背景
本文基于 SPARK 3.3.0
从一个unit test来探究SPARK Codegen的逻辑,
test("SortAggregate should be included in WholeStageCodegen") { val df = spark.range(10).agg(max(col("id")), avg(col("id"))) withSQLConf("spark.sql.test.forceApplySortAggregate" -> "true") { val plan = df.queryExecution.executedPlan assert(plan.exists(p => p.isInstanceOf[WholeStageCodegenExec] && p.asInstanceOf[WholeStageCodegenExec].child.isInstanceOf[SortAggregateExec])) assert(df.collect() === Array(Row(9, 4.5))) } }
该sql形成的执行计划第二部分的全代码生成部分如下:
WholeStageCodegen *(2) SortAggregate(key=[], functions=[max(id#0L), avg(id#0L)], output=[max(id)#5L, avg(id)#6]) InputAdapter +- Exchange SinglePartition, ENSURE_REQUIREMENTS, [id=#13]
分析
第二阶段wholeStageCodegen
第二阶段的代码生成涉及到SortAggregateExec和ShuffleExchangeExec以及InputAdapter的produce和consume方法,这里一一来分析:
第二阶段wholeStageCodegen数据流如下:
WholeStageCodegenExec SortAggregateExec(Final) InputAdapter ShuffleExchangeExec ==================================================================================== -> execute() | doExecute() ---------> inputRDDs() -----------------> inputRDDs() -------> execute() | | doCodeGen() doExecute() | | +-----------------> produce() ShuffledRowRDD | doProduce() | doProduceWithoutKeys() -------> produce() | doProduce() | doConsume() <------------------- consume() | doConsumeWithoutKeys() |并不是doConsumeWithoutKeys调用consume,而是由doProduceWithoutKeys调用 doConsume() <-------- consume()
SortAggregateExec(Final) 的doProduce
这里只列出和SortAggregateExec(Partial)的不同的部分:
val (resultVars, genResult) = if (modes.contains(Final) || modes.contains(Complete)) { // evaluate aggregate results ctx.currentVars = flatBufVars val aggResults = bindReferences( functions.map(_.evaluateExpression), aggregateBufferAttributes).map(_.genCode(ctx)) val evaluateAggResults = evaluateVariables(aggResults) // evaluate result expressions ctx.currentVars = aggResults val resultVars = bindReferences(resultExpressions, aggregateAttributes).map(_.genCode(ctx)) (resultVars, s""" |$evaluateAggResults |${evaluateVariables(resultVars)} """.stripMargin)
因为我们这里是Final部分,所以我们的数据流和Partial是不同的
ctx.currentVars = flatBufVars
赋值currentVars为当前buffer变量,便于下面进行数据绑定,该buffer变量是全局变量
val aggResults = bindReferences
functions.map(_.evaluateExpression) 这是对最终输出结果的计算,对于SUM来说是Divide(sum.cast(resultType), count.cast(resultType), failOnError = false) ,生成的代码如下:
boolean sortAgg_isNull_6 = sortAgg_bufIsNull_2; double sortAgg_value_6 = -1.0; if (!sortAgg_bufIsNull_2) { sortAgg_value_6 = (double) sortAgg_bufValue_2; } boolean sortAgg_isNull_4 = false; double sortAgg_value_4 = -1.0; if (sortAgg_isNull_6 || sortAgg_value_6 == 0) { sortAgg_isNull_4 = true; } else { if (sortAgg_bufIsNull_1) { sortAgg_isNull_4 = true; } else { sortAgg_value_4 = (double)(sortAgg_bufValue_1 / sortAgg_value_6); } }
aggregateBufferAttributes 聚合函数的buffer属性值 sum :: count :: Nil
这样在绑定数据的变量数据的时候和currentVars是一一对应的
val evaluateAggResults = evaluateVariables(aggResults)
对聚合的结果进行最终的计算
ctx.currentVars = aggResults
把最终结果的变量赋值给currentVars,便于后面的数据绑定
val resultVars = bindReferences(resultExpressions, aggregateAttributes).map(_.genCode(ctx))
这一步是把聚合结果的变量绑定到聚合表达式中,
其中resultExpressions为List( avg(id#0L)#3 AS avg(id)#6) (这里我们只考虑AVG)
aggregateAttributes是resultExpression的AttributeReference的一种表达,便于在BoundReference的时候进行映射绑定
对应的ExprCode为ExprCode(,sortAgg_isNull_4,sortAgg_value_4))
InputAdaptor的 doProduce
InputAdaptor的主要作用是承上启下,用来适配不支持Codegen的物理计划,sql如下:
override def doProduce(ctx: CodegenContext): String = { // Inline mutable state since an InputRDDCodegen is used once in a task for WholeStageCodegen val input = ctx.addMutableState("scala.collection.Iterator", "input", v => s"$v = inputs[0];", forceInline = true) val row = ctx.freshName("row") val outputVars = if (createUnsafeProjection) { // creating the vars will make the parent consume add an unsafe projection. ctx.INPUT_ROW = row ctx.currentVars = null output.zipWithIndex.map { case (a, i) => BoundReference(i, a.dataType, a.nullable).genCode(ctx) } } else { null } val updateNumOutputRowsMetrics = if (metrics.contains("numOutputRows")) { val numOutputRows = metricTerm(ctx, "numOutputRows") s"$numOutputRows.add(1);" } else { "" } s""" | while ($limitNotReachedCond $input.hasNext()) { | InternalRow $row = (InternalRow) $input.next(); | ${updateNumOutputRowsMetrics} | ${consume(ctx, outputVars, if (createUnsafeProjection) null else row).trim} | ${shouldStopCheckCode} | } """.stripMargin }
val input = ctx.addMutableState(“scala.collection.Iterator”, “input”, v => s"$v = inputs[0];"
定义一个input变量用来接受sortaggregate(partial)的输出的InteralRow(unsafeRow),对应的初始化方法会在init方法中调用
val row = ctx.freshName(“row”)
定义一个临时变量用来接受input中的unsafe类型的InteralRow,便于进行迭代操作
val outputVars = if (createUnsafeProjection)
对于InputAdaptor来说createUnsafeProjection是 false, 所以这块返回的是null
val updateNumOutputRowsMetrics =
因为metrics不满足条件,所以这里也是返回空字符串
代码组装
s""" | while ($limitNotReachedCond $input.hasNext()) { | InternalRow $row = (InternalRow) $input.next(); | ${updateNumOutputRowsMetrics} | ${consume(ctx, outputVars, if (createUnsafeProjection) null else row).trim} | ${shouldStopCheckCode} | } """.stripMargin
对输入的每一行数据进行迭代操作, 之后再调用consume方法,
注意: 这里的consume传入的是row,是InteralRow类型,而不是在RangeExec中的Long类型的变量
InputAdaptor的 consume
我们这里只说明和之前不一样的部分,对应的sql如下:
final def consume(ctx: CodegenContext, outputVars: Seq[ExprCode], row: String = null): String =
注意这里的参数 outputVars
为null
row
为InteralRow
类型的变量
- val inputVarsCandidate =
val inputVarsCandidate = if (outputVars != null) { assert(outputVars.length == output.length) // outputVars will be used to generate the code for UnsafeRow, so we should copy them outputVars.map(_.copy()) } else { assert(row != null, "outputVars and row cannot both be null.") ctx.currentVars = null ctx.INPUT_ROW = row output.zipWithIndex.map { case (attr, i) => BoundReference(i, attr.dataType, attr.nullable).genCode(ctx) } }
这里的数据流向了 else
:
- ctx.INPUT_ROW = row
设置当前的INPUT_ROW
为row
BoundReference
的doGenCode
方法也是走向了另一个分支:
assert(ctx.INPUT_ROW != null, "INPUT_ROW and currentVars cannot both be null.") val javaType = JavaCode.javaType(dataType) val value = CodeGenerator.getValue(ctx.INPUT_ROW, dataType, ordinal.toString) if (nullable) { ev.copy(code = code""" |boolean ${ev.isNull} = ${ctx.INPUT_ROW}.isNullAt($ordinal); |$javaType ${ev.value} = ${ev.isNull} ? | ${CodeGenerator.defaultValue(dataType)} : ($value); """.stripMargin) } else { ev.copy(code = code"$javaType ${ev.value} = $value;", isNull = FalseLiteral) }
分析
val value = CodeGenerator.getValue(ctx.INPUT_ROW, dataType,ordinal.toString)
根据数据类型的不同,调用UnsafeRow的不同方法
if (nullable)
因为AttributeReference("sum", sumDataType)()和AttributeReference("count", LongType)()表达式 nullable 为 TRUE,所以生成的代码为:
boolean inputadapter_isNull_0 = inputadapter_row_0.isNullAt(0); long inputadapter_value_0 = inputadapter_isNull_0 ? -1L : (inputadapter_row_0.getLong(0)); boolean inputadapter_isNull_1 = inputadapter_row_0.isNullAt(1); double inputadapter_value_1 = inputadapter_isNull_1 ? -1.0 : (inputadapter_row_0.getDouble(1)); boolean inputadapter_isNull_2 = inputadapter_row_0.isNullAt(2); long inputadapter_value_2 = inputadapter_isNull_2 ? -1L : (inputadapter_row_0.getLong(2));
- constructDoConsumeFunction方法中inputVarsInFunc
这里会多一个名为inputadapter_row_0的InternalRow类型的实参