yunfengzhou-hub commented on code in PR #11015: URL: https://github.com/apache/paimon/pull/11015#discussion_r4537985557
########## paimon-flink/paimon-flink-common/src/main/java/org/apache/paimon/flink/sink/StripedBucketWriteChannelComputer.java: ########## @@ -0,0 +1,175 @@ +/* + * 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.types.DataTypeRoot; +import org.apache.paimon.types.RowType; +import org.apache.paimon.utils.MurmurHashUtils; + +import java.util.ArrayList; +import java.util.HashMap; +import java.util.List; +import java.util.Map; + +/** + * 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; + + // Positions of FLOAT/DOUBLE fields in the trimmed primary key, whose NaN payloads must be + // canonicalized before hashing so that keys equal under the key comparator hash alike. + private final int[] floatingKeyPositions; + private final boolean[] floatingKeyIsDouble; + + private transient int numStripes; + // Round-robin progress per (partition, bucket). A single shared counter would correlate with + // the bucket arrival pattern (e.g. alternating buckets always pick the same stripe), leaving + // each bucket on one writer even though more are available. + private transient Map<BinaryRow, int[]> stripeCounters; + + public StripedBucketWriteChannelComputer(TableSchema schema, int writesPerBucket) { + super(schema); + this.writesPerBucket = writesPerBucket; + this.numBuckets = new CoreOptions(schema.options()).bucket(); + this.hasPrimaryKey = !schema.primaryKeys().isEmpty(); + + List<Integer> positions = new ArrayList<>(); + List<Boolean> isDouble = new ArrayList<>(); + if (hasPrimaryKey) { + RowType keyType = schema.logicalTrimmedPrimaryKeysType(); + for (int i = 0; i < keyType.getFieldCount(); i++) { + DataTypeRoot root = keyType.getTypeAt(i).getTypeRoot(); + if (root == DataTypeRoot.DOUBLE) { + positions.add(i); + isDouble.add(true); + } else if (root == DataTypeRoot.FLOAT) { + positions.add(i); + isDouble.add(false); + } + } + } + this.floatingKeyPositions = new int[positions.size()]; + this.floatingKeyIsDouble = new boolean[positions.size()]; + for (int i = 0; i < positions.size(); i++) { + this.floatingKeyPositions[i] = positions.get(i); + this.floatingKeyIsDouble[i] = isDouble.get(i); + } + } + + @Override + public void setup(int numChannels) { + super.setup(numChannels); + this.numStripes = resolveNumStripes(); + this.stripeCounters = new HashMap<>(); + } + + 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; + if (hasPrimaryKey) { + stripe = keyStripe(extractor.trimmedPrimaryKey()); + } else { + int[] counters = stripeCounters.get(partition); + if (counters == null) { + counters = new int[numBuckets]; Review Comment: Fixed. The key change is replacing the per-(partition, bucket) round-robin counters with a stateless hash mod: the stripe is now the bucket key hashed mod the stripe count, so the router keeps no state at all rather than counters that still grow with the observed pairs. A single stripe short-circuits to stripe 0, and records sharing a bucket key still land on the same writer. testAppendHashesRecordStatelessly pins the stateless behavior and fails with a round-robin counter; testAppendBucketKeyHashFanOut and testPartitionedAppendFanOut cover fan-out, including when the stripe count divides the bucket count. -- 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]
