When you say that the largest difference is from metrics.collect, how are
you measuring that? Wouldn't that be the difference between
max(partitionT1) and sparkT1, not sparkT0 and sparkT1?
As for further places to look, what's happening in the logs during that
time? Are the number of messages per partition roughly equal (you should
be able to see that in the log line I mentioned earlier)? Is there a
reason one of your executors is handling half as many tasks as the others?
On Wed, Oct 7, 2015 at 5:19 AM, Gerard Maas <gerard.m...@gmail.com> wrote:
> Thanks for the feedback.
>
> Cassandra does not seem to be the issue. The time for writing to Cassandra
> is in the same order of magnitude (see below)
>
> The code structure is roughly as follows:
>
> dstream.filter(pred).foreachRDD{rdd =>
> val sparkT0 = currentTimeMs
> val metrics = rdd.mapPartitions{partition =>
> val partitionT0 = currentTimeMs
> partition.foreach{ transform andThen storeInCassandra _}
> val partitionT1 = currentTimeMs
> Seq(Metric( "local time", executor, partitionT1 - partitionT0,
> records)).iterator
> }
> //materialize the rdd
> val allMetrics = metrics.collect()
> val sparkT1 = currentTimeMs
> val totalizedMetrics = // group by and reduce with sum
> val sparkT2 = currentTimeMs
> totalizedMetrics.foreach{ metric => gmetric.report(metric)}
> }
>
> Relating this code with the time table presented before (time in ms):
>
> How measured?Slow TaskFast Taskexecutor local totalizedMetrics347.6281.53spark
> computationsparkT1 - sparkT06930263metric collectionsparkT2 - sparkT170138wall
> clock processsparkT2 - sparkT07000401total records processed
> totalizedMetrics42975002
>
> What we observe is that the largest difference comes from the
> materialization of the RDD. This pattern repeats cyclically one on, one off.
>
> Any ideas where to further look?
>
> kr, Gerard.
>
>
> On Wed, Oct 7, 2015 at 1:33 AM, Tathagata Das <t...@databricks.com> wrote:
>
>> Good point!
>>
>> On Tue, Oct 6, 2015 at 4:23 PM, Cody Koeninger <c...@koeninger.org>
>> wrote:
>>
>>> I agree getting cassandra out of the picture is a good first step.
>>>
>>> But if you just do foreachRDD { _.count } recent versions of direct
>>> stream shouldn't do any work at all on the executor (since the number of
>>> messages in the rdd is known already)
>>>
>>> do a foreachPartition and println or count the iterator manually.
>>>
>>> On Tue, Oct 6, 2015 at 6:02 PM, Tathagata Das <t...@databricks.com>
>>> wrote:
>>>
>>>> Are sure that this is not related to Cassandra inserts? Could you just
>>>> do foreachRDD { _.count } instead to keep Cassandra out of the picture and
>>>> then test this agian.
>>>>
>>>> On Tue, Oct 6, 2015 at 12:33 PM, Adrian Tanase <atan...@adobe.com>
>>>> wrote:
>>>>
>>>>> Also check if the Kafka cluster is still balanced. Maybe one of the
>>>>> brokers manages too many partitions, all the work will stay on that
>>>>> executor unless you repartition right after kakfka (and I'm not saying you
>>>>> should).
>>>>>
>>>>> Sent from my iPhone
>>>>>
>>>>> On 06 Oct 2015, at 22:17, Cody Koeninger <c...@koeninger.org> wrote:
>>>>>
>>>>> I'm not clear on what you're measuring. Can you post relevant code
>>>>> snippets including the measurement code?
>>>>>
>>>>> As far as kafka metrics, nothing currently. There is an info-level
>>>>> log message every time a kafka rdd iterator is instantiated,
>>>>>
>>>>> log.info(s"Computing topic ${part.topic}, partition
>>>>> ${part.partition} " +
>>>>>
>>>>> s"offsets ${part.fromOffset} -> ${part.untilOffset}")
>>>>>
>>>>>
>>>>> If you log once you're done with an iterator you should be able to see
>>>>> the delta.
>>>>>
>>>>> The other thing to try is reduce the number of parts involved in the
>>>>> job to isolate it ... first thing I'd do there is take cassandra out of
>>>>> the
>>>>> equation.
>>>>>
>>>>>
>>>>>
>>>>> On Tue, Oct 6, 2015 at 2:00 PM, Gerard Maas <gerard.m...@gmail.com>
>>>>> wrote:
>>>>>
>>>>>> Hi Cody,
>>>>>>
>>>>>> The job is doing ETL from Kafka records to Cassandra. After a
>>>>>> single filtering stage on Spark, the 'TL' part is done using the
>>>>>> dstream.foreachRDD{rdd.foreachPartition{...TL ...}} pattern.
>>>>>>
>>>>>> We have metrics on the executor work which we collect and add
>>>>>> together, indicated here by 'local computation'. As you can see, we also
>>>>>> measure how much it cost us to measure :-)
>>>>>> See how 'local work' times are comparable. What's not visible is
>>>>>> the task scheduling and consuming the data from Kafka which becomes part
>>>>>> of
>>>>>> the 'spark computation' part.
>>>>>>
>>>>>> The pattern we see is 1 fast, 1 slow, 1 fast,... zig...zag...
>>>>>>
>>>>>> Are there metrics available somehow on the Kafka reading time?
>>>>>>
>>>>>> Slow Task Fast Task local computation 347.6 281.53 spark computation
>>>>>> 6930 263 metric collection 70 138 wall clock process 7000 401 total
>>>>>> records processed 4297 5002
>>>>>>
>>>>>> (time in ms)
>>>>>>
>>>>>> kr, Gerard.
>>>>>>
>>>>>>
>>>>>> On Tue, Oct 6, 2015 at 8:01 PM, Cody Koeninger <c...@koeninger.org>
>>>>>> wrote:
>>>>>>
>>>>>>> Can you say anything more about what the job is doing?
>>>>>>>
>>>>>>> First thing I'd do is try to get some metrics on the time taken by
>>>>>>> your code on the executors (e.g. when processing the iterator) to see if
>>>>>>> it's consistent between the two situations.
>>>>>>>
>>>>>>> On Tue, Oct 6, 2015 at 11:45 AM, Gerard Maas <gerard.m...@gmail.com>
>>>>>>> wrote:
>>>>>>>
>>>>>>>> Hi,
>>>>>>>>
>>>>>>>> We recently migrated our streaming jobs to the direct kafka
>>>>>>>> receiver. Our initial migration went quite fine but now we are seeing a
>>>>>>>> weird zig-zag performance pattern we cannot explain.
>>>>>>>> In alternating fashion, one task takes about 1 second to finish and
>>>>>>>> the next takes 7sec for a stable streaming rate.
>>>>>>>>
>>>>>>>> Here are comparable metrics for two successive tasks:
>>>>>>>> *Slow*:
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>> Executor ID Address Task Time Total Tasks Failed Tasks Succeeded
>>>>>>>> Tasks 20151006-044141-2408867082-5050-21047-S0
>>>>>>>> dnode-3.hdfs.private:36863 22 s 3 0 3
>>>>>>>> 20151006-044141-2408867082-5050-21047-S1 dnode-0.hdfs.private:43812 40
>>>>>>>> s 11 0 11 20151006-044141-2408867082-5050-21047-S4
>>>>>>>> dnode-5.hdfs.private:59945 49 s 10 0 10
>>>>>>>> *Fast*:
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>> Executor ID Address Task Time Total Tasks Failed Tasks Succeeded
>>>>>>>> Tasks 20151006-044141-2408867082-5050-21047-S0
>>>>>>>> dnode-3.hdfs.private:36863 0.6 s 4 0 4
>>>>>>>> 20151006-044141-2408867082-5050-21047-S1 dnode-0.hdfs.private:43812 1
>>>>>>>> s 9 0 9 20151006-044141-2408867082-5050-21047-S4
>>>>>>>> dnode-5.hdfs.private:59945 1 s 11 0 11
>>>>>>>> We have some custom metrics that measure wall-clock time of
>>>>>>>> execution of certain blocks of the job, like the time it takes to do
>>>>>>>> the
>>>>>>>> local computations (RDD.foreachPartition closure) vs total time.
>>>>>>>> The difference between the slow and fast executing task is on the
>>>>>>>> 'spark computation time' which is wall-clock for the task scheduling
>>>>>>>> (DStream.foreachRDD closure)
>>>>>>>>
>>>>>>>> e.g.
>>>>>>>> Slow task:
>>>>>>>>
>>>>>>>> local computation time: 347.60968499999996, *spark computation
>>>>>>>> time: 6930*, metric collection: 70, total process:
>>>>>>>> 7000, total_records: 4297
>>>>>>>>
>>>>>>>> Fast task:
>>>>>>>> local computation time: 281.539042,* spark computation time: 263*,
>>>>>>>> metric collection: 138, total process: 401, total_records: 5002
>>>>>>>>
>>>>>>>> We are currently running Spark 1.4.1. The load and the work to be
>>>>>>>> done is stable -this is on a dev env with that stuff under control.
>>>>>>>>
>>>>>>>> Any ideas what this behavior could be?
>>>>>>>>
>>>>>>>> thanks in advance, Gerard.
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>
>>>>>>
>>>>>
>>>>
>>>
>>
>
