hequn8128 commented on a change in pull request #9907: [FLINK-14202][table][python] Optimize the execution plan for Python Calc when there is a condition URL: https://github.com/apache/flink/pull/9907#discussion_r337309944
########## File path: flink-table/flink-table-planner-blink/src/main/scala/org/apache/flink/table/planner/plan/rules/logical/PythonCalcSplitRule.scala ########## @@ -0,0 +1,321 @@ +/* + * 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.flink.table.planner.plan.rules.logical + +import org.apache.calcite.plan.RelOptRule.{any, operand} +import org.apache.calcite.plan.{RelOptRule, RelOptRuleCall} +import org.apache.calcite.rex.{RexCall, RexInputRef, RexNode, RexProgram} +import org.apache.calcite.sql.validate.SqlValidatorUtil +import org.apache.flink.table.functions.FunctionLanguage +import org.apache.flink.table.functions.ScalarFunction +import org.apache.flink.table.planner.functions.utils.ScalarSqlFunction +import org.apache.flink.table.planner.plan.nodes.logical.FlinkLogicalCalc +import org.apache.flink.table.planner.plan.utils.PythonUtil.containsFunctionOf +import org.apache.flink.table.planner.plan.utils.{InputRefVisitor, RexDefaultVisitor} + +import scala.collection.JavaConverters._ +import scala.collection.JavaConversions._ +import scala.collection.mutable + +/** + * Rule that splits [[FlinkLogicalCalc]] into multiple [[FlinkLogicalCalc]]s. + * This is to ensure that the Python [[ScalarFunction]]s which could be + * executed in a batch are grouped into the same [[FlinkLogicalCalc]] node. + */ +abstract class PythonCalcSplitRuleBase(description: String) + extends RelOptRule( + operand(classOf[FlinkLogicalCalc], any), + description) { + + override def onMatch(call: RelOptRuleCall): Unit = { + val calc: FlinkLogicalCalc = call.rel(0).asInstanceOf[FlinkLogicalCalc] + val input = calc.getInput + val rexBuilder = call.builder().getRexBuilder + val program = calc.getProgram + val extractedRexCalls = new mutable.ArrayBuffer[RexCall]() + + val extractedFunctionOffset = input.getRowType.getFieldCount + val splitter = new ScalarFunctionSplitter( + extractedFunctionOffset, + extractedRexCalls, + convertPythonFunction(program)) + + val (bottomCalcCondition, topCalcCondition, topCalcProjects) = split(program, splitter) + val accessedFields = extractRefInputFields( + topCalcProjects, topCalcCondition, extractedFunctionOffset) + + val bottomCalcProjects = + accessedFields.map(RexInputRef.of(_, input.getRowType)) ++ extractedRexCalls + val bottomCalcFieldNames = SqlValidatorUtil.uniquify( + accessedFields.map(i => input.getRowType.getFieldNames.get(i)).toSeq ++ + extractedRexCalls.indices.map("f" + _), + rexBuilder.getTypeFactory.getTypeSystem.isSchemaCaseSensitive) + + val bottomCalc = new FlinkLogicalCalc( + calc.getCluster, + calc.getTraitSet, + input, + RexProgram.create( + input.getRowType, + bottomCalcProjects.toList, + bottomCalcCondition.orNull, + bottomCalcFieldNames, + rexBuilder)) + + val inputRewriter = new ExtractedFunctionInputRewriter(extractedFunctionOffset, accessedFields) + val topCalc = new FlinkLogicalCalc( + calc.getCluster, + calc.getTraitSet, + bottomCalc, + RexProgram.create( + bottomCalc.getRowType, + topCalcProjects.map(_.accept(inputRewriter)), + topCalcCondition.map(_.accept(inputRewriter)).orNull, + calc.getRowType, + rexBuilder)) + + call.transformTo(topCalc) + } + + /** + * Extracts the indices of the input fields referred by the specified projects and condition. + */ + private def extractRefInputFields( + projects: Seq[RexNode], + condition: Option[RexNode], + inputFieldsCount: Int): Array[Int] = { + val visitor = new InputRefVisitor + + // extract referenced input fields from projections + projects.foreach(exp => exp.accept(visitor)) + + // extract referenced input fields from condition + condition.foreach(_.accept(visitor)) + + // fields of indexes greater than inputFieldsCount is the extracted functions and + // should be filtered as they are not from the original input + visitor.getFields.filter(_ < inputFieldsCount) + } + + /** + * Whether to convert Python functions. + */ + def convertPythonFunction(program: RexProgram): Boolean + + /** + * Splits the specified [[RexProgram]] using the specified [[ScalarFunctionSplitter]]. + * It returns a triple of (bottom calc condition, top calc condition, top calc projects) + * as the split result. + */ + def split(program: RexProgram, splitter: ScalarFunctionSplitter) + : (Option[RexNode], Option[RexNode], Seq[RexNode]) +} + +/** + * Rule that splits [[FlinkLogicalCalc]]s which contain Python functions in the condition + * into multiple [[FlinkLogicalCalc]]s. After this rule is applied, there will be no + * Python functions in the condition of the [[FlinkLogicalCalc]]s. + */ +object PythonCalcSplitFilterRule extends PythonCalcSplitRuleBase( + "PythonCalcSplitFilterRule") { + + override def matches(call: RelOptRuleCall): Boolean = { + val calc: FlinkLogicalCalc = call.rel(0).asInstanceOf[FlinkLogicalCalc] + val condition = Option(calc.getProgram.getCondition).map(calc.getProgram.expandLocalRef) + + // matches if it contains Python functions in condition + condition.nonEmpty && containsFunctionOf(condition.get, FunctionLanguage.PYTHON) + } + + override def convertPythonFunction(program: RexProgram): Boolean = true + + override def split(program: RexProgram, splitter: ScalarFunctionSplitter) + : (Option[RexNode], Option[RexNode], Seq[RexNode]) = { + (None, Option(program.getCondition).map(program.expandLocalRef(_).accept(splitter)), + program.getProjectList.map(program.expandLocalRef)) + } +} + +/** + * Rule that splits [[FlinkLogicalCalc]]s which contain both Java functions and Python functions + * in the projection into multiple [[FlinkLogicalCalc]]s. After this rule is applied, it will + * only contain Python functions or Java functions in the projection of each [[FlinkLogicalCalc]]. + */ +object PythonCalcSplitProjectionRule extends PythonCalcSplitRuleBase( + "PythonCalcSplitProjectionRule") { + + override def matches(call: RelOptRuleCall): Boolean = { + val calc: FlinkLogicalCalc = call.rel(0).asInstanceOf[FlinkLogicalCalc] + val condition = Option(calc.getProgram.getCondition).map(calc.getProgram.expandLocalRef) + val projects = calc.getProgram.getProjectList.map(calc.getProgram.expandLocalRef) + + // matches if all the following conditions hold true: + // 1) it doesn't contain Python function in condition + // 2) it contains both Python functions and Java functions in the projection + !condition.exists(containsFunctionOf(_, FunctionLanguage.PYTHON)) && + projects.exists(containsFunctionOf(_, FunctionLanguage.PYTHON)) && + projects.exists(containsFunctionOf(_, FunctionLanguage.JVM)) + } + + override def convertPythonFunction(program: RexProgram): Boolean = { + program.getProjectList + .map(program.expandLocalRef) + .exists(containsFunctionOf(_, FunctionLanguage.JVM, recursive = false)) + } + + override def split(program: RexProgram, splitter: ScalarFunctionSplitter) + : (Option[RexNode], Option[RexNode], Seq[RexNode]) = { + (Option(program.getCondition).map(program.expandLocalRef), None, + program.getProjectList.map(program.expandLocalRef(_).accept(splitter))) + } +} + +/** + * Rule that pushes the condition of [[FlinkLogicalCalc]]s before it for the + * [[FlinkLogicalCalc]]s which contain Python functions in the projection. + */ +object PythonCalcPushFilterRule extends PythonCalcSplitRuleBase( + "PythonCalcPushFilterRule") { + + override def matches(call: RelOptRuleCall): Boolean = { + val calc: FlinkLogicalCalc = call.rel(0).asInstanceOf[FlinkLogicalCalc] + val projects = calc.getProgram.getProjectList.map(calc.getProgram.expandLocalRef) + + // matches if all the following conditions hold true: + // 1) the condition is not null + // 2) it contains Python functions in the projection + calc.getProgram.getCondition != null && + projects.exists(containsFunctionOf(_, FunctionLanguage.PYTHON)) + } + + override def convertPythonFunction(program: RexProgram): Boolean = false + + override def split(program: RexProgram, splitter: ScalarFunctionSplitter) + : (Option[RexNode], Option[RexNode], Seq[RexNode]) = { + (Option(program.getCondition).map(program.expandLocalRef), None, + program.getProjectList.map(program.expandLocalRef)) + } +} + +/** + * Rule that ensures that it only contains [[RexInputRef]]s at the beginning of + * the project list and [[RexCall]]s at the end of the project list for [[FlinkLogicalCalc]]s + * which contain Python functions in the projection. This rule exists to keep + * DataStreamPythonCalc as simple as possible and ensures that it only needs to + * handle the Python function execution. + */ +object PythonCalcRewriteProjectRule extends PythonCalcSplitRuleBase( + "PythonCalcRewriteProjectRule") { + + override def matches(call: RelOptRuleCall): Boolean = { + val calc: FlinkLogicalCalc = call.rel(0).asInstanceOf[FlinkLogicalCalc] + val projects = calc.getProgram.getProjectList.map(calc.getProgram.expandLocalRef) + + // matches if all the following conditions hold true: + // 1) it contains Python functions in the projection + // 2) it contains RexNodes besides RexInputRef and RexCall or + // not all the RexCalls lying at the end of the project list + projects.exists(containsFunctionOf(_, FunctionLanguage.PYTHON)) && + (projects.exists(expr => !expr.isInstanceOf[RexCall] && !expr.isInstanceOf[RexInputRef]) || + projects.indexWhere(_.isInstanceOf[RexCall]) < + projects.lastIndexWhere(_.isInstanceOf[RexInputRef])) + } + + override def convertPythonFunction(program: RexProgram): Boolean = true + + override def split(program: RexProgram, splitter: ScalarFunctionSplitter) + : (Option[RexNode], Option[RexNode], Seq[RexNode]) = { + (None, None, program.getProjectList.map(program.expandLocalRef(_).accept(splitter))) + } +} + +private class ScalarFunctionSplitter( + extractedFunctionOffset: Int, + extractedRexCalls: mutable.ArrayBuffer[RexCall], + convertPythonFunction: Boolean) + extends RexDefaultVisitor[RexNode] { + + override def visitCall(call: RexCall): RexNode = { + call.getOperator match { + case sfc: ScalarSqlFunction if sfc.scalarFunction.getLanguage == + FunctionLanguage.PYTHON => + visit(convertPythonFunction, call) + + case _ => + visit(!convertPythonFunction, call) + } + } + + override def visitNode(rexNode: RexNode): RexNode = rexNode + + private def visit(needConvert: Boolean, call: RexCall): RexNode = { + if (needConvert) { + val newNode = new RexInputRef( + extractedFunctionOffset + extractedRexCalls.length, call.getType) + extractedRexCalls.append(call) + newNode + } else { + call.clone(call.getType, call.getOperands.asScala.map(_.accept(this))) + } + } +} + +/** + * Rewrite field accesses of a RexNode as not all the fields from the original input are forwarded: + * 1) Fields of index greater than or equal to extractedFunctionOffset refer to the + * extracted function. + * 2) Fields of index less than extractedFunctionOffset refer to the original input field. + * + * @param extractedFunctionOffset the original start offset of the extracted functions + * @param accessedFields the accessed fields which will be forwarded + */ +private class ExtractedFunctionInputRewriter( + extractedFunctionOffset: Int, + accessedFields: Array[Int]) + extends RexDefaultVisitor[RexNode] { + + /** old input fields ref index -> new input fields ref index mappings */ + private val fieldMap: Map[Int, Int] = accessedFields.zipWithIndex.toMap + + override def visitInputRef(inputRef: RexInputRef): RexNode = { + if (inputRef.getIndex >= extractedFunctionOffset) { + new RexInputRef( + inputRef.getIndex - extractedFunctionOffset + accessedFields.length, + inputRef.getType) + } else { + new RexInputRef( + fieldMap.getOrElse(inputRef.getIndex, + throw new IllegalArgumentException("input field contains invalid index")), + inputRef.getType) + } + } + + override def visitCall(call: RexCall): RexNode = { + call.clone(call.getType, call.getOperands.asScala.map(_.accept(this))) + } + + override def visitNode(rexNode: RexNode): RexNode = rexNode +} + +object PythonCalcSplitRule { + val SPLIT_FILTER: RelOptRule = PythonCalcSplitFilterRule + val SPLIT_PROJECT: RelOptRule = PythonCalcSplitProjectionRule + val PUSH_FILTER: RelOptRule = PythonCalcPushFilterRule + val REWRITE_PROJECT: RelOptRule = PythonCalcRewriteProjectRule Review comment: Ad comments that these rules should be called sequentially. ---------------------------------------------------------------- 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. For queries about this service, please contact Infrastructure at: [email protected] With regards, Apache Git Services
