spark-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From JF Chen <>
Subject Re: How to increase the parallelism of Spark Streaming application´╝č
Date Thu, 08 Nov 2018 01:41:38 GMT
I have test it on my production environment, and I find a strange thing.
After I set the kafka partition to 100, some tasks are executed very fast,
but some are slow. The slow ones cost double time than fast ones(from event
timeline). However, I have checked the consumer offsets, the data amount
for each task should be similar, so it should be no unbalance problem.
Any one have some good idea?

Junfeng Chen

On Thu, Nov 8, 2018 at 12:34 AM Shahbaz <> wrote:

> Hi ,
>    - Do you have adequate CPU cores allocated to handle increased
>    partitions ,generally if you have Kafka partitions >=(greater than or equal
>    to) CPU Cores Total (Number of Executor Instances * Per Executor Core)
>    ,gives increased task parallelism for reader phase.
>    - However if you have too many partitions but not enough cores ,it
>    would eventually slow down the reader (Ex: 100 Partitions and only 20 Total
>    Cores).
>    - Additionally ,the next set of transformation will have there own
>    partitions ,if its involving  shuffle ,sq.shuffle.partitions then defines
>    next level of parallelism ,if you are not having any data skew,then you
>    should get good performance.
> Regards,
> Shahbaz
> On Wed, Nov 7, 2018 at 12:58 PM JF Chen <> wrote:
>> I have a Spark Streaming application which reads data from kafka and save
>> the the transformation result to hdfs.
>> My original partition number of kafka topic is 8, and repartition the
>> data to 100 to increase the parallelism of spark job.
>> Now I am wondering if I increase the kafka partition number to 100
>> instead of setting repartition to 100, will the performance be enhanced? (I
>> know repartition action cost a lot cpu resource)
>> If I set the kafka partition number to 100, does it have any negative
>> efficiency?
>> I just have one production environment so it's not convenient for me to
>> do the test....
>> Thanks!
>> Regard,
>> Junfeng Chen

View raw message