Datasets provide a new API for manipulating data within Spark. These provide a more user friendly experience than pure Scala for common queries. The Spark Cassandra Connector provides an integrated Data Source to make creating Cassandra Datasets easy.
Spark Docs:
DataSources in Spark take a map of Options which define how the source should act. The Connector provides a CassandraSource which recognizes the following key/value pairs. Those followed with a default of N/A are required, all others are optional.
| Option Key | Controls | Values | Default |
|---|---|---|---|
| table | The Cassandra table to connect to | String | N/A |
| keyspace | The keyspace where table is looked for | String | N/A |
| cluster | The group of the Cluster Level Settings to inherit | String | “default” |
| pushdown | Enables pushing down predicates to C* when applicable | (true,false) | true |
| confirm.truncate | Confirm to truncate table when use Save.overwrite mode | (true,false) | false |
Any normal Spark Connector configuration options for Connecting, Reading or Writing can be passed through as Dataset options as well. When using the read command below these options should appear exactly the same as when set in the SparkConf. See Config Helpers for typed helpers for setting these options.
The connector also provides a way to describe the options which should be applied to all Datasets within a cluster or within a keyspace. When a property has been specified at the table level it will override the default keyspace or cluster property.
To add these properties add keys to your SparkConf use the helpers explained in the next section or by manually entering them in the format
clusterName:keyspaceName/propertyName
There are also some helper methods which simplify setting Spark Cassandra Connector related parameters. This makes it easier to set parameters without remembering the above syntax:
import org.apache.spark.sql.cassandra._ import com.datastax.spark.connector.cql.CassandraConnectorConf import com.datastax.spark.connector.rdd.ReadConf // set params for all clusters and keyspaces spark.setCassandraConf(CassandraConnectorConf.KeepAliveMillisParam.option(10000)) // set params for the particular cluster spark.setCassandraConf("Cluster1", CassandraConnectorConf.ConnectionHostParam.option("127.0.0.1") ++ CassandraConnectorConf.ConnectionPortParam.option(12345)) spark.setCassandraConf("Cluster2", CassandraConnectorConf.ConnectionHostParam.option("127.0.0.2")) // set params for the particular keyspace spark.setCassandraConf("Cluster1", "ks1", ReadConf.SplitSizeInMBParam.option(128)) spark.setCassandraConf("Cluster1", "ks2", ReadConf.SplitSizeInMBParam.option(64)) spark.setCassandraConf("Cluster2", "ks3", ReadConf.SplitSizeInMBParam.option(80))
spark.setCassandraConf("ClusterOne", "ks1", ReadConf.SplitSizeInMBParam.option(32)) spark.setCassandraConf("default", "test", ReadConf.SplitSizeInMBParam.option(128)) val df = spark .read .format("org.apache.spark.sql.cassandra") .options(Map( "table" -> "words", "keyspace" -> "test")) .load() // This Dataset will use a spark.cassandra.input.size of 128 val otherdf = spark .read .format("org.apache.spark.sql.cassandra") .options(Map( "table" -> "words", "keyspace" -> "test" , "cluster" -> "ClusterOne")) .load() // This Dataset will use a spark.cassandra.input.size of 32 val lastdf = spark .read .format("org.apache.spark.sql.cassandra") .options(Map( "table" -> "words", "keyspace" -> "test" , "cluster" -> "ClusterOne", "spark.cassandra.input.split.size_in_mb" -> 48 ) ).load() // This Dataset will use a spark.cassandra.input.split.size of 48
The most programmatic way to create a Dataset is to invoke a read command on the SparkSession. This will build a DataFrameReader. Specify format as org.apache.spark.sql.cassandra. You can then use options to give a map of Map[String,String] of options as described above. Then finish by calling load to actually get a Dataset. This code is all lazy and will not actually load any data until an action is called.
As well as specifying all these parameters manually, we offer a set of helper functions to make this easier as well.
val df = spark .read .format("org.apache.spark.sql.cassandra") .options(Map( "table" -> "words", "keyspace" -> "test" )) .load() df.show
word count cat 30 fox 40
There are also some helper methods which can make creating Datasets easier. They can be accessed after importing org.apache.spark.sql.cassandra package. In the following example, all the commands used to create the Dataset are equivalent:
import org.apache.spark.sql.cassandra._ val df = spark .read .cassandraFormat("words", "test") .load() //Loading an Dataset using a format helper and a option helper val df = spark .read .cassandraFormat("words", "test") .options(ReadConf.SplitSizeInMBParam.option(32)) .load()
Accessing Datasets using Spark SQL involves creating temporary views with the format as org.apache.spark.sql.cassandra. The OPTIONS passed to this table are used to establish a relation between the CassandraTable and the Spark catalog reference.
Create Relation with the Cassandra table test.words
val createDDL = """CREATE TEMPORARY VIEW words USING org.apache.spark.sql.cassandra OPTIONS ( table "words", keyspace "test", cluster "Test Cluster", pushdown "true")""" spark.sql(createDDL) // Creates Catalog Entry registering an existing Cassandra Table spark.sql("SELECT * FROM words").show spark.sql("SELECT * FROM words WHERE word = 'fox'").show
Datasets provide a save function which allows them to persist their data to another DataSource. The connector supports using this feature to persist a Dataset to a Cassandra table.
val df = spark .read .cassandraFormat("words", "test") .load() df.write .cassandraFormat("words_copy", "test") .save()
Similarly to reading Cassandra tables into Datasets, we have some helper methods for the write path which are provided by org.apache.spark.sql.cassandra package. In the following example, all the commands are equivalent:
import org.apache.spark.sql.cassandra._ df.write .format("org.apache.spark.sql.cassandra") .options(Map("table" -> "words_copy", "keyspace" -> "test", "cluster" -> "cluster_B")) .save() df.write .cassandraFormat("words_copy", "test", "cluster_B") .save()
Connector specific options can be set by invoking options method on either DataFrameReader or DataFrameWriter. There are several settings you may want to change in ReadConf, WriteConf, CassandraConnectorConf, AuthConf and others. Those settings are identified by instances of ConfigParameter case class which offers an easy way to apply the option which it represents to a DataFrameReader or DataFrameWriter.
Suppose we want to set spark.cassandra.read.timeout_ms to 7 seconds on some DataFrameReader, we can do this both ways:
option("spark.cassandra.read.timeout_ms", "7000")
Since this setting is represented by CassandraConnectorConf.ReadTimeoutParam we can simply do:
options(CassandraConnectorConf.ReadTimeoutParam.sqlOption("7000"))
Each parameter, that is, each instance of ConfigParameter allows to invoke apply method with a single parameter. That method returns a Map[String, String] (note that you need to use options instead of option) so setting multiple parameters can be chained:
options(CassandraConnectorConf.ReadTimeoutParam.sqlOption("7000") ++ ReadConf.TaskMetricParam.sqlOption(true))
Spark Cassandra Connector adds a method to Dataset that allows it to create a new Cassandra table from the StructType schema of the Dataset. This is convenient for persisting a Dataset to a new table, especially when the schema of the Dataset is not known (fully or at all) ahead of time (at compile time of your application). Once the new table is created, you can persist the Dataset to the new table using the save function described above.
The partition key and clustering key of the newly generated table can be set by passing in a list of names of columns which should be used as partition key and clustering key.
// Add spark connector specific methods to Dataset import com.datastax.spark.connector._ val df = spark .read .cassandraFormat("words", "test") .load() val renamed = df.withColumnRenamed("col1", "newcolumnname") renamed.createCassandraTable( "test", "renamed", partitionKeyColumns = Some(Seq("user")), clusteringKeyColumns = Some(Seq("newcolumnname"))) renamed.write .cassandraFormat("renamed", "test") .save()
The Dataset API will automatically pushdown valid “where” clauses to Cassandra as long as the pushdown option is enabled (default is enabled). The full list of predicate pushdown restrictions is enumerated after the examples.
Example Table
CREATE KEYSPACE test WITH replication = {'class': 'SimpleStrategy', 'replication_factor': 1 }; USE test; CREATE table words ( user TEXT, word TEXT, count INT, PRIMARY KEY (user, word)); INSERT INTO words (user, word, count ) VALUES ( 'Russ', 'dino', 10 ); INSERT INTO words (user, word, count ) VALUES ( 'Russ', 'fad', 5 ); INSERT INTO words (user, word, count ) VALUES ( 'Sam', 'alpha', 3 ); INSERT INTO words (user, word, count ) VALUES ( 'Zebra', 'zed', 100 );
First we can create a Dataset and see that it has no pushdown filters set in the log. This means all requests will go directly to Cassandra and we will require reading all of the data to show this Dataset.
val df = spark .read .cassandraFormat("words", "test") .load df.explain
15/07/06 09:21:21 INFO CassandraSourceRelation: filters: 15/07/06 09:21:21 INFO CassandraSourceRelation: pushdown filters: //ArrayBuffer() == Physical Plan == PhysicalRDD [user#0,word#1,count#2], MapPartitionsRDD[2] at explain //at <console>:22
df.show
15/07/06 09:26:03 INFO CassandraSourceRelation: filters: 15/07/06 09:26:03 INFO CassandraSourceRelation: pushdown filters: //ArrayBuffer() +-----+-----+-----+ | user| word|count| +-----+-----+-----+ |Zebra| zed| 100| | Russ| dino| 10| | Russ| fad| 5| | Sam|alpha| 3| +-----+-----+-----+
The example schema has a clustering key of “word” so we can pushdown filters on that column to Cassandra. We do this by applying a normal Dataset filter. The connector will automatically determine that the filter can be pushed down and will add it to pushdown filters. All of the elements of pushdown filters will be automatically added to the CQL requests made to Cassandra for the data from this table. The subsequent call will then only serialize data from Cassandra which passes the filter, reducing the load on Cassandra.
val dfWithPushdown = df.filter(df("word") > "ham") dfWithPushdown.explain
15/07/06 09:29:10 INFO CassandraSourceRelation: filters: GreaterThan(word,ham) 15/07/06 09:29:10 INFO CassandraSourceRelation: pushdown filters: ArrayBuffer(GreaterThan(word,ham)) == Physical Plan == Filter (word#1 > ham) PhysicalRDD [user#0,word#1,count#2], MapPartitionsRDD[18] at explain at <console>:24
dfWithPushdown.show
15/07/06 09:30:48 INFO CassandraSourceRelation: filters: GreaterThan(word,ham) 15/07/06 09:30:48 INFO CassandraSourceRelation: pushdown filters: ArrayBuffer(GreaterThan(word,ham)) +-----+----+-----+ | user|word|count| +-----+----+-----+ |Zebra| zed| 100| +-----+----+-----+
Example table
CREATE KEYSPACE IF NOT EXISTS pushdowns WITH replication = { 'class' : 'SimpleStrategy', 'replication_factor' : 3 }; USE pushdowns; CREATE TABLE pushdownexample ( partitionkey1 BIGINT, partitionkey2 BIGINT, partitionkey3 BIGINT, clusterkey1 BIGINT, clusterkey2 BIGINT, clusterkey3 BIGINT, regularcolumn BIGINT, PRIMARY KEY ((partitionkey1, partitionkey2, partitionkey3), clusterkey1, clusterkey2, clusterkey3) );
val df = spark .read .cassandraFormat("pushdownexample", "pushdowns") .load()
To push down partition keys, all of them must be included, but not more than one predicate per partition key, otherwise nothing is pushed down.
df.filter("partitionkey1 = 1 AND partitionkey2 = 1 AND partitionkey3 = 1").show()
INFO 2015-08-26 00:37:40 org.apache.spark.sql.cassandra.CassandraSourceRelation: filters: EqualTo(partitionkey1,1), EqualTo(partitionkey2,1), EqualTo(partitionkey3,1) INFO 2015-08-26 00:37:40 org.apache.spark.sql.cassandra.CassandraSourceRelation: pushdown filters: ArrayBuffer(EqualTo(partitionkey1,1), EqualTo(partitionkey2,1), EqualTo(partitionkey3,1))
One partition key left out:
df.filter("partitionkey1 = 1 AND partitionkey2 = 1").show()
INFO 2015-08-26 00:53:07 org.apache.spark.sql.cassandra.CassandraSourceRelation: filters: EqualTo(partitionkey1,1), EqualTo(partitionkey2,1) INFO 2015-08-26 00:53:07 org.apache.spark.sql.cassandra.CassandraSourceRelation: pushdown filters: ArrayBuffer()
More than one predicate for partitionkey3:
df.filter("partitionkey1 = 1 AND partitionkey2 = 1 AND partitionkey3 > 0 AND partitionkey3 < 5").show()
INFO 2015-08-26 00:54:03 org.apache.spark.sql.cassandra.CassandraSourceRelation: filters: EqualTo(partitionkey1,1), EqualTo(partitionkey2,1), GreaterThan(partitionkey3,0), LessThan(partitionkey3,5) INFO 2015-08-26 00:54:03 org.apache.spark.sql.cassandra.CassandraSourceRelation: pushdown filters: ArrayBuffer()
Clustering keys are more relaxed. But only the last predicate can be non-EQ, and if there is more than one predicate for a column, they must not be EQ or IN, otherwise only some predicates may be pushed down.
df.filter("clusterkey1 = 1 AND clusterkey2 > 0 AND clusterkey2 < 10").show()
INFO 2015-08-26 01:01:02 org.apache.spark.sql.cassandra.CassandraSourceRelation: filters: EqualTo(clusterkey1,1), GreaterThan(clusterkey2,0), LessThan(clusterkey2,10) INFO 2015-08-26 01:01:02 org.apache.spark.sql.cassandra.CassandraSourceRelation: pushdown filters: ArrayBuffer(EqualTo(clusterkey1,1), GreaterThan(clusterkey2,0), LessThan(clusterkey2,10))
First predicate not EQ:
df.filter("clusterkey1 > 1 AND clusterkey2 > 1").show()
INFO 2015-08-26 00:55:01 org.apache.spark.sql.cassandra.CassandraSourceRelation: filters: GreaterThan(clusterkey1,1), GreaterThan(clusterkey2,1) INFO 2015-08-26 00:55:01 org.apache.spark.sql.cassandra.CassandraSourceRelation: pushdown filters: ArrayBuffer(GreaterThan(clusterkey1,1))
clusterkey2 EQ predicate:
df.filter("clusterkey1 = 1 AND clusterkey2 = 1 AND clusterkey2 < 10").show()
INFO 2015-08-26 00:56:37 org.apache.spark.sql.cassandra.CassandraSourceRelation: filters: EqualTo(clusterkey1,1), EqualTo(clusterkey2,1), LessThan(clusterkey2,10) INFO 2015-08-26 00:56:37 org.apache.spark.sql.cassandra.CassandraSourceRelation: pushdown filters: ArrayBuffer(EqualTo(clusterkey1,1), EqualTo(clusterkey2,1))
In Spark 2.0 DataFrames are now just a specific case of the Dataset API. In particular a DataFrame is just an alias for Dataset[Row]. This means everything you know about DataFrames is also applicable to Datasets. A DataFrame is just a special Dataset that is made up of Row objects. Many texts and resources still use the two terms interchangeably.
1. Only push down no-partition key column predicates with =, >, <, >=, <= predicate
2. Only push down primary key column predicates with = or IN predicate.
3. If there are regular columns in the pushdown predicates, they should have
at least one EQ expression on an indexed column and no IN predicates.
4. All partition column predicates must be included in the predicates to be pushed down,
any part of the partition key can be an EQ or IN predicate. For each partition column,
only one predicate is allowed.
5. For cluster column predicates, only last predicate can be RANGE predicate
and preceding column predicates must be EQ or IN predicates.
If there is only one cluster column predicate, the predicates could be EQ or IN or RANGE predicate.
6. There is no pushdown predicates if there is any OR condition or NOT IN condition.
7. We're not allowed to push down multiple predicates for the same column if any of them
is equality or IN predicate.