beliefer commented on code in PR #12756: URL: https://github.com/apache/gluten/pull/12756#discussion_r3871397200
########## backends-velox/src/main/scala/org/apache/gluten/extension/RewriteSelfJoinInequalityToAggregate.scala: ########## @@ -0,0 +1,674 @@ +/* + * Licensed to the Apache Software Foundation (ASF) under one or more + * contributor license agreements. See the NOTICE file distributed with + * this work for additional information regarding copyright ownership. + * The ASF licenses this file to You under the Apache License, Version 2.0 + * (the "License"); you may not use this file except in compliance with + * the License. You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ +package org.apache.gluten.extension + +import org.apache.gluten.config.VeloxConfig + +import org.apache.spark.internal.Logging +import org.apache.spark.sql.SparkSession +import org.apache.spark.sql.catalyst.expressions._ +import org.apache.spark.sql.catalyst.expressions.RowOrdering +import org.apache.spark.sql.catalyst.expressions.aggregate._ +import org.apache.spark.sql.catalyst.plans._ +import org.apache.spark.sql.catalyst.plans.logical._ +import org.apache.spark.sql.catalyst.rules.Rule +import org.apache.spark.sql.execution.datasources.{HadoopFsRelation, LogicalRelation} +import org.apache.spark.sql.execution.datasources.parquet.ParquetFileFormat +import org.apache.spark.sql.types.LongType + +/** + * Rewrites self-join with inequality into GROUP BY + HAVING COUNT(DISTINCT) > 1. + * + * Targets the two uncorrelated InSubquery shapes exercised by TPC-DS Q95: + * + * - Pattern A': the subquery top-level InnerJoin is a direct self-join. + * - Pattern A2: the subquery contains an outer InnerJoin with a self-join child; only the + * self-join child is replaced with Aggregate and the outer join is preserved. + * + * Both patterns require an existence-only membership context so row-count multiplicity from the + * original self-join cross-product does not affect semantics. Correlated InSubquery expressions are + * intentionally fail-closed because the ExprId remapping performed here does not rewrite correlated + * predicates. + * + * Both patterns share: + * - [[buildAggregateHavingDistinctGt1]] to construct `Filter(cnt > 1, Aggregate)` + * - [[canonicalizeWrapper]] to rebuild a wrapping Project so every equi-key reference points to + * the sjLeft-side attribute, with **fresh exprIds** (Spark's SPARK-21835 style -- no reuse of + * original exprIds), returning an old->new attribute remap for downstream rewrite. + * + * Controlled by `spark.gluten.sql.rewrite.selfJoinInequality` (default false, opt-in). + */ +case class RewriteSelfJoinInequalityToAggregate(spark: SparkSession) + extends Rule[LogicalPlan] + with PredicateHelper + with Logging { + + private val CountDistinctAliasName = "_gluten_rw_selfjoin_cnt_distinct" + + override def apply(plan: LogicalPlan): LogicalPlan = { + if (!VeloxConfig.get.enableRewriteSelfJoinInequality) { + logDebug("RewriteSelfJoinInequalityToAggregate: disabled via config, skipping") + return plan + } + + // Pattern A' / A2: rewrite uncorrelated InSubquery plans. + // Correlated subqueries carry outer references / correlated join conditions in + // `SubqueryExpression.children`; fail closed because this rule does not remap them. + val rewritten = plan.transformAllExpressions { + case in @ InSubquery(_, lq: ListQuery) if lq.children.isEmpty => + rewriteSubqueryPlan(lq.plan) match { + case Some(newSub) => in.copy(query = lq.copy(plan = newSub)) + case None => in + } + } + if (!(rewritten eq plan)) { + logDebug( + "RewriteSelfJoinInequalityToAggregate: rewrote self-join to " + + "GROUP BY + HAVING COUNT(DISTINCT) > 1") + } + rewritten + } + // ============================================================================ + // Shared helpers + // ============================================================================ + + /** + * Build `Filter(cnt > 1, Aggregate(equiKeys, [equiKeys, cnt_alias], Filter(IsNotNull(equiKeys), + * child)))`. Returns the Filter node whose output is `equiKeys ++ [count_alias_attr]`. + * + * The extra `IsNotNull(equiKeys)` filter is essential to preserve the original equi-join's NULL + * semantics. Under SQL 3VL, `left.k = right.k` never matches when either side is NULL, so the + * original self-join drops rows with NULL equi-keys. Aggregate, in contrast, groups NULL keys + * together into a single "NULL group" -- if that group has >= 2 distinct non-null neq values, + * COUNT(DISTINCT) > 1 fires and injects NULL into the subquery output. That leaked NULL then + * turns `NOT IN` into a spurious empty result (Spark's null-aware anti-join uses + * `Or(equi, IsNull(equi))` which any NULL sub-row satisfies) and can flip IN/NOT IN outcomes. The + * neq column needs no such filter: `COUNT(DISTINCT col)` already ignores NULL. + */ + private def buildAggregateHavingDistinctGt1( + equiKeys: Seq[Attribute], + neqCol: Attribute, + child: LogicalPlan): LogicalPlan = { + val countExpr = AggregateExpression( + Count(Seq(neqCol)), Review Comment: The COUNT(DISTINCT neqCol) > 1 rewrite diverges from the original s1.neqCol <> s2.neqCol predicate for special float/double values: - -0.0 vs +0.0: SQL -0.0 = 0.0 is true, so -0.0 <> 0.0 is false. The original self-join produces no row. But Spark's COUNT(DISTINCT) uses bit-level/Java Double.equals which keeps -0.0 != 0.0, so COUNT(DISTINCT {-0.0, 0.0}) = 2. The rewrite fires and the row becomes IN when it should be OUT. - NaN: The reverse — two NaN rows satisfy NaN <> NaN in SQL (true), so the original matches; but Java NaN.equals(NaN) is true so COUNT(DISTINCT NaN) = 1. The rewrite doesn't fire when it should. -- 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]
