Spark SQL Join Improvement at Facebook

Spark SQL Join Improvement at Facebook
Cheng Su
Facebook

About Me
▪ Cheng Su
▪ Software Engineer at Facebook (Data Platform Team)
▪ Apache Spark Contributor (Spark SQL)
▪ Previously worked on Hive & Hadoop at Facebook

Agenda
▪ Overview for Spark SQL Joins
▪ Shuffled Hash Join Improvement (SPARK-32461)
▪ Leverage Bloom Filter for Join (under discussion with community)
▪ Stream-stream Join Improvement (SPARK-32883)
▪ Future Work

Overview for Spark SQL Joins
▪ SQL Join Physical Operators
▪ BroadcastHashJoinExec (broadcast hash join)
▪ ShuffledHashJoinExec (shuffled hash join)
▪ SortMergeJoinExec (sort merge join)

Overview for Spark SQL Joins
Data Source A Data Source B
. . . . . . . . . .JOIN

Spark Driver
- Scan source B
- Join with A
Broadcast hash join
- Ship smaller data source to all nodes
- Stream the other side
One side should be smaller
than spark.sql.autoBroadcastJoinThreshold
(default 10 MB).
Pros:
No shuffle and sort on both sides.
No skew.
Cons:
OOM on driver.
Broadcast
- Broadcast hash
relation on source A

Scan A Scan B
. . . . . . . . . .
ShuffleShuffleShuffle
Join
Sort merge join
- Shuffle both sides,
- Sort both sides, buffer one,
stream the bigger one
Pros:
Handle large data size well on
both sides.
Cons:
Need shuffle and sort.
Skew.
Sort
Join
Sort
Join
Sort

Shuffle hash join
- Hash smaller one, stream the
bigger one
Disable by default via
spark.sql.join.preferSortMergeJoin.
One side should be smaller than
(spark.sql.autoBroadcastJoinThreshold
* spark.sql.shuffle.partitions) (default
10MB * 200), and 3x smaller than the
other side
Pros:
Handle large data size well on one
side.
Cons:
Need shuffle.
Skew.
OOM on task for building hash table.
Scan A Scan B
. . . . . . . . . .
Build hash
table
Join
Build hash
table
Join

Shuffled Hash Join Improvement (SPARK-32461)
▪ Agenda
▪ Code-gen support
▪ Full outer join support
▪ Sort-based fallback mechanism

Code-gen Support for Shuffled Hash Join (SPARK-
32421)
▪ Why?
▪ Save compute resource, improve CPU (whole-stage code-gen)
▪ How?
▪ Refactor broadcast hash join code-gen logic
▪ Broadcast hash join already supports code-gen
▪ Refactor into common parent class for BHJ and SHJ - HashJoin.scala
▪ Performance Improvement
▪ 30% run-time improvement compared to non-code-gen for benchmark query
▪ PR Status
▪ Merged, will be available in Spark 3.1

Full Outer Shuffled Hash Join (SPARK-32399)
▪ Why?
▪ Save compute resource, improve CPU and IO
▪ Only sort merge join supports full outer, and sort is very expensive when table is large and needs
to spill to disk.
▪ Shuffled hash join does hash table lookup join, instead of sorting.
▪ How?
▪ Need to record non-matched rows from both sides
▪ Stream side: trivial
▪ Build side: non-trivial, need extra data structure (e.g. hash set for matched rows)

Full outer shuffle
hash join
- Hash smaller one, stream the
bigger one
- Hash set for build side to
record matched rows
- Iterate build side hash table
and output non-matched rows
Scan A Scan B
. . . . . . . . . .
Build hash
table
Join
Build hash
table
Join
Hash set Hash set

Full Outer Shuffled Hash Join (SPARK-32399)
▪ Performance Improvement
▪ 30% run-time improvement compared to full outer sort merge join for benchmark query
▪ PR Status

Sort-based fallback mechanism for SHJ (SPARK-32634)
▪ Why?
▪ Build side hash table out of memory
▪ No fallback, no spill, task failure, query failure
▪ Hard to enable shuffled hash join by default given OOM limitation
▪ How?
▪ Introduce fallback when building hash table
▪ Whenever fail to get memory to insert current row to hash table, stop build hash table. Sort both
sides and do sort merge join.
▪ PR Status
▪ WIP

Leverage Blook Filter for Join
▪ Why?
▪ Save compute resource, improve CPU and IO for shuffled hash join and sort merge join
▪ How?
▪ Build bloom filter on join key of smaller side
▪ Use bloom filter to filter out rows when scanning larger side
▪ Reduce amount of data to process in followed stages (less data to shuffle/sort/etc)
▪ PR Status
▪ Under discussion with community members, will submit JIRA later

Stream-stream Join Improvement (SPARK-32883)
▪ Agenda
▪ Left semi join support
▪ Full outer join support

Quick Refreshment for Stream-Stream Join
▪ Join Physical Operator
▪ StreamingSymmetricHashJoinExec (stream-stream join for Structured Streaming)

Scan A Scan B
. . . . . . . . . .
Stream-stream join
- Join both sides by looking
up from each state store
State
Store
B
Join
State
Store
A
Join
State
Store
B
Join
State
Store
A
Join
State
Store
B
Join
State
Store
A
Join

Left semi stream-stream join (SPARK-32862)
▪ Why?
▪ Left semi is even more popular than left outer (observation on some FB streaming workload)
▪ Get all ads impression (left side) which has ads click (right side), but do not care what those ads clicks are
▪ How?
▪ For left side input row, check if there's a match on right side state store
▪ If there's a match, output the left side row, but do not put the row in left side state store (no need to put in state store).
▪ If there's no match, output nothing, but put the row in left side state store (with "matched" field to set to false in state store).
▪ For right side input row, check if there's a match on left side state store.
▪ For all matched left rows in state store, output the rows with "matched" field as false. Set all left rows with "matched" field to
be true. Only output the left side rows matched for the first time to guarantee left semi join semantics.
▪ State store eviction: evict rows from left/right side state store below watermark, same as inner join.
▪ PR Status

Full outer stream-stream join (SPARK-32863)
▪ How?
▪ for left side input row, check if there's a match on right side state store. If there's a match, output all
matched rows. Put the row in left side state store.
▪ for right side input row, check if there's a match on left side state store. If there's a match, output all
matched rows and update left side rows state with "matched" field to set to true. Put the row in right
side state store.
▪ for left side row needs to be evicted from state store, output the row if "matched" field is false.
▪ for right side row needs to be evicted from state store, output the row if "matched" field is false.
▪ PR Status
▪ WIP

Future Work
▪ History-based optimization (HBO) to select best join strategy
▪ Decide between broadcast hash join, shuffled hash join, and sort merge join based on historical join
input size

Summary
▪ Shuffled Hash Join Improvement (SPARK-32461)
▪ Leverage Bloom Filter for Join (under discussion with community)
▪ Stream-stream Join Improvement (SPARK-32883)

Thank you!
Your feedback is important to us.
Don’t forget to rate
and review the sessions.

Spark SQL Join Improvement at Facebook

Recommended

Recommended

More Related Content

What's hot

What's hot (20)

Similar to Spark SQL Join Improvement at Facebook

Similar to Spark SQL Join Improvement at Facebook (20)

More from Databricks

More from Databricks (20)

Recently uploaded

Recently uploaded (20)

Spark SQL Join Improvement at Facebook