DataFrame join optimization - Broadcast Hash Join

Question

I am trying to effectively join two DataFrames, one of which is large and the second is a bit smaller.

Is there a way to avoid all this shuffling? I cannot set autoBroadCastJoinThreshold, because it supports only Integers - and the table I am trying to broadcast is slightly bigger than integer number of bytes.

Is there a way to force broadcast ignoring this variable?

Answer 1

Broadcast Hash Joins (similar to map side join or map-side combine in Mapreduce) :

In SparkSQL you can see the type of join being performed by calling queryExecution.executedPlan. As with core Spark, if one of the tables is much smaller than the other you may want a broadcast hash join. You can hint to Spark SQL that a given DF should be broadcast for join by calling method broadcast on the DataFrame before joining it

Example: largedataframe.join(broadcast(smalldataframe), "key")

in DWH terms, where largedataframe may be like fact
smalldataframe may be like dimension

As described by my fav book (HPS) pls. see below to have better understanding..

Note : Above broadcast is from import org.apache.spark.sql.functions.broadcast not from SparkContext

Spark also, automatically uses the spark.sql.conf.autoBroadcastJoinThreshold to determine if a table should be broadcast.

Tip : see DataFrame.explain() method

def
explain(): Unit
Prints the physical plan to the console for debugging purposes.

---

Is there a way to force broadcast ignoring this variable?

sqlContext.sql("SET spark.sql.autoBroadcastJoinThreshold = -1")

---

NOTE :

Another similar out of box note w.r.t. Hive (not spark) : Similar thing can be achieved using hive hint MAPJOIN like below...

Select /*+ MAPJOIN(b) */ a.key, a.value from a join b on a.key = b.key

hive> set hive.auto.convert.join=true;
hive> set hive.auto.convert.join.noconditionaltask.size=20971520
hive> set hive.auto.convert.join.noconditionaltask=true;
hive> set hive.auto.convert.join.use.nonstaged=true;
hive> set hive.mapjoin.smalltable.filesize = 30000000; // default 25 mb made it as 30mb

Further Reading : Please refer my article on BHJ, SHJ, SMJ

Answer 2

You can hint for a dataframe to be broadcasted by using left.join(broadcast(right), ...)

Answer 3

Setting spark.sql.autoBroadcastJoinThreshold = -1 will disable broadcast completely. See Other Configuration Options in Spark SQL, DataFrames and Datasets Guide.

Answer 4

This is a current limitation of spark, see SPARK-6235. The 2GB limit also applies for broadcast variables.

Are you sure there is no other good way to do this, e.g. different partitioning?

Otherwise you can hack your way around it by manually creating multiple broadcast variables which are each <2GB.

Answer 5

I found this code works for Broadcast Join in Spark 2.11 version 2.0.0.

import org.apache.spark.sql.functions.broadcast  

val employeesDF = employeesRDD.toDF
val departmentsDF = departmentsRDD.toDF

// materializing the department data
val tmpDepartments = broadcast(departmentsDF.as("departments"))

import context.implicits._

employeesDF.join(broadcast(tmpDepartments), 
   $"depId" === $"id",  // join by employees.depID == departments.id 
   "inner").show()

Here is the reference for the above code Henning Kropp Blog, Broadcast Join with Spark