cloud-fan commented on code in PR #58077:
URL: https://github.com/apache/spark/pull/58077#discussion_r3849443918
##########
sql/core/src/main/scala/org/apache/spark/sql/execution/subquery.scala:
##########
@@ -165,14 +186,173 @@ case class InSubqueryExec(
}
}
+ // Invariant schema/ordering data for the multi-column evaluator, computed
once after the result
+ // is available. @transient so that serialization (result=null) does not
trigger evaluation.
+ @transient private lazy val multiColFieldTypes: Array[DataType] =
+ plan.output.map(_.dataType).toArray
+ @transient private lazy val multiColFieldOrderings: Array[Ordering[Any]] =
+ multiColFieldTypes.map(TypeUtils.getInterpretedOrdering)
+ // Struct-level ordering used to index fully non-null result rows in a
TreeSet.
+ @transient private lazy val multiColRowOrdering: Ordering[InternalRow] =
+
TypeUtils.getInterpretedOrdering(child.dataType).asInstanceOf[Ordering[InternalRow]]
+
+ // Split collected rows into a sorted set of fully non-null rows (O(log n)
membership test)
+ // and an array of rows that contain at least one null field (must be
scanned linearly).
+ // Built once; the TreeSet uses the struct-level Catalyst ordering. See
SPARK-58481.
+ @transient private lazy val (multiColNonNullSet, multiColNullRows) = {
+ val withNull = Array.newBuilder[InternalRow]
+ val nonNull = TreeSet.newBuilder[InternalRow](multiColRowOrdering)
+ result.foreach { r =>
+ val row = r.asInstanceOf[InternalRow]
+ if (row.anyNull) withNull += row else nonNull += row
+ }
+ (nonNull.result(), withNull.result())
+ }
+
+ // Three-valued IN semantics for multi-column subqueries.
+ // Result rows are InternalRow objects; InSet's TreeSet uses Catalyst
ordering, but membership
+ // cannot distinguish a definitively-false candidate from an indeterminate
one.
+ //
+ // When the LHS struct has no null fields:
+ // Fast path: O(log n) TreeSet lookup against fully non-null result rows
for TRUE.
+ // Slow path: linear scan over null-containing result rows only for
potential UNKNOWN.
+ //
+ // When the LHS struct has at least one null field, the fast path cannot be
used (a null LHS
+ // field produces UNKNOWN against any non-null RHS row whose non-null fields
all match). In
+ // that case we scan all result rows linearly.
Review Comment:
**Nit:**
Please describe this as a linear scan that stops once UNKNOWN is
established. Both loops short-circuit on `!hasUnknown`, so 'all result rows' is
also inaccurate in the repeated comment at `subquery.scala:265`.
##########
sql/core/src/main/scala/org/apache/spark/sql/execution/subquery.scala:
##########
@@ -125,7 +129,24 @@ case class InSubqueryExec(
@transient private lazy val inSet = InSet(child, result.toSet)
- override def nullable: Boolean = child.nullable
+ // Mirror the logical InSubquery.nullable: nullable when any output column
is nullable
+ // (null in any column position produces UNKNOWN on a miss) or when any LHS
field is nullable.
Review Comment:
**Nit:**
Describe this as a possible outcome rather than an unconditional one.
`(1,1)` versus `(99,NULL)` is definitively FALSE because the first field
differs, even though another comparison is UNKNOWN.
```suggestion
// (a nullable RHS field can produce UNKNOWN on a miss) or when any LHS
field is nullable.
```
##########
sql/core/src/main/scala/org/apache/spark/sql/execution/subquery.scala:
##########
@@ -165,14 +186,173 @@ case class InSubqueryExec(
}
}
+ // Invariant schema/ordering data for the multi-column evaluator, computed
once after the result
+ // is available. @transient so that serialization (result=null) does not
trigger evaluation.
+ @transient private lazy val multiColFieldTypes: Array[DataType] =
+ plan.output.map(_.dataType).toArray
+ @transient private lazy val multiColFieldOrderings: Array[Ordering[Any]] =
+ multiColFieldTypes.map(TypeUtils.getInterpretedOrdering)
+ // Struct-level ordering used to index fully non-null result rows in a
TreeSet.
+ @transient private lazy val multiColRowOrdering: Ordering[InternalRow] =
+
TypeUtils.getInterpretedOrdering(child.dataType).asInstanceOf[Ordering[InternalRow]]
+
+ // Split collected rows into a sorted set of fully non-null rows (O(log n)
membership test)
+ // and an array of rows that contain at least one null field (must be
scanned linearly).
+ // Built once; the TreeSet uses the struct-level Catalyst ordering. See
SPARK-58481.
+ @transient private lazy val (multiColNonNullSet, multiColNullRows) = {
+ val withNull = Array.newBuilder[InternalRow]
+ val nonNull = TreeSet.newBuilder[InternalRow](multiColRowOrdering)
+ result.foreach { r =>
+ val row = r.asInstanceOf[InternalRow]
+ if (row.anyNull) withNull += row else nonNull += row
+ }
+ (nonNull.result(), withNull.result())
+ }
+
+ // Three-valued IN semantics for multi-column subqueries.
+ // Result rows are InternalRow objects; InSet's TreeSet uses Catalyst
ordering, but membership
+ // cannot distinguish a definitively-false candidate from an indeterminate
one.
+ //
+ // When the LHS struct has no null fields:
+ // Fast path: O(log n) TreeSet lookup against fully non-null result rows
for TRUE.
+ // Slow path: linear scan over null-containing result rows only for
potential UNKNOWN.
+ //
+ // When the LHS struct has at least one null field, the fast path cannot be
used (a null LHS
+ // field produces UNKNOWN against any non-null RHS row whose non-null fields
all match). In
+ // that case we scan all result rows linearly.
+ //
+ // Per-candidate three-valued logic: TRUE if every field matches; UNKNOWN if
no field is
+ // definitively unequal but at least one comparison involves null; FALSE
otherwise.
+ private def evalMultiColumn(inputRow: InternalRow): Any = {
+ val value = child.eval(inputRow)
+ if (value == null) return null
+ val inputStruct = value.asInstanceOf[InternalRow]
+ val fieldTypes = multiColFieldTypes
+ val orderings = multiColFieldOrderings
+ val numFields = fieldTypes.length
+
+ if (!inputStruct.anyNull) {
+ // Fast path: indexed lookup among fully non-null candidates.
+ if (multiColNonNullSet.contains(inputStruct)) return true
+ // Materialize LHS fields once before the candidate scans. For types
like BinaryType,
+ // InternalRow.get copies the payload on every call; caching avoids one
copy per
Review Comment:
**Nit:**
Please remove the BinaryType copy-avoidance rationale here and at
`subquery.scala:267`. The LHS is produced by `CreateNamedStruct` as a
`GenericInternalRow`, and `BaseGenericInternalRow.get` returns the stored field
reference rather than copying a binary payload.
##########
sql/core/src/test/scala/org/apache/spark/sql/SubquerySuite.scala:
##########
@@ -2678,4 +2678,164 @@ class SubquerySuite extends SharedSparkSession
assert(exposedAttribute.exprId == outerReferenceAttribute.exprId)
}
+
+ test("SPARK-58481: InSubqueryExec nullable correctly accounts for subquery
output nullability") {
+ // 5 NOT IN (99, NULL) is UNKNOWN, not TRUE or FALSE. A join condition
that is not TRUE
+ // matches no rows, so a FULL OUTER JOIN must emit null-padded rows for
every row in each
+ // side -- 3 + 3 = 6 null-padded rows -- not the full cross product (9
rows).
+ withTable("t0", "t1", "t3") {
+ sql("CREATE TABLE t0(c0 INT) USING PARQUET")
+ sql("INSERT INTO t0 VALUES (1), (2), (3)")
+ sql("CREATE TABLE t1(c0 INT) USING PARQUET")
+ sql("INSERT INTO t1 VALUES (10), (20), (30)")
+ sql("CREATE TABLE t3(c0 INT) USING PARQUET")
+ sql("INSERT INTO t3 VALUES (99), (CAST(NULL AS INT))")
+
+ // t1 rows are null-padded (no match on left), t0 rows are null-padded
(no match on right).
+ val expected = Seq(
+ Row(null, 10), Row(null, 20), Row(null, 30), // t0 side: null-padded
+ Row(1, null), Row(2, null), Row(3, null)) // t1 side: null-padded
+ checkAnswer(
+ sql("SELECT t0.c0, t1.c0 FROM t1 FULL OUTER JOIN t0 ON (5 NOT IN
(SELECT t3.c0 FROM t3))"),
+ expected)
+ }
+ }
+
+ test("SPARK-58481: multi-column IN subquery with nullable non-head output is
nullable") {
+ // Disable the optimizer's join-condition IN rewrite so the query
exercises InSubqueryExec.
+ // Covers both per-candidate cases:
+ // (1,1) vs (99,NULL): first field differs => definitely FALSE (not
UNKNOWN).
+ // (1,1) vs (1,NULL): first fields equal, second null => UNKNOWN.
+ // (2,2) vs either row: both are FALSE => NOT IN = TRUE.
+ // Expected: (1,1) gets UNKNOWN => null-padded; (2,2) gets TRUE => joined.
+ withSQLConf(
+
"spark.sql.optimizer.optimizeUncorrelatedInSubqueriesInJoinCondition.enabled"
-> "false"
+ ) {
+ withTable("lhs", "rhs") {
+ sql("CREATE TABLE lhs(a INT NOT NULL, b INT NOT NULL) USING PARQUET")
Review Comment:
**Non-blocking:**
Use non-nullable `VALUES`-derived inputs for the LHS and the RHS first
column, leaving only a later RHS column nullable. File-source analysis applies
`dataSchema.asNullable`, so both this LHS and the RHS head analyze as nullable
despite the DDL; the test therefore still passes with the old
`plan.output.head.nullable` check and does not isolate the behavior named by
the test.
##########
sql/core/src/test/scala/org/apache/spark/sql/SubquerySuite.scala:
##########
@@ -2678,4 +2678,164 @@ class SubquerySuite extends SharedSparkSession
assert(exposedAttribute.exprId == outerReferenceAttribute.exprId)
}
+
+ test("SPARK-58481: InSubqueryExec nullable correctly accounts for subquery
output nullability") {
+ // 5 NOT IN (99, NULL) is UNKNOWN, not TRUE or FALSE. A join condition
that is not TRUE
+ // matches no rows, so a FULL OUTER JOIN must emit null-padded rows for
every row in each
+ // side -- 3 + 3 = 6 null-padded rows -- not the full cross product (9
rows).
+ withTable("t0", "t1", "t3") {
+ sql("CREATE TABLE t0(c0 INT) USING PARQUET")
+ sql("INSERT INTO t0 VALUES (1), (2), (3)")
+ sql("CREATE TABLE t1(c0 INT) USING PARQUET")
+ sql("INSERT INTO t1 VALUES (10), (20), (30)")
+ sql("CREATE TABLE t3(c0 INT) USING PARQUET")
+ sql("INSERT INTO t3 VALUES (99), (CAST(NULL AS INT))")
+
+ // t1 rows are null-padded (no match on left), t0 rows are null-padded
(no match on right).
Review Comment:
**Nit:**
The relation labels are reversed: `Row(null, 10)` retains the unmatched t1
row and pads t0, while `Row(1, null)` retains t0 and pads t1.
```suggestion
// Unmatched t1 rows null-pad t0; unmatched t0 rows null-pad t1.
```
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
