Please keep the discussion on the mailing list rather than commenting on the wiki (wiki discussions get unwieldy fast).

[This FLIP proposal is a joint work between Ran Jinhao and Dong Lin ]

Motivation

As shown in the figure below, we might have a job that pre-processes records from a bounded source (i.e. inputA) using an operator (i.e. operatorA) which only emits results after its input has ended. Then operatorB needs to join records emitted by operatorA with records from an unbounded source, and emit results with low processing latency in real-time.

Currently, supporting the above use-case requires all operators to be deployed at the start of the job. This approach wastes slot and memory resources because operatorB can not do any useful work until operatorA's input has ended. Even worse, operatorB might use a lot of disk space only to cache and spill records received from the unbounded source to its local disk while it is waiting for operatorA's output.

In this FLIP, we propose to optimize task deployment and resource utilization for the above use-case by allowing an operator to explicitly specify whether it only emits records after all its inputs have ended. JM will leverage this information to optimize job scheduling such that the partition type of the results emitted by this operator, as well as the results emitted by its upstream operators, will all be blocking, which effectively let Flink schedule and execute this operator as well as its upstream operators in batch mode. Hybrid shuffle mode(FLIP-235: Hybrid Shuffle Mode) can be used in batch mode part to further improve the performance when there are sufficient slot resources.

In addition, this FLIP also adds EndOfStreamWindows that can be used with the DataStream API to specify whether the window will end only at the end of its inputs. DataStream program (e.g. coGroup and aggregate) can take advantage of this information to significantly increase throughput and reduce resource utilization.

Objectives

The APIs provided in this FLIP achieves the following objectives / benefits:

1) Improve throughput. For use-case that involves a mixture of bounded and unbounded workloads (e.g. the use-case specified in the example), the user can start a single Flink job to address such use-case where the performance of the bounded workload can be optimal (e.g. similar/higher than the corresponding performance in batch mode).

2) Reduce resource usage.  For use-case that involves an operatorA (with unbounded input) depending on the output of another operatorB, where operatorB only emits results at the end of its input, Flink will deploy operatorB after operatorA is finished. This approach reduces the unnecessary resource usage when operatorA is still processing its inputs.

3) Improve usability. For use-case that needs to invoke DataStream APIs (e.g. KeyedStream#window) with a window assigner that covers all input data, users can use the off-the-shelf EndOfStreamWindows provided in this FLIP, instead of writing tens of lines of code to define this WindowAssigner subclass. 

Public Interfaces

1) Add EndOfStreamWindows which is a subclass of WindowAssigner. This class allows users of the DataStream API to specify whether the computation (e.g. co-group, aggregate) should emit data only after end-of-input.

/**
 * This WindowAssigner assigns all elements to the same window that is fired iff the input
 * stream reaches EOF.
 */
@PublicEvolving
public class EndOfStreamWindows extends WindowAssigner<Object, TimeWindow> {
 
    private static final TimeWindow TIME_WINDOW_INSTANCE =
            new TimeWindow(Long.MIN_VALUE, Long.MAX_VALUE);
 
    private EndOfStreamWindows() {}
 
    public static EndOfStreamWindows get() {
        return INSTANCE;
    }
 
    @Override
    public Collection<TimeWindow> assignWindows(
            Object element, long timestamp, WindowAssignerContext context) {
        return Collections.singletonList(TIME_WINDOW_INSTANCE);
    }
 
    @Override
    public Trigger<Object, TimeWindow> getDefaultTrigger(StreamExecutionEnvironment env) {
        return new EndOfStreamTrigger();
    }
 
    @Override
    public boolean isEventTime() {
        return true;
    }
 
    private static class EndOfStreamTrigger extends Trigger<Object, TimeWindow> {
        @Override
        public TriggerResult onElement(
                Object element, long timestamp, TimeWindow window, TriggerContext ctx)
                throws Exception {
            return TriggerResult.CONTINUE;
        }
 
        @Override
        public TriggerResult onEventTime(long time, TimeWindow window, TriggerContext ctx) {
            return time == window.maxTimestamp() ? TriggerResult.FIRE : TriggerResult.CONTINUE;
        }
 
        @Override
        public TriggerResult onProcessingTime(long time, TimeWindow window, TriggerContext ctx) {
            return TriggerResult.CONTINUE;
        }
        ...
    }
}

2) Add OperatorAttributesBuilder and OperatorAttributes for operator developers to specify operator attributes that Flink runtime can use to optimize the job performance.

package org.apache.flink.streaming.api.operators;
 
/** The builder class for {@link OperatorAttributes}. */
@Experimental
public class OperatorAttributesBuilder {
    @Nullable private Boolean outputOnEOF = null;
 
    public OperatorAttributesBuilder() {...}
 
    public OperatorAttributesBuilder setOutputOnEOF(boolean outputOnEOF) {...}
   
    /**
     * If any operator attribute is null, we will log it at DEBUG level and use the following
     * default values.
     * - outputOnEOF defaults to false
     */
     public OperatorAttributes build() {...}
}

package org.apache.flink.streaming.api.operators;
 
/**
 * OperatorAttributes element provides Job Manager with information that can be
 * used to optimize the job performance.
 */
@Experimental
public class OperatorAttributes {    
   /**
     * Returns true iff the operator can only emit records after inputs have reached EOF.
     *
     * <p>Here are the implications when it is true:
     *
     * <ul>
     *   <li> The results of this operator as well as its upstream operators have blocking partition type.
     *   <li> This operator as well as its upstream operators will be executed in batch mode.
     * </ul>
     */
    public boolean isOutputOnEOF() {...}
}

3) Add the getOperatorAttributes() API to the StreamOperator and StreamOperatorFactory interfaces.

@Experimental
public interface StreamOperator<OUT> extends CheckpointListener, KeyContext, Serializable {
    ...
 
    default OperatorAttributes getOperatorAttributes() {
        return new OperatorAttributesBuilder().build();
    }
}
 
@Experimental
public interface StreamOperatorFactory<OUT> extends Serializable {
    ...
 
    default OperatorAttributes getOperatorAttributes() {
        return new OperatorAttributesBuilder().build();
    }
}

Proposed Changes

1) Add the APIs on Transformation interface to get the corresponding operator attributes.

@Internal
public abstract class Transformation<T> {
    public boolean isOutputOnEOF() {
        return false;
    }
}

2) Update Transformation subclasses (e.g. OneInputTransformation and TwoInputTransformation) to override the newly added methods using the OperatorAttributes obtained from the underlying Operator.

3) Update JM to make use of the following operator attributes when compiling the Transformation graph into the JobGraph.

If a Transformation has isOutputOnEOF == true:

• The process of operator chain can still be done. After that, the results of its operator chain as well as and its upstream operators will be set blocking (by default) or hybrid shuffle partition type which can be controlled by configuring ExecutionOptions.BATCH_SHUFFLE_MODE.
• This operator as well as its upstream operators will be executed in batch mode (e.g checkpoint is disabled when these operators are running).

4) A blocking input edge with pending records is same as a source with isBacklog=true when an operator determines its RecordAttributes for downstream nodes.

This is needed in order for this FLIP to work with FLIP-327. More specifically, once both FLIP-327 and FLIP-331 are accepted, we need a way to determine the backlog status for input with blocking edge type.

5) When DataStream#coGroup is invoked with EndOfStreamWindows as the window assigner, Flink should generate an operator with isOutputOnEOF = true.

In addition, after FLIP-327 is accepted, suppose this API is called without any explicitly-assigned trigger or evictor, the program should also have isInternalSorterSupported = true to achieve higher throughput by using the optimizations currently done in batch mode.

The following optimization will be used to achieve higher throughput than the existing DataStream#coGroup in both stream and batch mode:

• It will instantiate two internal sorter to sorts records from its two inputs separately. Then it can pull the sorted records from these two sorters. This can be done without wrapping input records with TaggedUnion<...>. In comparison, the existing DataStream#coGroup needs to wrap input records with TaggedUnion<...> before sorting them using one external sorter, which introduces higher overhead.
• It will not invoke WindowAssigner#assignWindows or triggerContext#onElement for input records. In comparison, the existing WindowOperator#processElement invokes these methods for every records.

6) When DataStream#aggregate is invoked with EndOfStreamWindows as the window assigner, Flink should generate an operator with isOutputOnEOF = true.

In addition, after FLIP-327 is accepted, suppose this API is called without any explicitly-assigned trigger or evictor, the program should also have isInternalSorterSupported = true to achieve higher throughput by using the optimizations currently done in batch mode.

More specifically, This operator will sort the input before aggregation, and avoid invoking window actions, which is similar to '5)'.

Benchmark results

To demonstrate our optimization improvements，we run each benchmark in different execution modes and configurations. Each result is measured by taking the average execution time across 5 runs with the given configuration.

We run benchmarks on a MacBook with the latest Flink 1.18-snapshot and parallelism=1. Default standalone cluster configuration is used except:

jobmanager.memory.process.size: 6400m
taskmanager.memory.process.size: 6912m

1) Execute DataStream#CoGroup

This benchmark uses DataStream#coGroup to process records from two bounded inputs. Each input will generate records with (key = i, value = i) for i from 1 to 8*10^7.

Below is the DataStream program code snippet.

DataStream<Tuple2<Integer, Double>> source1  = env.fromCollection(
                new DataGenerator(dataNum), Types.TUPLE(Types.INT, Types.DOUBLE));
DataStream<Tuple2<Integer, Double>> source2  = env.fromCollection(
                new DataGenerator(dataNum), Types.TUPLE(Types.INT, Types.DOUBLE));

source1.coGroup(source2)
    .where(tuple -> tuple.f0)
    .equalTo(tuple -> tuple.f0)
    .window(EndOfStreamWindows.get())
    .apply(new CustomCoGroupFunction())
    .addSink(...);

private static class CustomCoGroupFunction
            extends RichCoGroupFunction<Tuple2<Integer, Double>, Tuple2<Integer, Double>, Integer> {
        @Override
        public void coGroup(
                Iterable<Tuple2<Integer, Double>> iterableA,
                Iterable<Tuple2<Integer, Double>> iterableB,
                Collector<Integer> collector) {
            collector.collect(1);
        }
    }

The following result shows the throughput (records/sec) when the benchmark is executed in streaming mode, batch mode, optimized streaming mode after this PR, and optimized streaming mode with hybrid shuffle after this PR.

The result shows that DataStream#coGroup in optimized streaming mode can be 22X as fast as streaming mode and 3X as fast as batch mode. Hybrid shuffle can further improve throughput by 11%.

2) Execute DataStream#Aggregate

This benchmark uses DataStream#aggregate to process 8*10^7 records. These records are evenly distributed across 8*10^5 keys. More specifically, the source will generate records with (key = i, value = i) for i from 1 to 8*10^5, and repeat this process 100 times.

Below is the DataStream program code snippet.

DataStreamSource<Tuple2<Long, Double>> source  = env.fromCollection(
                        new DataGenerator(dataNum, keyNum), Types.TUPLE(Types.LONG, Types.DOUBLE));
source.keyBy(value -> value.f0)
      .window(EndOfStreamWindows.get())
      .aggregate(new Aggregator())
      .addSink(...);

public static class Aggregator implements AggregateFunction<Tuple2<Long, Double>, Tuple2<Long, Double>, Double> {
        @Override
        public Tuple2<Long, Double> createAccumulator() {
            return new Tuple2<Long, Double>(0L, 0.0);
        }

        @Override
        public Tuple2<Long, Double> add(Tuple2<Long, Double> myData, Tuple2<Long, Double> accData) {
            accData.f1 = accData.f1 + myData.f1;
            return accData;
        }

        @Override
        public Double getResult(Tuple2<Long, Double> result) {
            return result.f1;
        }

        @Override
        public Tuple2<Long, Double> merge(Tuple2<Long, Double> acc1, Tuple2<Long, Double> acc2) {
            acc1.f1 = acc1.f1 + acc2.f1;
            return acc1;
        }
    }

The following result shows the throughput (records/sec) when the benchmark is executed in streaming mode, batch mode, optimized streaming mode after this PR, and optimized streaming mode with hybrid shuffle after this PR.

The result shows that DataStream#aggregate in optimized streaming mode can be 10X as fast as streaming mode and 11% faster than batch mode. Hybrid shuffle can further improve throughput by 15%.

3) Execute a program that needs to fully process data from a bounded source before processing data from another unbounded source.

The following program demonstrates the scenario described in the motivation section. The program needs to pre-processes records from a bounded source (Source1) using an operator (Process1) which only emits results after its input has ended. Then anther operator(Process2) needs to process records emitted by Process1 with records from an unbounded source, and emit results with low processing latency in real-time.

 source1.keyBy(value -> value.f0)
  .window(EndOfStreamWindows.get())
  .aggregate(new MyAggregator()).name("Process1")
  .connect(source2.keyBy(value -> value.f0))
  .transform("Process2", Types.INT, new MyProcessOperator())
  .addSink(...); 

We can use this program to demonstrate that the program requires less slot resources. More specifically, suppose we configure the standalone cluster with taskmanager.numberOfTaskSlots = 2, and set the Source1，Process1, Source2 and Process2 in 4 different SlotSharingGroups, the program will fail to be deployed before this FLIP. And the program can be deployed successfully after this FLIP. This is because Source2 and Process2 can be deplyed after Source1，Process1 finished and released their slots.

Additionally, we can use this program to demonstrate that it can achieve higher performance because Process2 will not need to keep buffer records emitted by Source2 in its memory while Process1 has not reached EOF. More specifically, the program can fail with OOM before this FLIP when the number of records in inputs is high. And the program can finish successfully without OOM after this FLIP.

Compatibility, Deprecation, and Migration Plan

The changes made in this FLIP are backward compatible. No deprecation or migration plan is needed.

Future Work

It would be useful to add an ExecutionState (e.g. Finishing) to specify whether the task has reached EOF for all its inputs. This allows JM to deploy its downstream tasks and possibly apply hybrid shuffle to increase job throughput.