Github user rxin commented on a diff in the pull request: https://github.com/apache/spark/pull/16395#discussion_r95524281 --- Diff: sql/catalyst/src/main/scala/org/apache/spark/sql/catalyst/plans/logical/statsEstimation/FilterEstimation.scala --- @@ -0,0 +1,555 @@ +/* + * 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.immutable.{HashSet, Map} +import scala.collection.mutable + +import org.apache.spark.internal.Logging +import org.apache.spark.sql.catalyst.expressions._ +import org.apache.spark.sql.catalyst.plans.logical._ +import org.apache.spark.sql.catalyst.util.DateTimeUtils +import org.apache.spark.sql.types._ +import org.apache.spark.unsafe.types.UTF8String + + +class FilterEstimation extends Logging { + + /** + * We use a mutable colStats because we need to update the corresponding ColumnStat + * for a column after we apply a predicate condition. For example, A column c has + * [min, max] value as [0, 100]. In a range condition such as (c > 40 AND c <= 50), + * we need to set the column's [min, max] value to [40, 100] after we evaluate the + * first condition c > 40. We need to set the column's [min, max] value to [40, 50] + * after we evaluate the second condition c <= 50. + */ + private var mutableColStats: mutable.Map[ExprId, ColumnStat] = mutable.Map.empty + + /** + * Returns an option of Statistics for a Filter logical plan node. + * For a given compound expression condition, this method computes filter selectivity + * (or the percentage of rows meeting the filter condition), which + * is used to compute row count, size in bytes, and the updated statistics after a given + * predicated is applied. + * + * @param plan a LogicalPlan node that must be an instance of Filter. + * @return Option[Statistics] When there is no statistics collected, it returns None. + */ + def estimate(plan: Filter): Option[Statistics] = { + val stats: Statistics = plan.child.statistics + if (stats.rowCount.isEmpty) return None + + // save a mutable copy of colStats so that we can later change it recursively + val statsExprIdMap: Map[ExprId, ColumnStat] = + stats.attributeStats.map(kv => (kv._1.exprId, kv._2)) + mutableColStats = mutable.Map.empty ++= statsExprIdMap + + // estimate selectivity of this filter predicate + val percent: Double = calculateConditions(plan, plan.condition) + + // attributeStats has mapping Attribute-to-ColumnStat. + // mutableColStats has mapping ExprId-to-ColumnStat. + // We use an ExprId-to-Attribute map to facilitate the mapping Attribute-to-ColumnStat + val expridToAttrMap: Map[ExprId, Attribute] = + stats.attributeStats.map(kv => (kv._1.exprId, kv._1)) + // copy mutableColStats contents to an immutable AttributeMap. + val mutableAttributeStats: mutable.Map[Attribute, ColumnStat] = + mutableColStats.map(kv => expridToAttrMap(kv._1) -> kv._2) + val newColStats = AttributeMap(mutableAttributeStats.toSeq) + + val filteredRowCountValue: BigInt = + EstimationUtils.ceil(BigDecimal(stats.rowCount.get) * percent) + val avgRowSize = BigDecimal(EstimationUtils.getRowSize(plan.output, newColStats)) + val filteredSizeInBytes: BigInt = + EstimationUtils.ceil(BigDecimal(filteredRowCountValue) * avgRowSize) + + Some(stats.copy(sizeInBytes = filteredSizeInBytes, rowCount = Some(filteredRowCountValue), + attributeStats = newColStats)) + } + + /** + * Returns a percentage of rows meeting a compound condition in Filter node. + * A compound condition is depomposed into multiple single conditions linked with AND, OR, NOT. + * For logical AND conditions, we need to update stats after a condition estimation + * so that the stats will be more accurate for subsequent estimation. This is needed for + * range condition such as (c > 40 AND c <= 50) + * For logical OR conditions, we do not update stats after a condition estimation. + * + * @param plan the Filter LogicalPlan node + * @param condition the compound logical expression + * @param update a boolean flag to specify if we need to update ColumnStat of a column + * for subsequent conditions + * @return a doube value to show the percentage of rows meeting a given condition + */ + def calculateConditions( + plan: Filter, + condition: Expression, + update: Boolean = true) + : Double = { + + condition match { + case And(cond1, cond2) => + val p1 = calculateConditions(plan, cond1, update) + val p2 = calculateConditions(plan, cond2, update) + p1 * p2 + + case Or(cond1, cond2) => + val p1 = calculateConditions(plan, cond1, update = false) + val p2 = calculateConditions(plan, cond2, update = false) + math.min(1.0, p1 + p2 - (p1 * p2)) + + case Not(cond) => calculateSingleCondition(plan, cond, isNot = true, update = false) + case _ => calculateSingleCondition(plan, condition, isNot = false, update) + } + } + + /** + * Returns a percentage of rows meeting a single condition in Filter node. + * Currently we only support binary predicates where one side is a column, + * and the other is a literal. + * + * @param plan the Filter LogicalPlan node + * @param condition a single logical expression + * @param isNot set to true for "IS NULL" condition. set to false for "IS NOT NULL" condition + * @param update a boolean flag to specify if we need to update ColumnStat of a column + * for subsequent conditions + * @return a doube value to show the percentage of rows meeting a given condition + */ + def calculateSingleCondition( + plan: Filter, + condition: Expression, + isNot: Boolean, + update: Boolean) + : Double = { + var notSupported: Boolean = false + val percent: Double = condition match { + // For evaluateBinary method, we assume the literal on the right side of an operator. + // So we will change the order if not. + + // EqualTo does not care about the order + case op@EqualTo(ExtractAttr(ar), l: Literal) => + evaluateBinary(op, ar, l, update) + case op@EqualTo(l: Literal, ExtractAttr(ar)) => + evaluateBinary(op, ar, l, update) + + case op@LessThan(ExtractAttr(ar), l: Literal) => + evaluateBinary(op, ar, l, update) + case op@LessThan(l: Literal, ExtractAttr(ar)) => + evaluateBinary(GreaterThan(ar, l), ar, l, update) + + case op@LessThanOrEqual(ExtractAttr(ar), l: Literal) => + evaluateBinary(op, ar, l, update) + case op@LessThanOrEqual(l: Literal, ExtractAttr(ar)) => + evaluateBinary(GreaterThanOrEqual(ar, l), ar, l, update) + + case op@GreaterThan(ExtractAttr(ar), l: Literal) => + evaluateBinary(op, ar, l, update) + case op@GreaterThan(l: Literal, ExtractAttr(ar)) => + evaluateBinary(LessThan(ar, l), ar, l, update) + + case op@GreaterThanOrEqual(ExtractAttr(ar), l: Literal) => + evaluateBinary(op, ar, l, update) + case op@GreaterThanOrEqual(l: Literal, ExtractAttr(ar)) => + evaluateBinary(LessThanOrEqual(ar, l), ar, l, update) + + case In(ExtractAttr(ar), expList) if !expList.exists(!_.isInstanceOf[Literal]) => + // Expression [In (value, seq[Literal])] will be replaced with optimized version + // [InSet (value, HashSet[Literal])] in Optimizer, but only for list.size > 10. + // Here we convert In into InSet anyway, because they share the same processing logic. + val hSet = expList.map(e => e.eval()) + evaluateInSet(ar, HashSet() ++ hSet, update) + + case InSet(ExtractAttr(ar), set) => + evaluateInSet(ar, set, update) + + // It's difficult to estimate IsNull after outer joins. Hence, + // we support IsNull and IsNotNull only when the child is a leaf node (table). + case IsNull(ExtractAttr(ar)) => + if (plan.child.isInstanceOf[LeafNode ]) { + evaluateIsNull(plan, ar, true, update) + } + else 1.0 + + case IsNotNull(ExtractAttr(ar)) => + if (plan.child.isInstanceOf[LeafNode ]) { + evaluateIsNull(plan, ar, false, update) + } + else 1.0 + + case _ => + // TODO: it's difficult to support string operators without advanced statistics. + // Hence, these string operators Like(_, _) | Contains(_, _) | StartsWith(_, _) + // | EndsWith(_, _) are not supported yet + logDebug("[CBO] Unsupported filter condition: " + condition) + notSupported = true + 1.0 + } + if (notSupported) { + 1.0 + } else if (isNot) { + 1.0 - percent + } else { + percent + } + } + + /** + * Returns a percentage of rows meeting "IS NULL" or "IS NOT NULL" condition. + * + * @param plan the Filter LogicalPlan node + * @param attrRef an AttributeReference (or a column) + * @param isNull set to true for "IS NULL" condition. set to false for "IS NOT NULL" condition + * @param update a boolean flag to specify if we need to update ColumnStat of a given column + * for subsequent conditions + * @return a doube value to show the percentage of rows meeting a given condition + */ + def evaluateIsNull( + plan: Filter, + attrRef: AttributeReference, + isNull: Boolean, + update: Boolean) + : Double = { + if (!mutableColStats.contains(attrRef.exprId)) { + logDebug("[CBO] No statistics for " + attrRef) + return 1.0 + } + val aColStat = mutableColStats(attrRef.exprId) + val rowCountValue = plan.child.statistics.rowCount.get + val nullPercent: BigDecimal = + if (rowCountValue == 0) 0.0 + else BigDecimal(aColStat.nullCount)/BigDecimal(rowCountValue) + + if (update) { + val newStats = + if (isNull) aColStat.copy(distinctCount = 0, min = None, max = None) + else aColStat.copy(nullCount = 0) + + mutableColStats += (attrRef.exprId -> newStats) + } + + val percent = + if (isNull) nullPercent.toDouble + else { + /** ISNOTNULL(column) */ + 1.0 - nullPercent.toDouble + } + + percent + } + + /** + * Returns a percentage of rows meeting a binary comparison expression. + * + * @param op a binary comparison operator uch as =, <, <=, >, >= + * @param attrRef an AttributeReference (or a column) + * @param literal a literal value (or constant) + * @param update a boolean flag to specify if we need to update ColumnStat of a given column + * for subsequent conditions + * @return a doube value to show the percentage of rows meeting a given condition + */ + def evaluateBinary( + op: BinaryComparison, + attrRef: AttributeReference, + literal: Literal, + update: Boolean) + : Double = { + if (!mutableColStats.contains(attrRef.exprId)) { + logDebug("[CBO] No statistics for " + attrRef) + return 1.0 + } + + // Make sure that the Date/Timestamp literal is a valid one + attrRef.dataType match { + case DateType => + val dateLiteral = DateTimeUtils.stringToDate(literal.value.asInstanceOf[UTF8String]) + if (dateLiteral.isEmpty) { + logDebug("[CBO] Date literal is wrong, No statistics for " + attrRef) + return 1.0 + } + case TimestampType => + val tsLiteral = DateTimeUtils.stringToTimestamp(literal.value.asInstanceOf[UTF8String]) + if (tsLiteral.isEmpty) { + logDebug("[CBO] Timestamp literal is wrong, No statistics for " + attrRef) + return 1.0 + } + case _ => + } + + op match { + case EqualTo(l, r) => evaluateEqualTo(attrRef, literal, update) + case _ => + attrRef.dataType match { + case _: NumericType | DateType | TimestampType => + evaluateBinaryForNumeric(op, attrRef, literal, update) + case StringType | BinaryType => + + // TODO: It is difficult to support other binary comparisons for String/Binary + // type without min/max and advanced statistics like histogram. + + logDebug("[CBO] No statistics for String/Binary type " + attrRef) + return 1.0 + } + } + } + + /** + * This method converts a numeric or Literal value of numeric type to a BigDecimal value. + * If isNumeric is true, then it is a numeric value. Otherwise, it is a Literal value. --- End diff -- sorry i don't understand what you are referring to on "numeric" vs "literal".
--- 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 infrastruct...@apache.org or file a JIRA ticket with INFRA. --- --------------------------------------------------------------------- To unsubscribe, e-mail: reviews-unsubscr...@spark.apache.org For additional commands, e-mail: reviews-h...@spark.apache.org