pvary commented on code in PR #6382: URL: https://github.com/apache/iceberg/pull/6382#discussion_r1054537560
########## flink/v1.16/flink/src/main/java/org/apache/iceberg/flink/sink/shuffle/ShuffleOperator.java: ########## @@ -0,0 +1,140 @@ +/* + * 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.iceberg.flink.sink.shuffle; + +import java.io.Serializable; +import java.util.Map; +import org.apache.flink.annotation.Internal; +import org.apache.flink.api.common.state.ListState; +import org.apache.flink.api.common.state.ListStateDescriptor; +import org.apache.flink.api.common.typeinfo.TypeInformation; +import org.apache.flink.api.java.functions.KeySelector; +import org.apache.flink.api.java.typeutils.MapTypeInfo; +import org.apache.flink.runtime.operators.coordination.OperatorEvent; +import org.apache.flink.runtime.operators.coordination.OperatorEventGateway; +import org.apache.flink.runtime.operators.coordination.OperatorEventHandler; +import org.apache.flink.runtime.state.StateInitializationContext; +import org.apache.flink.runtime.state.StateSnapshotContext; +import org.apache.flink.streaming.api.operators.AbstractStreamOperator; +import org.apache.flink.streaming.api.operators.OneInputStreamOperator; +import org.apache.flink.streaming.runtime.streamrecord.StreamRecord; +import org.apache.iceberg.relocated.com.google.common.annotations.VisibleForTesting; +import org.apache.iceberg.relocated.com.google.common.collect.Maps; + +/** + * Shuffle operator can help to improve data clustering based on the key. + * + * <p>It collects the data statistics information, sends to coordinator and gets the global data + * distribution weight from coordinator. Then it will ingest the weight into data stream(wrap by a + * class{@link ShuffleRecordWrapper}) and send to partitioner. + */ +@Internal +public class ShuffleOperator<T, K extends Serializable> + extends AbstractStreamOperator<ShuffleRecordWrapper<T, K>> + implements OneInputStreamOperator<T, ShuffleRecordWrapper<T, K>>, OperatorEventHandler { + + private static final long serialVersionUID = 1L; + + private final KeySelector<T, K> keySelector; + // the type of the key to collect data statistics + private final TypeInformation<K> keyType; + private final OperatorEventGateway operatorEventGateway; + // key is generated by applying KeySelector to record + // value is the times key occurs + private transient Map<K, Long> localDataStatisticsMap; Review Comment: I think the data statistics interface will be very straightforward. I would expect something like this: ``` public interface DataStatistic<K> { long size(); void put(K key); void merge(DataStatistic<K> other); } ``` I do not think that the ShuffleOperator needs to know anything about the statistics and the partitioners used. My feeling is that the ShuffleOperator is just responsible to: - Collect statistics - Distribute the statistics to the Partitioners The `Partitioners` are responsible to decide on the distribution based on the collected `DataStatistics`. Here, I am not sure we can separate the `Partitioner` from the actual strategy (Bin Pack, or Range) used, since fetching a distribution based on a key (Iceberg partition based - Map), or fetching a distribution based on a range list (Iceberg sorted tables - TreeSet) has different performance on the hot code path, so we could not fall back to the more general solution (TreeSet) for performance reasons. I still feel that having an interface defined for `DataStatistics` would be a good direction even at this stage. We can implement the `MapStatistics` easily like this: ``` public class MapStatistic<K> implements DataStatistic<K> { private Map<K, Long> data = Maps.newHashMap(); @Override public long size() { return data.size(); } @Override public void put(K key) { data.put(key, data.getOrDefault(key, 0L) + 1); } @Override public void merge(DataStatistic<K> other) { Preconditions.checkArgument(other instanceof MapStatistic, "Can not merge this type of statistics: " + other); MapStatistic mapStatistic = (MapStatistic<K>) other; Map<K, Long> otherData = mapStatistic.data; otherData.forEach((key, count) -> data.put(key, data.getOrDefault(key, 0L) + count)); } } ``` WTYT? -- 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: issues-unsubscr...@iceberg.apache.org For queries about this service, please contact Infrastructure at: us...@infra.apache.org --------------------------------------------------------------------- To unsubscribe, e-mail: issues-unsubscr...@iceberg.apache.org For additional commands, e-mail: issues-h...@iceberg.apache.org
