You are viewing a plain text version of this content. The canonical link for it is here.
Posted to user@spark.apache.org by gaurav sharma <sh...@gmail.com> on 2015/07/12 13:30:26 UTC
Re: createDirectStream and Stats
Hi guys,
I too am facing similar challenge with directstream.
I have 3 Kafka Partitions.
and running spark on 18 cores, with parallelism level set to 48.
I am running simple map-reduce job on incoming stream.
Though the reduce stage takes milliseconds-seconds for around 15 million
packets, the Map stage takes around 4-5 minutes, since it creates only 3
tasks for Map stage(I believe 3 tasks because I have 3 kafka partitions and
the
JavaPairDStream<String, String> kafkaStream =
KafkaConnector.getKafkaStream(jsc);
kafkaStream that i create in code is the parent Rdd for Map Job, so it
would create only 3 tasks)
I have 10 such jobs similar to above one working on same KafkaStream i
create
Could you guys please advise, if repartitioning the KafkaStream (taking
into account the rechuffle at repartition stage) would optimize my overall
batch processing time.
On Sat, Jun 20, 2015 at 7:24 PM, Silvio Fiorito <
silvio.fiorito@granturing.com> wrote:
> Are you sure you were using all 100 executors even with the receiver
> model? Because in receiver mode, the number of partitions is dependent on
> the batch duration and block interval. It may not necessarily map directly
> to the number of executors in your app unless you've adjusted the block
> interval and batch duration.
>
> *From:* Tim Smith <se...@gmail.com>
> *Sent:* Friday, June 19, 2015 10:36 PM
> *To:* user@spark.apache.org
>
> I did try without repartition, initially, but that was even more
> horrible because instead of the allocated 100 executors, only 30 (which is
> the number of kafka partitions) would have to do the work. The "MyFunc" is
> a CPU bound task so adding more memory per executor wouldn't help and I saw
> that each of the 30 executors was only using one thread/core on each Spark
> box. I could go and play with threading in MyFunc but I don't want to mess
> with threading with all the parallelism already involved and I don't think
> in-app threading outside of what the framework does is really desirable.
>
> With repartition, there is shuffle involved, but at least the
> computation load spreads across all 100 executors instead of just 30.
>
>
>
>
> On Fri, Jun 19, 2015 at 7:14 PM, Cody Koeninger <co...@koeninger.org>
> wrote:
>
>> If that's the case, you're still only using as many read executors as
>> there are kafka partitions.
>>
>> I'd remove the repartition. If you weren't doing any shuffles in the
>> old job, and are doing a shuffle in the new job, it's not really comparable.
>>
>> On Fri, Jun 19, 2015 at 8:16 PM, Tim Smith <se...@gmail.com> wrote:
>>
>>> On Fri, Jun 19, 2015 at 5:15 PM, Tathagata Das <td...@databricks.com>
>>> wrote:
>>>
>>>> Also, can you find from the spark UI the break up of the stages in each
>>>> batch's jobs, and find which stage is taking more time after a while?
>>>>
>>>
>>> Sure, will try to debug/troubleshoot. Are there enhancements to this
>>> specific API between 1.3 and 1.4 that can substantially change it's
>>> behaviour?
>>>
>>>
>>>> On Fri, Jun 19, 2015 at 4:51 PM, Cody Koeninger <co...@koeninger.org>
>>>> wrote:
>>>>
>>>>> when you say your old version was
>>>>>
>>>>> k = createStream .....
>>>>>
>>>>> were you manually creating multiple receivers? Because otherwise
>>>>> you're only using one receiver on one executor...
>>>>>
>>>>
>>> Yes, sorry, the earlier/stable version was more like:
>>> kInStreams = (1 to n).map{_ => KafkaUtils.createStream ............ //
>>> n being the number of kafka partitions, 1 receiver per partition
>>> val k = ssc.union(kInStreams)
>>> val dataout = k.map(x=>myFunc(x._2,someParams))
>>> dataout.foreachRDD ( rdd => rdd.foreachPartition(rec => {
>>> myOutputFunc.write(rec) })
>>>
>>> Thanks,
>>>
>>> Tim
>>>
>>>
>>>
>>>
>>>
>>>>
>>>>> If that's the case I'd try direct stream without the repartitioning.
>>>>>
>>>>>
>>>>> On Fri, Jun 19, 2015 at 6:43 PM, Tim Smith <se...@gmail.com> wrote:
>>>>>
>>>>>> Essentially, I went from:
>>>>>> k = createStream .....
>>>>>> val dataout = k.map(x=>myFunc(x._2,someParams))
>>>>>> dataout.foreachRDD ( rdd => rdd.foreachPartition(rec => {
>>>>>> myOutputFunc.write(rec) })
>>>>>>
>>>>>> To:
>>>>>> kIn = createDirectStream .....
>>>>>> k = kIn.repartition(numberOfExecutors) //since #kafka partitions <
>>>>>> #spark-executors
>>>>>> val dataout = k.map(x=>myFunc(x._2,someParams))
>>>>>> dataout.foreachRDD ( rdd => rdd.foreachPartition(rec => {
>>>>>> myOutputFunc.write(rec) })
>>>>>>
>>>>>> With the new API, the app starts up and works fine for a while but
>>>>>> I guess starts to deteriorate after a while. With the existing API
>>>>>> "createStream", the app does deteriorate but over a much longer period,
>>>>>> hours vs days.
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>>> On Fri, Jun 19, 2015 at 1:40 PM, Tathagata Das <td...@databricks.com>
>>>>>> wrote:
>>>>>>
>>>>>>> Yes, please tell us what operation are you using.
>>>>>>>
>>>>>>> TD
>>>>>>>
>>>>>>> On Fri, Jun 19, 2015 at 11:42 AM, Cody Koeninger <cody@koeninger.org
>>>>>>> > wrote:
>>>>>>>
>>>>>>>> Is there any more info you can provide / relevant code?
>>>>>>>>
>>>>>>>> On Fri, Jun 19, 2015 at 1:23 PM, Tim Smith <se...@gmail.com>
>>>>>>>> wrote:
>>>>>>>>
>>>>>>>>> Update on performance of the new API: the new code using the
>>>>>>>>> createDirectStream API ran overnight and when I checked the app state in
>>>>>>>>> the morning, there were massive scheduling delays :(
>>>>>>>>>
>>>>>>>>> Not sure why and haven't investigated a whole lot. For now,
>>>>>>>>> switched back to the createStream API build of my app. Yes, for the record,
>>>>>>>>> this is with CDH 5.4.1 and Spark 1.3.
>>>>>>>>>
>>>>>>>>>
>>>>>>>>>
>>>>>>>>> On Thu, Jun 18, 2015 at 7:05 PM, Tim Smith <se...@gmail.com>
>>>>>>>>> wrote:
>>>>>>>>>
>>>>>>>>>> Thanks for the super-fast response, TD :)
>>>>>>>>>>
>>>>>>>>>> I will now go bug my hadoop vendor to upgrade from 1.3 to 1.4.
>>>>>>>>>> Cloudera, are you listening? :D
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>> On Thu, Jun 18, 2015 at 7:02 PM, Tathagata Das <
>>>>>>>>>> tathagata.das1565@gmail.com> wrote:
>>>>>>>>>>
>>>>>>>>>>> Are you using Spark 1.3.x ? That explains. This issue has been
>>>>>>>>>>> fixed in Spark 1.4.0. Bonus you get a fancy new streaming UI with more
>>>>>>>>>>> awesome stats. :)
>>>>>>>>>>>
>>>>>>>>>>> On Thu, Jun 18, 2015 at 7:01 PM, Tim Smith <se...@gmail.com>
>>>>>>>>>>> wrote:
>>>>>>>>>>>
>>>>>>>>>>>> Hi,
>>>>>>>>>>>>
>>>>>>>>>>>> I just switched from "createStream" to the
>>>>>>>>>>>> "createDirectStream" API for kafka and while things otherwise seem happy,
>>>>>>>>>>>> the first thing I noticed is that stream/receiver stats are gone from the
>>>>>>>>>>>> Spark UI :( Those stats were very handy for keeping an eye on health of the
>>>>>>>>>>>> app.
>>>>>>>>>>>>
>>>>>>>>>>>> What's the best way to re-create those in the Spark UI?
>>>>>>>>>>>> Maintain Accumulators? Would be really nice to get back receiver-like stats
>>>>>>>>>>>> even though I understand that "createDirectStream" is a receiver-less
>>>>>>>>>>>> design.
>>>>>>>>>>>>
>>>>>>>>>>>> Thanks,
>>>>>>>>>>>>
>>>>>>>>>>>> Tim
>>>>>>>>>>>>
>>>>>>>>>>>>
>>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>
>>>>>>>>
>>>>>>>
>>>>>>
>>>>>
>>>>
>>>
>>
>
Re: createDirectStream and Stats
Posted by Cody Koeninger <co...@koeninger.org>.
Reading from kafka is always going to be bounded by the number of kafka
partitions you have, regardless of what you're using to read it.
If most of your time is coming from calculation, not reading, then yes a
spark repartition will help. If most of your time is coming just from
reading, you need more kafka partitions.
On Sun, Jul 12, 2015 at 6:30 AM, gaurav sharma <sh...@gmail.com>
wrote:
> Hi guys,
>
> I too am facing similar challenge with directstream.
>
> I have 3 Kafka Partitions.
>
> and running spark on 18 cores, with parallelism level set to 48.
>
> I am running simple map-reduce job on incoming stream.
>
> Though the reduce stage takes milliseconds-seconds for around 15 million
> packets, the Map stage takes around 4-5 minutes, since it creates only 3
> tasks for Map stage(I believe 3 tasks because I have 3 kafka partitions and
> the
>
> JavaPairDStream<String, String> kafkaStream =
> KafkaConnector.getKafkaStream(jsc);
>
> kafkaStream that i create in code is the parent Rdd for Map Job, so it
> would create only 3 tasks)
>
>
> I have 10 such jobs similar to above one working on same KafkaStream i
> create
>
> Could you guys please advise, if repartitioning the KafkaStream (taking
> into account the rechuffle at repartition stage) would optimize my overall
> batch processing time.
>
>
> On Sat, Jun 20, 2015 at 7:24 PM, Silvio Fiorito <
> silvio.fiorito@granturing.com> wrote:
>
>> Are you sure you were using all 100 executors even with the receiver
>> model? Because in receiver mode, the number of partitions is dependent on
>> the batch duration and block interval. It may not necessarily map directly
>> to the number of executors in your app unless you've adjusted the block
>> interval and batch duration.
>>
>> *From:* Tim Smith <se...@gmail.com>
>> *Sent:* Friday, June 19, 2015 10:36 PM
>> *To:* user@spark.apache.org
>>
>> I did try without repartition, initially, but that was even more
>> horrible because instead of the allocated 100 executors, only 30 (which is
>> the number of kafka partitions) would have to do the work. The "MyFunc" is
>> a CPU bound task so adding more memory per executor wouldn't help and I saw
>> that each of the 30 executors was only using one thread/core on each Spark
>> box. I could go and play with threading in MyFunc but I don't want to mess
>> with threading with all the parallelism already involved and I don't think
>> in-app threading outside of what the framework does is really desirable.
>>
>> With repartition, there is shuffle involved, but at least the
>> computation load spreads across all 100 executors instead of just 30.
>>
>>
>>
>>
>> On Fri, Jun 19, 2015 at 7:14 PM, Cody Koeninger <co...@koeninger.org>
>> wrote:
>>
>>> If that's the case, you're still only using as many read executors as
>>> there are kafka partitions.
>>>
>>> I'd remove the repartition. If you weren't doing any shuffles in the
>>> old job, and are doing a shuffle in the new job, it's not really comparable.
>>>
>>> On Fri, Jun 19, 2015 at 8:16 PM, Tim Smith <se...@gmail.com> wrote:
>>>
>>>> On Fri, Jun 19, 2015 at 5:15 PM, Tathagata Das <td...@databricks.com>
>>>> wrote:
>>>>
>>>>> Also, can you find from the spark UI the break up of the stages in
>>>>> each batch's jobs, and find which stage is taking more time after a while?
>>>>>
>>>>
>>>> Sure, will try to debug/troubleshoot. Are there enhancements to this
>>>> specific API between 1.3 and 1.4 that can substantially change it's
>>>> behaviour?
>>>>
>>>>
>>>>> On Fri, Jun 19, 2015 at 4:51 PM, Cody Koeninger <co...@koeninger.org>
>>>>> wrote:
>>>>>
>>>>>> when you say your old version was
>>>>>>
>>>>>> k = createStream .....
>>>>>>
>>>>>> were you manually creating multiple receivers? Because otherwise
>>>>>> you're only using one receiver on one executor...
>>>>>>
>>>>>
>>>> Yes, sorry, the earlier/stable version was more like:
>>>> kInStreams = (1 to n).map{_ => KafkaUtils.createStream ............
>>>> // n being the number of kafka partitions, 1 receiver per partition
>>>> val k = ssc.union(kInStreams)
>>>> val dataout = k.map(x=>myFunc(x._2,someParams))
>>>> dataout.foreachRDD ( rdd => rdd.foreachPartition(rec => {
>>>> myOutputFunc.write(rec) })
>>>>
>>>> Thanks,
>>>>
>>>> Tim
>>>>
>>>>
>>>>
>>>>
>>>>
>>>>>
>>>>>> If that's the case I'd try direct stream without the repartitioning.
>>>>>>
>>>>>>
>>>>>> On Fri, Jun 19, 2015 at 6:43 PM, Tim Smith <se...@gmail.com> wrote:
>>>>>>
>>>>>>> Essentially, I went from:
>>>>>>> k = createStream .....
>>>>>>> val dataout = k.map(x=>myFunc(x._2,someParams))
>>>>>>> dataout.foreachRDD ( rdd => rdd.foreachPartition(rec => {
>>>>>>> myOutputFunc.write(rec) })
>>>>>>>
>>>>>>> To:
>>>>>>> kIn = createDirectStream .....
>>>>>>> k = kIn.repartition(numberOfExecutors) //since #kafka partitions <
>>>>>>> #spark-executors
>>>>>>> val dataout = k.map(x=>myFunc(x._2,someParams))
>>>>>>> dataout.foreachRDD ( rdd => rdd.foreachPartition(rec => {
>>>>>>> myOutputFunc.write(rec) })
>>>>>>>
>>>>>>> With the new API, the app starts up and works fine for a while but
>>>>>>> I guess starts to deteriorate after a while. With the existing API
>>>>>>> "createStream", the app does deteriorate but over a much longer period,
>>>>>>> hours vs days.
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> On Fri, Jun 19, 2015 at 1:40 PM, Tathagata Das <td...@databricks.com>
>>>>>>> wrote:
>>>>>>>
>>>>>>>> Yes, please tell us what operation are you using.
>>>>>>>>
>>>>>>>> TD
>>>>>>>>
>>>>>>>> On Fri, Jun 19, 2015 at 11:42 AM, Cody Koeninger <
>>>>>>>> cody@koeninger.org> wrote:
>>>>>>>>
>>>>>>>>> Is there any more info you can provide / relevant code?
>>>>>>>>>
>>>>>>>>> On Fri, Jun 19, 2015 at 1:23 PM, Tim Smith <se...@gmail.com>
>>>>>>>>> wrote:
>>>>>>>>>
>>>>>>>>>> Update on performance of the new API: the new code using the
>>>>>>>>>> createDirectStream API ran overnight and when I checked the app state in
>>>>>>>>>> the morning, there were massive scheduling delays :(
>>>>>>>>>>
>>>>>>>>>> Not sure why and haven't investigated a whole lot. For now,
>>>>>>>>>> switched back to the createStream API build of my app. Yes, for the record,
>>>>>>>>>> this is with CDH 5.4.1 and Spark 1.3.
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>> On Thu, Jun 18, 2015 at 7:05 PM, Tim Smith <se...@gmail.com>
>>>>>>>>>> wrote:
>>>>>>>>>>
>>>>>>>>>>> Thanks for the super-fast response, TD :)
>>>>>>>>>>>
>>>>>>>>>>> I will now go bug my hadoop vendor to upgrade from 1.3 to 1.4.
>>>>>>>>>>> Cloudera, are you listening? :D
>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>> On Thu, Jun 18, 2015 at 7:02 PM, Tathagata Das <
>>>>>>>>>>> tathagata.das1565@gmail.com> wrote:
>>>>>>>>>>>
>>>>>>>>>>>> Are you using Spark 1.3.x ? That explains. This issue has been
>>>>>>>>>>>> fixed in Spark 1.4.0. Bonus you get a fancy new streaming UI with more
>>>>>>>>>>>> awesome stats. :)
>>>>>>>>>>>>
>>>>>>>>>>>> On Thu, Jun 18, 2015 at 7:01 PM, Tim Smith <se...@gmail.com>
>>>>>>>>>>>> wrote:
>>>>>>>>>>>>
>>>>>>>>>>>>> Hi,
>>>>>>>>>>>>>
>>>>>>>>>>>>> I just switched from "createStream" to the
>>>>>>>>>>>>> "createDirectStream" API for kafka and while things otherwise seem happy,
>>>>>>>>>>>>> the first thing I noticed is that stream/receiver stats are gone from the
>>>>>>>>>>>>> Spark UI :( Those stats were very handy for keeping an eye on health of the
>>>>>>>>>>>>> app.
>>>>>>>>>>>>>
>>>>>>>>>>>>> What's the best way to re-create those in the Spark UI?
>>>>>>>>>>>>> Maintain Accumulators? Would be really nice to get back receiver-like stats
>>>>>>>>>>>>> even though I understand that "createDirectStream" is a receiver-less
>>>>>>>>>>>>> design.
>>>>>>>>>>>>>
>>>>>>>>>>>>> Thanks,
>>>>>>>>>>>>>
>>>>>>>>>>>>> Tim
>>>>>>>>>>>>>
>>>>>>>>>>>>>
>>>>>>>>>>>>>
>>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>
>>>>>>>>
>>>>>>>
>>>>>>
>>>>>
>>>>
>>>
>>
>