Spark Bloom Filter Join

1 综述

1.1 目的

  Bloom Filter Join，或者说Row-level Runtime Filtering（还额外有一条Semi-Join分支），是Spark 3.3对运行时过滤的一个最新补充
  之前运行时过滤主要有两个：动态分区裁剪DPP（开源实现）、动态文件裁剪DFP（Databricks实现），两者都能有效减少数据源层面的Scan IO
  Bloom Filter Join的主要优化点是在shuffle层，通过在join shuffle前对表进行过滤从而提高运行效率

1.2 场景

• 普通的shuffle join
  

• Broadcast join并且子结构中存在shuffle
  

1.3 基础过程

  将存在过滤条件的小表端称为Filter Creation Side，另一层称为Filter Application Side
  对于如下的SQL：SELECT * FROM R JOIN S ON R.r_sk = S.s_sk where S.x = 5
  首先Creation端进行bloomFilter创建，简单来说就是对小表创建一个bloomFilter的过滤数据集合

SELECT BloomFilterAggregate(XxHash64(S.s_sk), n_items, n_bits)
FROM S where S.x = 5

  之后Application端进行重写（实际是整个查询重写），就是把小表的bloomFilter数据集合拿来对大表的数据进行过滤
  根据上面的场景图看，其实小表Creation端在整个SQL树上并没有变化，只改变了大表端的树结构

SELECT *
FROM R JOIN S ON R.r_sk = S.s_sk
WHERE S.x=5 AND BloomFilterMightContain( 
(
  SELECT BloomFilterAggregate(XxHash64(S.s_sk), n_items, n_bits) bloom_filter
  		  FROM S where S.x = 5 ),     -- Bloom filter creation
	          XxHash64(R.r_sk))       -- Bloom filter application

1.4 触发条件

  设计文档中写的触发条件

1. 小表在broadcast join当中（存疑）
2. 小表有过滤器
3. 小表是Scan (-> Project) -> Filter的建档形式，否则依赖流增加可能延长查询时间
4. 小表是确定性的
5. 大表端有shuffle，小表可以通过shuffl传送bloomFilter结果
6. join的列上没有应用DPP

2 InjectRuntimeFilter

  InjectRuntimeFilter是Spark源码中对应的优化器类，只执行一次（FixedPoint(1)和Once的差异是Once强制幂等）

Batch("InjectRuntimeFilter", FixedPoint(1),
  InjectRuntimeFilter) :+

  apply中定义了规则的整体流程，前面是两个条件判断

//  相关子查询不支持，相关子查询的子查询结果依赖于主查询，不能应用
case s: Subquery if s.correlated => plan
//  相关的配置开关是否开启
case _ if !conf.runtimeFilterSemiJoinReductionEnabled &&
  !conf.runtimeFilterBloomFilterEnabled => plan
case _ =>
  //  应用优化规则，尝试注入运行时过滤器
  val newPlan = tryInjectRuntimeFilter(plan)
  //  semi join配置未开或者规则应用后无变化，不处理
  if (conf.runtimeFilterSemiJoinReductionEnabled && !plan.fastEquals(newPlan)) {
  //  子查询重写成semi/anti join
    RewritePredicateSubquery(newPlan)
  } else {
    newPlan
  }

  相关的配置为，默认bloomFilter开启了，Semi join关闭的

val RUNTIME_FILTER_SEMI_JOIN_REDUCTION_ENABLED =
  buildConf("spark.sql.optimizer.runtimeFilter.semiJoinReduction.enabled")
    .doc("When true and if one side of a shuffle join has a selective predicate, we attempt " +
      "to insert a semi join in the other side to reduce the amount of shuffle data.")
    .version("3.3.0")
    .booleanConf
    .createWithDefault(false)
    
val RUNTIME_BLOOM_FILTER_ENABLED =
  buildConf("spark.sql.optimizer.runtime.bloomFilter.enabled")
    .doc("When true and if one side of a shuffle join has a selective predicate, we attempt " +
      "to insert a bloom filter in the other side to reduce the amount of shuffle data.")
    .version("3.3.0")
    .booleanConf
    .createWithDefault(true)

2.1 tryInjectRuntimeFilter

  tryInjectRuntimeFilter使用核心的处理流程，尝试应用Runtime Filter，整体代码如下

private def tryInjectRuntimeFilter(plan: LogicalPlan): LogicalPlan = {
  var filterCounter = 0
  val numFilterThreshold = conf.getConf(SQLConf.RUNTIME_FILTER_NUMBER_THRESHOLD)
  plan transformUp {
    case join @ ExtractEquiJoinKeys(joinType, leftKeys, rightKeys, _, _, left, right, hint) =>
      var newLeft = left
      var newRight = right
      (leftKeys, rightKeys).zipped.foreach((l, r) => {
        // Check if:
        // 1. There is already a DPP filter on the key
        // 2. There is already a runtime filter (Bloom filter or IN subquery) on the key
        // 3. The keys are simple cheap expressions
        if (filterCounter < numFilterThreshold &&
          !hasDynamicPruningSubquery(left, right, l, r) &&
          !hasRuntimeFilter(newLeft, newRight, l, r) &&
          isSimpleExpression(l) && isSimpleExpression(r)) {
          val oldLeft = newLeft
          val oldRight = newRight
          if (canPruneLeft(joinType) && filteringHasBenefit(left, right, l, hint)) {
            newLeft = injectFilter(l, newLeft, r, right)
          }
          // Did we actually inject on the left? If not, try on the right
          if (newLeft.fastEquals(oldLeft) && canPruneRight(joinType) &&
            filteringHasBenefit(right, left, r, hint)) {
            newRight = injectFilter(r, newRight, l, left)
          }
          if (!newLeft.fastEquals(oldLeft) || !newRight.fastEquals(oldRight)) {
            filterCounter = filterCounter + 1
          }
        }
      })
      join.withNewChildren(Seq(newLeft, newRight))
  }
}

  过程中有很多的条件判断，应用Runtime Filter的基本条件：

1. 插入的Runtime Filter没超过阈值（默认10）
2. 等值条件的Key上不能有DPP、Runtime Filter
3. 等值条件的Key是一个简单表达式（即没有套上UDF等）

  之后根据条件，选择将Runtime Filter应用到左子树还是右子树，条件为

1. Join类型支持下推（比如RightOuter只能用于左子树）
2. Application端支持通过joins、aggregates、windows下推过滤条件
3. Creation端有过滤条件
4. 当前join是shuffle join或者是一个子结构中包含shuffle的broadcast join
5. Application端的扫描数据大于阈值（默认10G）

  提到的两个阈值的配置项

val RUNTIME_FILTER_NUMBER_THRESHOLD =
  buildConf("spark.sql.optimizer.runtimeFilter.number.threshold")
    .doc("The total number of injected runtime filters (non-DPP) for a single " +
      "query. This is to prevent driver OOMs with too many Bloom filters.")
    .version("3.3.0")
    .intConf
    .checkValue(threshold => threshold >= 0, "The threshold should be >= 0")
    .createWithDefault(10)

val RUNTIME_BLOOM_FILTER_APPLICATION_SIDE_SCAN_SIZE_THRESHOLD =
  buildConf("spark.sql.optimizer.runtime.bloomFilter.applicationSideScanSizeThreshold")
    .doc("Byte size threshold of the Bloom filter application side plan's aggregated scan " +
      "size. Aggregated scan byte size of the Bloom filter application side needs to be over " +
      "this value to inject a bloom filter.")
    .version("3.3.0")
    .bytesConf(ByteUnit.BYTE)
    .createWithDefaultString("10GB")

2.2 injectFilter

  injectFilter是核心进行Runtime Filter规则应用的地方，在此处，bloomFilter和Semi Join是互斥的，只能有一个执行

if (conf.runtimeFilterBloomFilterEnabled) {
  injectBloomFilter(
    filterApplicationSideExp,
    filterApplicationSidePlan,
    filterCreationSideExp,
    filterCreationSidePlan
  )
} else {
  injectInSubqueryFilter(
    filterApplicationSideExp,
    filterApplicationSidePlan,
    filterCreationSideExp,
    filterCreationSidePlan
  )

2.3 injectBloomFilter

2.3.1 执行条件

  首先进行一个判断，在Creation端的数据不能大于阈值（Creation端数据量大会导致bloomFilter的误判率高，最终过滤效果差）

// Skip if the filter creation side is too big
if (filterCreationSidePlan.stats.sizeInBytes > conf.runtimeFilterCreationSideThreshold) {
  return filterApplicationSidePlan
}

  阈值配置默认10M

val RUNTIME_BLOOM_FILTER_CREATION_SIDE_THRESHOLD =
  buildConf("spark.sql.optimizer.runtime.bloomFilter.creationSideThreshold")
    .doc("Size threshold of the bloom filter creation side plan. Estimated size needs to be " +
      "under this value to try to inject bloom filter.")
    .version("3.3.0")
    .bytesConf(ByteUnit.BYTE)
    .createWithDefaultString("10MB")

  Creation端的数据是一个预估数据，是LogicalPlan中的属性LogicalPlanStats获取的，分是否开启CBO，具体获取方式待研究

def stats: Statistics = statsCache.getOrElse {
  if (conf.cboEnabled) {
    statsCache = Option(BasicStatsPlanVisitor.visit(self))
  } else {
    statsCache = Option(SizeInBytesOnlyStatsPlanVisitor.visit(self))
  }
  statsCache.get
}

2.3.2 创建Creation端的聚合

  就是创建一个bloomFilter的聚合函数BloomFilterAggregate，是AggregateFunction的子类，属于Expression。根据统计信息中是否存在行数，会传入不同的参数

val rowCount = filterCreationSidePlan.stats.rowCount
val bloomFilterAgg =
  if (rowCount.isDefined && rowCount.get.longValue > 0L) {
    new BloomFilterAggregate(new XxHash64(Seq(filterCreationSideExp)), rowCount.get.longValue)
  } else {
    new BloomFilterAggregate(new XxHash64(Seq(filterCreationSideExp)))
  }

2.3.3 创建Application端的过滤条件

  根据1.3中的描述，此处就是把上节中Creation端创建的bloomFilter过滤条件构建成Application端的条件
  Alias就是一个别名的效果；ColumnPruning就是进行列裁剪，后续不需要的列不读取；ConstantFolding就是进行常量折叠；ScalarSubquery是标量子查询，标量子查询的查询结果是一行一列的值（单一值）
  BloomFilterMightContain就是一个内部标量函数，检查数据是否由bloomFilter包含，继承自Predicate，返回boolean值

val alias = Alias(bloomFilterAgg.toAggregateExpression(), "bloomFilter")()
val aggregate =
  ConstantFolding(ColumnPruning(Aggregate(Nil, Seq(alias), filterCreationSidePlan)))
val bloomFilterSubquery = ScalarSubquery(aggregate, Nil)
val filter = BloomFilterMightContain(bloomFilterSubquery,
  new XxHash64(Seq(filterApplicationSideExp)))

  最终结果是在原Application端的计划树上加一个filter，如下就是最终的返回结果

Filter(filter, filterApplicationSidePlan)

2.4 injectInSubqueryFilter

  injectInSubqueryFilter整体流程与injectBloomFilter差不多，差异应该是在Application端生成的过滤条件变成in

val actualFilterKeyExpr = mayWrapWithHash(filterCreationSideExp)
val alias = Alias(actualFilterKeyExpr, actualFilterKeyExpr.toString)()
val aggregate =
  ColumnPruning(Aggregate(Seq(filterCreationSideExp), Seq(alias), filterCreationSidePlan))
if (!canBroadcastBySize(aggregate, conf)) {
  // Skip the InSubquery filter if the size of `aggregate` is beyond broadcast join threshold,
  // i.e., the semi-join will be a shuffled join, which is not worthwhile.
  return filterApplicationSidePlan
}
val filter = InSubquery(Seq(mayWrapWithHash(filterApplicationSideExp)),
  ListQuery(aggregate, childOutputs = aggregate.output))
Filter(filter, filterApplicationSidePlan)

  这里有一个小优化就是mayWrapWithHash，当数据类型的大小超过int时，就是把数据转为hash

// Wraps `expr` with a hash function if its byte size is larger than an integer.
private def mayWrapWithHash(expr: Expression): Expression = {
  if (expr.dataType.defaultSize > IntegerType.defaultSize) {
    new Murmur3Hash(Seq(expr))
  } else {
    expr
  }
}

3 BloomFilterAggregate

  类有三个核心参数：

1. child：子表达式，就是InjectRuntimeFilter里传的XxHash64，目前看起来数据先经过XxHash64处理成long再放入BloomFilter
2. estimatedNumItemsExpression：估计的数据量，如果InjectRuntimeFilter没拿到统计信息，就用配置的默认值
3. numBitsExpression：要使用的bit数

case class BloomFilterAggregate(
    child: Expression,
    estimatedNumItemsExpression: Expression,
    numBitsExpression: Expression,

  estimatedNumItemsExpression和numBitsExpression对应的配置如下

val RUNTIME_BLOOM_FILTER_EXPECTED_NUM_ITEMS =
  buildConf("spark.sql.optimizer.runtime.bloomFilter.expectedNumItems")
    .doc("The default number of expected items for the runtime bloomfilter")
    .version("3.3.0")
    .longConf
    .createWithDefault(1000000L)
    
val RUNTIME_BLOOM_FILTER_NUM_BITS =
  buildConf("spark.sql.optimizer.runtime.bloomFilter.numBits")
    .doc("The default number of bits to use for the runtime bloom filter")
    .version("3.3.0")
    .longConf
    .createWithDefault(8388608L)

  BloomFilter用的是Spark自己实现的一个类BloomFilterImpl，BloomFilterAggregate的createAggregationBuffer接口中创建

override def createAggregationBuffer(): BloomFilter = {
  BloomFilter.create(estimatedNumItems, numBits)
}

  参数就是前面的estimatedNumItemsExpression和numBitsExpression，是懒加载的参数（应该在处理过程会被改变，所以实际跟前面的值之间还加了一层与默认值的比较赋值）

// Mark as lazy so that `estimatedNumItems` is not evaluated during tree transformation.
private lazy val estimatedNumItems: Long =
  Math.min(estimatedNumItemsExpression.eval().asInstanceOf[Number].longValue,
    SQLConf.get.getConf(RUNTIME_BLOOM_FILTER_MAX_NUM_ITEMS))

  处理数据的接口应该是update，把数据用XxHash64处理后加入BloomFilter

override def update(buffer: BloomFilter, inputRow: InternalRow): BloomFilter = {
  val value = child.eval(inputRow)
  // Ignore null values.
  if (value == null) {
    return buffer
  }
  buffer.putLong(value.asInstanceOf[Long])
  buffer
}

  对象BloomFilterAggregate有对应的序列化和反序列化接口

object BloomFilterAggregate {
  final def serialize(obj: BloomFilter): Array[Byte] = {
    // BloomFilterImpl.writeTo() writes 2 integers (version number and num hash functions), hence
    // the +8
    val size = (obj.bitSize() / 8) + 8
    require(size <= Integer.MAX_VALUE, s"actual number of bits is too large $size")
    val out = new ByteArrayOutputStream(size.intValue())
    obj.writeTo(out)
    out.close()
    out.toByteArray
  }

  final def deserialize(bytes: Array[Byte]): BloomFilter = {
    val in = new ByteArrayInputStream(bytes)
    val bloomFilter = BloomFilter.readFrom(in)
    in.close()
    bloomFilter
  }
}

4 BloomFilterMightContain

  有两个参数

1. bloomFilterExpression：是上节BloomFilter的二进制数据
2. valueExpression：应该跟上节的child一致，对输入数据做处理的表达式，XxHash64

case class BloomFilterMightContain(
    bloomFilterExpression: Expression,
    valueExpression: Expression)

  bloomFilter通过反序列化获取

// The bloom filter created from `bloomFilterExpression`.
@transient private lazy val bloomFilter = {
  val bytes = bloomFilterExpression.eval().asInstanceOf[Array[Byte]]
  if (bytes == null) null else deserialize(bytes)
}

  做数据判断的应该是eval，就是调用的BloomFilter的接口进行判断。eval应该就是Spark中Expression表达式的执行接口

override def eval(input: InternalRow): Any = {
  if (bloomFilter == null) {
    null
  } else {
    val value = valueExpression.eval(input)
    if (value == null) null else bloomFilter.mightContainLong(value.asInstanceOf[Long])
  }
}

  也有doGenCode接口用来生成代码

override def doGenCode(ctx: CodegenContext, ev: ExprCode): ExprCode = {
  if (bloomFilter == null) {
    ev.copy(isNull = TrueLiteral, value = JavaCode.defaultLiteral(dataType))
  } else {
    val bf = ctx.addReferenceObj("bloomFilter", bloomFilter, classOf[BloomFilter].getName)
    val valueEval = valueExpression.genCode(ctx)
    ev.copy(code = code"""
    ${valueEval.code}
    boolean ${ev.isNull} = ${valueEval.isNull};
    ${CodeGenerator.javaType(dataType)} ${ev.value} = ${CodeGenerator.defaultValue(dataType)};
    if (!${ev.isNull}) {
      ${ev.value} = $bf.mightContainLong((Long)${valueEval.value});
    }""")
  }
}

5 计划变更

  取Spark单元测试的样例（InjectRuntimeFilterSuite）：select * from bf1 join bf2 on bf1.c1 = bf2.c2 where bf2.a2 = 62

• 规则前的plan

GlobalLimit 21
+- LocalLimit 21
   +- Project [cast(a1#38430 as string) AS a1#38468, cast(b1#38431 as string) AS b1#38469, cast(c1#38432 as string) AS c1#38470, cast(d1#38433 as string) AS d1#38471, cast(e1#38434 as string) AS e1#38472, cast(f1#38435 as string) AS f1#38473, cast(a2#38436 as string) AS a2#38474, cast(b2#38437 as string) AS b2#38475, cast(c2#38438 as string) AS c2#38476, cast(d2#38439 as string) AS d2#38477, cast(e2#38440 as string) AS e2#38478, cast(f2#38441 as string) AS f2#38479]
      +- Join Inner, (c1#38432 = c2#38438)
         :- Filter isnotnull(c1#38432)
         :  +- Relation spark_catalog.default.bf1[a1#38430,b1#38431,c1#38432,d1#38433,e1#38434,f1#38435] parquet
         +- Filter ((isnotnull(a2#38436) AND (a2#38436 = 62)) AND isnotnull(c2#38438))
            +- Relation spark_catalog.default.bf2[a2#38436,b2#38437,c2#38438,d2#38439,e2#38440,f2#38441] parquet

• 规则后的plan

GlobalLimit 21
+- LocalLimit 21
   +- Project [cast(a1#38430 as string) AS a1#38468, cast(b1#38431 as string) AS b1#38469, cast(c1#38432 as string) AS c1#38470, cast(d1#38433 as string) AS d1#38471, cast(e1#38434 as string) AS e1#38472, cast(f1#38435 as string) AS f1#38473, cast(a2#38436 as string) AS a2#38474, cast(b2#38437 as string) AS b2#38475, cast(c2#38438 as string) AS c2#38476, cast(d2#38439 as string) AS d2#38477, cast(e2#38440 as string) AS e2#38478, cast(f2#38441 as string) AS f2#38479]
      +- Join Inner, (c1#38432 = c2#38438)
         :- Filter might_contain(scalar-subquery#38494 [], xxhash64(c1#38432, 42))
         :  :  +- Aggregate [bloom_filter_agg(xxhash64(c2#38438, 42), 1000000, 8388608, 0, 0) AS bloomFilter#38493]
         :  :     +- Project [c2#38438]
         :  :        +- Filter ((isnotnull(a2#38436) AND (a2#38436 = 62)) AND isnotnull(c2#38438))
         :  :           +- Relation spark_catalog.default.bf2[a2#38436,b2#38437,c2#38438,d2#38439,e2#38440,f2#38441] parquet
         :  +- Filter isnotnull(c1#38432)
         :     +- Relation spark_catalog.default.bf1[a1#38430,b1#38431,c1#38432,d1#38433,e1#38434,f1#38435] parquet
         +- Filter ((isnotnull(a2#38436) AND (a2#38436 = 62)) AND isnotnull(c2#38438))
            +- Relation spark_catalog.default.bf2[a2#38436,b2#38437,c2#38438,d2#38439,e2#38440,f2#38441] parquet