spark-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Malcolm Lockyer <>
Subject Spark + Kafka processing trouble
Date Tue, 31 May 2016 01:45:36 GMT
Hopefully this is not off topic for this list, but I am hoping to
reach some people who have used Kafka + Spark before.

We are new to Spark and are setting up our first production
environment and hitting a speed issue that maybe configuration related
- and we have little experience in configuring Spark environments.

So we've got a Spark streaming job that seems to take an inordinate
amount of time to process. I realize that without specifics, it is
difficult to trace - however the most basic primitives in Spark are
performing horribly. The lazy nature of Spark is making it difficult
for me to understand what is happening - any suggestions are very much

Environment is MBP 2.2 i7. Spark master is "local[*]". We are using
Kafka and PostgreSQL, both local. The job is designed to:

a) grab some data from Kafka
b) correlate with existing data in PostgreSQL
c) output data to Kafka

I am isolating timings by calling System.nanoTime() before and after
something that forces calculation, for example .count() on a
DataFrame. It seems like every operation has a MASSIVE fixed overhead
and that is stacking up making each iteration on the RDD extremely
slow. Slow operations include pulling a single item from the Kafka
queue, running a simple query against PostgresSQL, and running a Spark
aggregation on a RDD with a handful of rows.

The machine is not maxing out on memory, disk or CPU. The machine
seems to be doing nothing for a high percentage of the execution time.
We have reproduced this behavior on two other machines. So we're
suspecting a configuration issue

As a concrete example, we have a DataFrame produced by running a JDBC
query by mapping over an RDD from Kafka. Calling count() (I guess
forcing execution) on this DataFrame when there is *1* item/row (Note:
SQL database is EMPTY at this point so this is not a factor) takes 4.5
seconds, calling count when there are 10,000 items takes 7 seconds.

Can anybody offer experience of something like this happening for
them? Any suggestions on how to understand what is going wrong?

I have tried tuning the number of Kafka partitions - increasing this
seems to increase the concurrency and ultimately number of things
processed per minute, but to get something half decent, I'm going to
need running with 1024 or more partitions. Is 1024 partitions a
reasonable number? What do you use in you environments?

I've tried different options for batchDuration. The calculation seems
to be batchDuration * Kafka partitions for number of items per
iteration, but this is always still extremely slow (many per iteration
vs. very few doesn't seem to really improve things). Can you suggest a
list of the Spark configuration parameters related to speed that you
think are key - preferably with the values you use for those

I'd really really appreciate any help or suggestions as I've been
working on this speed issue for 3 days without success and my head is
starting to hurt. Thanks in advance.



Malcolm Lockyer

To unsubscribe, e-mail:
For additional commands, e-mail:

View raw message