Re 1), Kafka Streams intentionally close all underlying clients before closing processors since some of closing the processors require shutting down its processor state managers, for example we need to make sure producer's message sends // have all been acked before the state manager records // changelog sent offsets. To complement it we trigger commitAll() before closing the clients.
Guozhang On Sun, Apr 10, 2016 at 9:17 AM, Jay Kreps <j...@confluent.io> wrote: > Also, I wonder if this issue is related: > https://issues.apache.org/jira/browse/KAFKA-3135 > > -Jay > > On Sun, Apr 10, 2016 at 8:58 AM, Jay Kreps <j...@confluent.io> wrote: > > > Two things: > > 1. Caching data in the processor is a bit dangerous since it will be lost > > on failure. Nonetheless, I think you have a point that we should ideally > > close the processors first, then commit in case they send any messages on > > close. > > 2. The issue you describe shouldn't happen for the reason you describe. > > Both the broker and the consumer handle batches of messages so fetching a > > single 1 MB message versus 1024 1KB messages should be the same. The > > proposed max.poll.messages would just effect how many records are handed > > out they will have been fetched and be in memory in the consumer no > matter > > what. I wonder if you could help us trace down what's happening for > > you--maybe provide a simple test case that reproduces the problem? > > > > > > On Sun, Apr 10, 2016 at 6:13 AM, Michael D. Coon < > > mdco...@yahoo.com.invalid> wrote: > > > >> Guozhang, > >> Yes, I'm merging message contents into larger messages before sending > >> to the producer. We have demonstrated that many tiny messages of < 1K > >> causes tremendous slow down on the down stream consumers. Not because of > >> memory contention but because of the broker filling up the max fetch > >> request size by adding hundreds of thousands of tiny messages to the > fetch > >> response. The consumer then has to deal with those messages and it > causes > >> huge latency problems….the broker has to add those hundreds of > thousands of > >> messages to the response. It takes > 5 seconds per fetch to return from > the > >> broker in most cases. In contrast, when I merge messages into bundled > >> single-messages with larger payloads, we get excellent throughput > because > >> there is less polling and the number of messages is reduced. > >> I'm locked into a battle between fetch size constraints and max > >> message size constraints…my max message size can actually spike over 5MB > >> for a single message (non-merged) but most of the time it's < 1K. That's > >> just the kind of data set we're dealing with. So I can't set fetch size > too > >> low or one of these larger messages will come in and break the consumer > >> from being able to process anything. > >> So we either need a way to tell the broker not to fill the max fetch > >> size before returning (max.poll.messages) or I need a way to flush to > the > >> producer when it's about to close my producer. The latter offers the > >> benefit of flushing data that may be the results of processing input > data > >> whose offsets were already committed asynchronously. > >> Mike > >> > >> On Saturday, April 9, 2016 2:27 PM, Guozhang Wang < > wangg...@gmail.com> > >> wrote: > >> > >> > >> Mike, > >> > >> Not clear what do you mean by "buffering up the contents". Producer > itself > >> already did some buffering and batching when sending to Kafka. Did you > >> actually "merge" multiple small messages into one large message before > >> giving it to the producer in the app code? In either case, I am not sure > >> how it will help the downstream consumer memory pressure issue? > >> > >> About bounding the consumer memory usage, we already have some thoughts > >> about that issue and plan to add the memory bounding feature like the > >> producer does in the near future ( > >> https://issues.apache.org/jira/browse/KAFKA-2045), so it won't be a > >> problem > >> for long. And for the "max.poll.messages" config and 0.10.0, just FYI we > >> are shooting to have it released end of this month. > >> > >> Guozhang > >> > >> > >> On Sat, Apr 9, 2016 at 5:59 AM, Michael D. Coon > <mdco...@yahoo.com.invalid > >> > > >> wrote: > >> > >> > Guozhang, > >> > In my processor, I'm buffering up contents of the final messages in > >> > order to make them larger. This is to optimize throughput and avoid > tiny > >> > messages from being injected downstream. So nothing is being pushed to > >> the > >> > producer until my configured thresholds are met in the buffering > >> mechanism. > >> > So as it stands, these messages are left dangling after the producer > >> closes > >> > and, even worse, if periodic commits are happening behind the scenes, > >> the > >> > data is lost on restart. > >> > What we need is a way to notify the processors that everything is > >> > "about" to close so that I can properly flush what I have in memory > out > >> to > >> > the producer. Otherwise, I'm stuck with always sending tiny messages > >> into > >> > kafka--which I know for certain causes problems on down stream > consumers > >> > (where they set a high fetch memory size and it causes hundreds of > >> > thousands of messages to be retrieved at a time…and thus bogs down the > >> > consumer). I think the "max.poll.messages" setting we discussed before > >> > would help here but if it's not available until 0.10, I'm kind of > stuck. > >> > Another option might be to disable periodic commits and only commit > >> > when the processor requests it. This would mitigate some data loss and > >> is > >> > better than nothing. There is still a chance that data in RecordQueue > >> not > >> > yet sent to my processor would be committed but never processed in > this > >> > case. > >> > Another thought I had was to reduce the max fetch size; however, > some > >> > messages can be very large (i.e. data spikes periodically). In this > >> case, > >> > the messages size would exceed my lower max fetch size causing the > >> consumer > >> > to simply stop consuming. So I'm stuck. So either we need to roll in > the > >> > max.poll.messages sooner than 0.10 or maybe a callback mechanism > >> letting me > >> > know that the producer is about to close so I can clear my buffers. > >> > Ideas? > >> > Mike > >> > > >> > On Friday, April 8, 2016 8:24 PM, Guozhang Wang < > wangg...@gmail.com> > >> > wrote: > >> > > >> > > >> > Hi Michael, > >> > > >> > When you call KafkaStreams.close(), it will first trigger a > commitAll() > >> > function, which will 1) flush local state store if necessary; 2) flush > >> > messages buffered in producer; 3) commit offsets on consumer. Then it > >> will > >> > close the producer / consumer clients and shutdown the tasks. So when > >> you > >> > see processor's "close" function triggered, any buffered messages in > the > >> > producer should already been flushed. > >> > > >> > Did you see a different behavior than the above described? > >> > > >> > Guozhang > >> > > >> > > >> > On Fri, Apr 8, 2016 at 12:23 PM, Michael D. Coon > >> <mdco...@yahoo.com.invalid > >> > > > >> > wrote: > >> > > >> > > All, > >> > > I'm seeing my processor's "close" method being called AFTER my > >> > > downstream producer has been closed. I had assumed that on close I > >> would > >> > be > >> > > able to flush whatever I had been buffering up to send to kafka > >> topic. In > >> > > other words, we've seen significant performance differences in > >> building > >> > > flows with small messages and large messages in/out of kafka. So my > >> > > processor buffers up messages to a threshold and flushes those as a > >> > > composite message bundle to improve downstream processing. But if > this > >> > > close method is called AFTER the producer has already been closed, I > >> > would > >> > > have no way to actually flush the final composite bundles to my > topic > >> on > >> > > shutdown. Is there some way to get a call BEFORE producer shutdown > >> > occurs? > >> > > Mike > >> > > > >> > > > >> > > >> > > >> > -- > >> > -- Guozhang > >> > > >> > > >> > > >> > > >> > >> > >> > >> -- > >> -- Guozhang > >> > >> > >> > > > > > -- -- Guozhang
