xumingming commented on code in PR #58229: URL: https://github.com/apache/spark/pull/58229#discussion_r4482126423
########## sql/catalyst/src/main/scala/org/apache/spark/sql/catalyst/optimizer/ConvertViewToMaterializedCTE.scala: ########## @@ -0,0 +1,297 @@ +/* + * 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.spark.sql.catalyst.optimizer + +import scala.collection.mutable + +import org.apache.spark.sql.catalyst.TableIdentifier +import org.apache.spark.sql.catalyst.expressions.{Alias, Attribute} +import org.apache.spark.sql.catalyst.plans.logical._ +import org.apache.spark.sql.catalyst.rules.Rule +import org.apache.spark.sql.internal.SQLConf + +/** + * Rewrites multiple references to the same view into a single `CTERelationDef` with multiple + * `CTERelationRef`s, so that the view's underlying plan is computed once (through exchange + * reuse at the physical layer) instead of once per reference. + * + * The rule runs first in `FinishAnalysis`, before `EliminateResolvedHint`, + * `EliminateSubqueryAliases` and `EliminatePipeOperators` strip analysis-only nodes from + * the plan. Those rules are applied to the main plan before its subquery plans, so a + * definition body taken from a first-visited subquery occurrence may still hold such + * nodes; running the conversion first lets the cleanup rules reach the body at the top + * level, wherever it was taken from. `EliminateView` runs after the cleanup rules: no + * `View` nodes remain and every reference site holds an independent copy of the view's + * plan. + * + * A join hint written at the top of a view body binds to the join(s) the view takes part + * in (see `SubqueryHintPropagationSuite`): `EliminateResolvedHint` extracts hints from + * Join children by walking down unary nodes, and processes the body inline today so the + * hint reaches the consuming join. After this rule the body lives inside a `CTERelationDef` + * the walk cannot reach, so the rule strips those top-of-body hints itself and re-attaches + * them above every reference site. Hints deeper in the body bind to the body's own joins + * and are left in place. + * + * A converted definition keeps `forceSkipInline = true` only when at least two references + * survive the rewrite; otherwise `InlineCTE` would immediately flatten it back into + * duplicated subtrees (the definition body is deterministic in every case we convert), + * making the rule a no-op. Definitions with fewer than two surviving references keep the + * flag off and are inlined: their occurrences do not all survive the rewrite (a later + * occurrence's body is dropped, taking references to nested converted views with it), and + * some references may sit inside unreferenced user-written CTEs - materializing either + * way would add a shuffle that exchange reuse cannot deduplicate. + * + * Only deterministic, batch views are eligible: a multi-reference CTE guarantees that its + * definition is evaluated exactly once (even for non-deterministic definitions), while + * multiple references to a non-deterministic view are evaluated independently today. + * Converting such views would change query results. + * + * A view body may contain correlated subqueries whose outer references resolve to relations + * inside the same body (e.g. `t WHERE x IN (SELECT y FROM s WHERE s.k = t.k)`). The view is + * analyzed standalone when it is created, so an outer reference that does not resolve inside + * the body fails view analysis and can never escape to the outer query. The converted + * definition contains the whole body, so internal correlations resolve within it and these + * bodies are safe to convert. `InlineCTE`'s rejection of boundary-crossing outer references + * is only a generic safety net for non-view `forceSkipInline` producers. + * + * Each reference site gains a shuffle boundary added by `ReplaceCTERefWithRepartition` and + * deduplicated by exchange reuse, so the conversion trades recomputation for a + * shuffle plus reuse; it is therefore gated behind + * [[SQLConf.CONVERT_VIEW_TO_MATERIALIZED_CTE]] and off by default. + */ +object ConvertViewToMaterializedCTE extends Rule[LogicalPlan] { + + override def apply(plan: LogicalPlan): LogicalPlan = { + if (!SQLConf.get.getConf(SQLConf.CONVERT_VIEW_TO_MATERIALIZED_CTE)) return plan + // FinishAnalysis re-runs the rule batch on Subquery roots and requires the result + // to stay a Subquery; the top-level pass already covers views inside subqueries. + if (plan.isInstanceOf[Subquery]) return plan + val occurrences = plan.collectWithSubqueries { case v: View => v } + if (occurrences.length < 2) return plan + + // Admit an identifier only when every occurrence of it forms one qualifying group. + // Grouping is by identifier first because the rewrite below matches occurrences by + // identifier: an identifier carried by occurrences with divergent bodies (e.g. one + // occurrence resolved against a view definition that was replaced after another + // occurrence captured it), or whose occurrences fail qualification for any other + // reason, must not convert at all - otherwise the rewrite would rebind those + // occurrences against a definition their plan does not match, silently changing + // query results. + // + // Inner CTE definition ids are normalized before the bodies are compared: a view + // body is re-analyzed per occurrence, and the re-analysis re-substitutes the body's + // inner CTEs, minting a fresh `CTERelationDef` id each time (e.g. for a SQL view + // whose body has a WITH clause over a persistent table). The re-minted ids are the + // only difference between such occurrences' bodies, so the normalization lets them + // group together; bodies that differ in anything else still stay apart. + val qualifiedIdentifiers = occurrences + .groupBy(_.desc.identifier) + .collect { + case (identifier, occs) + if qualifies(occs) && occs.map(occ => + normalizeCteIds(occ.child).canonicalized).distinct.length == 1 => identifier + } + .toSet + if (qualifiedIdentifiers.isEmpty) return plan + + // Bottom-up rewrite: nested views are visited (and their definitions appended) before + // the views containing them, so a referenced definition always precedes its referrer + // in `cteDefs`. The first occurrence of a group creates the definition, replaced by a + // bare reference; later occurrences are wrapped in a `Project` re-minting the + // occurrence's ids from the definition output, so consumers above need no rewriting. + val cteDefs = mutable.ArrayBuffer.empty[CTERelationDef] + val defByGroup = mutable.HashMap.empty[TableIdentifier, (CTERelationDef, Seq[HintInfo])] + + val rewritten = plan.transformUpWithSubqueries { + case v: View if qualifiedIdentifiers.contains(v.desc.identifier) => + // The body's top-of-node hints must not stay in the definition, where neither + // this rule nor `EliminateResolvedHint` can hoist them into the consuming joins; + // build the definition from the hint-free body and re-apply the hints above each + // reference site instead. All occurrences share one canonicalized body, so the + // hints extracted from the first one describe every reference site uniformly. + defByGroup.get(v.desc.identifier) match { + case Some((cteDef, bodyHints)) => + // Later occurrence: re-bind the reference output to this occurrence's + // attributes positionally. The group qualification has already asserted that + // name, type and nullability align element-wise. + val ref = CTERelationRef( + cteDef.id, + _resolved = true, + output = cteDef.output, + isStreaming = false, + maxRows = cteDef.maxRows) + withBodyHints(bodyHints, + Project(rebindingProjectList(v.output, cteDef.output), ref)) + + case None => + // First occurrence: consumers above already reference this occurrence's + // expression ids, which are exactly the definition output, so the bare + // reference is output-compatible. `forceSkipInline` is decided after the + // rewrite, by the number of surviving references. + val (hintFreeBody, bodyHints) = + EliminateResolvedHint.extractHintsFromPlan(v.child) + val cteDef = CTERelationDef(hintFreeBody) + defByGroup.put(v.desc.identifier, (cteDef, bodyHints)) + cteDefs += cteDef + withBodyHints(bodyHints, CTERelationRef( + cteDef.id, + _resolved = true, + output = v.child.output, + isStreaming = false, + maxRows = cteDef.maxRows)) + } + } + + if (cteDefs.isEmpty) { + plan + } else { + // Not every occurrence survives the rewrite as a reference: the body of a later + // occurrence is dropped (taking references to nested converted views with it), and + // some references sit inside unreferenced user-written CTEs. Materializing a + // definition with fewer than two live references would add a shuffle that exchange + // reuse cannot deduplicate, so re-count after the rewrite and leave such definitions + // inlinable - InlineCTE flattens them back into their consumers. + val newDefIds = cteDefs.map(_.id).toSet + val withDefs = attachDefs(rewritten, cteDefs.toSeq) + val liveRefCounts = countLiveRefs(withDefs) + withDefs.transformUpWithSubqueries { + case d: CTERelationDef if newDefIds.contains(d.id) => + d.copy(forceSkipInline = liveRefCounts.getOrElse(d.id, 0) >= 2) + } + } + } + + // Counts, per definition id, the references reachable from the query root: a reference + // inside a definition counts only when the definition itself is live, mirroring how + // `InlineCTE` discounts references held by unreferenced definitions. Recursive CTEs are + // handled by the visited-set worklist. + private def countLiveRefs(plan: LogicalPlan): Map[Long, Int] = { + // The region of a reference is its innermost enclosing definition's id, or + // ROOT_REGION at the query root. + val ROOT_REGION = -1L + val outgoing = mutable.HashMap.empty[Long, mutable.ArrayBuffer[Long]] + + def walk(p: LogicalPlan, region: Long): Unit = { + p match { + case d: CTERelationDef => walk(d.child, d.id) + case r: CTERelationRef => + outgoing.getOrElseUpdate(region, mutable.ArrayBuffer.empty) += r.cteId + case _ => p.children.foreach(walk(_, region)) + } + p.subqueries.foreach(walk(_, region)) + } + walk(plan, ROOT_REGION) + + val liveRefCounts = mutable.HashMap.empty[Long, Int].withDefaultValue(0) + val visitedDefs = mutable.HashSet.empty[Long] + val pending = mutable.Queue.empty[Long] ++= + outgoing.getOrElse(ROOT_REGION, mutable.ArrayBuffer.empty) + while (pending.nonEmpty) { + val target = pending.dequeue() + liveRefCounts(target) += 1 + if (visitedDefs.add(target)) { + pending ++= outgoing.getOrElse(target, mutable.ArrayBuffer.empty) + } + } + liveRefCounts.toMap + } + + // Group by view identity: the rule dedupes references of the SAME view, not distinct + // views with coincidentally equal bodies, and all occurrences of one view resolve + // through the same (db-qualified) identifier. The canonicalized body comparison is a + // precondition guard in the identifier admission above: if occurrences of one view + // ever diverge structurally, we skip conversion instead of building a wrong shared + // definition. + // + // Rewrites every inner CTE definition id and reference to an appearance ordinal, so + // that bodies differing only in re-minted inner CTE ids canonicalize equal. Same tree + // shapes are visited in the same order, so equal bodies map to equal ordinals, while + // genuinely different structures (e.g. two sibling identical CTEs) still produce + // distinct ordinals and stay distinguishable. + private def normalizeCteIds(plan: LogicalPlan): LogicalPlan = { + val ids = mutable.HashMap.empty[Long, Long] + plan.transformUpWithSubqueries { + case d: CTERelationDef => d.copy(id = ids.getOrElseUpdate(d.id, ids.size)) + case r: CTERelationRef => r.copy(cteId = ids.getOrElseUpdate(r.cteId, ids.size)) + } + } + + private def qualifies(occs: Seq[View]): Boolean = { + val first = occs.head + occs.length >= 2 && occs.forall { v => + v.resolved && + v.child.deterministic && Review Comment: Thanks for the repro. The bug is real, and the existing `TABLESAMPLE` test misses it because `hasMainRegionSort` rejects the `ORDER BY` inner view first. I would rather not special-case `Sample` in `qualifies`. `v.child.deterministic` is the check `InlineCTE`, `MergeSubplans`, and `CreateSQLFunctionCommand` already use, so a guard here leaves the same hole at every other call site. `aes_encrypt` without an IV has the same shape: its children are deterministic, so the plan reports `deterministic`, while `ExpressionImplUtils.aesEncrypt` draws a random IV from `SecureRandom`. A `Sample`-only check would still convert that view and share one ciphertext. The fix belongs on the source: `Sample` should be non-deterministic when `seed` is empty, and `AesEncrypt` when it generates the IV. I'll file a JIRA and handle both in a dedicated PR. `REPEATABLE` keeps its seed on the operator, so it stays deterministic and this rule can still convert it. How do you think? -- 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]
