spark-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Cody Koeninger <c...@koeninger.org>
Subject Re: MapWithState partitioning
Date Mon, 31 Oct 2016 13:45:34 GMT
If you call a transformation on an rdd using the same partitioner as that
rdd, no shuffle will occur.  KafkaRDD doesn't have a partitioner, there's
no consistent partitioning scheme that works for all kafka uses. You can
wrap each kafkardd with an rdd that has a custom partitioner that you write
to match your kafka partitioning scheme, and avoid a shuffle.

The danger there is if you have any misbehaving producers, or translate the
partitioning wrongly, you'll get bad results. It's safer just to shuffle.

On Oct 31, 2016 04:31, "Andrii Biletskyi"
<andrii.biletskyi@yahoo.com.invalid> wrote:

Hi all,

I'm using Spark Streaming mapWithState operation to do a stateful operation
on my Kafka stream (though I think similar arguments would apply for any
source).

Trying to understand a way to control mapWithState's partitioning schema.

My transformations are simple:

1) create KafkaDStream
2) mapPartitions to get a key-value stream where `key` corresponds to Kafka
message key
3) apply mapWithState operation on key-value stream, the state stream
shares keys with the original stream, the resulting streams doesn't change
keys either

The problem is that, as I understand, mapWithState stream has a different
partitioning schema and thus I see shuffles in Spark Web UI.

>From the mapWithState implementation I see that:
mapwithState uses Partitioner if specified, otherwise partitions data with
HashPartitioner(<default-parallelism-conf>). The thing is that original
KafkaDStream has a specific partitioning schema: Kafka partitions
correspond Spark RDD partitions.

Question: is there a way for mapWithState stream to inherit partitioning
schema from the original stream (i.e. correspond to Kafka partitions).

Thanks,
Andrii

Mime
View raw message