Hi, I changed auto.offset.reset to largest. The result 30, 50, 40, 40, 35, 30 seconds... Instead of 10 seconds. It looks like attempt to react on backpressure but very slow. In any case it is far from any real time tasks including soft real time. And ok, I agreed with Spark usage with data flows without peaks and with hot reserves of hardware.
If it is interesting for you I added Flink test with same logic. Just run it by ./gradlew test_flink. Just as a reference. https://github.com/rssdev10/spark-kafka-streaming Cheers 2016-07-06 20:12 GMT+02:00 Cody Koeninger <c...@koeninger.org>: > > Yes and I sent you results. It is appropriate only with known parameters > of input data stream. > > No, as far as I can tell from your posts in this thread and your > linked project, you only tested with auto.offset.reset smallest and a > large backlog. That's not what I advised you to do. Don't draw > inaccurate conclusions about Spark DStreams from that test. The > reason you need to specify maxRatePerPartition is because you're > starting with a large backlog and thus a large first batch. If you > were testing an ongoing stream with auto.offset.reset largest, > backpressure alone should be sufficient. > > > > On Wed, Jul 6, 2016 at 12:23 PM, rss rss <rssde...@gmail.com> wrote: > >> If you aren't processing messages as fast as you receive them, you're > >> going to run out of kafka retention regardless of whether you're using > >> Spark or Flink. Again, physics. It's just a question of what > >> compromises you choose. > > > > > > Yes. I wrote about it. But in case of Flink you will have output strictly > > after specified time. If it is impossible to process 1000 messages per 1 > > second but possible process 500, then Flink makes an output for 500. If > only > > 1 message processed, Flink produced an output for one only but after 1 > > second. At the same time Spark processes all 1000 but much longer that 1 > > second in this case. > > > >> that's what backpressure > >> and maxRatePerPartition are for. As long as those are set reasonably, > >> you'll have a reasonably fixed output interval. Have you actually > >> tested this in the way I suggested? > > > > > > Yes and I sent you results. It is appropriate only with known parameters > of > > input data stream. I'm not able to estimate bounds of Sparks usage in > > general. And I'm not about it. I'm about Spark has these limitations. And > > most problem is when a calculation time depends on input data. That is if > > you want to have a stable period of output data generation you have to > use > > computational system with free resources in hot reserve. > > > > In any case thanks, now I understand how to use Spark. > > > > PS: I will continue work with Spark but to minimize emails stream I plan > to > > unsubscribe from this mail list > > > > 2016-07-06 18:55 GMT+02:00 Cody Koeninger <c...@koeninger.org>: > >> > >> If you aren't processing messages as fast as you receive them, you're > >> going to run out of kafka retention regardless of whether you're using > >> Spark or Flink. Again, physics. It's just a question of what > >> compromises you choose. > >> > >> If by "growing of a processing window time of Spark" you mean a > >> processing time that exceeds batch time... that's what backpressure > >> and maxRatePerPartition are for. As long as those are set reasonably, > >> you'll have a reasonably fixed output interval. Have you actually > >> tested this in the way I suggested? > >> > >> On Wed, Jul 6, 2016 at 11:38 AM, rss rss <rssde...@gmail.com> wrote: > >> > Ok, thanks. But really this is not full decision. In case of growing > of > >> > processing time I will have growing of window time. That is really > with > >> > Spark I have 2 points of a latency growing. First is a delay of > >> > processing > >> > of messages in Kafka queue due to physical limitation of a computer > >> > system. > >> > And second one is growing of a processing window time of Spark. In > case > >> > of > >> > Flink there is only first point of delay but time intervals of output > >> > data > >> > are fixed. It is really looks like limitation of Spark. That is if > >> > dataflow > >> > is stable, all is ok. If there are peaks of loading more than > >> > possibility of > >> > computational system or data dependent time of calculation, Spark is > not > >> > able to provide a periodically stable results output. Sometimes this > is > >> > appropriate but sometime this is not appropriate. > >> > > >> > 2016-07-06 18:11 GMT+02:00 Cody Koeninger <c...@koeninger.org>: > >> >> > >> >> Then double the upper limit you have set until the processing time > >> >> approaches the batch time. > >> >> > >> >> On Wed, Jul 6, 2016 at 11:06 AM, rss rss <rssde...@gmail.com> wrote: > >> >> > Ok, with: > >> >> > > >> >> > .set("spark.streaming.backpressure.enabled","true") > >> >> > .set("spark.streaming.receiver.maxRate", "10000") > >> >> > .set("spark.streaming.kafka.maxRatePerPartition", "10000") > >> >> > > >> >> > I have something like > >> >> > > >> >> > > >> >> > > >> >> > > *************************************************************************** > >> >> > Processing time: 5626 > >> >> > Expected time: 10000 > >> >> > Processed messages: 100000 > >> >> > Message example: {"message": 950002, > >> >> > "uid":"81e2d447-69f2-4ce6-a13d-50a1a8b569a0"} > >> >> > Recovered json: > >> >> > {"message":950002,"uid":"81e2d447-69f2-4ce6-a13d-50a1a8b569a0"} > >> >> > > >> >> > That is yes, it works but throughput is much less than without > >> >> > limitations > >> >> > because of this is an absolute upper limit. And time of processing > is > >> >> > half > >> >> > of available. > >> >> > > >> >> > Regarding Spark 2.0 structured streaming I will look it some later. > >> >> > Now > >> >> > I > >> >> > don't know how to strictly measure throughput and latency of this > >> >> > high > >> >> > level > >> >> > API. My aim now is to compare streaming processors. > >> >> > > >> >> > > >> >> > 2016-07-06 17:41 GMT+02:00 Cody Koeninger <c...@koeninger.org>: > >> >> >> > >> >> >> The configuration you set is spark.streaming.receiver.maxRate. > The > >> >> >> direct stream is not a receiver. As I said in my first message in > >> >> >> this thread, and as the pages at > >> >> >> > >> >> >> > >> >> >> > >> >> >> > http://spark.apache.org/docs/latest/streaming-kafka-integration.html#approach-2-direct-approach-no-receivers > >> >> >> and > >> >> >> > >> >> >> > http://spark.apache.org/docs/latest/configuration.html#spark-streaming > >> >> >> also say, use maxRatePerPartition for the direct stream. > >> >> >> > >> >> >> Bottom line, if you have more information than your system can > >> >> >> process > >> >> >> in X amount of time, after X amount of time you can either give > the > >> >> >> wrong answer, or take longer to process. Flink can't violate the > >> >> >> laws > >> >> >> of physics. If the tradeoffs that Flink make are better for your > >> >> >> use > >> >> >> case than the tradeoffs that DStreams make, you may be better off > >> >> >> using Flink (or testing out spark 2.0 structured streaming, > although > >> >> >> there's no kafka integration available for that yet) > >> >> >> > >> >> >> On Wed, Jul 6, 2016 at 10:25 AM, rss rss <rssde...@gmail.com> > wrote: > >> >> >> > ok, thanks. I tried to set minimum max rate for beginning. > >> >> >> > However > >> >> >> > in > >> >> >> > general I don't know initial throughput. BTW it would be very > >> >> >> > useful > >> >> >> > to > >> >> >> > explain it in > >> >> >> > > >> >> >> > > >> >> >> > > >> >> >> > > https://spark.apache.org/docs/latest/streaming-programming-guide.html#performance-tuning > >> >> >> > > >> >> >> > And really with > >> >> >> > > >> >> >> > .set("spark.streaming.backpressure.enabled","true") > >> >> >> > .set("spark.streaming.receiver.maxRate", "10000") > >> >> >> > > >> >> >> > I have same problem: > >> >> >> > > >> >> >> > > >> >> >> > > >> >> >> > > *************************************************************************** > >> >> >> > Processing time: 36994 > >> >> >> > Expected time: 10000 > >> >> >> > Processed messages: 3015830 > >> >> >> > Message example: {"message": 1, > >> >> >> > "uid":"dde09b16-248b-4a2b-8936-109c72eb64cc"} > >> >> >> > Recovered json: > >> >> >> > {"message":1,"uid":"dde09b16-248b-4a2b-8936-109c72eb64cc"} > >> >> >> > > >> >> >> > > >> >> >> > Regarding auto.offset.reset smallest, now it is because of a > test > >> >> >> > and > >> >> >> > I > >> >> >> > want > >> >> >> > to get same messages for each run. But in any case I expect to > >> >> >> > read > >> >> >> > all > >> >> >> > new > >> >> >> > messages from queue. > >> >> >> > > >> >> >> > Regarding backpressure detection. What is to do then a process > >> >> >> > time > >> >> >> > is > >> >> >> > much > >> >> >> > more then input rate? Now I see growing time of processing > instead > >> >> >> > of > >> >> >> > stable > >> >> >> > 10 second and decreasing number of processed messages. Where is > a > >> >> >> > limit > >> >> >> > of > >> >> >> > of backpressure algorithm? > >> >> >> > > >> >> >> > Regarding Flink. I don't know how works core of Flink but you > can > >> >> >> > check > >> >> >> > self > >> >> >> > that Flink will strictly terminate processing of messages by > time. > >> >> >> > Deviation > >> >> >> > of the time window from 10 seconds to several minutes is > >> >> >> > impossible. > >> >> >> > > >> >> >> > PS: I prepared this example to make possible easy observe the > >> >> >> > problem > >> >> >> > and > >> >> >> > fix it if it is a bug. For me it is obvious. May I ask you to be > >> >> >> > near > >> >> >> > to > >> >> >> > this simple source code? In other case I have to think that this > >> >> >> > is a > >> >> >> > technical limitation of Spark to work with unstable data flows. > >> >> >> > > >> >> >> > Cheers > >> >> >> > > >> >> >> > 2016-07-06 16:40 GMT+02:00 Cody Koeninger <c...@koeninger.org>: > >> >> >> >> > >> >> >> >> The direct stream determines batch sizes on the driver, in > >> >> >> >> advance > >> >> >> >> of > >> >> >> >> processing. If you haven't specified a maximum batch size, how > >> >> >> >> would > >> >> >> >> you suggest the backpressure code determine how to limit the > >> >> >> >> first > >> >> >> >> batch? It has no data on throughput until at least one batch > is > >> >> >> >> completed. > >> >> >> >> > >> >> >> >> Again, this is why I'm saying test by producing messages into > >> >> >> >> kafka > >> >> >> >> at > >> >> >> >> a rate comparable to production, not loading a ton of messages > in > >> >> >> >> and > >> >> >> >> starting from auto.offset.reset smallest. > >> >> >> >> > >> >> >> >> Even if you're unwilling to take that advice for some reason, > and > >> >> >> >> unwilling to empirically determine a reasonable maximum > partition > >> >> >> >> size, you should be able to estimate an upper bound such that > the > >> >> >> >> first batch does not encompass your entire kafka retention. > >> >> >> >> Backpressure will kick in once it has some information to work > >> >> >> >> with. > >> >> >> >> > >> >> >> >> On Wed, Jul 6, 2016 at 8:45 AM, rss rss <rssde...@gmail.com> > >> >> >> >> wrote: > >> >> >> >> > Hello, > >> >> >> >> > > >> >> >> >> > thanks, I tried to > >> >> >> >> > .set("spark.streaming.backpressure.enabled","true") > >> >> >> >> > but > >> >> >> >> > result is negative. Therefore I have prepared small test > >> >> >> >> > https://github.com/rssdev10/spark-kafka-streaming > >> >> >> >> > > >> >> >> >> > How to run: > >> >> >> >> > git clone > >> >> >> >> > https://github.com/rssdev10/spark-kafka-streaming.git > >> >> >> >> > cd spark-kafka-streaming > >> >> >> >> > > >> >> >> >> > # downloads kafka and zookeeper > >> >> >> >> > ./gradlew setup > >> >> >> >> > > >> >> >> >> > # run zookeeper, kafka, and run messages generation > >> >> >> >> > ./gradlew test_data_prepare > >> >> >> >> > > >> >> >> >> > And in other console just run: > >> >> >> >> > ./gradlew test_spark > >> >> >> >> > > >> >> >> >> > It is easy to break data generation by CTRL-C. And continue > by > >> >> >> >> > same > >> >> >> >> > command > >> >> >> >> > ./gradlew test_data_prepare > >> >> >> >> > > >> >> >> >> > To stop all run: > >> >> >> >> > ./gradlew stop_all > >> >> >> >> > > >> >> >> >> > Spark test must generate messages each 10 seconds like: > >> >> >> >> > > >> >> >> >> > > >> >> >> >> > > >> >> >> >> > > >> >> >> >> > > *************************************************************************** > >> >> >> >> > Processing time: 33477 > >> >> >> >> > Expected time: 10000 > >> >> >> >> > Processed messages: 2897866 > >> >> >> >> > Message example: {"message": 1, > >> >> >> >> > "uid":"dde09b16-248b-4a2b-8936-109c72eb64cc"} > >> >> >> >> > Recovered json: > >> >> >> >> > {"message":1,"uid":"dde09b16-248b-4a2b-8936-109c72eb64cc"} > >> >> >> >> > > >> >> >> >> > > >> >> >> >> > message is number of fist message in the window. Time values > >> >> >> >> > are > >> >> >> >> > in > >> >> >> >> > milliseconds. > >> >> >> >> > > >> >> >> >> > Brief results: > >> >> >> >> > > >> >> >> >> > Spark always reads all messaged from Kafka after first > >> >> >> >> > connection > >> >> >> >> > independently on dstream or window size time. It looks like a > >> >> >> >> > bug. > >> >> >> >> > When processing speed in Spark's app is near to speed of data > >> >> >> >> > generation > >> >> >> >> > all > >> >> >> >> > is ok. > >> >> >> >> > I added delayFactor in > >> >> >> >> > > >> >> >> >> > > >> >> >> >> > > >> >> >> >> > > >> >> >> >> > > https://github.com/rssdev10/spark-kafka-streaming/blob/master/src/main/java/SparkStreamingConsumer.java > >> >> >> >> > to emulate slow processing. And streaming process is in > >> >> >> >> > degradation. > >> >> >> >> > When > >> >> >> >> > delayFactor=0 it looks like stable process. > >> >> >> >> > > >> >> >> >> > > >> >> >> >> > Cheers > >> >> >> >> > > >> >> >> >> > > >> >> >> >> > 2016-07-05 17:51 GMT+02:00 Cody Koeninger < > c...@koeninger.org>: > >> >> >> >> >> > >> >> >> >> >> Test by producing messages into kafka at a rate comparable > to > >> >> >> >> >> what > >> >> >> >> >> you > >> >> >> >> >> expect in production. > >> >> >> >> >> > >> >> >> >> >> Test with backpressure turned on, it doesn't require you to > >> >> >> >> >> specify > >> >> >> >> >> a > >> >> >> >> >> fixed limit on number of messages and will do its best to > >> >> >> >> >> maintain > >> >> >> >> >> batch timing. Or you could empirically determine a > reasonable > >> >> >> >> >> fixed > >> >> >> >> >> limit. > >> >> >> >> >> > >> >> >> >> >> Setting up a kafka topic with way more static messages in it > >> >> >> >> >> than > >> >> >> >> >> your > >> >> >> >> >> system can handle in one batch, and then starting a stream > >> >> >> >> >> from > >> >> >> >> >> the > >> >> >> >> >> beginning of it without turning on backpressure or limiting > >> >> >> >> >> the > >> >> >> >> >> number > >> >> >> >> >> of messages... isn't a reasonable way to test steady state > >> >> >> >> >> performance. Flink can't magically give you a correct > answer > >> >> >> >> >> under > >> >> >> >> >> those circumstances either. > >> >> >> >> >> > >> >> >> >> >> On Tue, Jul 5, 2016 at 10:41 AM, rss rss < > rssde...@gmail.com> > >> >> >> >> >> wrote: > >> >> >> >> >> > Hi, thanks. > >> >> >> >> >> > > >> >> >> >> >> > I know about possibility to limit number of messages. > But > >> >> >> >> >> > the > >> >> >> >> >> > problem > >> >> >> >> >> > is > >> >> >> >> >> > I don't know number of messages which the system able to > >> >> >> >> >> > process. > >> >> >> >> >> > It > >> >> >> >> >> > depends > >> >> >> >> >> > on data. The example is very simple. I need a strict > >> >> >> >> >> > response > >> >> >> >> >> > after > >> >> >> >> >> > specified time. Something like soft real time. In case of > >> >> >> >> >> > Flink > >> >> >> >> >> > I > >> >> >> >> >> > able > >> >> >> >> >> > to > >> >> >> >> >> > setup strict time of processing like this: > >> >> >> >> >> > > >> >> >> >> >> > KeyedStream<Event, Integer> keyed = > >> >> >> >> >> > eventStream.keyBy(event.userId.getBytes()[0] % partNum); > >> >> >> >> >> > WindowedStream<Event, Integer, TimeWindow> uniqUsersWin = > >> >> >> >> >> > keyed.timeWindow( > >> >> >> >> >> > Time.seconds(10) ); > >> >> >> >> >> > DataStream<Aggregator> uniqUsers = > >> >> >> >> >> > uniq.trigger(ProcessingTimeTrigger.create()) > >> >> >> >> >> > .fold(new Aggregator(), new FoldFunction<Event, > >> >> >> >> >> > Aggregator>() > >> >> >> >> >> > { > >> >> >> >> >> > @Override > >> >> >> >> >> > public Aggregator fold(Aggregator accumulator, > >> >> >> >> >> > Event > >> >> >> >> >> > value) > >> >> >> >> >> > throws Exception { > >> >> >> >> >> > accumulator.add(event.userId); > >> >> >> >> >> > return accumulator; > >> >> >> >> >> > } > >> >> >> >> >> > }); > >> >> >> >> >> > > >> >> >> >> >> > uniq.print(); > >> >> >> >> >> > > >> >> >> >> >> > And I can see results every 10 seconds independently on > >> >> >> >> >> > input > >> >> >> >> >> > data > >> >> >> >> >> > stream. > >> >> >> >> >> > Is it possible something in Spark? > >> >> >> >> >> > > >> >> >> >> >> > Regarding zeros in my example the reason I have prepared > >> >> >> >> >> > message > >> >> >> >> >> > queue > >> >> >> >> >> > in > >> >> >> >> >> > Kafka for the tests. If I add some messages after I able > to > >> >> >> >> >> > see > >> >> >> >> >> > new > >> >> >> >> >> > messages. But in any case I need first response after 10 > >> >> >> >> >> > second. > >> >> >> >> >> > Not > >> >> >> >> >> > minutes > >> >> >> >> >> > or hours after. > >> >> >> >> >> > > >> >> >> >> >> > Thanks. > >> >> >> >> >> > > >> >> >> >> >> > > >> >> >> >> >> > > >> >> >> >> >> > 2016-07-05 17:12 GMT+02:00 Cody Koeninger > >> >> >> >> >> > <c...@koeninger.org>: > >> >> >> >> >> >> > >> >> >> >> >> >> If you're talking about limiting the number of messages > per > >> >> >> >> >> >> batch > >> >> >> >> >> >> to > >> >> >> >> >> >> try and keep from exceeding batch time, see > >> >> >> >> >> >> > >> >> >> >> >> >> http://spark.apache.org/docs/latest/configuration.html > >> >> >> >> >> >> > >> >> >> >> >> >> look for backpressure and maxRatePerParition > >> >> >> >> >> >> > >> >> >> >> >> >> > >> >> >> >> >> >> But if you're only seeing zeros after your job runs for a > >> >> >> >> >> >> minute, > >> >> >> >> >> >> it > >> >> >> >> >> >> sounds like something else is wrong. > >> >> >> >> >> >> > >> >> >> >> >> >> > >> >> >> >> >> >> On Tue, Jul 5, 2016 at 10:02 AM, rss rss > >> >> >> >> >> >> <rssde...@gmail.com> > >> >> >> >> >> >> wrote: > >> >> >> >> >> >> > Hello, > >> >> >> >> >> >> > > >> >> >> >> >> >> > I'm trying to organize processing of messages from > >> >> >> >> >> >> > Kafka. > >> >> >> >> >> >> > And > >> >> >> >> >> >> > there > >> >> >> >> >> >> > is > >> >> >> >> >> >> > a > >> >> >> >> >> >> > typical case when a number of messages in kafka's queue > >> >> >> >> >> >> > is > >> >> >> >> >> >> > more > >> >> >> >> >> >> > then > >> >> >> >> >> >> > Spark > >> >> >> >> >> >> > app's possibilities to process. But I need a strong > time > >> >> >> >> >> >> > limit > >> >> >> >> >> >> > to > >> >> >> >> >> >> > prepare > >> >> >> >> >> >> > result for at least for a part of data. > >> >> >> >> >> >> > > >> >> >> >> >> >> > Code example: > >> >> >> >> >> >> > > >> >> >> >> >> >> > SparkConf sparkConf = new SparkConf() > >> >> >> >> >> >> > .setAppName("Spark") > >> >> >> >> >> >> > .setMaster("local"); > >> >> >> >> >> >> > > >> >> >> >> >> >> > JavaStreamingContext jssc = new > >> >> >> >> >> >> > JavaStreamingContext(sparkConf, > >> >> >> >> >> >> > Milliseconds.apply(5000)); > >> >> >> >> >> >> > > >> >> >> >> >> >> > jssc.checkpoint("/tmp/spark_checkpoint"); > >> >> >> >> >> >> > > >> >> >> >> >> >> > Set<String> topicMap = new > >> >> >> >> >> >> > HashSet<>(Arrays.asList(topicList.split(","))); > >> >> >> >> >> >> > Map<String, String> kafkaParams = new > >> >> >> >> >> >> > HashMap<String, > >> >> >> >> >> >> > String>() > >> >> >> >> >> >> > { > >> >> >> >> >> >> > { > >> >> >> >> >> >> > put("metadata.broker.list", > >> >> >> >> >> >> > bootstrapServers); > >> >> >> >> >> >> > put("auto.offset.reset", "smallest"); > >> >> >> >> >> >> > } > >> >> >> >> >> >> > }; > >> >> >> >> >> >> > > >> >> >> >> >> >> > JavaPairInputDStream<String, String> messages = > >> >> >> >> >> >> > KafkaUtils.createDirectStream(jssc, > >> >> >> >> >> >> > String.class, > >> >> >> >> >> >> > String.class, > >> >> >> >> >> >> > StringDecoder.class, > >> >> >> >> >> >> > StringDecoder.class, > >> >> >> >> >> >> > kafkaParams, > >> >> >> >> >> >> > topicMap); > >> >> >> >> >> >> > > >> >> >> >> >> >> > messages.countByWindow(Seconds.apply(10), > >> >> >> >> >> >> > Milliseconds.apply(5000)) > >> >> >> >> >> >> > .map(x -> {System.out.println(x); > return > >> >> >> >> >> >> > x;}) > >> >> >> >> >> >> > > .dstream().saveAsTextFiles("/tmp/spark", > >> >> >> >> >> >> > "spark-streaming"); > >> >> >> >> >> >> > > >> >> >> >> >> >> > > >> >> >> >> >> >> > I need to see a result of window operation each 10 > >> >> >> >> >> >> > seconds > >> >> >> >> >> >> > (this > >> >> >> >> >> >> > is > >> >> >> >> >> >> > only > >> >> >> >> >> >> > simplest example). But really with my test data ~10M > >> >> >> >> >> >> > messages I > >> >> >> >> >> >> > have > >> >> >> >> >> >> > first > >> >> >> >> >> >> > result a minute after and further I see only zeros. Is > a > >> >> >> >> >> >> > way > >> >> >> >> >> >> > to > >> >> >> >> >> >> > limit > >> >> >> >> >> >> > processing time to guarantee a response in specified > time > >> >> >> >> >> >> > like > >> >> >> >> >> >> > Apache > >> >> >> >> >> >> > Flink's triggers? > >> >> >> >> >> >> > > >> >> >> >> >> >> > Thanks. > >> >> >> >> >> > > >> >> >> >> >> > > >> >> >> >> > > >> >> >> >> > > >> >> >> > > >> >> >> > > >> >> > > >> >> > > >> > > >> > > > > > >
