Improving Spark SQL Performance_

Transcript

1. IMPROVING SPARK SQL PERFORMANCE
   Keiji Yoshida - LINE Corporation
   

   View Slide

2. OASIS
   

   View Slide

3. Agenda 1. Issue
   2. Gathering Statistics
   3. Injecting Custom Optimization Rules
   4. Cost-based Optimizer
   

   View Slide

4. Agenda 1. Issue
   2. Gathering Statistics
   3. Injecting Custom Optimization Rules
   4. Cost-based Optimizer
   

   View Slide

5. Users / Day
   1,300+
   Spark Apps / Day
   7,000+
   Spark Jobs / Day
   200+
   USAGE
   

   View Slide

6. SPARK VERSION AND CONFIGURATION
   • 2.4.0
   • spark.sql.cbo.enabled = false
   

   View Slide

7. SLOW QUERIES
   Rank Query Count
   Average Time
   (min)
   Total Time
   (min)
   1 Query A 335 20 6,842
   2 Query B 28 85 2,380
   3 Query C 109 20 2,182
   4 Query D 4028 0 1,543
   5 Query E 14 105 1,480
   

   View Slide

8. IMPACTS
   • Speed of data analysis and decision making
   • Usage efficiency of computing resources
   

   View Slide

9. CHALLENGES FOR END USERS
   • Acquiring knowledge and skills of data engineering
   • Taking care of queries’ performances
   • Changing queries and tables in the production environment
   

   View Slide

10. SOLUTION
    • Improving queries’ performances on the platform side
    

    View Slide

11. Agenda 1. Issue
    2. Gathering Statistics
    3. Injecting Custom Optimization Rules
    4. Cost-based Optimizer
    

    View Slide

12. GATHERING STATISTICS
    • 1,500 sec → 60 sec
    

    View Slide

13. QUERY
    select
    t.dt, m.name, count(1)
    from
    (
    select
    dt, key
    from
    db.transaction
    where
    dt between '20190211' and '20190217'
    ) t
    inner join
    db.master m
    on
    m.key = t.key
    group by
    t.dt, m.name
    

    View Slide

14. QUERY
    select
    t.dt, m.name, count(1)
    from
    (
    select
    dt, key
    from
    db.transaction
    where
    dt between '20190211' and '20190217'
    ) t
    inner join
    db.master m
    on
    m.key = t.key
    group by
    t.dt, m.name
    1,800,000,000 records
    200 records
    

    View Slide

15. PHYSICAL PLAN
    *(6) HashAggregate(keys=[dt#958, name#970], functions=[count(1)])
    +- Exchange hashpartitioning(dt#958, name#970, 200)
    +- *(5) HashAggregate(keys=[dt#958, name#970], functions=[partial_count(1)])
    +- *(5) Project [dt#958, name#970]
    +- *(5) SortMergeJoin [key#937], [key#960], Inner
    :- *(2) Sort [key#937 ASC NULLS FIRST], false, 0
    : +- Exchange hashpartitioning(key#937, 200)
    : +- *(1) Project [dt#958, key#937]
    : +- *(1) Filter isnotnull(key#937)
    : +- *(1) FileScan orc transaction[key#937,dt#958] Batched: true, Format: ORC, Location:…
    +- *(4) Sort [key#960 ASC NULLS FIRST], false, 0
    +- Exchange hashpartitioning(key#960, 200)
    +- *(3) Project [key#960, name#970]
    +- *(3) Filter isnotnull(key#960)
    +- *(3) FileScan orc master[key#960,name#970] Batched: true, Format: ORC, Location:…
    

    View Slide

16. SORT MERGE JOIN
    

    View Slide

17. BROADCAST HASH JOIN
    

    View Slide

18. AUTO BROADCASTJOIN THRESHOLD
    • spark.sql.autoBroadcastJoinThreshold = 10 MB
    

    View Slide

19. QUERY
    select
    t.dt, m.name, count(1)
    from
    (
    select
    dt, key
    from
    db.transaction
    where
    dt between '20190211' and '20190217'
    ) t
    inner join
    db.master m
    on
    m.key = t.key
    group by
    t.dt, m.name
    1,800,000,000 records
    200 records
    

    View Slide

20. PHYSICAL PLAN
    *(6) HashAggregate(keys=[dt#958, name#970], functions=[count(1)])
    +- Exchange hashpartitioning(dt#958, name#970, 200)
    +- *(5) HashAggregate(keys=[dt#958, name#970], functions=[partial_count(1)])
    +- *(5) Project [dt#958, name#970]
    +- *(5) SortMergeJoin [key#937], [key#960], Inner
    :- *(2) Sort [key#937 ASC NULLS FIRST], false, 0
    : +- Exchange hashpartitioning(key#937, 200)
    : +- *(1) Project [dt#958, key#937]
    : +- *(1) Filter isnotnull(key#937)
    : +- *(1) FileScan orc db.transaction[key#937,dt#958] Batched: true, Format: ORC
    +- *(4) Sort [key#960 ASC NULLS FIRST], false, 0
    +- Exchange hashpartitioning(key#960, 200)
    +- *(3) Project [key#960, name#970]
    +- *(3) Filter isnotnull(key#960)
    +- *(3) FileScan orc db.master[key#960,name#970] Batched: true, Format: ORC
    

    View Slide

21. STATS
    val q = "select * from db.master”
    spark.sql(q).queryExecution.optimizedPlan.stats.sizeInBytes
    → 9223372036854775807
    

    View Slide

22. OASIS.INSERTOVERWRITE
    spark.sql(q).repartition(filesNum).write().mode(“overwrite”).insertInto(“db.master”)
    

    View Slide

23. OASIS.INSERTOVERWRITE
    spark.sql(q).repartition(filesNum).write().mode(“overwrite”).insertInto(“db.master”)
    spark.sql(“analyze table db.master compute statistics noscan”)
    

    View Slide

24. STATS
    

    View Slide

25. STATS
    val q = "select * from db.master”
    spark.sql(q).queryExecution.optimizedPlan.stats.sizeInBytes
    → 8239
    

    View Slide

26. PHYSICAL PLAN
    *(3) HashAggregate(keys=[dt#149, name#161], functions=[count(1)])
    +- Exchange hashpartitioning(dt#149, name#161, 200)
    +- *(2) HashAggregate(keys=[dt#149, name#161], functions=[partial_count(1)])
    +- *(2) Project [dt#149, name#161]
    +- *(2) BroadcastHashJoin [key#128], [key#151], Inner, BuildRight
    :- *(2) Project [dt#149, key#128]
    : +- *(2) Filter isnotnull(key#128)
    : +- *(2) FileScan orc db.transaction[key#128,dt#149] Batched: true, Format:…
    +- BroadcastExchange HashedRelationBroadcastMode(List(input[0, string, true]))
    +- *(1) Project [key#151, name#161]
    +- *(1) Filter isnotnull(key#151)
    +- *(1) FileScan orc db.master[key#151,name#161] Batched: true, Format:…
    

    View Slide

27. RESULT
    • 1,500 sec → 60 sec
    

    View Slide

28. Agenda 1. Issue
    2. Gathering Statistics
    3. Injecting Custom Optimization Rules
    4. Cost-based Optimizer
    

    View Slide

29. INJECTING CUSTOM OPTIMIZATION RULES
    • 6,000 sec → 200 sec
    

    View Slide

30. QUERY
    select
    t.dt, a.area_name, count(distinct uid)
    from
    (
    select
    dt, area_code, uid
    from
    db.transaction
    where
    dt between '20180818' and '20190217'
    ) t
    inner join
    (select * from db.area where dt = ‘20190217’) a
    on
    a.area_code = t.area_code
    group by
    t.dt, a.area_name
    

    View Slide

31. QUERY
    select
    t.dt, a.area_name, count(distinct uid)
    from
    (
    select
    dt, area_code, uid
    from
    db.transaction
    where
    dt between '20180818' and '20190217'
    ) t
    inner join
    (select * from db.area where dt = ‘20190217’) a
    on
    a.area_code = t.area_code
    group by
    t.dt, a.area_name
    4,500,000,000 records
    90 records
    

    View Slide

32. PHYSICAL PLAN
    +- *(5) HashAggregate(...)
    +- *(5) HashAggregate(...)
    +- *(5) HashAggregate(...)
    +- *(5) Project [...]
    +- *(5) SortMergeJoin [...], Inner
    :- *(2) Sort [...], false, 0
    : +- Exchange hashpartitioning(...)
    : +- *(1) Project [...]
    : +- *(1) Filter (...)
    : +- *(1) FileScan orc db.transaction[...] Batched: true, Format: ORC
    +- *(4) Sort [...], false, 0
    +- Exchange hashpartitioning(...)
    +- *(3) Filter (...)
    +- Scan hive db.area [...], HiveTableRelation
    

    View Slide

33. HIVE.STATS.AUTOGATHER
    

    View Slide

34. BROADCAST HINT
    select /*+ broadcast(r) */ * from records r join src s on r.key = s.key
    

    View Slide

35. CUSTOMIZE OPTIMIZATION
    spark.experimental.extraOptimizations = MyOptimization :: Nil
    

    View Slide

36. CUSTOMIZE OPTIMIZATION
    object InsertBroadcastHints extends Rule[LogicalPlan] {
    val targets = Map(("db", "area") -> "dt")
    override def apply(plan: LogicalPlan): LogicalPlan = plan transform {
    case join: Join if isSelectFromTarget(join.left) => join.copy(left = ResolvedHint(join.left, HintInfo(true)))
    case join: Join if isSelectFromTarget(join.right) => join.copy(right = ResolvedHint(join.right, HintInfo(true)))
    }
    private def isSelectFromTarget(plan: LogicalPlan): Boolean = plan match {
    case p: Project => p.child match {
    case f: Filter => f.child match {
    case rel: HiveTableRelation => targets.get(rel.tableMeta.database, rel.tableMeta.identifier.table) match {
    case Some(pt) => isFilteredByLiteral(f.condition, pt)
    case None => false
    }
    …
    }
    private def isFilteredByLiteral(expr: Expression, pt: String): Boolean = ...
    }
    spark.experimental.extraOptimizations = InsertBroadcastHints :: Nil
    

    View Slide

37. OPTIMIZED LOGICAL PLAN
    Aggregate [...]
    +- Project [...]
    +- Join Inner, (...)
    :- Project [...]
    : +- Filter (...)
    : +- Relation[...] orc
    +- ResolvedHint (broadcast)
    +- Project [...]
    +- Filter (...)
    +- HiveTableRelation `db`.`area`, ...
    

    View Slide

38. PHYSICAL PLAN
    *(4) HashAggregate(...)
    +- Exchange hashpartitioning(...)
    +- *(3) HashAggregate(...)
    +- *(3) HashAggregate(...)
    +- Exchange hashpartitioning(...)
    +- *(2) HashAggregate(...)
    +- *(2) Project [...]
    +- *(2) BroadcastHashJoin [...], [...], Inner, BuildRight
    :- *(2) Project [...]
    : +- *(2) Filter (...)
    : +- *(2) FileScan orc db.transaction[...] Batched: true, Format: ORC
    +- BroadcastExchange HashedRelationBroadcastMode(...)
    +- *(1) Filter (...)
    +- Scan hive db.area [...], HiveTableRelation
    

    View Slide

39. RESULT
    • 6,000 sec → 200 sec
    

    View Slide

40. Agenda 1. Issue
    2. Gathering Statistics
    3. Injecting Custom Optimization Rules
    4. Cost-based Optimizer
    

    View Slide

41. COST-BASED OPTIMIZATION
    • Apache Spark 2.2.0~
    • Statistics, Cost
    • Advanced Optimizations
    1. Broadcast Hash Join Based on Column Statistics
    2. Join Reordering
    

    View Slide

42. COST-BASED OPTIMIZATION
    • Apache Spark 2.2.0~
    • Statistics, Cost
    • Advanced Optimizations
    1. Broadcast Hash Join Based on Column Statistics
    2. Join Reordering
    

    View Slide

43. QUERY
    select
    a.*
    from
    db.a
    inner join
    db.b
    on
    a.id = b.id
    inner join
    db.c
    on
    b.key = c.key
    

    View Slide

44. QUERY
    select
    a.*
    from
    db.a
    inner join
    db.b
    on
    a.id = b.id
    inner join
    db.c
    on
    b.key = c.key
    1,500,000,000 records
    400,000,000 records
    1 records
    

    View Slide

45. JOIN ORDER
    db.a
    (1,500,000,000 records)
    db.b
    (400,000,000 records)
    db.c (1 record)
    

    View Slide

46. db.a db.b
    db.c
    

    View Slide

47. CBO
    • set spark.sql.cbo.enabled = true
    • set spark.sql.cbo.joinReorder.enabled = true
    

    View Slide

48. db.b db.c
    db.a
    

    View Slide

49. RESULT
    • 1,000 sec → 100 sec
    

    View Slide

50. COST
    • cost = weight * numOfRows + (1.0 - weight) * dataSize
    

    View Slide

51. THANK YOU
    

    View Slide