[SQL][SPARK-2212]Hash Outer Join

[chenghao-intel]

This patch is to support the hash based outer join. Currently, outer join for big relations are resort to BoradcastNestedLoopJoin, which is super slow. This PR will create 2 hash tables for both relations in the same partition, which greatly reduce the table scans.

Here is the testing code that I used:

package org.apache.spark.sql.hive
import org.apache.spark.SparkContext
import org.apache.spark.SparkConf
import org.apache.spark.sql._
case class Record(key: String, value: String)
object JoinTablePrepare extends App {
import TestHive2._
val rdd = sparkContext.parallelize((1 to 3000000).map(i => Record(s"${i % 828193}", s"val_$i")))
runSqlHive("SHOW TABLES")
runSqlHive("DROP TABLE if exists a")
runSqlHive("DROP TABLE if exists b")
runSqlHive("DROP TABLE if exists result")
rdd.registerAsTable("records")
runSqlHive("""CREATE TABLE a (key STRING, value STRING)
| ROW FORMAT SERDE | 'org.apache.hadoop.hive.serde2.columnar.LazyBinaryColumnarSerDe' | STORED AS RCFILE
""".stripMargin)
runSqlHive("""CREATE TABLE b (key STRING, value STRING)
| ROW FORMAT SERDE | 'org.apache.hadoop.hive.serde2.columnar.LazyBinaryColumnarSerDe' | STORED AS RCFILE
""".stripMargin)
runSqlHive("""CREATE TABLE result (key STRING, value STRING)
| ROW FORMAT SERDE | 'org.apache.hadoop.hive.serde2.columnar.LazyBinaryColumnarSerDe' | STORED AS RCFILE
""".stripMargin)
hql(s"""from records | insert into table a
| select key, value
""".stripMargin)
hql(s"""from records | insert into table b select key + 100000, value
""".stripMargin)
}
object JoinTablePerformanceTest extends App {
import TestHive2._
hql("SHOW TABLES")
hql("set spark.sql.shuffle.partitions=20")
val leftOuterJoin = "insert overwrite table result select a.key, b.value from a left outer join b on a.key=b.key"
val rightOuterJoin = "insert overwrite table result select a.key, b.value from a right outer join b on a.key=b.key"
val fullOuterJoin = "insert overwrite table result select a.key, b.value from a full outer join b on a.key=b.key"
val results = ("LeftOuterJoin", benchmark(leftOuterJoin)) :: ("LeftOuterJoin", benchmark(leftOuterJoin)) :: ("RightOuterJoin", benchmark(rightOuterJoin)) :: ("RightOuterJoin", benchmark(rightOuterJoin)) :: ("FullOuterJoin", benchmark(fullOuterJoin)) :: ("FullOuterJoin", benchmark(fullOuterJoin)) :: Nil
val explains = hql(s"explain $leftOuterJoin").collect ++ hql(s"explain $rightOuterJoin").collect ++ hql(s"explain $fullOuterJoin").collect
println(explains.mkString(",\n"))
results.foreach { case (prompt, result) => {
println(s"$prompt: took ${result._1} ms (${result._2} records)")
}
}
def benchmark(cmd: String) = {
val begin = System.currentTimeMillis()
val result = hql(cmd)
val end = System.currentTimeMillis()
val count = hql("select count(1) from result").collect.mkString("")
((end - begin), count)
}
}

And the result as shown below:

[Physical execution plan:],
[InsertIntoHiveTable (MetastoreRelation default, result, None), Map(), true],
[ Project [key#95,value#98]],
[ HashOuterJoin [key#95], [key#97], LeftOuter, None],
[ Exchange (HashPartitioning [key#95], 20)],
[ HiveTableScan [key#95], (MetastoreRelation default, a, None), None],
[ Exchange (HashPartitioning [key#97], 20)],
[ HiveTableScan [key#97,value#98], (MetastoreRelation default, b, None), None],
[Physical execution plan:],
[InsertIntoHiveTable (MetastoreRelation default, result, None), Map(), true],
[ Project [key#102,value#105]],
[ HashOuterJoin [key#102], [key#104], RightOuter, None],
[ Exchange (HashPartitioning [key#102], 20)],
[ HiveTableScan [key#102], (MetastoreRelation default, a, None), None],
[ Exchange (HashPartitioning [key#104], 20)],
[ HiveTableScan [key#104,value#105], (MetastoreRelation default, b, None), None],
[Physical execution plan:],
[InsertIntoHiveTable (MetastoreRelation default, result, None), Map(), true],
[ Project [key#109,value#112]],
[ HashOuterJoin [key#109], [key#111], FullOuter, None],
[ Exchange (HashPartitioning [key#109], 20)],
[ HiveTableScan [key#109], (MetastoreRelation default, a, None), None],
[ Exchange (HashPartitioning [key#111], 20)],
[ HiveTableScan [key#111,value#112], (MetastoreRelation default, b, None), None]
LeftOuterJoin: took 16072 ms ([3000000] records)
LeftOuterJoin: took 14394 ms ([3000000] records)
RightOuterJoin: took 14802 ms ([3000000] records)
RightOuterJoin: took 14747 ms ([3000000] records)
FullOuterJoin: took 17715 ms ([6000000] records)
FullOuterJoin: took 17629 ms ([6000000] records)

Without this PR, the benchmark will run seems never end.

[AmplabJenkins]

Merged build triggered.

[AmplabJenkins]

Merged build started.

[AmplabJenkins]

Merged build finished.

[AmplabJenkins]

Refer to this link for build results: https://amplab.cs.berkeley.edu/jenkins/job/SparkPullRequestBuilder/15942/

[chenghao-intel]

Remove the join filter push down stuff, as we have the PushPredicateThroughJoin in Optimizer(Catalyst),

[marmbrus]

I think it would be much better if this PR just added support for LeftOuter (and maybe RightOuter too?) to HashJoin instead of rewriting all of our join planning code.

[chenghao-intel]

Thank you all for the comments, I will changed some of the code accordingly.
This PR actually contains 2 relevant parts:

• Code Re-factor for Join
  • Removed FilteredOperation from the patterns.scala, cause the filters(WHERE CONDITION & JOIN CONDITION) has been pushed down via the PushPredicateThroughJoin in logical.Optimizer.scala already. Discard the combination of filters(where and join condition) seems make the join pattern match more clean and simple.
  • Pattern matching order is actually very critical for the Join Operator Selection in SparkStrategies.scala, hence I merged the 3 Join Strategies into 1.
  • The trait BinaryJoinNode, which can be utilized by HashJoin / SortMergeJoin(will implement soon) / CartesionProduct(InnerJoin) / MapSide Join (Left/Inner/LeftSemi, assume the right table is the build table) for all of the join types; and if we want to add code gen for join condition, only we need to modify is the trait BinaryJoinNode.
• Add Outer Join Support for HashJoin
  • With BinaryJoinNode, add hash based outer join support is easy.

Sorry, this PR changes lots of code, and make the whole logic not easy to understand.

[chenghao-intel]

I am planning to split the PR into 3 smaller ones, 2 of those (#1187 & #1190) are done , and I will update this PR as the 3rd one once the previous have been merged.

[chenghao-intel]

BTW, it will be cool if #1163 is merged before I start working on this updating.

[AmplabJenkins]

Merged build triggered.

[AmplabJenkins]

Merged build started.

[AmplabJenkins]

Merged build finished. All automated tests passed.

[AmplabJenkins]

All automated tests passed.
Refer to this link for build results: https://amplab.cs.berkeley.edu/jenkins/job/SparkPullRequestBuilder/16189/

[chenghao-intel]

Thanks @marmbrus merged some of the dependent PRs. I've also updated both code and description for this PR accordingly, some of the issues were listed in the PR description, probably we can discuss those during the code review.

[SparkQA]

QA tests have started for PR 1147. This patch merges cleanly.
View progress: https://amplab.cs.berkeley.edu/jenkins/job/SparkPullRequestBuilder/17551/consoleFull

[chenghao-intel]

I've updated the code by rebasing the latest master, this PR will greatly improve the outer join performance for big tables. See the local benchmark in the description.

[marmbrus]

I think we can remove left, right, beforeEach, afterEach.

[marmbrus]

Thanks a lot for working on this! I agree that it would be great to merge this in before 1.1. I'm a little worried about how much memory this is going to require. However, the current implementation is so bad and since I think what you have done here is strictly better, we can probably just merge this in and then revisit.

[marmbrus]

I also really like how isolated this change is and the inclusion of the benchmark :)

[chenghao-intel]

Thank you @marmbrus I've updated the code as suggested.

[SparkQA]

QA tests have started for PR 1147. This patch merges cleanly.
View progress: https://amplab.cs.berkeley.edu/jenkins/job/SparkPullRequestBuilder/17663/consoleFull

[SparkQA]

QA results for PR 1147:
- This patch PASSES unit tests.
- This patch merges cleanly
- This patch adds the following public classes (experimental):
case class HashOuterJoin(

For more information see test ouptut:
https://amplab.cs.berkeley.edu/jenkins/job/SparkPullRequestBuilder/17663/consoleFull

[marmbrus]

Please use Seq.empty[Row] inline instead of this variable.

[marmbrus]

I'm going to go ahead and merge this so we can have it in 1.1. It would be great if you could address the final readability concerns in a follow up PR.

Thanks again for implementing this :) We have had a few questions about it on the mailing list!

[chenghao-intel]

Thank you very much @marmbrus , I will create follow up for the improvement. :)