >> its > >> >> > data into Kafka. If it did not, Y wants to tell X to buffer data > >> locally > >> >> > and resend it later. > >> >> > > >> >> > If producer is ONLY async, Y can't easily do that. Or maybe Y > would > >> just > >> >> > need to wait for the Future to come back and only then send the > >> response > >> >> > back to X? If so, I'm guessing the delay would be more or less the > >> same > >> >> as > >> >> > if the Producer was using SYNC mode? > >> >> > > >> >> > Thanks, > >> >> > Otis > >> >> > -- > >> >> > Monitoring * Alerting * Anomaly Detection * Centralized Log > Management > >> >> > Solr & Elasticsearch Support * http://sematext.com/ > >> >> > > >> >> > > >> >> > On Mon, Feb 2, 2015 at 3:13 PM, Jay Kreps <jay.kr...@gmail.com> > >> wrote: > >> >> > > >> >> >> Yeah as Gwen says there is no sync/async mode anymore. There is a > new > >> >> >> configuration which does a lot of what async did in terms of > allowing > >> >> >> batching: > >> >> >> > >> >> >> batch.size - This is the target amount of data per partition the > >> server > >> >> >> will attempt to batch together. > >> >> >> linger.ms - This is the time the producer will wait for more data > >> to be > >> >> >> sent to better batch up writes. The default is 0 (send > immediately). > >> So > >> >> if > >> >> >> you set this to 50 ms the client will send immediately if it has > >> already > >> >> >> filled up its batch, otherwise it will wait to accumulate the > number > >> of > >> >> >> bytes given by batch.size. > >> >> >> > >> >> >> To send asynchronously you do > >> >> >> producer.send(record) > >> >> >> whereas to block on a response you do > >> >> >> producer.send(record).get(); > >> >> >> which will wait for acknowledgement from the server. > >> >> >> > >> >> >> One advantage of this model is that the client will do it's best > to > >> >> batch > >> >> >> under the covers even if linger.ms=0. It will do this by batching > >> any > >> >> data > >> >> >> that arrives while another send is in progress into a single > >> >> >> request--giving a kind of "group commit" effect. > >> >> >> > >> >> >> The hope is that this will be both simpler to understand (a single > >> api > >> >> that > >> >> >> always works the same) and more powerful (you always get a > response > >> with > >> >> >> error and offset information whether or not you choose to use it). > >> >> >> > >> >> >> -Jay > >> >> >> > >> >> >> > >> >> >> On Mon, Feb 2, 2015 at 11:15 AM, Gwen Shapira < > gshap...@cloudera.com > >> > > >> >> >> wrote: > >> >> >> > >> >> >> > If you want to emulate the old sync producer behavior, you need > to > >> set > >> >> >> > the batch size to 1 (in producer config) and wait on the future > >> you > >> >> >> > get from Send (i.e. future.get) > >> >> >> > > >> >> >> > I can't think of good reasons to do so, though. > >> >> >> > > >> >> >> > Gwen > >> >> >> > > >> >> >> > > >> >> >> > On Mon, Feb 2, 2015 at 11:08 AM, Otis Gospodnetic > >> >> >> > <otis.gospodne...@gmail.com> wrote: > >> >> >> > > Hi, > >> >> >> > > > >> >> >> > > Is the plan for New Producer to have ONLY async mode? I'm > asking > >> >> >> because > >> >> >> > > of this info from the Wiki: > >> >> >> > > > >> >> >> > > > >> >> >> > > - The producer will always attempt to batch data and will > >> always > >> >> >> > > immediately return a SendResponse which acts as a Future to > >> allow > >> >> >> the > >> >> >> > > client to await the completion of the request. > >> >> >> > > > >> >> >> > > > >> >> >> > > The word "always" makes me think there will be no sync mode. > >> >> >> > > > >> >> >> > > Thanks, > >> >> >> > > Otis > >> >> >> > > -- > >> >> >> > > Monitoring * Alerting * Anomaly Detection * Centralized Log > >> >> Management > >> >> >> > > Solr & Elasticsearch Support * http://sematext.com/ > >> >> >> > > >> >> >> > >> >> > >> >
