Spark Bloom Filter Join_bloomfilteraggregate(1),2024年最新腾讯架构师首发

先自我介绍一下,小编浙江大学毕业,去过华为、字节跳动等大厂,目前阿里P7

深知大多数程序员,想要提升技能,往往是自己摸索成长,但自己不成体系的自学效果低效又漫长,而且极易碰到天花板技术停滞不前!

因此收集整理了一份《2024年最新大数据全套学习资料》,初衷也很简单,就是希望能够帮助到想自学提升又不知道该从何学起的朋友。
img
img
img
img
img

既有适合小白学习的零基础资料,也有适合3年以上经验的小伙伴深入学习提升的进阶课程,涵盖了95%以上大数据知识点,真正体系化!

由于文件比较多,这里只是将部分目录截图出来,全套包含大厂面经、学习笔记、源码讲义、实战项目、大纲路线、讲解视频,并且后续会持续更新

如果你需要这些资料,可以添加V获取:vip204888 (备注大数据)
img

正文

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

网上学习资料一大堆,但如果学到的知识不成体系,遇到问题时只是浅尝辄止,不再深入研究,那么很难做到真正的技术提升。

需要这份系统化的资料的朋友,可以添加V获取:vip204888 (备注大数据)
img

一个人可以走的很快,但一群人才能走的更远!不论你是正从事IT行业的老鸟或是对IT行业感兴趣的新人,都欢迎加入我们的的圈子(技术交流、学习资源、职场吐槽、大厂内推、面试辅导),让我们一起学习成长!

Contain

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

网上学习资料一大堆,但如果学到的知识不成体系,遇到问题时只是浅尝辄止,不再深入研究,那么很难做到真正的技术提升。

需要这份系统化的资料的朋友,可以添加V获取:vip204888 (备注大数据)
[外链图片转存中…(img-EoDXGnSv-1713153091791)]

一个人可以走的很快,但一群人才能走的更远!不论你是正从事IT行业的老鸟或是对IT行业感兴趣的新人,都欢迎加入我们的的圈子(技术交流、学习资源、职场吐槽、大厂内推、面试辅导),让我们一起学习成长!

评论
添加红包

请填写红包祝福语或标题

红包个数最小为10个

红包金额最低5元

当前余额3.43前往充值 >
需支付:10.00
成就一亿技术人!
领取后你会自动成为博主和红包主的粉丝 规则
hope_wisdom
发出的红包
实付
使用余额支付
点击重新获取
扫码支付
钱包余额 0

抵扣说明:

1.余额是钱包充值的虚拟货币,按照1:1的比例进行支付金额的抵扣。
2.余额无法直接购买下载,可以购买VIP、付费专栏及课程。

余额充值