raoraoxiong commented on code in PR #28998: URL: https://github.com/apache/flink/pull/28998#discussion_r3920895517
########## flink-table/flink-table-planner/src/main/java/org/apache/flink/table/planner/plan/nodes/exec/common/PythonCallDeduplicator.java: ########## @@ -0,0 +1,128 @@ +/* + * 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.nodes.exec.common; + +import org.apache.flink.annotation.Internal; +import org.apache.flink.annotation.VisibleForTesting; +import org.apache.flink.table.planner.plan.utils.PythonUtil; +import org.apache.flink.table.planner.utils.ShortcutUtils; + +import org.apache.calcite.rex.RexCall; +import org.apache.calcite.rex.RexNode; + +import java.util.ArrayList; +import java.util.LinkedHashMap; +import java.util.List; + +/** + * Utility for Python UDF Common Sub-expression Elimination (CSE) of nested calls. + * + * <p>Nested Python UDF call trees are flattened in post-order and deduplicated by structural + * equivalence, so that a sub-expression shared between calls is evaluated only once by the Python + * worker. Duplicates between whole projection entries are already removed in the planner by {@code + * RemoteCalcProjectionCseRule}, so this only concerns sub-expressions. + */ +@Internal +public class PythonCallDeduplicator { + + /** + * Recursively collects all deterministic Python UDF calls from a call tree in DFS post-order. + * + * <p>Post-order ensures child results are computed before parents that reference them via + * refIndex. Non-deterministic children are NOT flattened to prevent incorrect sharing. + */ + @VisibleForTesting + static List<RexCall> collectAllPythonUdfCalls(RexCall root) { + List<RexCall> result = new ArrayList<>(); + for (RexNode operand : root.getOperands()) { + if (operand instanceof RexCall && PythonUtil.isPythonCall((RexCall) operand)) { + RexCall childCall = (RexCall) operand; + // Only flatten deterministic child calls for CSE. + // Non-deterministic calls must remain nested to avoid incorrect sharing. + if (ShortcutUtils.isDeterministicThroughProgram(childCall, null)) { + result.addAll(collectAllPythonUdfCalls(childCall)); + } + } + } + result.add(root); + return result; + } + + /** + * Flattens the given Python UDF call trees and deduplicates the resulting calls. + * + * <p>Flattening all trees into a single list enables cross-subtree reuse: e.g. in {@code SELECT + * udf1(x), udf2(udf1(x))}, the inner {@code udf1(x)} is evaluated only once and {@code udf2} + * receives its result by reference. + */ + public static PythonCallCseResult deduplicate(List<RexCall> pythonRexCalls) { + // Flatten: collect all Python UDF calls from all projection trees in post-order, so a + // nested sub-expression is always evaluated before the call referencing it. The root of + // each tree is the last element of its own sub-list. + List<RexCall> allCalls = new ArrayList<>(); + int[] rootPositions = new int[pythonRexCalls.size()]; + for (int i = 0; i < pythonRexCalls.size(); i++) { + List<RexCall> subtreeCalls = collectAllPythonUdfCalls(pythonRexCalls.get(i)); + rootPositions[i] = allCalls.size() + subtreeCalls.size() - 1; + allCalls.addAll(subtreeCalls); + } + + // Deduplicate the flattened list by structural equivalence, preserving post-order. + LinkedHashMap<RexCall, Integer> callToIndex = new LinkedHashMap<>(); + List<RexCall> uniqueCalls = new ArrayList<>(); + int[] allToUnique = new int[allCalls.size()]; + for (int i = 0; i < allCalls.size(); i++) { + RexCall call = allCalls.get(i); + boolean canReuse = ShortcutUtils.isDeterministicThroughProgram(call, null); + Integer existing = canReuse ? callToIndex.get(call) : null; + if (existing != null) { + allToUnique[i] = existing; + continue; + } + int newPos = uniqueCalls.size(); + if (canReuse) { + callToIndex.put(call, newPos); + } + uniqueCalls.add(call); + allToUnique[i] = newPos; + } + + // Flattening adds entries for nested sub-expressions, and post-order means a top-level + // result is not necessarily last, so record where each projection entry ended up. Only + // those positions form the operator output. + int[] outputIndices = new int[pythonRexCalls.size()]; + for (int i = 0; i < pythonRexCalls.size(); i++) { + outputIndices[i] = allToUnique[rootPositions[i]]; + } + + // Build refMap for sub-expression cross-referencing. putIfAbsent preserves the first + // occurrence index, ensuring a parent references its own child rather than a later + // structurally-equal duplicate. + LinkedHashMap<RexCall, Integer> refMap = new LinkedHashMap<>(); + for (int i = 0; i < uniqueCalls.size(); i++) { Review Comment: **`refMap` is now derived inside `PythonCallCseResult`** — good catch, it was redundant state. Since the list is already deduplicated, the first occurrence of a structurally equal call is the one computing it, so the constructor only takes the deduplicated calls and the output indices now. -- 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. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
