Github user cloud-fan commented on a diff in the pull request:
https://github.com/apache/spark/pull/16228#discussion_r101177911
--- Diff:
sql/catalyst/src/main/scala/org/apache/spark/sql/catalyst/plans/logical/statsEstimation/JoinEstimation.scala
---
@@ -0,0 +1,316 @@
+/*
+ * 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.plans.logical.statsEstimation
+
+import scala.collection.mutable
+import scala.collection.mutable.ArrayBuffer
+
+import org.apache.spark.internal.Logging
+import org.apache.spark.sql.catalyst.CatalystConf
+import org.apache.spark.sql.catalyst.expressions.{Attribute, AttributeMap,
AttributeReference, Expression}
+import org.apache.spark.sql.catalyst.planning.ExtractEquiJoinKeys
+import org.apache.spark.sql.catalyst.plans._
+import org.apache.spark.sql.catalyst.plans.logical.{ColumnStat, Join,
Statistics}
+import
org.apache.spark.sql.catalyst.plans.logical.statsEstimation.EstimationUtils._
+
+
+object JoinEstimation extends Logging {
+ /**
+ * Estimate statistics after join. Return `None` if the join type is not
supported, or we don't
+ * have enough statistics for estimation.
+ */
+ def estimate(conf: CatalystConf, join: Join): Option[Statistics] = {
+ join.joinType match {
+ case Inner | Cross | LeftOuter | RightOuter | FullOuter =>
+ InnerOuterEstimation(conf, join).doEstimate()
+ case LeftSemi | LeftAnti =>
+ LeftSemiAntiEstimation(conf, join).doEstimate()
+ case _ =>
+ logDebug(s"[CBO] Unsupported join type: ${join.joinType}")
+ None
+ }
+ }
+}
+
+case class InnerOuterEstimation(conf: CatalystConf, join: Join) extends
Logging {
+
+ private val leftStats = join.left.stats(conf)
+ private val rightStats = join.right.stats(conf)
+
+ /**
+ * Estimate output size and number of rows after a join operator, and
update output column stats.
+ */
+ def doEstimate(): Option[Statistics] = join match {
+ case _ if !rowCountsExist(conf, join.left, join.right) =>
+ None
+
+ case ExtractEquiJoinKeys(joinType, leftKeys, rightKeys, condition,
left, right) =>
+ // 1. Compute join selectivity
+ val joinKeyPairs = extractJoinKeysWithColStats(leftKeys, rightKeys)
+ val selectivity = joinSelectivity(joinKeyPairs)
+
+ // 2. Estimate the number of output rows
+ val leftRows = leftStats.rowCount.get
+ val rightRows = rightStats.rowCount.get
+ val innerJoinedRows = ceil(BigDecimal(leftRows * rightRows) *
selectivity)
+
+ // Make sure outputRows won't be too small based on join type.
+ val outputRows = joinType match {
+ case LeftOuter =>
+ // All rows from left side should be in the result.
+ leftRows.max(innerJoinedRows)
+ case RightOuter =>
+ // All rows from right side should be in the result.
+ rightRows.max(innerJoinedRows)
+ case FullOuter =>
+ // T(A FOJ B) = T(A LOJ B) + T(A ROJ B) - T(A IJ B)
+ leftRows.max(innerJoinedRows) + rightRows.max(innerJoinedRows) -
innerJoinedRows
+ case _ =>
+ // Don't change for inner or cross join
+ innerJoinedRows
+ }
+
+ // 3. Update statistics based on the output of join
+ val inputAttrStats = AttributeMap(
+ leftStats.attributeStats.toSeq ++ rightStats.attributeStats.toSeq)
+ val attributesWithStat = join.output.filter(a =>
inputAttrStats.contains(a))
+ val (fromLeft, fromRight) =
attributesWithStat.partition(join.left.outputSet.contains(_))
+
+ val outputStats: Seq[(Attribute, ColumnStat)] = if (innerJoinedRows
== 0) {
+ joinType match {
+ // For outer joins, if the inner join part is empty, the number
of output rows is the
+ // same as that of the outer side. And column stats of join keys
from the outer side
+ // keep unchanged, while column stats of join keys from the
other side should be updated
+ // based on added null values.
+ case LeftOuter =>
+ fromLeft.map(a => (a, inputAttrStats(a))) ++
+ fromRight.map(a => (a, nullColumnStat(a.dataType, leftRows)))
+ case RightOuter =>
+ fromRight.map(a => (a, inputAttrStats(a))) ++
+ fromLeft.map(a => (a, nullColumnStat(a.dataType, rightRows)))
+ case FullOuter =>
+ fromLeft.map { a =>
+ val oriColStat = inputAttrStats(a)
+ (a, oriColStat.copy(nullCount = oriColStat.nullCount +
rightRows))
+ } ++ fromRight.map { a =>
+ val oriColStat = inputAttrStats(a)
+ (a, oriColStat.copy(nullCount = oriColStat.nullCount +
leftRows))
+ }
+ case _ =>
+ // For inner join, since the output is empty, we don't need to
keep column stats.
+ Nil
+ }
+ } else {
+ val joinKeyStats = getIntersectedStats(joinKeyPairs)
+ join.joinType match {
+ // For outer joins, don't update column stats from the outer
side.
+ case LeftOuter =>
+ fromLeft.map(a => (a, inputAttrStats(a))) ++
+ updateAttrStats(outputRows, fromRight, inputAttrStats,
joinKeyStats)
+ case RightOuter =>
+ updateAttrStats(outputRows, fromLeft, inputAttrStats,
joinKeyStats) ++
+ fromRight.map(a => (a, inputAttrStats(a)))
+ case FullOuter =>
+ attributesWithStat.map(a => (a, inputAttrStats(a)))
+ case _ =>
+ // Update column stats from both sides for inner or cross join.
+ updateAttrStats(outputRows, attributesWithStat,
inputAttrStats, joinKeyStats)
+ }
+ }
+
+ val outputAttrStats = AttributeMap(outputStats)
+ Some(Statistics(
+ sizeInBytes = getOutputSize(join.output, outputRows,
outputAttrStats),
+ rowCount = Some(outputRows),
+ attributeStats = outputAttrStats,
+ isBroadcastable = false))
+
+ case _ =>
+ // When there is no equi-join condition, we do estimation like
cartesian product.
+ val inputAttrStats = AttributeMap(
+ leftStats.attributeStats.toSeq ++ rightStats.attributeStats.toSeq)
+ // Propagate the original column stats
+ val outputAttrStats = getOutputMap(inputAttrStats, join.output)
+ val outputRows = leftStats.rowCount.get * rightStats.rowCount.get
+ Some(Statistics(
+ sizeInBytes = getOutputSize(join.output, outputRows,
outputAttrStats),
+ rowCount = Some(outputRows),
+ attributeStats = outputAttrStats,
+ isBroadcastable = false))
+ }
+
+ // scalastyle:off
+ /**
+ * The number of rows of A inner join B on A.k1 = B.k1 is estimated by
this basic formula:
+ * T(A IJ B) = T(A) * T(B) / max(V(A.k1), V(B.k1)), where V is the
number of distinct values of
+ * that column. The underlying assumption for this formula is: each
value of the smaller domain
+ * is included in the larger domain.
+ * Generally, inner join with multiple join keys can also be estimated
based on the above
+ * formula:
+ * T(A IJ B) = T(A) * T(B) / (max(V(A.k1), V(B.k1)) * max(V(A.k2),
V(B.k2)) * ... * max(V(A.kn), V(B.kn)))
+ * However, the denominator can become very large and excessively reduce
the result, so we use a
+ * conservative strategy to take only the largest max(V(A.ki), V(B.ki))
as the denominator.
+ */
+ // scalastyle:on
+ def joinSelectivity(joinKeyPairs: Seq[(AttributeReference,
AttributeReference)]): BigDecimal = {
+ var ndvDenom: BigInt = -1
+ var i = 0
+ while(i < joinKeyPairs.length && ndvDenom != 0) {
+ val (leftKey, rightKey) = joinKeyPairs(i)
+ // Check if the two sides are disjoint
+ val leftKeyStats = leftStats.attributeStats(leftKey)
+ val rightKeyStats = rightStats.attributeStats(rightKey)
+ val lRange = Range(leftKeyStats.min, leftKeyStats.max,
leftKey.dataType)
+ val rRange = Range(rightKeyStats.min, rightKeyStats.max,
rightKey.dataType)
+ if (Range.isIntersected(lRange, rRange)) {
+ // Get the largest ndv among pairs of join keys
+ val maxNdv =
leftKeyStats.distinctCount.max(rightKeyStats.distinctCount)
+ if (maxNdv > ndvDenom) ndvDenom = maxNdv
+ } else {
+ // Set ndvDenom to zero to indicate that this join should have no
output
+ ndvDenom = 0
+ }
+ i += 1
+ }
+
+ if (ndvDenom < 0) {
+ // There isn't join keys or column stats for any of the join key
pairs, we do estimation like
+ // cartesian product.
+ 1
+ } else if (ndvDenom == 0) {
+ // One of the join key pairs is disjoint, thus the two sides of join
is disjoint.
+ 0
+ } else {
+ 1 / BigDecimal(ndvDenom)
+ }
+ }
+
+ /**
+ * Propagate or update column stats for output attributes.
+ * 1. For empty output, we don't need to keep any column stats.
+ * 2. For cartesian product, all values are preserved, so there's no
need to change column stats.
+ * 3. For other cases, a) update max/min of join keys based on their
intersected range. b) update
+ * distinct count of other attributes based on output rows after join.
+ */
+ private def updateAttrStats(
+ outputRows: BigInt,
+ attributes: Seq[Attribute],
+ oldAttrStats: AttributeMap[ColumnStat],
+ joinKeyStats: AttributeMap[ColumnStat]): Seq[(Attribute,
ColumnStat)] = {
+ val outputAttrStats = new ArrayBuffer[(Attribute, ColumnStat)]()
+ val leftRows = leftStats.rowCount.get
+ val rightRows = rightStats.rowCount.get
+ if (outputRows == leftRows * rightRows) {
+ // Cartesian product, just propagate the original column stats
+ attributes.foreach(a => outputAttrStats += a -> oldAttrStats(a))
+ } else if (outputRows != 0) {
+ val leftRatio =
+ if (leftRows != 0) BigDecimal(outputRows) / BigDecimal(leftRows)
else BigDecimal(0)
+ val rightRatio =
+ if (rightRows != 0) BigDecimal(outputRows) / BigDecimal(rightRows)
else BigDecimal(0)
+ attributes.foreach { a =>
+ // check if this attribute is a join key
+ if (joinKeyStats.contains(a)) {
+ outputAttrStats += a -> joinKeyStats(a)
+ } else {
+ val oldCS = oldAttrStats(a)
--- End diff --
nit: `oldColumnStats`
---
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]
