rovboyko commented on code in PR #26313: URL: https://github.com/apache/flink/pull/26313#discussion_r2106737750
########## flink-table/flink-table-runtime/src/main/java/org/apache/flink/table/runtime/operators/join/stream/state/MultiJoinStateViews.java: ########## @@ -0,0 +1,359 @@ +/* + * 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.runtime.operators.join.stream.state; + +import org.apache.flink.api.common.functions.RuntimeContext; +import org.apache.flink.api.common.state.MapState; +import org.apache.flink.api.common.state.MapStateDescriptor; +import org.apache.flink.api.common.state.StateTtlConfig; +import org.apache.flink.api.common.typeinfo.TypeInformation; +import org.apache.flink.api.common.typeinfo.Types; +import org.apache.flink.api.java.functions.KeySelector; +import org.apache.flink.table.data.RowData; +import org.apache.flink.table.runtime.operators.join.stream.utils.JoinInputSideSpec; +import org.apache.flink.table.runtime.typeutils.InternalTypeInfo; +import org.apache.flink.types.RowKind; +import org.apache.flink.util.IterableIterator; + +import java.util.ArrayList; +import java.util.Collections; +import java.util.HashMap; +import java.util.Iterator; +import java.util.List; +import java.util.Map; +import java.util.NoSuchElementException; + +import static org.apache.flink.table.runtime.util.StateConfigUtil.createTtlConfig; +import static org.apache.flink.util.Preconditions.checkNotNull; + +/** + * Factory class to create different implementations of {@link MultiJoinStateView} based on the + * characteristics described in {@link JoinInputSideSpec}. + * + * <p>Each state view uses a {@link MapState} where the primary key is the `mapKey` derived from the + * join conditions (via {@link + * org.apache.flink.table.runtime.operators.join.stream.keyselector.JoinKeyExtractor}). The value + * stored within this map depends on whether the input side has a unique key and how it relates to + * the join key, optimizing storage and access patterns. + */ +public final class MultiJoinStateViews { + + /** Creates a {@link MultiJoinStateView} depends on {@link JoinInputSideSpec}. */ + public static MultiJoinStateView create( + RuntimeContext ctx, + String stateName, + JoinInputSideSpec inputSideSpec, + InternalTypeInfo<RowData> mapKeyType, // Type info for the outer map key + InternalTypeInfo<RowData> recordType, + long retentionTime) { + StateTtlConfig ttlConfig = createTtlConfig(retentionTime); + + if (inputSideSpec.hasUniqueKey()) { + if (inputSideSpec.joinKeyContainsUniqueKey()) { + return new JoinKeyContainsUniqueKey( + ctx, stateName, mapKeyType, recordType, ttlConfig); + } else { + return new InputSideHasUniqueKey( + ctx, + stateName, + mapKeyType, + recordType, + inputSideSpec.getUniqueKeyType(), + inputSideSpec.getUniqueKeySelector(), + ttlConfig); + } + } else { + return new InputSideHasNoUniqueKey(ctx, stateName, mapKeyType, recordType, ttlConfig); + } + } + + /** + * Creates a {@link MapStateDescriptor} with the given parameters and applies TTL configuration. + * + * @param <K> Key type + * @param <V> Value type + * @param stateName Unique name for the state + * @param keyTypeInfo Type information for the key + * @param valueTypeInfo Type information for the value + * @param ttlConfig State TTL configuration + * @return Configured MapStateDescriptor + */ + private static <K, V> MapStateDescriptor<K, V> createStateDescriptor( + String stateName, + TypeInformation<K> keyTypeInfo, + TypeInformation<V> valueTypeInfo, + StateTtlConfig ttlConfig) { + MapStateDescriptor<K, V> descriptor = + new MapStateDescriptor<>(stateName, keyTypeInfo, valueTypeInfo); + if (ttlConfig.isEnabled()) { + descriptor.enableTimeToLive(ttlConfig); + } + return descriptor; + } + + // ------------------------------------------------------------------------------------ + // Multi Join State View Implementations + // ------------------------------------------------------------------------------------ + + /** + * State view for input sides where the unique key is fully contained within the join key. + * + * <p>Stores data as {@code MapState<MapKey, Record>}. + */ + private static final class JoinKeyContainsUniqueKey implements MultiJoinStateView { + + // stores record in the mapping <MapKey, Record> + private final MapState<RowData, RowData> recordState; + private final List<RowData> reusedList; + + private JoinKeyContainsUniqueKey( + RuntimeContext ctx, + final String stateName, + final InternalTypeInfo<RowData> mapKeyType, + final InternalTypeInfo<RowData> recordType, + final StateTtlConfig ttlConfig) { + + MapStateDescriptor<RowData, RowData> recordStateDesc = + createStateDescriptor(stateName, mapKeyType, recordType, ttlConfig); + + this.recordState = ctx.getMapState(recordStateDesc); + // the result records always not more than 1 per mapKey + this.reusedList = new ArrayList<>(1); + } + + @Override + public void addRecord(RowData mapKey, RowData record) throws Exception { + recordState.put(mapKey, record); + } + + @Override + public void retractRecord(RowData mapKey, RowData record) throws Exception { + // Only one record is kept per mapKey, remove it directly. + recordState.remove(mapKey); + } + + @Override + public Iterable<RowData> getRecords(RowData mapKey) throws Exception { + reusedList.clear(); + RowData record = recordState.get(mapKey); + if (record != null) { + reusedList.add(record); + } + return reusedList; + } + + @Override + public void cleanup(RowData mapKey) throws Exception { + recordState.remove(mapKey); + } + } + + /** + * State view for input sides that have a unique key, but it differs from the join key. + * + * <p>Stores data as {@code MapState<MapKey, Map<UK, Record>>}. + */ + private static final class InputSideHasUniqueKey implements MultiJoinStateView { + + // stores map in the mapping <MapKey, Map<UK, Record>> + private final MapState<RowData, Map<RowData, RowData>> recordState; + private final KeySelector<RowData, RowData> uniqueKeySelector; + + private InputSideHasUniqueKey( + RuntimeContext ctx, + final String stateName, + final InternalTypeInfo<RowData> mapKeyType, + final InternalTypeInfo<RowData> recordType, + final InternalTypeInfo<RowData> uniqueKeyType, + final KeySelector<RowData, RowData> uniqueKeySelector, + final StateTtlConfig ttlConfig) { + checkNotNull(uniqueKeyType); + checkNotNull(uniqueKeySelector); + this.uniqueKeySelector = uniqueKeySelector; + + TypeInformation<Map<RowData, RowData>> mapValueTypeInfo = + Types.MAP(uniqueKeyType, recordType); // UK is the key in the inner map + + MapStateDescriptor<RowData, Map<RowData, RowData>> recordStateDesc = + createStateDescriptor(stateName, mapKeyType, mapValueTypeInfo, ttlConfig); + + this.recordState = ctx.getMapState(recordStateDesc); + } + + @Override + public void addRecord(RowData mapKey, RowData record) throws Exception { + RowData uniqueKey = uniqueKeySelector.getKey(record); + Map<RowData, RowData> uniqueKeyToRecordMap = recordState.get(mapKey); + if (uniqueKeyToRecordMap == null) { + uniqueKeyToRecordMap = new HashMap<>(); + } + uniqueKeyToRecordMap.put(uniqueKey, record); + recordState.put(mapKey, uniqueKeyToRecordMap); + } + + @Override + public void retractRecord(RowData mapKey, RowData record) throws Exception { + RowData uniqueKey = uniqueKeySelector.getKey(record); + Map<RowData, RowData> uniqueKeyToRecordMap = recordState.get(mapKey); + if (uniqueKeyToRecordMap != null) { + uniqueKeyToRecordMap.remove(uniqueKey); + if (uniqueKeyToRecordMap.isEmpty()) { + // Clean up the entry for mapKey if the inner map becomes empty + recordState.remove(mapKey); + } else { + recordState.put(mapKey, uniqueKeyToRecordMap); + } + } + // ignore uniqueKeyToRecordMap == null + } + + @Override + public Iterable<RowData> getRecords(RowData mapKey) throws Exception { + Map<RowData, RowData> uniqueKeyToRecordMap = recordState.get(mapKey); + if (uniqueKeyToRecordMap == null) { + return Collections.emptyList(); + } else { + // Return the values (records) from the inner map + return uniqueKeyToRecordMap.values(); + } + } + + @Override + public void cleanup(RowData mapKey) throws Exception { + recordState.remove(mapKey); + } + } + + /** + * State view for input sides that do not have a unique key (multi-set semantics). + * + * <p>Stores data as {@code MapState<MapKey, Map<Record, Count>>}. + */ + private static final class InputSideHasNoUniqueKey implements MultiJoinStateView { + + // stores map in the mapping <MapKey, Map<Record, Count>> + private final MapState<RowData, Map<RowData, Integer>> recordState; Review Comment: Hi @gustavodemorais @SteveStevenpoor ! Maybe my comment is too late but I think I need to say a couple of words about using CommonJoinKey as a subset of JoinKey. I agree with Gustavo - using the CommonJoinKey helps us to get all needed records on correct subtask thus the data consistency in not broken. But in other hand we'll have the following example `SELECT * FROM T1 JOIN T2 on T1.k1 = T2.k1 AND **T1.k98 = T2.k98** JOIN T3 on T2.k1 = T3.k1 AND T2.k99 == T3.k99`. And if we receive the input data from T1 (with particular values of T1.k1 and T1.k98) we need first to get all data from T2 and then request from T3 the whole set of T2.k1 and T2.k99 which we found on previous step. So there would be as many requests to T3 as many different values of k99 were found in T2. And the performance would decrease **exponentially** depending on input tables number. And more over, the future optimization which might be used for MJ is changing state order in runtime would not be possible to use if the JoinKey != CommonKey. Or we can use parallel requests to different states in future. But using CommonJoinKey will not allow such optimizations. So I'm tending to agree with Stepan's idea in comments above. Where he offered to not use the CommonJoinKey instead of JoinKey. I think such approach will slightly reduce the use cases but will bring much better performance (we won't need an exponential traverse across N states, just one iteration through each state) and, which is more important, the ability for future optimizations. -- 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...@flink.apache.org For queries about this service, please contact Infrastructure at: us...@infra.apache.org
