WencongLiu commented on code in PR #23362: URL: https://github.com/apache/flink/pull/23362#discussion_r1360163929
########## docs/content/docs/dev/datastream/dataset_migration.md: ########## @@ -0,0 +1,758 @@ +--- +title: "How To Migrate From DataSet to DataStream" +weight: 302 +type: docs +--- +<!-- +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. +--> + +# How to Migrate from DataSet to DataStream + +The DataSet API has been formally deprecated and will no longer receive active maintenance and support. It will be removed in the +Flink 2.0 version. Flink users are recommended to migrate from the DataSet API to the DataStream API, Table API and SQL for their +data processing requirements. + +For the most of DataSet APIs, the users can utilize the DataStream API to get the same calculation result in the batch jobs. However, +different DataSet API can be implemented by DataStream API with various difference on semantic and behavior. All DataSet APIs can be +categorized into four types: + +Category 1: These DataSet APIs can be implemented by DataStream APIs with same semantic and same calculation behavior. + +Category 2: These DataSet APIs can be implemented by DataStream APIs with different semantic but same calculation behavior. This will +make the job code more complex. + +Category 3: These DataSet APIs can be implemented by DataStream APIs with different semantic and different calculation behavior. This +will involve additional computation and shuffle costs. + +Category 4: These DataSet APIs are not supported by DataStream APIs. + +The subsequent sections will first introduce how to set the execution environment and provide detailed explanations on how to implement +each category of DataSet APIs using the DataStream APIs, highlighting the specific considerations and challenges associated with each +category. + + +## Setting the execution environment + +To execute a DataSet pipeline by DataStream API, we should first start by replacing ExecutionEnvironment with StreamExecutionEnvironment. + +<table> + <tr> + <th>DataSet</th> + <th>DataStream</th> + </tr> + <tr> + <td> + <code style="white-space: pre-line;">// Create the execution environment + ExecutionEnvironment.getExecutionEnvironment(); + // Create the local execution environment + ExecutionEnvironment.createLocalEnvironment(); + // Create the collection environment + new CollectionEnvironment(); + // Create the remote environment + ExecutionEnvironment.createRemoteEnvironment(host, port, jarFile); + </code> + </td> + <td> + <code style="white-space: pre-line;">// Create the execution environment + StreamExecutionEnvironment.getExecutionEnvironment(); + // Create the local execution environment + StreamExecutionEnvironment.createLocalEnvironment(); + // The collection environment is not supported. + // Create the remote environment + StreamExecutionEnvironment.createRemoteEnvironment(host, port, jarFile); + </code> + </td> + </tr> +</table> + + +As the source of DataSet is always bounded, the execution mode must be set to RuntimeMode.BATCH to make Flink execute in batch mode. + +```java +StreamExecutionEnvironment executionEnvironment = StreamExecutionEnvironment.getExecutionEnvironment(); +executionEnvironment.setRuntimeMode(RuntimeExecutionMode.BATCH); +``` + +## Using the streaming sources and sinks + +Sources: The DataStream API uses `DataStreamSource` to read records from external system, while the DataSet API uses the +`DataSource`. + +Sinks: The DataStream API uses the implementations of `SinkFunction` and `Sink` to write records to external system, while the +DataSet API uses the `FileOutputFormat`. + +If you are looking for pre-defined source and sink connectors of DataStream, please check the [Connector Docs]({{< ref "docs/connectors/datastream/overview" >}}) + + +## Implement the DataSet API by DataStream + +#### Category 1 + +For Category 1, the usage of the API in DataStream is almost identical to that in DataSet. This means that implementing these +DataSet APIs by the DataStream API is relatively straightforward and does not require significant modifications or complexity +in the job code. + +### Map + +<table> + <tr> + <th>DataSet</th> + <th>DataStream</th> + </tr> + <tr> + <td> + <code style="white-space: pre-line;">dataSet.map(new MapFunction(){ + // implement user-defined map logic +}); + </code> + </td> + <td> + <code style="white-space: pre-line;">dataStream.map(new MapFunction(){ + // implement user-defined map logic +}); + </code> + </td> + </tr> +</table> + + +### FlatMap + +<table> + <tr> + <th>DataSet</th> + <th>DataStream</th> + </tr> + <tr> + <td> + <code style="white-space: pre-line;">dataSet.flatMap(new FlatMapFunction(){ + // implement user-defined flatmap logic +}); + </code> + </td> + <td> + <code style="white-space: pre-line;">dataStream.flatMap(new FlatMapFunction(){ + // implement user-defined flatmap logic +}); + </code> + </td> + </tr> +</table> + +### Filter + +<table> + <tr> + <th>DataSet</th> + <th>DataStream</th> + </tr> + <tr> + <td> + <code style="white-space: pre-line;">dataSet.filter(new FilterFunction(){ + // implement user-defined filter logic +}); + </code> + </td> + <td> + <code style="white-space: pre-line;">dataStream.filter(new FilterFunction(){ + // implement user-defined filter logic +}); + </code> + </td> + </tr> +</table> + +### Union + +<table> + <tr> + <th>DataSet</th> + <th>DataStream</th> + </tr> + <tr> + <td> + <code style="white-space: pre-line;">dataSet1.union(dataSet2); + </code> + </td> + <td> + <code style="white-space: pre-line;">dataStream1.union(dataStream2); + </code> + </td> + </tr> +</table> + +### Rebalance + +<table> + <tr> + <th>DataSet</th> + <th>DataStream</th> + </tr> + <tr> + <td> + <code style="white-space: pre-line;">dataSet.rebalance(); + </code> + </td> + <td> + <code style="white-space: pre-line;">dataStream.rebalance(); + </code> + </td> + </tr> +</table> + +### Project + +<table> + <tr> + <th>DataSet</th> + <th>DataStream</th> + </tr> + <tr> + <td> + <code style="white-space: pre-line;">DataSet<Tuple3<Integer, Double, String>> dataSet = // [...] +dataSet.project(2,0); + </code> + </td> + <td> + <code style="white-space: pre-line;">DataStream<Tuple3<Integer, Double, String>> dataStream = // [...] +dataStream.project(2,0); + </code> + </td> + </tr> +</table> + +### Reduce on Grouped DataSet + +<table> + <tr> + <th>DataSet</th> + <th>DataStream</th> + </tr> + <tr> + <td> + <code style="white-space: pre-line;">DataSet<Tuple2<String, Integer>> dataSet = // [...] +dataSet.groupBy(value -> value.f0) + .reduce(new ReduceFunction(){ + // implement user-defined reduce logic + }); + </code> + </td> + <td> + <code style="white-space: pre-line;">DataStream<Tuple2<String, Integer>> dataStream = // [...] +dataStream.keyBy(value -> value.f0) + .reduce(new ReduceFunction(){ + // implement user-defined reduce logic + }); + </code> + </td> + </tr> +</table> + +### Aggregate on Grouped DataSet + +<table> + <tr> + <th>DataSet</th> + <th>DataStream</th> + </tr> + <tr> + <td> + <code style="white-space: pre-line;">DataSet<Tuple2<String, Integer>> dataSet = // [...] +// compute sum of the second field +dataSet.groupBy(value -> value.f0).aggregate(SUM, 1); +// compute min of the second field +dataSet.groupBy(value -> value.f0).aggregate(MIN, 1); +// compute max of the second field +dataSet.groupBy(value -> value.f0).aggregate(MAX, 1); + </code> + </td> + <td> + <code style="white-space: pre-line;">DataStream<Tuple2<String, Integer>> dataStream = // [...] +// compute sum of the second field +dataStream.keyBy(value -> value.f0).sum(1); +// compute min of the second field +dataStream.keyBy(value -> value.f0).min(1); +// compute max of the second field +dataStream.keyBy(value -> value.f0).max(1); + </code> + </td> + </tr> +</table> + +#### Category 2 + +For category 2, these DataSet APIs can be implemented by DataStream APIs with different semantic but same calculation behavior. +The developers need to adapt their code to accommodate these variations, which introduces additional complexity. + +### Distinct + +<table> + <tr> + <th>DataSet</th> + <th>DataStream</th> + </tr> + <tr> + <td> + <code style="white-space: pre-line;">DataSet<Integer> dataSet = // [...] +dataSet.distinct(); + </code> + </td> + <td> + <code style="white-space: pre-line;">DataStream<Integer> dataStream = // [...] +dataStream + .keyBy(value -> value) + .reduce((value1, value2) -> value1); + </code> + </td> + </tr> +</table> + +### Hash-Partition + +<table> + <tr> + <th>DataSet</th> + <th>DataStream</th> + </tr> + <tr> + <td> + <code style="white-space: pre-line;">DataSet<Tuple2<String, Integer>> dataSet = // [...] +dataSet.partitionByHash(value -> value.f0); + </code> + </td> + <td> + <code style="white-space: pre-line;">DataStream<Tuple2<String, Integer>> dataStream = // [...] +// partition by the hashcode of key +dataStream.partitionCustom((key, numSubpartition) -> key.hashCode() % numSubpartition, value -> value.f0); + </code> + </td> + </tr> +</table> + +### Reduce on Full DataSet + +If developers want to compute data of full datastream, `GlobalWindow` could be used to collect all records of datastream. +However, a special trigger is also required to trigger the computation of `GlobalWindow` at the end of its inputs. Here is an example +code snippet of the trigger. + +```java +public class EOFTrigger extends Trigger<Object, GlobalWindow> { + + private boolean hasRegistered; + + @Override + public TriggerResult onElement( + Object element, long timestamp, GlobalWindow window, TriggerContext ctx) { + if (!hasRegistered) { + ctx.registerEventTimeTimer(Long.MAX_VALUE); + hasRegistered = true; + } + return TriggerResult.CONTINUE; + } + + @Override + public TriggerResult onEventTime(long time, GlobalWindow window, TriggerContext ctx) { + return TriggerResult.FIRE_AND_PURGE; + } + + @Override + public TriggerResult onProcessingTime(long time, GlobalWindow window, TriggerContext ctx) { + return TriggerResult.CONTINUE; + } + + @Override + public void clear(GlobalWindow window, TriggerContext ctx) throws Exception {} +} +``` +Then the reduce operation on full datastream could be performed by `EOFTrigger`. + +<table> + <tr> + <th>DataSet</th> + <th>DataStream</th> + </tr> + <tr> + <td> + <code style="white-space: pre-line;">DataSet<String> dataSet = // [...] +dataSet.reduce(new ReduceFunction(){ + // implement user-defined reduce logic + }); + </code> + </td> + <td> + <code style="white-space: pre-line;">DataStream<String> dataStream = // [...] +dataStream.windowAll(GlobalWindows.create()).trigger(new EOFTrigger()) + .reduce(new ReduceFunction(){ + // implement user-defined reduce logic + }); + </code> + </td> + </tr> +</table> + + +### Aggregate on Full DataSet + +<table> + <tr> + <th>DataSet</th> + <th>DataStream</th> + </tr> + <tr> + <td> + <code style="white-space: pre-line;">DataSet<Tuple2<Integer, Integer>> dataSet = // [...] +// compute sum of the second field +dataSet.aggregate(SUM, 1); +// compute min of the second field +dataSet.aggregate(MIN, 1); +// compute max of the second field +dataSet.aggregate(MAX, 1); + </code> + </td> + <td> + <code style="white-space: pre-line;">DataStream<Tuple2<Integer, Integer>> dataStream = // [...] +// compute sum of the second field +dataStream.windowAll(GlobalWindows.create()).trigger(new EOFTrigger()).sum(1); +// compute min of the second field +dataStream.windowAll(GlobalWindows.create()).trigger(new EOFTrigger()).min(1); +// compute max of the second field +dataStream.windowAll(GlobalWindows.create()).trigger(new EOFTrigger()).max(1); + </code> + </td> + </tr> +</table> + +### GroupReduce on Full DataSet + +<table> + <tr> + <th>DataSet</th> + <th>DataStream</th> + </tr> + <tr> + <td> + <code style="white-space: pre-line;">DataSet<Integer> dataSet = // [...] +dataSet.reduceGroup(new GroupReduceFunction(){ + // implement user-defined group reduce logic + }); + </code> + </td> + <td> + <code style="white-space: pre-line;">DataStream<Integer> dataStream = // [...] +dataStream.windowAll(GlobalWindows.create()).trigger(new EOFTrigger()) + .apply(new WindowFunction(){ + // implement user-defined group reduce logic + }); + </code> + </td> + </tr> +</table> + +### GroupReduce on Grouped DataSet + +<table> + <tr> + <th>DataSet</th> + <th>DataStream</th> + </tr> + <tr> + <td> + <code style="white-space: pre-line;">DataSet<Tuple2<Integer, String>> dataSet = // [...] +dataSet.groupBy(value -> value.f0) + .reduceGroup(new GroupReduceFunction(){ + // implement user-defined group reduce logic + }); + </code> + </td> + <td> + <code style="white-space: pre-line;">DataStream<Tuple2<String, Integer>> dataStream = // [...] +dataStream.keyBy(value -> value.f0) + .window(GlobalWindows.create()) + .trigger(new EOFTrigger()) + .apply(new WindowFunction(){ + // implement user-defined group reduce logic + }); + </code> + </td> + </tr> +</table> + +### First-n + +<table> + <tr> + <th>DataSet</th> + <th>DataStream</th> + </tr> + <tr> + <td> + <code style="white-space: pre-line;">dataSet.first(n) + </code> + </td> + <td> + <code style="white-space: pre-line;">dataStream.windowAll(GlobalWindows.create()) +.trigger(new EOFTrigger()) +.apply(new AllWindowFunction(){ +// implement first-n logic +}); + </code> + </td> + </tr> +</table> + + +### Join + +<table> + <tr> + <th>DataSet</th> + <th>DataStream</th> + </tr> + <tr> + <td> + <code style="white-space: pre-line;">DataSet<Tuple2<String, Integer>> dataSet1 = // [...] +DataSet<Tuple2<String, Integer>> dataSet2 = // [...] +dataSet1.join(dataSet2) + .where(data -> data.f0) + .equalTo(data -> data.f0) + .with(new JoinFunction(){ + // implement user-defined join logic + }); + </code> + </td> + <td> + <code style="white-space: pre-line;">DataStream<Tuple2<String, Integer>> dataStream1 = // [...] +DataStream<Tuple2<String, Integer>> dataStream2 = // [...] +dataStream1.join(dataStream2) + .where(value -> value.f0) + .equalTo(value -> value.f0) + .window(GlobalWindows.create()) + .trigger(new EOFTrigger()) + .apply(new JoinFunction(){ + // implement user-defined join logic + }); + </code> + </td> + </tr> +</table> + +### CoGroup + +<table> + <tr> + <th>DataSet</th> + <th>DataStream</th> + </tr> + <tr> + <td> + <code style="white-space: pre-line;">DataSet<Tuple2<String, Integer>> dataSet1 = // [...] +DataSet<Tuple2<String, Integer>> dataSet2 = // [...] +dataSet1.coGroup(dataSet2).where(value -> value.f0) + .equalTo(value -> value.f0) + .with(new CoGroupFunction(){ + // implement user-defined co group logic + }); + </code> + </td> + <td> + <code style="white-space: pre-line;">DataStream<Tuple2<String, Integer>> dataStream1 = // [...] +DataStream<Tuple2<String, Integer>> dataStream2 = // [...] +dataStream1.coGroup(dataStream2) + .where(value -> value.f0) + .equalTo(value -> value.f0) + .window(GlobalWindows.create()) + .trigger(new EOFTrigger()) + .apply(new CoGroupFunction(){ + // implement user-defined co group logic + }); + </code> + </td> + </tr> +</table> + +### OuterJoin + +To implement the OuterJoin operation between two DataStreams, datastream1 and datastream2, follow these steps: + +1. Co-group the two idatastreams in a `GlobalWindow` with `EOFTrigger` to collect elements with the specific key of each datastream. + +2. Implement left/right outer join in the user-defined CoGroupFunction. + +<table> + <tr> + <th>DataSet</th> + <th>DataStream</th> + </tr> + <tr> + <td> + <code style="white-space: pre-line;">DataSet<Tuple2<String, Integer>> dataSet1 = // [...] +DataSet<Tuple2<String, Integer>> dataSet2 = // [...] +// left outer join +dataSet1.leftOuterJoin(dataSet2).where(dataSet1.f0) + .equalTo(dataSet2.f0) + .with(new JoinFunction(){ + // implement user-defined left outer join logic + }); +// right outer join +dataSet1.rightOuterJoin(dataSet2).where(dataSet1.f0) + .equalTo(dataSet2.f0) + .with(new JoinFunction(){ + // implement user-defined right outer join logic + }); + </code> + </td> + <td> + <code style="white-space: pre-line;">DataStream<Tuple2<String, Integer>> dataStream1 = // [...] +DataStream<Tuple2<String, Integer>> dataStream2 = // [...] +// left outer join +dataStream1.coGroup(dataStream2) + .where(value -> value.f0) + .equalTo(value -> value.f0) + .window(GlobalWindows.create()) + .trigger(new EOFTrigger()) + .apply((leftIterable, rightInterable, collector) -> { + if(!rightInterable.iterator().hasNext()){ + // implement user-defined left outer join logic + } + }); +// right outer join +dataStream1.coGroup(dataStream2) + .where(value -> value.f0) + .equalTo(value -> value.f0) + .window(GlobalWindows.create()) + .trigger(new EOFTrigger()) + .apply((leftIterable, rightInterable, collector) -> { + if(!leftIterable.iterator().hasNext()){ + // implement user-defined right outer join logic + } + }); + </code> + </td> + </tr> +</table> + +#### Category 3 + +For category 3, these DataSet APIs can be implemented by DataStream APIs with different semantic and different calculation behavior. Additional +calculation steps will be added. + +To collect records from each subtask, every record needs to be assigned a unique subtask ID and grouped accordingly within the window. +This additional step of assigning the subtask ID and performing a groupby operation introduces shuffle costs. Here is an example code +snippet showing how to assign a subtask ID to each record. The function assignSubtaskID is used to explain the detailed behavior and will +be utilized in the subsequent sections: +```java +// assign subtask ID to all records +DataStream<Tuple2<String, Integer>> assignSubtaskID(DataStream<Integer> dataStream) { + return dataStream.map(new RichMapFunction<Integer, Tuple2<String, Integer>>() { Review Comment: Good idea. I'd prefer to name the function class as `AddSubtaskIDMapFunction`. -- 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]
