pratham76 commented on code in PR #58656:
URL: https://github.com/apache/spark/pull/58656#discussion_r4286926784
##########
sql/catalyst/src/main/scala/org/apache/spark/sql/catalyst/optimizer/subquery.scala:
##########
@@ -398,62 +406,476 @@ object RewritePredicateSubquery extends
Rule[LogicalPlan] with PredicateHelper {
(newExpr, newPlan)
}
+ /**
+ * Same as [[rewriteExistentialExpr]], but it also returns the newly
introduced attributes, and
+ * it only rewrites the existential sub-queries for which `canRewrite`
returns true. A sub-query
+ * that is not rewritten stays in the returned expression as it is, and is
not descended into:
+ * rewriting an existential sub-query nested in its join condition would
graft an existence join
+ * onto the plan for an `exists` reference that the sub-query left in place
may never evaluate.
+ */
private def rewriteExistentialExprWithAttrs(
exprs: Seq[Expression],
- plan: LogicalPlan): (Option[Expression], LogicalPlan, Seq[Attribute]) = {
+ plan: LogicalPlan,
+ canRewrite: Expression => Boolean = _ => true): (Option[Expression],
LogicalPlan,
+ Seq[Attribute]) = {
var newPlan = plan
val introducedAttrs = ArrayBuffer.empty[Attribute]
- val newExprs = exprs.map { e =>
- e.transformDownWithPruning(_.containsAnyPattern(EXISTS_SUBQUERY,
IN_SUBQUERY)) {
- case Exists(sub, _, _, conditions, subHint) =>
- val exists = AttributeReference("exists", BooleanType, nullable =
false)()
- val existenceJoin = ExistenceJoin(exists)
- val newCondition = conditions.reduceLeftOption(And)
- newPlan =
- buildJoin(newPlan, rewriteDomainJoinsIfPresent(newPlan, sub,
newCondition),
- existenceJoin, newCondition, subHint)
- introducedAttrs += exists
- exists
- case Not(InSubquery(values, ListQuery(sub, _, _, _, conditions,
subHint))) =>
- val exists = AttributeReference("exists", BooleanType, nullable =
false)()
- // Deduplicate conflicting attributes if any.
- val newSub = dedupSubqueryOnSelfJoin(newPlan, sub, Some(values))
- val inConditions = values.zip(newSub.output).map(EqualTo.tupled)
- // To handle a null-aware predicate not-in-subquery in nested
conditions
- // (e.g., `v > 0 OR t1.id NOT IN (SELECT id FROM t2)`), we transform
- // `inCondition` (t1.id=t2.id) into `(inCondition) OR
ISNULL(inCondition)`.
- //
- // For example, `SELECT * FROM t1 WHERE v > 0 OR t1.id NOT IN
(SELECT id FROM t2)`
- // is transformed into a plan below;
- // == Optimized Logical Plan ==
- // Project [id#78, v#79]
- // +- Filter ((v#79 > 0) OR NOT exists#83)
- // +- Join ExistenceJoin(exists#83), ((id#78 = id#80) OR
isnull((id#78 = id#80)))
- // :- Relation[id#78,v#79] parquet
- // +- Relation[id#80] parquet
- val nullAwareJoinConds = inConditions.map(c => Or(c, IsNull(c)))
- val finalJoinCond = (nullAwareJoinConds ++
conditions).reduceLeft(And)
- val joinHint = JoinHint(None, subHint)
- newPlan = Join(newPlan,
- rewriteDomainJoinsIfPresent(newPlan, newSub, Some(finalJoinCond)),
- ExistenceJoin(exists), Some(finalJoinCond), joinHint)
- introducedAttrs += exists
- Not(exists)
- case InSubquery(values, ListQuery(sub, _, _, _, conditions, subHint))
=>
- val exists = AttributeReference("exists", BooleanType, nullable =
false)()
- // Deduplicate conflicting attributes if any.
- val newSub = dedupSubqueryOnSelfJoin(newPlan, sub, Some(values))
- val inConditions = values.zip(newSub.output).map(EqualTo.tupled)
- val newConditions = (inConditions ++
conditions).reduceLeftOption(And)
- val joinHint = JoinHint(None, subHint)
- newPlan = Join(newPlan, rewriteDomainJoinsIfPresent(newPlan, newSub,
newConditions),
- ExistenceJoin(exists), newConditions, joinHint)
- introducedAttrs += exists
- exists
+ // Builds the replacement of a sub-query expression under that
expression's origin and with
+ // its tags, as TreeNode.transformDownWithPruning does for the nodes a
rule replaces. Without
+ // it every node built below, the `exists` attribute and the join
conditions included, would
+ // take the origin of the enclosing operator, and a runtime error raised
from the rewritten
+ // condition would quote that operator rather than the sub-query fragment.
+ def replacing(sq: Expression)(replacement: => Expression): Expression = {
+ val newExpr = CurrentOrigin.withOrigin(sq.origin)(replacement)
+ newExpr.copyTagsFrom(sq)
+ newExpr
+ }
+ def rewrite(expr: Expression): Expression = {
+ if (!expr.containsAnyPattern(EXISTS_SUBQUERY, IN_SUBQUERY)) {
+ return expr
+ }
+ expr match {
+ case sq @ Exists(sub, _, _, conditions, subHint) if canRewrite(sq) =>
+ replacing(sq) {
+ val exists = AttributeReference("exists", BooleanType, nullable =
false)()
+ val existenceJoin = ExistenceJoin(exists)
+ val newCondition = conditions.reduceLeftOption(And)
+ newPlan =
+ buildJoin(newPlan, rewriteDomainJoinsIfPresent(newPlan, sub,
newCondition),
+ existenceJoin, newCondition, subHint)
+ introducedAttrs += exists
+ exists
+ }
+ case sq @ Not(InSubquery(values, ListQuery(sub, _, _, _, conditions,
subHint)))
+ if canRewrite(sq) =>
+ replacing(sq) {
+ val exists = AttributeReference("exists", BooleanType, nullable =
false)()
+ // Deduplicate conflicting attributes if any.
+ val newSub = dedupSubqueryOnSelfJoin(newPlan, sub, Some(values))
+ val inConditions = values.zip(newSub.output).map(EqualTo.tupled)
+ // To handle a null-aware predicate not-in-subquery in nested
conditions
+ // (e.g., `v > 0 OR t1.id NOT IN (SELECT id FROM t2)`), we
transform
+ // `inCondition` (t1.id=t2.id) into `(inCondition) OR
ISNULL(inCondition)`.
+ //
+ // For example, `SELECT * FROM t1 WHERE v > 0 OR t1.id NOT IN
(SELECT id FROM t2)`
+ // is transformed into a plan below;
+ // == Optimized Logical Plan ==
+ // Project [id#78, v#79]
+ // +- Filter ((v#79 > 0) OR NOT exists#83)
+ // +- Join ExistenceJoin(exists#83), ((id#78 = id#80) OR
isnull((id#78 = id#80)))
+ // :- Relation[id#78,v#79] parquet
+ // +- Relation[id#80] parquet
+ val nullAwareJoinConds = inConditions.map(c => Or(c, IsNull(c)))
+ val finalJoinCond = (nullAwareJoinConds ++
conditions).reduceLeft(And)
+ val joinHint = JoinHint(None, subHint)
+ newPlan = Join(newPlan,
+ rewriteDomainJoinsIfPresent(newPlan, newSub,
Some(finalJoinCond)),
+ ExistenceJoin(exists), Some(finalJoinCond), joinHint)
+ introducedAttrs += exists
+ Not(exists)
+ }
+ case sq @ InSubquery(values, ListQuery(sub, _, _, _, conditions,
subHint))
+ if canRewrite(sq) =>
+ replacing(sq) {
+ val exists = AttributeReference("exists", BooleanType, nullable =
false)()
+ // Deduplicate conflicting attributes if any.
+ val newSub = dedupSubqueryOnSelfJoin(newPlan, sub, Some(values))
+ val inConditions = values.zip(newSub.output).map(EqualTo.tupled)
+ val newConditions = (inConditions ++
conditions).reduceLeftOption(And)
+ val joinHint = JoinHint(None, subHint)
+ newPlan = Join(newPlan, rewriteDomainJoinsIfPresent(newPlan,
newSub, newConditions),
+ ExistenceJoin(exists), newConditions, joinHint)
+ introducedAttrs += exists
+ exists
+ }
+ // A sub-query that `canRewrite` declined is left as it is, children
included.
+ case sq @ (_: Exists | Not(_: InSubquery) | _: InSubquery) => sq
+ case other => other.mapChildren(rewrite)
}
}
+ val newExprs = exprs.map(rewrite)
(newExprs.reduceOption(And), newPlan, introducedAttrs.toSeq)
}
+
+ /**
+ * Returns true if `e` references any of the attributes produced by `plan`.
+ */
+ private def referencesPlan(e: Expression, plan: LogicalPlan): Boolean = {
+ // Iterates the references and probes the output set, rather than
materialising the
+ // intersection: AttributeSet.intersect builds a set from its argument, so
the operand order
+ // there would decide whether this walks the references or the whole
output of `plan`. The
+ // membership probes still construct wrappers, so this is cheaper, not
allocation-free.
+ e.references.exists(plan.outputSet.contains)
+ }
+
+ /**
+ * Returns true if `e` effectively references any of the attributes produced
by `plan`, see
+ * [[effectiveReferences]].
+ */
+ private def effectivelyReferencesPlan(e: Expression, plan: LogicalPlan):
Boolean = {
+ effectiveReferences(e).exists(plan.outputSet.contains)
+ }
+
+ /**
+ * Which of an outer plan and a sub-query plan a set of references can be
evaluated by, as
+ * [[planReferences]] determines it.
+ */
+ private case class PlanReferences(subPlanOnly: Boolean, bothPlans: Boolean)
+
+ /**
+ * Classifies `references` against `outerPlan` and `subPlan` in a single
pass: whether they can
+ * only be evaluated by `subPlan`, and whether they need both plans.
+ *
+ * An attribute that both plans produce is attributed to `subPlan` rather
than to `outerPlan`,
+ * which is what the `subPlan` probe coming first does. The two plans share
no ExprId in any
+ * shape reached by the suite: `DeduplicateRelations` has given the
sub-query its own, and the
+ * EXISTS arms classify against the raw sub-query plan while the IN arms
classify against the
+ * plan `dedupSubqueryOnSelfJoin` has already separated from the outer plan.
The order is
+ * therefore defensive rather than something a query exercises, and it is
the safe one: a
+ * nested sub-query resolves in the scope of the sub-query it is nested in,
so an attribute
+ * both plans produce is one of that sub-query's own, and attributing it to
`outerPlan` instead
+ * would let a single shared attribute satisfy both sides of the
classification and have a
+ * nested sub-query referencing only the inner query rejected as referencing
both.
+ *
+ * An attribute neither plan produces leaves both answers false, which
declines the rewrite
+ * without rejecting the sub-query, as a sub-query referencing neither plan
can be left in the
+ * join condition unharmed.
+ */
+ private def planReferences(
+ references: AttributeSet,
+ outerPlan: LogicalPlan,
+ subPlan: LogicalPlan): PlanReferences = {
+ var subPlanReference = false
+ var outerPlanOnlyReference = false
+ val attrs = references.iterator
+ // Stops as soon as both plans are known to be referenced, which is all
the two answers need.
+ while (attrs.hasNext && !(subPlanReference && outerPlanOnlyReference)) {
+ val attr = attrs.next()
+ if (subPlan.outputSet.contains(attr)) {
+ subPlanReference = true
+ } else if (outerPlan.outputSet.contains(attr)) {
+ outerPlanOnlyReference = true
+ }
+ }
+ PlanReferences(
+ subPlanOnly = subPlanReference && !outerPlanOnlyReference,
+ bothPlans = subPlanReference && outerPlanOnlyReference)
+ }
+
+ /**
+ * The attributes that `e` references and can still evaluate. For an
existential sub-query
+ * expression these are the ones of the values it compares and of the
correlated condition that
+ * [[PullupCorrelatedPredicates]] hoisted into it, rather than the ones of
`references`, which
+ * also include the outer attributes the pull-up retains for idempotency.
Those outlive the
+ * condition that referenced them once BooleanSimplification has eliminated
it, as in
+ * `a IN (SELECT col1 FROM t3 WHERE false AND col1 = c1)`, which keeps `c1`
as an outer
+ * attribute of a sub-query that no longer reads it, and which must still be
rewritten against
+ * the outer plan rather than treated as referencing the sub-query plan.
+ *
+ * The sets are merged with `AttributeSet.fromAttributeSets` rather than
built from a flattened
+ * sequence of attributes: `AttributeSet.apply` would take one allocation
per attribute of a
+ * sequence as long as the total number of references, which a row-valued IN
over a wide row
+ * makes unnecessarily large.
+ */
+ private def effectiveReferences(e: Expression): AttributeSet = e match {
+ case Exists(_, _, _, joinCond, _) =>
+ AttributeSet.fromAttributeSets(joinCond.map(_.references))
+ case Not(in: InSubquery) => effectiveReferences(in)
+ case InSubquery(values, ListQuery(_, _, _, _, joinCond, _)) =>
+ AttributeSet.fromAttributeSets(values.map(_.references) ++
joinCond.map(_.references))
+ case _ => e.references
+ }
+
+ /**
+ * Rewrites the existential sub-queries that are nested in the correlated
predicates which
+ * [[PullupCorrelatedPredicates]] hoisted out of a predicate sub-query, and
which therefore end
+ * up in the condition of the semi/anti join that replaces that sub-query.
+ *
+ * A hoisted predicate can carry a nested existential sub-query out of the
sub-query plan,
+ * because a correlated predicate is hoisted as a whole when it is a
disjunction. For example
+ *
+ * SELECT * FROM t1 WHERE EXISTS (
+ * SELECT 1 FROM t2 WHERE t1.a = t2.c1 OR t2.c1 IN (SELECT col1 FROM t3))
+ *
+ * hoists `a = c1 OR c1 IN (SELECT col1 FROM t3)` into the join condition.
Such a nested
+ * sub-query must be rewritten against the plan that produces the attributes
it references:
+ * the one above references `c1`, which is produced by the sub-query plan
and not by the outer
+ * plan, so its existence join has to be built on top of the sub-query plan.
Building it on top
+ * of the outer plan instead yields a join whose condition references an
attribute that neither
+ * of its children can produce (SPARK-59351).
+ *
+ * A nested sub-query that references both plans cannot be rewritten into an
existence join on
+ * either side, since neither alone can evaluate it, so it has to stay in
the join condition,
+ * where both plans are in scope. Leaving it there is only correct while it
is uncorrelated: the
+ * join condition of a correlated one is dropped when it is planned as an
in-subquery filter,
+ * which would silently change the result, so a correlated one is reported
as unsupported
+ * instead. Note that such a sub-query can be
+ * correlated only to the sub-query plan, as being correlated to the outer
plan as well would
+ * require two levels of correlation, which the Analyzer rejects.
+ *
+ * A sub-query left here is further subject to the rewrite of predicate
sub-queries in join
+ * conditions, which rejects one referencing both plans under the default
configuration. It
+ * survives that rewrite when
+ *
`spark.sql.optimizer.decorrelatePredicateSubqueriesInJoinPredicate.enabled` is
disabled, and
+ * also when it is an uncorrelated IN sub-query and
+ *
`spark.sql.optimizer.optimizeUncorrelatedInSubqueriesInJoinCondition.enabled`
is disabled,
+ * since it is then not among the sub-queries that rewrite considers.
+ *
+ * An existence join yields only whether a row matched, so its `exists`
attribute cannot tell
+ * FALSE from unknown, while `IN` is three-valued. The two are
indistinguishable while the value
+ * only feeds a predicate, which is what a hoisted condition normally does,
but not when it
+ * reaches something else, e.g. `(c1 IN (SELECT col1 FROM t3)) <=> false`,
which is FALSE for a
+ * NULL that matches nothing and TRUE for the `exists` attribute. An IN
sub-query whose row
+ * comparison can evaluate to unknown is therefore rejected in that position
rather than
+ * rewritten. NOT IN is rewritten with a null-aware join condition of its
own, which is equally
+ * two-valued, so it is rejected there too.
+ *
+ * Returns the rewritten condition along with the updated outer and
sub-query plans.
+ */
+ private def rewriteExistentialExprInJoinCondition(
+ conditions: Seq[Expression],
+ outerPlan: LogicalPlan,
+ subPlan: LogicalPlan): (Option[Expression], LogicalPlan, LogicalPlan) = {
+ val (subCond, newSubPlan) =
+ rewriteExistentialExprInSubqueryPlan(conditions, outerPlan, subPlan)
+ // The sub-queries that do not reference the sub-query plan are rewritten
against the outer
+ // plan, as they only reference attributes of the outer plan, if any. This
accepts on the
+ // negation of the test above rather than on a positive one, which rests
on the invariant that
+ // `effectiveReferences` sees every attribute the rewrite can put in a
join condition: one it
+ // cannot see would be rewritten here and end up in a condition the outer
plan cannot evaluate.
+ // A positive test is not available without also refusing the sub-queries
whose retained outer
+ // attributes no longer correspond to a condition, which must keep being
rewritten here.
+ val (newCond, newOuterPlan, _) = rewriteExistentialExprWithAttrs(
+ subCond.toSeq, outerPlan, e => !effectivelyReferencesPlan(e, subPlan))
+ (newCond, newOuterPlan, newSubPlan)
+ }
+
+ /**
+ * Rewrites the existential sub-queries in `conditions` that can only be
evaluated by the
+ * sub-query plan, that is those referencing the sub-query plan but not the
outer plan, into
+ * existence joins on top of the sub-query plan. See
+ * [[rewriteExistentialExprInJoinCondition]] for details.
+ *
+ * Returns the rewritten condition along with the updated sub-query plan.
+ */
+ private def rewriteExistentialExprInSubqueryPlan(
+ conditions: Seq[Expression],
+ outerPlan: LogicalPlan,
+ subPlan: LogicalPlan): (Option[Expression], LogicalPlan) = {
+ // The rejection collector below is a plain `collect`, which visits every
node of the hoisted
+ // conditions whether or not it can hold a sub-query, so skip the
classification entirely when
+ // they hold none at all, which is the common case for a correlated EXISTS
or IN.
+ // [[collectUnknownSensitiveInSubqueries]] prunes subquery-free subtrees
on its own, so this
+ // guard only saves it the top-level call.
+ if (!conditions.exists(_.containsAnyPattern(EXISTS_SUBQUERY,
IN_SUBQUERY))) {
+ return (conditions.reduceOption(And), subPlan)
+ }
+ // A sub-query left in the join condition loses its own join condition
when it is planned
+ // there, so a correlated one would silently return a wrong result: reject
it instead. The
+ // same holds for a correlated sub-query nested below one that is left
there, since the
+ // rewrite leaves a declined sub-query whole, children included, and
nothing else will reach
+ // it: see carriesCorrelatedSubquery. Note that this walks the whole
expression, including the
+ // join condition of a sub-query the rewrite declines to descend into,
which is deliberate for
+ // the same reason.
+ //
+ // Which plan a candidate references is asked of it up to three times
below, once for the
+ // rejection of a sub-query referencing both plans, once to keep only the
unknown-sensitive
+ // ones the sub-query plan owns, and once more by the rewrite itself. Both
answers come from
+ // the same reference set in a single pass over it, and the result is
memoised per candidate,
+ // so neither the set nor the scan is repeated. The memo is keyed by
identity: the candidates
+ // are nodes of the expressions being classified, and comparing two of
them structurally would
+ // compare the sub-query plans they hold.
+ val classified = new java.util.IdentityHashMap[Expression,
PlanReferences]()
+ def classify(sq: Expression): PlanReferences = {
+ var references = classified.get(sq)
+ if (references == null) {
+ references = planReferences(effectiveReferences(sq), outerPlan,
subPlan)
+ classified.put(sq, references)
+ }
+ references
+ }
+ def ownedBySubPlan(sq: Expression): Boolean = classify(sq).subPlanOnly
+ def referencesBothPlans(sq: Expression): Boolean = classify(sq).bothPlans
+ val referencingBothPlans = conditions.flatMap(_.collect {
+ case sq @ (_: Exists | _: InSubquery)
+ if referencesBothPlans(sq) &&
Review Comment:
Thanks for your inputs @cloud-fan, I fixed by routing the carried descendant
against the newly built join, rather than by rejecting it or by making the
validation configuration-dependent. (option 1)
`rewriteCarriedExistentialExprs` now runs on every existence join this rule
builds, in all three arms. If the grafted condition carries a correlated
existential subquery, it is rewritten against whichever of that join's own
inputs can evaluate it; the routing recurses, since the joins it builds go
through it in turn. An uncorrelated one is deliberately left alone -- it has no
hoisted condition to lose and `InSubqueryExec` evaluates it three-valued, which
is also what `r4281437494` is about. One that would need both inputs, or whose
unknown is observable, is rejected.
Both of your sub-cases are fixed, and the two configurations now agree:
| | `decorrelate...InJoinCondition=false` | `=true` |
|---|---|---|
| master `6e5c61fb` | `INTERNAL_ERROR_ATTRIBUTE_NOT_FOUND` |
`INTERNAL_ERROR_ATTRIBUTE_NOT_FOUND` |
| before this commit | `1, 2, 3` | `1` |
| now | `1` | `1` |
DuckDB returns `1`. The plan under the disabled configuration is now the one
the enabled path already produced:
```
+- Join ExistenceJoin(exists#39), ((c1#12 = col1#23) AND ((k#24 = c1#12) OR
exists#40))
:- t2
+- Join ExistenceJoin(exists#40), ((col1#23 = col1#32) AND (col1#32 =
k#24))
:- t4
+- t5
```
That is the point of doing it here rather than relying on the join-condition
handling: that handling is gated on the configuration, which is exactly why the
two disagreed.
New test `existential subquery carried by a rewritten nested subquery` runs
the shape under both settings; it fails with the routing disabled, returning
`1, 2, 3`. The scaladoc on `planReferences` no longer claims a subquery
referencing neither plan "can be left in the join condition unharmed", which
was the assumption behind the defect; it now points at the carried-routing
instead.
--
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]
