Pyspark Cassandra Save

PySpark Cassandra brings back the fun in working with Cassandra data in PySpark.

Project README

PySpark Cassandra

This PySpark Cassandra repository is no longer maintained. Please check this repository for Spark 2.0+ support:

Build Status Codacy Badge

PySpark Cassandra brings back the fun in working with Cassandra data in PySpark.

This module provides python support for Apache Spark's Resillient Distributed Datasets from Apache Cassandra CQL rows using Cassandra Spark Connector within PySpark, both in the interactive shell and in python programmes submitted with spark-submit.

This project was initially forked from, but in order to submit it to, a plain old repository was created.



Feedback on (in-)compatibility is much appreciated.


The current version of PySpark Cassandra is succesfully used with Spark version 1.5 and 1.6. Use older versions for Spark 1.2, 1.3 or 1.4.


PySpark Cassandra is compatible with Cassandra:

  • 2.1.5 and higher
  • 2.2
  • 3


PySpark Cassandra is used with python 2.7, python 3.3 and 3.4.


PySpark Cassandra is currently only packaged for Scala 2.10

Using with PySpark

With Spark Packages

Pyspark Cassandra is published at Spark Packages. This allows easy usage with Spark through:

spark-submit \
	--packages TargetHolding/pyspark-cassandra:<version> \

Without Spark Packages

spark-submit \
	--jars /path/to/pyspark-cassandra-assembly-<version>.jar \
	--driver-class-path /path/to/pyspark-cassandra-assembly-<version>.jar \
	--py-files /path/to/pyspark-cassandra-assembly-<version>.jar \
	--conf,cassandra,node,names \
	--master spark://spark-master:7077 \

(note that the the --driver-class-path due to SPARK-5185) (also not that the assembly will include the python source files, quite similar to a python source distribution)

Using with PySpark shell

Replace spark-submit with pyspark to start the interactive shell and don't provide a script as argument and then import PySpark Cassandra. Note that when performing this import the sc variable in pyspark is augmented with the cassandraTable(...) method.

import pyspark_cassandra


For Spark Packages Pyspark Cassandra can be published using:

sbt compile

The package can be published locally with:

sbt spPublishLocal

The package can be published to Spark Packages with (requires authentication and authorization):

make publish

For local testing / without Spark Packages

A Java / JVM library as well as a python library is required to use PySpark Cassandra. They can be built with:

make dist

This creates a fat jar with the Spark Cassandra Connector and additional classes for bridging Spark and PySpark for Cassandra data and the .py source files at: target/scala-2.10/pyspark-cassandra-assembly-<version>.jar


The PySpark Cassandra API aims to stay close to the Cassandra Spark Connector API. Reading its documentation is a good place to start.


The primary representation of CQL rows in PySpark Cassandra is the ROW format. However sc.cassandraTable(...) supports the row_format argument which can be any of the constants from RowFormat:

  • DICT: The default layout, a CQL row is represented as a python dict with the CQL row columns as keys.
  • TUPLE: A CQL row is represented as a python tuple with the values in CQL table column order / the order of the selected columns.
  • ROW: A pyspark_cassandra.Row object representing a CQL row.

Column values are related between CQL and python as follows:

CQL python
ascii unicode string
bigint long
blob bytearray
boolean boolean
counter int, long
decimal decimal
double float
float float
inet str
int int
map dict
set set
list list
text unicode string
timestamp datetime.datetime
timeuuid uuid.UUID
varchar unicode string
varint long
uuid uuid.UUID
UDT pyspark_cassandra.UDT


This is the default type to which CQL rows are mapped. It is directly compatible with pyspark.sql.Row but is (correctly) mutable and provides some other improvements.


This type is structurally identical to pyspark_cassandra.Row but serves user defined types. Mapping to custom python types (e.g. via CQLEngine) is not yet supported.


A CassandraSparkContext is very similar to a regular SparkContext. It is created in the same way, can be used to read files, parallelize local data, broadcast a variable, etc. See the Spark Programming Guide for more details. But it exposes one additional method:

  • cassandraTable(keyspace, table, ...): Returns a CassandraRDD for the given keyspace and table. Additional arguments which can be provided:

    • row_format can be set to any of the pyspark_cassandra.RowFormat values (defaults to ROW)
    • split_size sets the size in the number of CQL rows in each partition (defaults to 100000)
    • fetch_size sets the number of rows to fetch per request from Cassandra (defaults to 1000)
    • consistency_level sets with which consistency level to read the data (defaults to LOCAL_ONE)


PySpark Cassandra supports saving arbitrary RDD's to Cassandra using:

  • rdd.saveToCassandra(keyspace, table, ...): Saves an RDD to Cassandra. The RDD is expected to contain dicts with keys mapping to CQL columns. Additional arguments which can be supplied are:

    • columns(iterable): The columns to save, i.e. which keys to take from the dicts in the RDD.
    • batch_size(int): The size in bytes to batch up in an unlogged batch of CQL inserts.
    • batch_buffer_size(int): The maximum number of batches which are 'pending'.
    • batch_grouping_key(string): The way batches are formed (defaults to "partition"):
      • all: any row can be added to any batch
      • replicaset: rows are batched for replica sets
      • partition: rows are batched by their partition key
    • consistency_level(cassandra.ConsistencyLevel): The consistency level used in writing to Cassandra.
    • parallelism_level(int): The maximum number of batches written in parallel.
    • throughput_mibps: Maximum write throughput allowed per single core in MB/s.
    • ttl(int or timedelta): The time to live as milliseconds or timedelta to use for the values.
    • timestamp(int, date or datetime): The timestamp in milliseconds, date or datetime to use for the values.
    • metrics_enabled(bool): Whether to enable task metrics updates.


A CassandraRDD is very similar to a regular RDD in pyspark. It is extended with the following methods:

  • select(*columns): Creates a CassandraRDD with the select clause applied.
  • where(clause, *args): Creates a CassandraRDD with a CQL where clause applied. The clause can contain ? markers with the arguments supplied as *args.
  • limit(num): Creates a CassandraRDD with the limit clause applied.
  • take(num): Takes at most num records from the Cassandra table. Note that if limit() was invoked before take() a normal pyspark take() is performed. Otherwise, first limit is set and then a take() is performed.
  • cassandraCount(): Lets Cassandra perform a count, instead of loading the data to Spark first.
  • saveToCassandra(...): As above, but the keyspace and/or table may be omitted to save to the same keyspace and/or table.
  • spanBy(*columns): Groups rows by the given columns without shuffling.
  • joinWithCassandraTable(keyspace, table): Join an RDD with a Cassandra table on the partition key. Use .on(...) to specifiy other columns to join on. .select(...), .where(...) and .limit(...) can be used as well.


When importing pyspark_cassandra.streaming the method ``saveToCassandra(...)``` is made available on DStreams. Also support for joining with a Cassandra table is added:

  • joinWithCassandraTable(keyspace, table, selected_columns, join_columns):


Creating a SparkContext with Cassandra support

import pyspark_cassandra

conf = SparkConf() \
	.setAppName("PySpark Cassandra Test") \
	.setMaster("spark://spark-master:7077") \
	.set("", "cas-1")

sc = CassandraSparkContext(conf=conf)

Using select and where to narrow the data in an RDD and then filter, map, reduce and collect it::

sc \
	.cassandraTable("keyspace", "table") \
	.select("col-a", "col-b") \
	.where("key=?", "x") \
	.filter(lambda r: r["col-b"].contains("foo")) \
	.map(lambda r: (r["col-a"], 1)
	.reduceByKey(lambda a, b: a + b)

Storing data in Cassandra::

rdd = sc.parallelize([{
	"key": k,
	"val": random() * 10,
	"tags": ["a", "b", "c"],
	"options": {
		"foo": "bar",
		"baz": "qux",
} for k in ["x", "y", "z"]])


Create a streaming context, convert every line to a generater of words which are saved to cassandra. Through this example all unique words are stored in Cassandra.

The words are wrapped as a tuple so that they are in a format which can be stored. A dict or a pyspark_cassandra.Row object would have worked as well.

from pyspark.streaming import StreamingContext
from pyspark_cassandra import streaming

ssc = StreamingContext(sc, 2)

ssc \
    .socketTextStream("localhost", 9999) \
    .flatMap(lambda l: ((w,) for w in (l,))) \
    .saveToCassandra('keyspace', 'words')


Joining with Cassandra:

joined = rdd \
    .joinWithCassandraTable('keyspace', 'accounts') \
    .on('id') \
    .select('e-mail', 'followers')

for left, right in joined:

Or with a DStream:

joined = dstream.joinWithCassandraTable(self.keyspace, self.table, ['e-mail', 'followers'], ['id'])
Open Source Agenda is not affiliated with "Pyspark Cassandra" Project. README Source: TargetHolding/pyspark-cassandra
Open Issues
Last Commit
6 years ago

Open Source Agenda Badge

Open Source Agenda Rating