peter-toth commented on code in PR #58229:
URL: https://github.com/apache/spark/pull/58229#discussion_r4093602457


##########
sql/catalyst/src/main/scala/org/apache/spark/sql/catalyst/optimizer/ConvertViewToMaterializedCTE.scala:
##########
@@ -0,0 +1,219 @@
+/*
+ * 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 in `FinishAnalysis`, immediately before `EliminateView`: 
after `EliminateView`
+ * no `View` nodes remain and every reference site holds an independent copy 
of the view's plan.
+ *
+ * A converted definition always sets `forceSkipInline = true`; 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.
+ *
+ * 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]
+
+    val rewritten = plan.transformUpWithSubqueries {
+      case v: View if qualifiedIdentifiers.contains(v.desc.identifier) =>
+        defByGroup.get(v.desc.identifier) match {
+          case Some(cteDef) =>
+            // 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)
+            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.
+            val cteDef = CTERelationDef(v.child, forceSkipInline = true)
+            defByGroup.put(v.desc.identifier, cteDef)
+            cteDefs += cteDef
+            CTERelationRef(
+              cteDef.id,
+              _resolved = true,
+              output = v.child.output,
+              isStreaming = false,
+              maxRows = cteDef.maxRows)
+        }
+    }
+
+    if (cteDefs.isEmpty) {
+      plan
+    } else {
+      attachDefs(rewritten, cteDefs.toSeq)
+    }
+  }
+
+  // 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 &&
+        !v.child.isStreaming &&
+        !hasTopLevelSort(v.child) &&
+        v.desc.viewSQLConfigs == first.desc.viewSQLConfigs &&
+        schemasAlign(first, v)
+    }
+  }
+
+  // The per-reference shuffle boundary is added above the definition, so a
+  // top-level ORDER BY in the view body would be destroyed by it.
+  private def hasTopLevelSort(plan: LogicalPlan): Boolean = plan match {
+    case _: Sort => true
+    case Project(_, child) => hasTopLevelSort(child)
+    case Filter(_, child) => hasTopLevelSort(child)
+    case SubqueryAlias(_, child) => hasTopLevelSort(child)
+    case GlobalLimit(_, child) => hasTopLevelSort(child)
+    case LocalLimit(_, child) => hasTopLevelSort(child)

Review Comment:
   **Finding 6.** `hasTopLevelSort` looks through `Project`, `Filter`, 
`SubqueryAlias` and the two limits, but not `Offset`. `ORDER BY id OFFSET 1` 
analyzes to `Offset(1, Sort(...))`, and `ORDER BY id LIMIT 50 OFFSET 1` to 
`GlobalLimit(LocalLimit(Offset(Sort(...))))`. Both reach `case _ => false`, so 
the view converts.
   
   I ran your `view with top-level ORDER BY is not converted` test with those 
two bodies on this head:
   
   | view body | repartitions | rows, conversion off | rows, conversion on |
   |---|---|---|---|
   | `SELECT id FROM range(100) ORDER BY id OFFSET 1` | 2 | `1, 2, 3` | `19, 
61, 14` |
   | `SELECT id FROM range(100) ORDER BY id LIMIT 50 OFFSET 1` | 2 | `1, 2, 3` 
| `26, 20, 3` |
   
   That is the change the guard is there to prevent, and the config doc says 
these views are never converted.
   
   ```suggestion
       case LocalLimit(_, child) => hasTopLevelSort(child)
       case Offset(_, child) => hasTopLevelSort(child)
   ```
   
   With it applied, both bodies give 0 repartitions and `1, 2, 3`, and 
`ConvertViewToMaterializedCTEQuerySuite` stays green. Adding the two bodies to 
the ORDER BY test would pin it.
   



##########
sql/catalyst/src/main/scala/org/apache/spark/sql/catalyst/optimizer/ConvertViewToMaterializedCTE.scala:
##########
@@ -0,0 +1,219 @@
+/*
+ * 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 in `FinishAnalysis`, immediately before `EliminateView`: 
after `EliminateView`
+ * no `View` nodes remain and every reference site holds an independent copy 
of the view's plan.
+ *
+ * A converted definition always sets `forceSkipInline = true`; 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.
+ *
+ * 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]
+
+    val rewritten = plan.transformUpWithSubqueries {
+      case v: View if qualifiedIdentifiers.contains(v.desc.identifier) =>
+        defByGroup.get(v.desc.identifier) match {
+          case Some(cteDef) =>
+            // 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)
+            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.
+            val cteDef = CTERelationDef(v.child, forceSkipInline = true)

Review Comment:
   **Finding 7.** The rule decides from the number of `View` occurrences, but 
not every occurrence survives the rewrite as a reference. Then a definition 
ends up with one live reference, and `forceSkipInline` still gives it a shuffle 
that nothing reuses. I measured two shapes on this head:
   
   - **Nested views.** In `nested view is converted together with the view it 
references`, `v1` occurs once inside each `v2`, so both views qualify. The 
second `v2` occurrence is replaced by `Project(..., ref)` at line 114, which 
drops its body together with its `v1` occurrence. That leaves `v1`'s definition 
with one reference, inside `v2`'s definition. The executed plan (AQE on and 
off, broadcast off) has 2 round-robin shuffles and 1 reused exchange. One is 
`v2`'s shared shuffle. The other is `v1`'s shuffle underneath it, which runs 
once and saves nothing. The test's 4 repartitions is that shuffle counted once 
per `v2` site.
   - **A view that a dead CTE also mentions.** `WITH c AS (SELECT id FROM v) 
SELECT id FROM v WHERE id < 10` reads `v` once. But the unreferenced `c` still 
holds a second occurrence when this rule runs, so the rule converts it. The 
executed plan gains a round-robin shuffle over a single reference.
   
   Both costs are opt-in. But nested views are common in the ETL setups the 
description targets. To cover both shapes, count the references to each new 
definition after the rewrite, and set `forceSkipInline` only where there are at 
least two. `InlineCTE` already does this kind of counting: it discounts 
references that come from unreferenced definitions (`InlineCTE.scala:231`). 
With `forceSkipInline = false`, `InlineCTE` would then flatten the 
single-reference ones back.
   



-- 
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]

Reply via email to