Github user wzhfy commented on a diff in the pull request:
https://github.com/apache/spark/pull/17138#discussion_r104172223
--- Diff:
sql/catalyst/src/main/scala/org/apache/spark/sql/catalyst/optimizer/CostBasedJoinReorder.scala
---
@@ -0,0 +1,274 @@
+/*
+ * 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.CatalystConf
+import org.apache.spark.sql.catalyst.expressions.{And, AttributeSet,
Expression, PredicateHelper}
+import org.apache.spark.sql.catalyst.plans.{Inner, InnerLike}
+import org.apache.spark.sql.catalyst.plans.logical.{Join, LogicalPlan,
Project}
+import org.apache.spark.sql.catalyst.rules.Rule
+
+
+/**
+ * Cost-based join reorder.
+ * We may have several join reorder algorithms in the future. This class
is the entry of these
+ * algorithms, and chooses which one to use.
+ */
+case class CostBasedJoinReorder(conf: CatalystConf) extends
Rule[LogicalPlan] with PredicateHelper {
+ def apply(plan: LogicalPlan): LogicalPlan = {
+ if (!conf.cboEnabled || !conf.joinReorderEnabled) {
+ plan
+ } else {
+ plan transform {
+ case p @ Project(projectList, j @ Join(_, _, _: InnerLike, _)) if
!j.ordered =>
+ reorder(j, p.outputSet)
+ case j @ Join(_, _, _: InnerLike, _) if !j.ordered =>
+ reorder(j, j.outputSet)
+ }
+ }
+ }
+
+ def reorder(plan: LogicalPlan, output: AttributeSet): LogicalPlan = {
+ val (items, conditions) = extractInnerJoins(plan)
+ val result =
+ if (items.size > 2 && items.size <= conf.joinReorderDPThreshold &&
conditions.nonEmpty) {
+ JoinReorderDP(conf, items, conditions,
output).search().getOrElse(plan)
+ } else {
+ plan
+ }
+ // Set all inside joins ordered.
+ setOrdered(result)
+ result
+ }
+
+ /**
+ * Extract inner joinable items and join conditions.
+ * This method works for bushy trees and left/right deep trees.
+ */
+ def extractInnerJoins(plan: LogicalPlan): (Seq[LogicalPlan],
Set[Expression]) = plan match {
+ case j @ Join(left, right, _: InnerLike, cond) =>
+ val (leftPlans, leftConditions) = extractInnerJoins(left)
+ val (rightPlans, rightConditions) = extractInnerJoins(right)
+ (leftPlans ++ rightPlans,
cond.map(splitConjunctivePredicates).getOrElse(Nil).toSet ++
+ leftConditions ++ rightConditions)
+ case Project(_, j @ Join(left, right, _: InnerLike, cond)) =>
+ val (leftPlans, leftConditions) = extractInnerJoins(left)
+ val (rightPlans, rightConditions) = extractInnerJoins(right)
+ (leftPlans ++ rightPlans,
cond.map(splitConjunctivePredicates).getOrElse(Nil).toSet ++
+ leftConditions ++ rightConditions)
+ case _ =>
+ (Seq(plan), Set())
+ }
+
+ def setOrdered(plan: LogicalPlan): Unit = plan match {
+ case j @ Join(left, right, _: InnerLike, cond) =>
+ j.ordered = true
+ setOrdered(left)
+ setOrdered(right)
+ case Project(_, j @ Join(left, right, _: InnerLike, cond)) =>
+ j.ordered = true
+ setOrdered(left)
+ setOrdered(right)
+ case _ =>
+ }
+}
+
+/**
+ * Reorder the joins using a dynamic programming algorithm:
+ * First we put all items (basic joined nodes) into level 1, then we build
all two-way joins
+ * at level 2 from plans at level 1 (single items), then build all 3-way
joins from plans
+ * at previous levels (two-way joins and single items), then 4-way joins
... etc, until we
+ * build all n-way joins and pick the best plan among them.
+ *
+ * When building m-way joins, we only keep the best plan (with the lowest
cost) for the same set
+ * of m items. E.g., for 3-way joins, we keep only the best plan for items
{A, B, C} among
+ * plans (A J B) J C, (A J C) J B and (B J C) J A.
+ *
+ * Thus the plans maintained for each level when reordering four items A,
B, C, D are as follows:
+ * level 1: p({A}), p({B}), p({C}), p({D})
+ * level 2: p({A, B}), p({A, C}), p({A, D}), p({B, C}), p({B, D}), p({C,
D})
+ * level 3: p({A, B, C}), p({A, B, D}), p({A, C, D}), p({B, C, D})
+ * level 4: p({A, B, C, D})
+ * where p({A, B, C, D}) is the final output plan.
+ *
+ * For cost evaluation, since physical costs for operators are not
available currently, we use
--- End diff --
This is a good point. Ideally, choices of join order, join algorithm and
enforcer like distribute and sort are made in the same algorithm, but two
issues make this difficult to do in Spark.
1. Catalyst separates logical optimization, physical implementation
(strategies) and enforcer (EnsureRequirements) into different stages.
2. If we want to consider choosing algorithms and enforcers, we need
physical cost evaluation to guide us, while we don't have this now.
So in the current stage, join reordering is logical optimization. In the
future after we support physical cost evaluation and maybe modify catalyst's
structure, we can have a more mature algorithm. The DP framework will still
work, the difference should be the logic of plan cost comparison.
---
If your project is set up for it, you can reply to this email and have your
reply appear on GitHub as well. If your project does not have this feature
enabled and wishes so, or if the feature is enabled but not working, please
contact infrastructure at [email protected] or file a JIRA ticket
with INFRA.
---
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
