This is an automated email from the ASF dual-hosted git repository.
cloud-fan pushed a commit to branch branch-4.2
in repository https://gitbox.apache.org/repos/asf/spark.git
The following commit(s) were added to refs/heads/branch-4.2 by this push:
new aa564a5e0455 [SPARK-56482][SQL][FOLLOWUP] Simplify UnionExec codegen
and narrow partition-index gate
aa564a5e0455 is described below
commit aa564a5e045545b074ba1aa5703b97d3707a7eb7
Author: Wenchen Fan <[email protected]>
AuthorDate: Fri May 8 17:37:20 2026 +0800
[SPARK-56482][SQL][FOLLOWUP] Simplify UnionExec codegen and narrow
partition-index gate
### What changes were proposed in this pull request?
Followup to SPARK-56482 (#55425). Two groups of changes to `UnionExec`'s
whole-stage codegen path.
**Code cleanness:**
- Hoist `metricTerm("numOutputRows")` to `doProduce` and store it on the
instance. `doConsume` runs once per child during emission, so the previous code
registered the same metric N times in `references[]` for an N-child Union; now
once.
- Drop the dead `assert` in `perChildProjections` and the duplicate
`allChildOutputDataTypesMatch` lazy val. The dataType comparison now has a
single source of truth in the `type-mismatch` branch of the gate.
- Inline the one-shot `hasAnyPartitionIndexDependentDescendant` lazy val.
- Drop the unreachable `case other` in the `UnionPartition` match and
replace with `asInstanceOf`. `unionedInputRDD` is built as `new UnionRDD(...)`
two lines up, and `getPartitions` only ever returns `UnionPartition[_]`.
- Factor `isPlainUnion` helper used by the gate and `doExecute` so the
invariant "codegen path matches `sparkContext.union` semantics" lives in one
place.
- Bind `currentPartitionIndexVar` to the array-deref expression `((int[])
refs[K])[partitionIndex]` directly. An earlier revision hoisted this to a
`childLocalIdx` local at helper entry, but `SampleExec.doConsume` reads
`currentPartitionIndexVar` from inside an `addMutableState` initializer, which
is emitted into the state-init function — outside the per-child helper — so the
local was not in scope and the generated code failed to compile. The expression
form resolves in any emission [...]
- Drop the `try/finally` around codegen state restoration. Codegen failure
aborts the whole stage, so the restoration is unreachable.
**Gate narrowing:**
- Narrow `hasPartitionIndexDependentCodegen` to exclude `InputFileName`,
`InputFileBlockStart`, and `InputFileBlockLength`. These are `Nondeterministic`
but read from `InputFileBlockHolder` (a per-task thread-local) and do not embed
`partitionIndex`, so they are safe under fusion. Queries like `SELECT
input_file_name() FROM a UNION ALL SELECT input_file_name() FROM b` now fuse.
### Why are the changes needed?
The cleanups remove accidental complexity in the fused code path: an N-fold
metric reference, two duplicated dataType comparisons, an unreachable defensive
guard, and a `try/finally` that protects against an unreachable case. The gate
narrowing turns a missed optimization (file-scan unions) into a fused plan.
### Does this PR introduce _any_ user-facing change?
No. `spark.sql.codegen.wholeStage.union.enabled` remains off by default;
when on, the new behavior fuses additional plans (file-scan unions with
`input_file_name()`) that the previous gate over-rejected.
### How was this patch tested?
`UnionCodegenSuite`, `UnionCodegenAnsiSuite`, `UnionCodegenAqeSuite`, and
the relevant `SQLMetricsSuite` test all pass. Three tests added:
- `partitioning-aware union falls back to non-codegen` — covers a
`supportCodegenFailureReason` branch that lacked explicit coverage.
- `input_file_name child fuses (Nondeterministic but partition-index-free)`
— validates the gate narrowing.
- `union with sample children fuses (or falls back) without crashing` —
regression test for the `currentPartitionIndexVar` binding (caught by
LuciferYang in review).
The `columnar` fallback branch is not covered by a new test: reliably
constructing a plan where `Union.supportsColumnar` is true via the user-facing
API turned out to be brittle, since `ApplyColumnarRulesAndInsertTransitions`
aggressively rebalances columnar/row transitions.
### Was this patch authored or co-authored using generative AI tooling?
Generated-by: Claude Code
Closes #55719 from cloud-fan/SPARK-56482-followup.
Authored-by: Wenchen Fan <[email protected]>
Signed-off-by: Wenchen Fan <[email protected]>
(cherry picked from commit d905e7378b6dbdf04b3671323583536f662dea19)
Signed-off-by: Wenchen Fan <[email protected]>
---
.../sql/execution/basicPhysicalOperators.scala | 159 ++++++++++-----------
.../spark/sql/execution/UnionCodegenSuite.scala | 51 +++++++
2 files changed, 128 insertions(+), 82 deletions(-)
diff --git
a/sql/core/src/main/scala/org/apache/spark/sql/execution/basicPhysicalOperators.scala
b/sql/core/src/main/scala/org/apache/spark/sql/execution/basicPhysicalOperators.scala
index 9d27f39f6111..92cf3f59d575 100644
---
a/sql/core/src/main/scala/org/apache/spark/sql/execution/basicPhysicalOperators.scala
+++
b/sql/core/src/main/scala/org/apache/spark/sql/execution/basicPhysicalOperators.scala
@@ -901,19 +901,19 @@ case class UnionExec(children: Seq[SparkPlan]) extends
SparkPlan with CodegenSup
}
}
- // `WidenSetOperationTypes` inserts a `Project(Cast)` above each child whose
- // dataType differs from the widened set type, so on the codegen path
- // `src.dataType == tgt.dataType` holds. The Alias only remaps each child
- // attribute onto the union's output exprId/name/metadata. Mismatched cases
- // are gated upstream by `allChildOutputDataTypesMatch`, so the assert is a
- // defensive guard.
+ // True when the codegen path applies: `outputPartitioning` is
`UnknownPartitioning`,
+ // and `unionedInputRDD` matches the semantics of `sparkContext.union(...)`
in `doExecute`.
+ private def isPlainUnion: Boolean =
outputPartitioning.isInstanceOf[UnknownPartitioning]
+
+ // Per-child projection from the child's output to the union's output. The
wrapped
+ // child is always the source `Attribute` (deterministic by construction);
the Alias
+ // only remaps the exprId/name/metadata. `WidenSetOperationTypes` aligns
top-level
+ // dataTypes, but nested nullability differences bypass it; those cases are
caught
+ // by the `type-mismatch` gate below, which is the single source of truth
for the
+ // `src.dataType == tgt.dataType` invariant `doConsume` relies on.
@transient private lazy val perChildProjections:
IndexedSeq[Seq[NamedExpression]] =
children.toIndexedSeq.map { child =>
child.output.zip(output).map { case (src, tgt) =>
- assert(src.dataType == tgt.dataType,
- s"UnionExec child output dataType ${src.dataType} does not match " +
- s"union output dataType ${tgt.dataType}; supportCodegen should " +
- "have returned false via the 'type-mismatch' reason.")
Alias(src, tgt.name)(
exprId = tgt.exprId,
qualifier = tgt.qualifier,
@@ -921,21 +921,6 @@ case class UnionExec(children: Seq[SparkPlan]) extends
SparkPlan with CodegenSup
}
}
- // True iff every child output dataType matches the corresponding union
- // output dataType, including all nested nullabilities.
- // `Union.allChildrenCompatible` ignores nested nullability, so children
- // differing only there bypass `WidenSetOperationTypes`; `UnionExec.output`
- // then merges those flags via `StructType.unionLikeMerge`, leaving src/tgt
- // mismatched.
- @transient private lazy val allChildOutputDataTypesMatch: Boolean =
- children.forall { c =>
- c.output.zip(output).forall { case (src, tgt) => src.dataType ==
tgt.dataType }
- }
-
- // Memoized: `supportCodegen` is called multiple times during planning.
- @transient private lazy val hasAnyPartitionIndexDependentDescendant: Boolean
=
- children.exists(UnionExec.hasPartitionIndexDependentCodegen)
-
// Memoized: consulted by `supportCodegen` (called multiple times by
// `CollapseCodegenStages`) and by `metrics`. Conf and children are stable
// for a given UnionExec instance; cross-plan staleness is impossible since
@@ -943,19 +928,20 @@ case class UnionExec(children: Seq[SparkPlan]) extends
SparkPlan with CodegenSup
@transient private lazy val supportCodegenFailureReason: Option[String] = {
if (!conf.getConf(SQLConf.WHOLESTAGE_UNION_CODEGEN_ENABLED)) {
Some("union-codegen-disabled")
- } else if (!outputPartitioning.isInstanceOf[UnknownPartitioning]) {
+ } else if (!isPlainUnion) {
Some("partitioning-aware")
} else if (children.exists(_.exists(_.isInstanceOf[UnionExec]))) {
Some("nested-union")
} else if
(children.exists(_.exists(UnionExec.isKnownMultiInputRDDCodegen))) {
Some("multi-rdd-child")
- } else if (hasAnyPartitionIndexDependentDescendant) {
+ } else if (children.exists(UnionExec.hasPartitionIndexDependentCodegen)) {
Some("partition-index-dependent-child")
} else if (children.size >
conf.getConf(SQLConf.WHOLESTAGE_UNION_MAX_CHILDREN)) {
Some("max-children-exceeded")
} else if (supportsColumnar) {
Some("columnar")
- } else if (!allChildOutputDataTypesMatch) {
+ } else if (children.exists(c =>
+ c.output.zip(output).exists { case (src, tgt) => src.dataType !=
tgt.dataType })) {
Some("type-mismatch")
} else {
None
@@ -1002,61 +988,66 @@ case class UnionExec(children: Seq[SparkPlan]) extends
SparkPlan with CodegenSup
override def inputRDDs(): Seq[RDD[InternalRow]] = Seq(unionedInputRDD)
- // Driver-side cursor written by `doProduce` and read by `doConsume` during
- // single-threaded code emission; resets to -1 once emission completes.
+ // Set in `doProduce`, read in `doConsume` during single-threaded code
+ // emission. `numOutputRowsTerm` is registered once per stage so the
+ // metric appears in `references[]` exactly once instead of once per
+ // child. `currentEmittingChild` tells `doConsume` which child's
+ // projection to bind.
+ @transient private var numOutputRowsTerm: String = _
@transient private var currentEmittingChild: Int = -1
override protected def doProduce(ctx: CodegenContext): String = {
+ numOutputRowsTerm = metricTerm(ctx, "numOutputRows")
+
// For each partition of the unioned RDD, record its owning child and its
// index within that child's RDD. Read both fields directly off the
// `UnionPartition` so the lookup arrays do not assume `UnionRDD` lays
// partitions out in child order.
- val (partitionToChild, partitionToLocalIdx) =
- unionedInputRDD.partitions.map {
- case up: UnionPartition[_] => (up.parentRddIndex,
up.parentPartition.index)
- case other =>
- throw SparkException.internalError(
- s"UnionExec: Unexpected partition type ${other.getClass.getName}")
- }.unzip
+ val (partitionToChild, partitionToLocalIdx) =
unionedInputRDD.partitions.map { p =>
+ val up = p.asInstanceOf[UnionPartition[_]]
+ (up.parentRddIndex, up.parentPartition.index)
+ }.unzip
val p2cRef = ctx.addReferenceObj("partitionToChild", partitionToChild)
val p2lRef = ctx.addReferenceObj("partitionToLocalIdx",
partitionToLocalIdx)
val childIndexVar = ctx.freshName("unionChildIdx")
- // Each child's produce output is wrapped in its own helper method. The
- // outer `switch` in `doProduce`'s return value dispatches to the helper.
+ // Each child's produced code is wrapped in its own helper method.
// Without this, the fused method's bytecode grows linearly with the
// number of children and quickly exceeds HotSpot's per-method limit,
// forcing the whole stage to run interpreted.
//
- // `partitionIndex` is passed as a parameter (shadowing the superclass
- // field) rather than read from the enclosing scope. `addNewFunction` may
- // spill helpers into a nested class when the outer class fills up, and a
- // nested class cannot access the protected
- // `BufferedRowIterator.partitionIndex` field. Using the parameter name
- // `partitionIndex` keeps any child-emitted reference to that identifier
- // resolving locally.
+ // The helper takes `int partitionIndex` as a parameter; `addNewFunction`
+ // may spill helpers into a nested class once the outer class fills up,
+ // and a nested class cannot access the protected
+ // `BufferedRowIterator.partitionIndex` field.
+ //
+ // `currentPartitionIndexVar` is rebound to an array-deref expression
+ // (rather than a local) so leaf operators (`RangeExec`, `SampleExec`)
+ // see the child-local index regardless of where their code is emitted.
+ // `SampleExec.doConsume` uses `addMutableState`, whose initializer is
+ // emitted into the state-init function, not the helper - a local in
+ // the helper would not be in scope there. The expression resolves
+ // against `partitionIndex` (the helper parameter inside the helper,
+ // and the `BufferedRowIterator` field elsewhere) in every context.
val savedPartIdxVar = ctx.currentPartitionIndexVar
- val cases = try {
- children.zipWithIndex.map { case (c, i) =>
- currentEmittingChild = i
- ctx.currentPartitionIndexVar = s"((int[]) $p2lRef)[partitionIndex]"
- val producedCode = c.asInstanceOf[CodegenSupport].produce(ctx, this)
- val helper = ctx.freshName("unionChildProcess")
- val qualifiedHelper = ctx.addNewFunction(helper,
- s"""
- |private void $helper(int partitionIndex) throws
java.io.IOException {
- | $producedCode
- |}
- """.stripMargin)
- s"""case $i: {
- | $qualifiedHelper(partitionIndex);
- | break;
- |}""".stripMargin
- }
- } finally {
- currentEmittingChild = -1
- ctx.currentPartitionIndexVar = savedPartIdxVar
+ ctx.currentPartitionIndexVar = s"((int[]) $p2lRef)[partitionIndex]"
+ val cases = children.zipWithIndex.map { case (c, i) =>
+ currentEmittingChild = i
+ val producedCode = c.asInstanceOf[CodegenSupport].produce(ctx, this)
+ val helper = ctx.freshName("unionChildProcess")
+ val qualifiedHelper = ctx.addNewFunction(helper,
+ s"""
+ |private void $helper(int partitionIndex) throws
java.io.IOException {
+ | $producedCode
+ |}
+ """.stripMargin)
+ s"""case $i: {
+ | $qualifiedHelper(partitionIndex);
+ | break;
+ |}""".stripMargin
}
+ currentEmittingChild = -1
+ ctx.currentPartitionIndexVar = savedPartIdxVar
s"""
|int $childIndexVar = ((int[]) $p2cRef)[partitionIndex];
@@ -1071,24 +1062,17 @@ case class UnionExec(children: Seq[SparkPlan]) extends
SparkPlan with CodegenSup
override def doConsume(
ctx: CodegenContext, input: Seq[ExprCode], row: ExprCode): String = {
- require(currentEmittingChild >= 0,
- "UnionExec.doConsume invoked outside doProduce emission window")
val i = currentEmittingChild
- // The wrapped child in each `perChildProjections(i)` element is always an
- // `Attribute`, which is deterministic by definition; no
- // `evaluateRequiredVariables` call is needed to force single-evaluation
- // of non-deterministic expressions.
- val bound = BindReferences.bindReferences(perChildProjections(i),
children(i).output)
-
+ require(i >= 0, "UnionExec.doConsume invoked outside doProduce emission
window")
// Route BoundReference reads through `currentVars` (the incoming row is
// delivered as variables under WSCG, not via ctx.INPUT_ROW).
+ val bound = BindReferences.bindReferences(perChildProjections(i),
children(i).output)
ctx.currentVars = input
ctx.INPUT_ROW = null
val projectedExprCodes = bound.map(_.genCode(ctx))
- val numOutput = metricTerm(ctx, "numOutputRows")
s"""
- |$numOutput.add(1L);
+ |$numOutputRowsTerm.add(1L);
|${consume(ctx, projectedExprCodes)}
""".stripMargin
}
@@ -1103,7 +1087,7 @@ case class UnionExec(children: Seq[SparkPlan]) extends
SparkPlan with CodegenSup
override def usedInputs: AttributeSet = AttributeSet.empty
protected override def doExecute(): RDD[InternalRow] = {
- if (outputPartitioning.isInstanceOf[UnknownPartitioning]) {
+ if (isPlainUnion) {
sparkContext.union(children.map(_.execute()))
} else {
// This union has a known partitioning, i.e., its children have the same
partitioning
@@ -1138,13 +1122,24 @@ object UnionExec {
}
/**
- * True if any expression in the subtree is [[Nondeterministic]]. Such
- * expressions may embed the raw `partitionIndex` field via
- * `addPartitionInitializationStatement`, which would read the global
+ * True if any expression in the subtree embeds the raw `partitionIndex`
field
+ * via `addPartitionInitializationStatement`, which would read the global
* UnionRDD index instead of the child-local one under fusion.
+ *
+ * The check uses [[Nondeterministic]] as the proxy: every catalyst
expression
+ * that calls `addPartitionInitializationStatement` referencing
`partitionIndex`
+ * is `Nondeterministic`. The `InputFile*` expressions are `Nondeterministic`
+ * but read from `InputFileBlockHolder` (a per-task thread-local) and do not
+ * embed `partitionIndex`, so they are safe under fusion.
*/
- def hasPartitionIndexDependentCodegen(p: SparkPlan): Boolean = p.exists {
- plan => plan.expressions.exists(_.exists(_.isInstanceOf[Nondeterministic]))
+ def hasPartitionIndexDependentCodegen(p: SparkPlan): Boolean = p.exists {
plan =>
+ plan.expressions.exists(_.exists {
+ case _: InputFileName => false
+ case _: InputFileBlockStart => false
+ case _: InputFileBlockLength => false
+ case _: Nondeterministic => true
+ case _ => false
+ })
}
}
diff --git
a/sql/core/src/test/scala/org/apache/spark/sql/execution/UnionCodegenSuite.scala
b/sql/core/src/test/scala/org/apache/spark/sql/execution/UnionCodegenSuite.scala
index fc049044d92e..fe6ef1bc5e25 100644
---
a/sql/core/src/test/scala/org/apache/spark/sql/execution/UnionCodegenSuite.scala
+++
b/sql/core/src/test/scala/org/apache/spark/sql/execution/UnionCodegenSuite.scala
@@ -602,6 +602,57 @@ class UnionCodegenSuite extends QueryTest with
SharedSparkSession {
"numOutputRows should be 0 for all-empty union")
}
}
+
+ test("SPARK-56482: partitioning-aware union falls back to non-codegen") {
+ // After repartition, both children expose a `HashPartitioning` on the
same key,
+ // so `UnionExec.outputPartitioning` is non-Unknown and the codegen path
is denied.
+ // AQE is disabled here so the executedPlan exposes the UnionExec directly
+ // (under AQE the plan is wrapped in `AdaptiveSparkPlanExec`, which does
not
+ // surface its inputPlan via `children`).
+ withSQLConf(
+ SQLConf.UNION_OUTPUT_PARTITIONING.key -> "true",
+ SQLConf.ADAPTIVE_EXECUTION_ENABLED.key -> "false") {
+ val a = rangeDF(100).repartition(4, col("id"))
+ val b = rangeDF(100, 200).repartition(4, col("id"))
+ val df = a.union(b)
+ assert(!unionInsideWSCG(df),
+ "Partitioning-aware union must not fuse into WSCG")
+ val unionExec = df.queryExecution.executedPlan.collectFirst {
+ case u: UnionExec => u
+ }.get
+ assert(!unionExec.metrics.contains("numOutputRows"),
+ "numOutputRows metric must not be registered on the partitioning-aware
path")
+ assertFlagParity(() => a.union(b).orderBy("id"))
+ }
+ }
+
+ test("SPARK-56482: input_file_name child fuses (Nondeterministic but
partition-index-free)") {
+ // `InputFileName` is `Nondeterministic` but reads from
`InputFileBlockHolder`
+ // (a per-task thread-local) and does not embed `partitionIndex`. The
gate's
+ // narrow check should let this fuse.
+ withTempPath { dir =>
+ val path = dir.getCanonicalPath
+ rangeDF(20).write.parquet(path)
+ val a = spark.read.parquet(path).select(col("id"),
input_file_name().as("f"))
+ val b = spark.read.parquet(path).select(col("id"),
input_file_name().as("f"))
+ val df = a.union(b).filter(col("id") > 0)
+ assert(unionInsideWSCG(df),
+ "Union with input_file_name child should fuse into WSCG")
+ assertFlagParity(() => a.union(b).orderBy("id", "f"))
+ }
+ }
+
+ test("SPARK-56482: union with sample children fuses (or falls back) without
crashing") {
+ // `SampleExec.doConsume` reads `currentPartitionIndexVar` from inside an
+ // `addMutableState` initializer, which is emitted into the state-init
+ // function rather than the per-child helper. The bound expression must
+ // therefore resolve in any emission scope, not just inside the helper.
+ val a = rangeDF(20).sample(false, 0.5, 1L)
+ val b = rangeDF(20).sample(false, 0.5, 1L)
+ val df = a.union(b).filter(col("id") > 0)
+ df.collect()
+ assertFlagParity(() => a.union(b).orderBy("id"))
+ }
}
/** Runs [[UnionCodegenSuite]] with ANSI mode enabled. */
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
