spark-reviews mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From hvanhovell <...@git.apache.org>
Subject [GitHub] spark pull request #17138: [SPARK-17080] [SQL] join reorder
Date Fri, 03 Mar 2017 13:46:57 GMT
Github user hvanhovell commented on a diff in the pull request:

    https://github.com/apache/spark/pull/17138#discussion_r104153912
  
    --- 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
    + * cardinalities and sizes to compute costs.
    + */
    +case class JoinReorderDP(
    +    conf: CatalystConf,
    +    items: Seq[LogicalPlan],
    +    conditions: Set[Expression],
    +    topOutput: AttributeSet) extends PredicateHelper{
    +
    +  /** Level i maintains all found plans for sets of i joinable items. */
    +  val foundPlans = new Array[mutable.Map[Set[Int], JoinPlan]](items.length + 1)
    +  for (i <- 1 to items.length) foundPlans(i) = mutable.Map.empty
    +
    +  def search(): Option[LogicalPlan] = {
    +    // Start from the first level: each plan is a single item with zero cost.
    +    val itemIndex = items.zipWithIndex
    +    foundPlans(1) ++=
    +      itemIndex.map { case (item, id) => Set(id) -> JoinPlan(Set(id), item, Cost(0,
0)) }
    +
    +    for (lev <- 2 to items.length) {
    +      searchForLevel(lev)
    +    }
    +
    +    val plansLastLevel = foundPlans(items.length)
    +    if (plansLastLevel.isEmpty) {
    +      // Failed to find a plan, fall back to the original plan
    +      None
    +    } else {
    +      // There must be only one plan at the last level, which contains all items.
    +      assert(plansLastLevel.size == 1 && plansLastLevel.head._1.size == items.length)
    +      Some(plansLastLevel.head._2.plan)
    +    }
    +  }
    +
    +  /** Find all possible plans in one level, based on previous levels. */
    +  private def searchForLevel(level: Int): Unit = {
    +    val foundPlansCurLevel = foundPlans(level)
    +    var k = 1
    +    var continue = true
    +    while (continue) {
    +      val otherLevel = level - k
    +      if (k > otherLevel) {
    +        // We can break from here, because when building a join from A and B, both A
J B and B J A
    +        // are handled.
    +        continue = false
    +      } else {
    +        val joinPlansLevelK = foundPlans(k).values.toSeq
    +        for (i <- joinPlansLevelK.indices) {
    +          val curJoinPlan = joinPlansLevelK(i)
    +
    +          val joinPlansOtherLevel = if (k == otherLevel) {
    +            // Both sides of a join are at the same level, no need to repeat for previous
ones.
    +            joinPlansLevelK.drop(i)
    +          } else {
    +            foundPlans(otherLevel).values.toSeq
    +          }
    +
    +          joinPlansOtherLevel.foreach { otherJoinPlan =>
    +            // Should not join two overlapping item sets.
    +            if (curJoinPlan.itemIds.intersect(otherJoinPlan.itemIds).isEmpty) {
    +              val joinPlan = buildJoin(curJoinPlan, otherJoinPlan)
    +              if (joinPlan.nonEmpty) {
    +                // Check if it's the first plan for the item set, or it's a better plan
than
    +                // the existing one due to lower cost.
    +                val existingPlan = foundPlansCurLevel.get(joinPlan.get.itemIds)
    +                if (existingPlan.isEmpty || joinPlan.get.cost < existingPlan.get.cost)
{
    +                  foundPlansCurLevel.update(joinPlan.get.itemIds, joinPlan.get)
    +                }
    +              }
    +            }
    +          }
    +        }
    +
    +        k += 1
    +      }
    +    }
    +  }
    +
    +  /** Build a new join node. */
    +  private def buildJoin(curJoinPlan: JoinPlan, otherJoinPlan: JoinPlan): Option[JoinPlan]
= {
    +    // Check if these two nodes are inner joinable. We consider cartesian product very
    +    // costly, thus exclude such plans. This also helps us to reduce the search space.
    +    val curPlan = curJoinPlan.plan
    +    val otherPlan = otherJoinPlan.plan
    +    val joinCond = conditions
    +      .filterNot(l => canEvaluate(l, curPlan))
    +      .filterNot(r => canEvaluate(r, otherPlan))
    +      .filter(e => e.references.subsetOf(curPlan.outputSet ++ otherPlan.outputSet))
    +
    +    if (joinCond.nonEmpty) {
    +      val curPlanStats = curPlan.stats(conf)
    +      val otherPlanStats = otherPlan.stats(conf)
    +      if (curPlanStats.rowCount.nonEmpty && otherPlanStats.rowCount.nonEmpty)
{
    --- End diff --
    
    When does this happen?


---
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 infrastructure@apache.org or file a JIRA ticket
with INFRA.
---

---------------------------------------------------------------------
To unsubscribe, e-mail: reviews-unsubscribe@spark.apache.org
For additional commands, e-mail: reviews-help@spark.apache.org


Mime
View raw message