cshuo commented on code in PR #19307: URL: https://github.com/apache/hudi/pull/19307#discussion_r3629871917
########## hudi-flink-datasource/hudi-flink/src/main/java/org/apache/hudi/sink/utils/RecordKeySortKeyComputer.java: ########## @@ -0,0 +1,111 @@ +/* + * 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.hudi.sink.utils; + +import org.apache.hudi.sink.bulk.RowDataKeyGen; + +import org.apache.flink.api.common.typeutils.base.StringComparator; +import org.apache.flink.core.memory.MemorySegment; +import org.apache.flink.table.data.RowData; +import org.apache.flink.table.runtime.generated.NormalizedKeyComputer; + +/** + * Computes a normalized prefix for sorting rows by their encoded Hudi record key. + * + * <p>The common LSM reader compares encoded record keys with {@link String#compareTo(String)}, + * whose order is based on UTF-16 code units. This computer therefore uses Flink's normalized-key + * encoding for Java strings, which preserves that order while still using only one byte for the + * common ASCII case. + * + * <p>The normalized prefix is eight bytes for one record-key field and sixteen bytes for two or + * more fields. It is packed into one or two big-endian {@code long}s so the sort hot path can + * compare and swap the prefix with long operations instead of processing every byte separately. + * Short prefixes are padded with zeroes. Truncated prefixes, as well as collisions caused by zero + * padding, are resolved by {@link RecordKeySortComparator}; therefore the normalized key never + * fully determines the result. + */ +public class RecordKeySortKeyComputer implements NormalizedKeyComputer { Review Comment: The following **microbenchmark** compares the encoded Hoodie record-key computer/comparator with Flink code-generated sorting. Each scenario sorts 300,000 records using `BinaryInMemorySortBuffer`, with 3 warm-up rounds and 7 measured rounds; the table reports the median. <table> <thead> <tr> <th rowspan="2">Key Scenario</th> <th colspan="2">Write (ms)</th> <th colspan="2">Sort (ms)</th> <th colspan="2">Total (ms)</th> </tr> <tr> <th>Codegen</th> <th>Encoded</th> <th>Codegen</th> <th>Encoded</th> <th>Codegen</th> <th>Encoded</th> </tr> </thead> <tbody> <tr> <td>LONG</td> <td>24.2</td> <td>54.4</td> <td>198.7</td> <td>195.6</td> <td>222.9</td> <td>250.0</td> </tr> <tr> <td>Composite String, no fallback</td> <td>67.7</td> <td>167.2</td> <td>261.0</td> <td>159.5</td> <td>328.8</td> <td>326.7</td> </tr> <tr> <td>Composite String, 50% fallback</td> <td>35.1</td> <td>94.1</td> <td>574.8</td> <td>2,134.8</td> <td>609.9</td> <td>2,228.9</td> </tr> </tbody> </table> The encoded path has a higher write cost because it constructs the encoded Hoodie record key and computes its normalized prefix for every record. For LONG keys, its sort cost is comparable to codegen, resulting in an approximately 12% higher total cost. For composite strings (two keys) without normalized-key collisions, the encoded sort is about 39% faster than codegen because comparisons are resolved using the packed 16-byte prefix. This offsets the additional write cost, making the total costs effectively equal. Performance degrades when normalized-key collisions require the full comparator. With an observed fallback rate of approximately 49%, the encoded sort is about 3.7x slower and the total cost is about 3.65x higher. The main overhead is repeatedly constructing and comparing complete encoded Hoodie record keys in the `O(N log N)` sorting path. These numbers are from a focused microbenchmark and should be used to compare the relative cost of the sorting implementations, rather than as end-to-end Flink write-performance measurements. -- 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]
