yunfengzhou-hub commented on code in PR #11015: URL: https://github.com/apache/paimon/pull/11015#discussion_r4501141135
########## paimon-flink/paimon-flink-common/src/main/java/org/apache/paimon/flink/sink/StripedBucketWriteChannelComputer.java: ########## @@ -0,0 +1,104 @@ +/* + * 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.paimon.flink.sink; + +import org.apache.paimon.CoreOptions; +import org.apache.paimon.data.BinaryRow; +import org.apache.paimon.data.InternalRow; +import org.apache.paimon.schema.TableSchema; +import org.apache.paimon.table.sink.ChannelComputer; +import org.apache.paimon.utils.MurmurHashUtils; + +/** + * A {@link RowDataChannelComputer} that stripes each bucket's writes across multiple writer + * channels, so that write parallelism can exceed the bucket count without rescaling the table. + * + * <p>Each bucket is routed to several channels. For a primary-key table the record's primary key + * picks the channel, so records sharing a key still land on the same channel; for an append-only + * table records are distributed round-robin. A single writer per bucket is equivalent to {@link + * RowDataChannelComputer}. + */ +public class StripedBucketWriteChannelComputer extends RowDataChannelComputer { + + private static final long serialVersionUID = 1L; + + /** Derive the number of stripes from the channel count and the bucket count. */ + public static final int AUTO = 0; + + /** Route every bucket to all channels, so every writer writes every bucket. */ + public static final int MATCH_PARALLELISM = -1; + + private final int writesPerBucket; + private final int numBuckets; + private final boolean hasPrimaryKey; + + private transient int numStripes; + private transient int nextStripe; + + public StripedBucketWriteChannelComputer(TableSchema schema, int writesPerBucket) { + super(schema); + this.writesPerBucket = writesPerBucket; + this.numBuckets = new CoreOptions(schema.options()).bucket(); + this.hasPrimaryKey = !schema.primaryKeys().isEmpty(); + } + + @Override + public void setup(int numChannels) { + super.setup(numChannels); + this.numStripes = resolveNumStripes(); + } + + private int resolveNumStripes() { + if (writesPerBucket > 0) { + return writesPerBucket; + } + if (writesPerBucket == MATCH_PARALLELISM) { + return numChannels; + } + // AUTO: ceil(numChannels / numBuckets), but never less than one writer per bucket + return Math.max(1, (numChannels + numBuckets - 1) / numBuckets); + } + + @Override + public int channel(InternalRow record) { + extractor.setRecord(record); + BinaryRow partition = extractor.partition(); + int bucket = extractor.bucket(); + int stripe = + hasPrimaryKey + ? keyStripe(extractor.trimmedPrimaryKey()) + : Math.floorMod(nextStripe++, numStripes); + // Widen to long so that bucket * numStripes cannot overflow for large bucket counts + long bucketSlot = (long) bucket * numStripes + stripe; + int startChannel = ChannelComputer.startChannel(partition, numChannels); + return (int) ((startChannel + bucketSlot) % numChannels); + } + + private int keyStripe(BinaryRow key) { + // Avalanche the key hash so the stripe is independent of the bucket modulo, which also + // derives from the key hash; otherwise shared factors between numStripes and numBuckets + // skew it. + return Math.floorMod(MurmurHashUtils.fmix(key.hashCode()), numStripes); Review Comment: keyStripe now canonicalizes NaN payloads on a copy of the key before hashing, matching the key comparator's equality (Double.compare/Float.compare treat all NaNs as equal); the written record is unchanged. Keys without FLOAT/DOUBLE fields keep the raw-hash fast path. testEqualNaNKeysRouteToSameChannel routes two equal quiet-NaN-payload keys and asserts the same channel across stripe counts 2-5; it fails on the raw-bit hash and passes after canonicalization. This normalizes the stripe sub-hash only; the upstream bucket assignment still hashes raw bytes (DefaultBucketFunction), which is NaN-free in this repro since bucket-key=tenant. -- 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]
