pyspark joinWithCassandraTable refactor without maps - apache-spark

Question

Im new to using spark/scala here and im having trouble with a refactor of some of my code here. Im running Scala 2.11 using pyspark and in a spark/yarn setup. The following is working but id like to clean it up, and to get the max performance out of this. I read elsewhere that pyspark udf and lambdas can cause huge performance impact so im trying to reduce or remove them were possible.
# Reduce ingest df1 data by joining on allowed table df2
to_process = df2\
.join(
sf.broadcast(df1),
df2.secondary_id == df1.secondary_id,
how="inner")\
.rdd\
.map(lambda r: Row(tag=r['tag_id'], user_uuid=r['user_uuid']))
# Type column fixed to type=2, and tag==key
ready_to_join = to_process.map(lambda r: (r[0], 2, r[1]))
# Join with cassandra table to find matches
exists_in_cass = ready_to_join\
.joinWithCassandraTable(keyspace, table3)\
.on("user_uuid", "type")\
.select("user_uuid")
log.error(f"TEST PRINT - [{exists_in_cass.count()}]")
the cassandra table is such that
CREATE TABLE keyspace.table3 (
user_uuid uuid,
type int,
key text,
value text,
PRIMARY KEY (user_uuid, type, key)
) WITH CLUSTERING ORDER BY (type ASC, key ASC)
currently ive got
to_process = df2\
.join(
sf.broadcast(df1),
df2.secondary_id == df1.secondary_id,
how="inner")\
.select(col("user_uuid"), col("tag_id").alias("tag"))
ready_to_join = to_process\
.withColumn("type", sf.lit(2))\
.select('user_uuid', 'type', col('tag').alias("key"))\
.rdd\
.map(lambda x: Row(x))
# planning on using repartitionByCassandraReplica here after I get it logically working
exists_in_cass = ready_to_join\
.joinWithCassandraTable(keyspace, table3)\
.on("user_uuid", "type")\
.select("user_uuid")
log.error(f"TEST PRINT - [{exists_in_cass.count()}]")
but im getting errors like
2020-10-30 15:10:42 WARN TaskSetManager:66 - Lost task 148.0 in stage 22.0 (TID ----, ---, executor 9): net.razorvine.pickle.PickleException: expected zero arguments for construction of ClassDict (for pyspark.sql.types._create_row)
at net.razorvine.pickle.objects.ClassDictConstructor.construct(ClassDictConstructor.java:23)
looking help from any spark gurus out there to point me to anything stupid I am doing here.
Update
Thanks to Alex's suggestion using the spark-cassandra-connector v2.5+ gives the ability for dataframes to join directly. I updated my code to use this instead.
to_process = df2\
.join(
sf.broadcast(df1),
df2.secondary_id == df1.secondary_id,
how="inner")\
.select(col("user_uuid"), col("tag_id").alias("tag"))
ready_to_join = to_process\
.withColumn("type", sf.lit(2))\
.select(col('user_uuid').alias('c1_user_uuid'), 'type', col('tag').alias("key"))\
cass_table = spark_session
.read \
.format("org.apache.spark.sql.cassandra") \
.options(table=config.table, keyspace=config.keyspace) \
.load()
exists_in_cass = ready_to_join\
.join(
cass_table,
[(cass_table["user_uuid"] == ready_to_join["c1_user_uuid"]) &
(cass_table["key"] == ready_to_join["key"]) &
(cass_table["type"] == ready_to_join["type"])])\
.select(col("c1_user_uuid").alias("user_uuid"))
exists_in_cass.explain()
log.error(f"TEST PRINT - [{exists_in_cass.count()}]")
As far as I know, in theory this should be alot faster ! But im getting errors during runtime with the database timing out.
WARN TaskSetManager:66 - Lost task 827.0 in stage 12.0 (TID 9946, , executor 4): java.io.IOException: Exception during execution of SELECT "user_uuid", "key" FROM "keyspace"."table3" WHERE token("user_uuid") > ? AND token("user_uuid") <= ? AND "type" = ? ALLOW FILTERING: Query timed out after PT2M
TaskSetManager:66 - Lost task 125.0 in stage 12.0 (TID 9215, , executor 7): com.datastax.oss.driver.api.core.DriverTimeoutException: Query timed out after PT2M
etc
I have the config for spark setup to allow for the spark extensions
--packages mysql:mysql-connector-java:5.1.47,com.datastax.spark:spark-cassandra-connector_2.11:2.5.1 \
--conf spark.sql.extensions=com.datastax.spark.connector.CassandraSparkExtensions \
The DAG from spark shows all nodes completely maxed out. Should I be partitioning my data before running my join here?
The explain for this also doesnt show a direct join (explain has more code than snippet above)
== Physical Plan ==
*(6) Project [c1_user_uuid#124 AS user_uuid#158]
+- *(6) SortMergeJoin [c1_user_uuid#124, key#125L], [user_uuid#129, cast(key#131 as bigint)], Inner
:- *(3) Sort [c1_user_uuid#124 ASC NULLS FIRST, key#125L ASC NULLS FIRST], false, 0
: +- Exchange hashpartitioning(c1_user_uuid#124, key#125L, 200)
: +- *(2) Project [id#0 AS c1_user_uuid#124, tag_id#101L AS key#125L]
: +- *(2) BroadcastHashJoin [secondary_id#60], [secondary_id#100], Inner, BuildRight
: :- *(2) Filter (isnotnull(secondary_id#60) && isnotnull(id#0))
: : +- InMemoryTableScan [secondary_id#60, id#0], [isnotnull(secondary_id#60), isnotnull(id#0)]
: : +- InMemoryRelation [secondary_id#60, id#0], StorageLevel(disk, memory, deserialized, 1 replicas)
: : +- *(7) Project [secondary_id#60, id#0]
: : +- Generate explode(split(secondary_ids#1, \|)), [id#0], false, [secondary_id#60]
: : +- *(6) Project [id#0, secondary_ids#1]
: : +- *(6) SortMergeJoin [id#0], [guid#46], Inner
: : :- *(2) Sort [id#0 ASC NULLS FIRST], false, 0
: : : +- Exchange hashpartitioning(id#0, 200)
: : : +- *(1) Filter (isnotnull(id#0) && id#0 RLIKE [0-9a-fA-F]{8}\-[0-9a-fA-F]{4}\-[0-9a-fA-F]{4}\-[0-9a-fA-F]{4}\-[0-9a-fA-F]{12})
: : : +- InMemoryTableScan [id#0, secondary_ids#1], [isnotnull(id#0), id#0 RLIKE [0-9a-fA-F]{8}\-[0-9a-fA-F]{4}\-[0-9a-fA-F]{4}\-[0-9a-fA-F]{4}\-[0-9a-fA-F]{12}]
: : : +- InMemoryRelation [id#0, secondary_ids#1], StorageLevel(disk, memory, deserialized, 1 replicas)
: : : +- Exchange RoundRobinPartitioning(3840)
: : : +- *(1) Filter AtLeastNNulls(n, id#0,secondary_ids#1)
: : : +- *(1) FileScan csv [id#0,secondary_ids#1] Batched: false, Format: CSV, Location: InMemoryFileIndex[inputdata_file, PartitionFilters: [], PushedFilters: [], ReadSchema: struct<id:string,secondary_ids:string>
: : +- *(5) Sort [guid#46 ASC NULLS FIRST], false, 0
: : +- Exchange hashpartitioning(guid#46, 200)
: : +- *(4) Filter (guid#46 RLIKE [0-9a-fA-F]{8}\-[0-9a-fA-F]{4}\-[0-9a-fA-F]{4}\-[0-9a-fA-F]{4}\-[0-9a-fA-F]{12} && isnotnull(guid#46))
: : +- Generate explode(set_guid#36), false, [guid#46]
: : +- *(3) Project [set_guid#36]
: : +- *(3) Filter (isnotnull(allowed#39) && (allowed#39 = 1))
: : +- *(3) FileScan orc whitelist.whitelist1[set_guid#36,region#39,timestamp#43] Batched: false, Format: ORC, Location: PrunedInMemoryFileIndex[hdfs://file, PartitionCount: 1, PartitionFilters: [isnotnull(timestamp#43), (timestamp#43 = 18567)], PushedFilters: [IsNotNull(region), EqualTo(region,1)], ReadSchema: struct<set_guid:array<string>,region:int>
: +- BroadcastExchange HashedRelationBroadcastMode(List(input[0, string, true]))
FROM TAG as T
JOIN MAP as M
ON T.tag_id = M.tag_id
WHERE (expire >= NOW() OR expire IS NULL)
ORDER BY T.tag_id) AS subset) [numPartitions=1] [secondary_id#100,tag_id#101L] PushedFilters: [*IsNotNull(secondary_id), *IsNotNull(tag_id)], ReadSchema: struct<secondary_id:string,tag_id:bigint>
+- *(5) Sort [user_uuid#129 ASC NULLS FIRST, cast(key#131 as bigint) ASC NULLS FIRST], false, 0
+- Exchange hashpartitioning(user_uuid#129, cast(key#131 as bigint), 200)
+- *(4) Project [user_uuid#129, key#131]
+- *(4) Scan org.apache.spark.sql.cassandra.CassandraSourceRelation [user_uuid#129,key#131] PushedFilters: [*EqualTo(type,2)], ReadSchema: struct<user_uuid:string,key:string>
Im not getting the direct joins working which is causing time outs.
Update 2
I think this isnt resolving to direct joins as my datatypes in the dataframes are off. Specifically the uuid type

Answer 1

Instead of using RDD API with PySpark, I suggest to take Spark Cassandra Connector (SCC) 2.5.x or 3.0.x (release announcement) that contain the implementation of the join of Dataframe with Cassandra - in this case you won't need to go down to RDDs, but just use normal Dataframe API joins.
Please note that this is not enabled by default, so you will need to start your pyspark or spark-submit with special configuration, like this:
pyspark --packages com.datastax.spark:spark-cassandra-connector_2.11:2.5.1 \
--conf spark.sql.extensions=com.datastax.spark.connector.CassandraSparkExtensions
You can find more about joins with Cassandra in my recent blog post on this topic (although it uses Scala, Dataframe part should be translated almost one to one to PySpark)