阿里巴巴开源的Asynchronous I/O Design and Implementation

Motivation

I/O access, for the most case, is a time-consuming process, making the TPS for single operator much lower than in-memory computing, particularly for streaming job, when low latency is a big concern for users. Starting multiple threads may be an option to handle this problem, but the drawbacks are obvious: The programming model for end users may become more complicated as they have to implement thread model in the operator. Furthermore, they have to pay attention to coordinate with checkpointing.

Terms

AsyncFunction: Async I/O will be triggered in AsyncFunction.

AsyncWaitOperator: An StreamOperator which will invoke AsyncFunction.

AsyncCollector: For each input streaming record, an AsyncCollector will be created and passed into user's callback to get the async i/o result.

AsyncCollectorBuffer: A buffer to keep all AsyncCollectors.

Emitter Thread: A working thread in AsyncCollectorBuffer, being signalled while some of AsyncCollectors have finished async i/o and emitting results to the following opeartors.

Public Interfaces

A helper class, named AsyncDataStream, is added to provide methods to add AsyncFunction, which will do async i/o operation, into FLINK streaming job.

AsyncDataStream.java

public class AsyncDataStream {
 /**
  * Add an AsyncWaitOperator. The order of output stream records may be reordered.
  *
  * @param in Input data stream
  * @param func AsyncFunction
  * @bufSize The max number of async i/o operation that can be triggered
  * @return A new DataStream.
  */
 public static DataStream<OUT> unorderedWait(DataStream<IN> in, AsyncFunction<IN, OUT> func, int bufSize);
 public static DataStream<OUT> unorderedWait(DataStream<IN> in, AsyncFunction<IN, OUT> func);
  
 /**
  * Add an AsyncWaitOperator. The order of output stream records is guaranteed to be the same as input ones.
  *
  * @param func AsyncWaitFunction
  * @param func AsyncFunction
  * @bufSize The max number of async i/o operation that can be triggered
  * @return A new DataStream.
  */
 public static DataStream<OUT> orderedWait(DataStream<IN> in, AsyncFunction<IN, OUT> func, int bufSize);
 public static DataStream<OUT> orderedWait(DataStream<IN> in, AsyncFunction<IN, OUT> func);
}

Proposed Changes

Overview

The following diagram illustrates how the streaming records are processed while

• arriving at AsyncWaitOperator
• recovering from task failover
• snapshotting state
• being emitted by Emitter Thread

Sequence Diagram

AsyncFunction

AsyncFunction works as a user function in AsyncWaitOperator which looks like StreamFlatMap operator, having open()/processElement(StreamRecord<IN> record)/processWatermark(Watermark mark).

For user’s concrete AsyncFunction, the asyncInvoke(IN input, AsyncCollector<OUT> collector) has to be overriden to supply codes to start an async operation.

AsyncFunction.java

public interface AsyncFunction<IN, OUT> extends Function, Serializable {
  /**
   * Trigger async operation for each stream input.
   * The AsyncCollector should be registered into async client.
   *
   * @param input Stream Input
   * @param collector AsyncCollector
   */
  void asyncInvoke(IN input, AsyncCollector<OUT> collector) throws Exception;
}
  
public abstract class RichAsyncFunction<IN, OUT> extends AbstractRichFunction
    implements AsyncFunction<IN, OUT> {
  @Override
  public abstract void asyncInvoke(IN input, AsyncCollector<OUT> collector) throws Exception;
}

For each input stream record of AsyncWaitOperator, they will be processed by AsyncFunction.asyncInvoke(IN input, AsyncCollector<OUT> cb). Then AsyncCollector will be append into AsyncCollectorBuffer. We will cover AsyncCollector and AsyncCollectorBuffer later. 

AsyncCollector 

AsyncCollector is created by AsyncWaitOperator, and passed into AsyncFunction, where it should be added into user’s callback. It acts as a role to get results or errors from user codes and notify the AsyncCollectorBuffer to emit results.

 The functions specific for the user is the collect, and they should be called when async operation is done or errors are thrown out.

AsyncCollector.java

public class AsyncCollector<OUT> {
  private List<OUT> result;
  private Throwable error;
  private AsyncCollectorBuffer<OUT> buffer;
 
  /**
   * Set result
   * @param result A list of results.
   */
  public void collect(List<OUT> result) {
    this.result = result;
    buffer.mark(this);
  }
 
  /**
   * Set error
   * @param error A Throwable object.
   */
  public void collect(Throwable error) {
    this.error = error;
    buffer.mark(this);
  }
 
  /**
   * Get result. Throw RuntimeException while encountering an error.
   * @return A List of result.
   * @throws RuntimeException RuntimeException wrapping errors from user codes.
   */
  public List<OUT> getResult() throws RuntimeException { ... }
}

How is it used

Before calling AsyncFunction.asyncInvoke(IN input, AsyncCollector<OUT> collector), AsyncWaitOperator will try to get an instance of AsyncCollector from AsyncCollectorBuffer. Then it will be taken into user’s callback function. If the buffer is full, it will wait until some of ongoing callbacks has finished.

Once async operation has done, the AsyncCollector.collect() will take results or errors and AsyncCollectorBuffer will be notified.

AsyncCollector is implemented by FLINK.

AsyncCollectorBuffer

AsyncCollectorBuffer keeps all AsyncCollectors, and emit results to the next nodes.

When AsyncCollector.collect() is called, a mark will be placed in AsyncCollectorBuffer, indicating finished AsyncCollectors. A working thread, named Emitter, will also be signalledonce a AsyncCollector gets results, and then try to emit results depending on the ordered or unordered setting.

For simplicity, we will refer task to AsycnCollector in the AsyncCollectorBuffer in the following text.

 

Ordered and Unordered

Based on the user configuration, the order of output elements will or will not be guaranteed. If not guaranteed, the finished AsyncCollectors coming later will be emitted earlier.

Emitter Thread

The Emitter Thread will wait for finished AsyncCollectors. When it is signalled, it will process tasks in the buffer as follow:

• Ordered Mode

If the first task in the buffer is finished, then Emitter will collect its results, and then proceed to the second task. If the first task is not finished yet, just wait for it again.

• Unordered Mode

Check all finished tasks in the buffer, and collect results from those tasks which are prior to the oldest Watermark in the buffer.

The Emitter Thread and Task Thread will access exclusively by acquiring/releasing the lock.

Signal Task Thread when all tasks have finished to notify it that all data has been processed, and it is OK to close the operator.

Signal Task Thread after removing some tasks from the buffer.

Propagate Exceptions to Task Thread.

Task Thread

Access AsyncCollectorBuffer exclusively against the Emitter Thread.

Get and Add a new AsyncCollector to the buffer, wait while buffer is full.

Watermark

All watermarks will also be kept in AsyncCollectorBuffer. A watermark will be emitted if and only if after all AsyncCollectors coming before current watermark have been emitted. 

State, Failover and Checkpoint

State and Checkpoint

All input StreamRecords will be kept in state. Instead of storing each input stream records into state one by one while processing, AsyncWaitOperator will put all input stream records in AsyncCollectorBuffer into state while snapshotting operator state. Old data in the state will be cleared before persisting those records.

When all barriers have arrived at the operator, checkpoint can be carried out immediately.

Failover

While restoring the operator’s state, the operator will scan all elements in the state, get AsyncCollectors, call AsyncFunction.asyncInvoke() and insert them back into AsyncCollectorBuffer.

 

Notes

Async Resource Sharing

For the case to share async resources(like connection to hbase, netty connections) among different slots(task workers) in the same TaskManager(a.k.a the same JVM), we can make the connection static so that all threads in the same process can share the same instance.

Of course, please pay attention to thread safety while using those resources.

Example

For callback

Example.java

/***  ***/
public class HBaseAsyncFunction implements AsyncFunction<String, String> {
  // initialize it while reading object
  transient Connection connection;
 
  @Override
  public void asyncInvoke(String val, AsyncCollector<String> c) {
    Get get = new Get(Bytes.toBytes(val));
    Table ht = connection.getTable(TableName.valueOf(Bytes.toBytes("test")));
    // UserCallback is from user’s async client.
    ((AsyncableHTableInterface) ht).asyncGet(get, new UserCallback(c));
  }
}
 
// create data stream
public void createHBaseAsyncTestStream(StreamExecutionEnvironment env) {
  DataStream<String> source = getDataStream(env);
  DataStream<String> stream = AsyncDataStream.unorderedWait(source, new HBaseAsyncFunction());
  stream.print();
}

For ListenableFuture

Example2.java

import com.google.common.util.concurrent.FutureCallback;
import com.google.common.util.concurrent.Futures;
import com.google.common.util.concurrent.MoreExecutors;
import com.google.common.util.concurrent.ListenableFuture;
 
public class HBaseAsyncFunction implements AsyncFunction<String, String> {
  // initialize it while reading object
  transient Connection connection;
 
  @Override
  public void asyncInvoke(String val, AsyncCollector<String> c) {
    Get get = new Get(Bytes.toBytes(val));
    Table ht = connection.getTable(TableName.valueOf(Bytes.toBytes("test")));
 
    ListenableFuture<Result> future = ht.asyncGet(get);
    Futures.addCallback(future,
      new FutureCallback<Result>() {
        @Override public void onSuccess(Result result) {
          List ret = new ArrayList<String>();
          ret.add(result.get(...));
          c.collect(ret);
        }
 
        @Override public void onFailure(Throwable t) {
          c.collect(t);
        }
      },
      MoreExecutors.newDirectExecutorService()
    );
  }
}
  
// create data stream
public void createHBaseAsyncTestStream(StreamExecutionEnvironment env) {
  DataStream<String> source = getDataStream(env);
  DataStream<String> stream = AsyncDataStream.unorderedWait(source, new HBaseAsyncFunction());
  stream.print();
}
 

源码完整DEMO

flinkflink-examplesflink-examples-streamingsrcmainjavaorgapacheflinkstreamingexamplesasyncAsyncIOExample.java

/*
 * 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.streaming.examples.async;

import org.apache.flink.api.common.functions.FlatMapFunction;
import org.apache.flink.api.java.tuple.Tuple2;
import org.apache.flink.api.java.utils.ParameterTool;
import org.apache.flink.configuration.Configuration;
import org.apache.flink.runtime.state.filesystem.FsStateBackend;
import org.apache.flink.streaming.api.CheckpointingMode;
import org.apache.flink.streaming.api.TimeCharacteristic;
import org.apache.flink.streaming.api.checkpoint.ListCheckpointed;
import org.apache.flink.streaming.api.datastream.AsyncDataStream;
import org.apache.flink.streaming.api.datastream.DataStream;
import org.apache.flink.streaming.api.environment.StreamExecutionEnvironment;
import org.apache.flink.streaming.api.functions.async.AsyncFunction;
import org.apache.flink.streaming.api.functions.async.ResultFuture;
import org.apache.flink.streaming.api.functions.async.RichAsyncFunction;
import org.apache.flink.streaming.api.functions.source.SourceFunction;
import org.apache.flink.util.Collector;

import org.slf4j.Logger;
import org.slf4j.LoggerFactory;

import java.util.ArrayList;
import java.util.Collections;
import java.util.List;
import java.util.Random;
import java.util.concurrent.ExecutorService;
import java.util.concurrent.Executors;
import java.util.concurrent.TimeUnit;

/**
 * Example to illustrates how to use {@link AsyncFunction}.
 */
public class AsyncIOExample {

    private static final Logger LOG = LoggerFactory.getLogger(AsyncIOExample.class);

    private static final String EXACTLY_ONCE_MODE = "exactly_once";
    private static final String EVENT_TIME = "EventTime";
    private static final String INGESTION_TIME = "IngestionTime";
    private static final String ORDERED = "ordered";

    /**
     * A checkpointed source.
     */
    private static class SimpleSource implements SourceFunction<Integer>, ListCheckpointed<Integer> {
        private static final long serialVersionUID = 1L;

        private volatile boolean isRunning = true;
        private int counter = 0;
        private int start = 0;

        @Override
        public List<Integer> snapshotState(long checkpointId, long timestamp) throws Exception {
            return Collections.singletonList(start);
        }

        @Override
        public void restoreState(List<Integer> state) throws Exception {
            for (Integer i : state) {
                this.start = i;
            }
        }

        public SimpleSource(int maxNum) {
            this.counter = maxNum;
        }

        @Override
        public void run(SourceContext<Integer> ctx) throws Exception {
            while ((start < counter || counter == -1) && isRunning) {
                synchronized (ctx.getCheckpointLock()) {
                    ctx.collect(start);
                    ++start;

                    // loop back to 0
                    if (start == Integer.MAX_VALUE) {
                        start = 0;
                    }
                }
                Thread.sleep(10L);
            }
        }

        @Override
        public void cancel() {
            isRunning = false;
        }
    }


    /**
     * An sample of {@link AsyncFunction} using a thread pool and executing working threads
     * to simulate multiple async operations.
     *
     * <p>For the real use case in production environment, the thread pool may stay in the
     * async client.
     */
    private static class SampleAsyncFunction extends RichAsyncFunction<Integer, String> {
        private static final long serialVersionUID = 2098635244857937717L;

        private static ExecutorService executorService;
        private static Random random;

        private int counter;

        /**
         * The result of multiplying sleepFactor with a random float is used to pause
         * the working thread in the thread pool, simulating a time consuming async operation.
         */
        private final long sleepFactor;

        /**
         * The ratio to generate an exception to simulate an async error. For example, the error
         * may be a TimeoutException while visiting HBase.
         */
        private final float failRatio;

        private final long shutdownWaitTS;

        SampleAsyncFunction(long sleepFactor, float failRatio, long shutdownWaitTS) {
            this.sleepFactor = sleepFactor;
            this.failRatio = failRatio;
            this.shutdownWaitTS = shutdownWaitTS;
        }

        @Override
        public void open(Configuration parameters) throws Exception {
            super.open(parameters);

            synchronized (SampleAsyncFunction.class) {
                if (counter == 0) {
                    executorService = Executors.newFixedThreadPool(30);

                    random = new Random();
                }

                ++counter;
            }
        }

        @Override
        public void close() throws Exception {
            super.close();

            synchronized (SampleAsyncFunction.class) {
                --counter;

                if (counter == 0) {
                    executorService.shutdown();

                    try {
                        if (!executorService.awaitTermination(shutdownWaitTS, TimeUnit.MILLISECONDS)) {
                            executorService.shutdownNow();
                        }
                    } catch (InterruptedException e) {
                        executorService.shutdownNow();
                    }
                }
            }
        }

        @Override
        public void asyncInvoke(final Integer input, final ResultFuture<String> resultFuture) throws Exception {
            this.executorService.submit(new Runnable() {
                @Override
                public void run() {
                    // wait for while to simulate async operation here
                    long sleep = (long) (random.nextFloat() * sleepFactor);
                    try {
                        Thread.sleep(sleep);

                        if (random.nextFloat() < failRatio) {
                            resultFuture.completeExceptionally(new Exception("wahahahaha..."));
                        } else {
                            resultFuture.complete(
                                Collections.singletonList("key-" + (input % 10)));
                        }
                    } catch (InterruptedException e) {
                        resultFuture.complete(new ArrayList<String>(0));
                    }
                }
            });
        }
    }

    private static void printUsage() {
        System.out.println("To customize example, use: AsyncIOExample [--fsStatePath <path to fs state>] " +
                "[--checkpointMode <exactly_once or at_least_once>] " +
                "[--maxCount <max number of input from source, -1 for infinite input>] " +
                "[--sleepFactor <interval to sleep for each stream element>] [--failRatio <possibility to throw exception>] " +
                "[--waitMode <ordered or unordered>] [--waitOperatorParallelism <parallelism for async wait operator>] " +
                "[--eventType <EventTime or IngestionTime>] [--shutdownWaitTS <milli sec to wait for thread pool>]" +
                "[--timeout <Timeout for the asynchronous operations>]");
    }

    public static void main(String[] args) throws Exception {

        // obtain execution environment
        StreamExecutionEnvironment env = StreamExecutionEnvironment.getExecutionEnvironment();

        // parse parameters
        final ParameterTool params = ParameterTool.fromArgs(args);

        final String statePath;
        final String cpMode;
        final int maxCount;
        final long sleepFactor;
        final float failRatio;
        final String mode;
        final int taskNum;
        final String timeType;
        final long shutdownWaitTS;
        final long timeout;

        try {
            // check the configuration for the job
            statePath = params.get("fsStatePath", null);
            cpMode = params.get("checkpointMode", "exactly_once");
            maxCount = params.getInt("maxCount", 100000);
            sleepFactor = params.getLong("sleepFactor", 100);
            failRatio = params.getFloat("failRatio", 0.001f);
            mode = params.get("waitMode", "ordered");
            taskNum = params.getInt("waitOperatorParallelism", 1);
            timeType = params.get("eventType", "EventTime");
            shutdownWaitTS = params.getLong("shutdownWaitTS", 20000);
            timeout = params.getLong("timeout", 10000L);
        } catch (Exception e) {
            printUsage();

            throw e;
        }

        StringBuilder configStringBuilder = new StringBuilder();

        final String lineSeparator = System.getProperty("line.separator");

        configStringBuilder
            .append("Job configuration").append(lineSeparator)
            .append("FS state path=").append(statePath).append(lineSeparator)
            .append("Checkpoint mode=").append(cpMode).append(lineSeparator)
            .append("Max count of input from source=").append(maxCount).append(lineSeparator)
            .append("Sleep factor=").append(sleepFactor).append(lineSeparator)
            .append("Fail ratio=").append(failRatio).append(lineSeparator)
            .append("Waiting mode=").append(mode).append(lineSeparator)
            .append("Parallelism for async wait operator=").append(taskNum).append(lineSeparator)
            .append("Event type=").append(timeType).append(lineSeparator)
            .append("Shutdown wait timestamp=").append(shutdownWaitTS);

        LOG.info(configStringBuilder.toString());

        if (statePath != null) {
            // setup state and checkpoint mode
            env.setStateBackend(new FsStateBackend(statePath));
        }

        if (EXACTLY_ONCE_MODE.equals(cpMode)) {
            env.enableCheckpointing(1000L, CheckpointingMode.EXACTLY_ONCE);
        }
        else {
            env.enableCheckpointing(1000L, CheckpointingMode.AT_LEAST_ONCE);
        }
        env.getConfig().setTracingMetricsEnabled(true);

        // enable watermark or not
        if (EVENT_TIME.equals(timeType)) {
            env.setStreamTimeCharacteristic(TimeCharacteristic.EventTime);
        }
        else if (INGESTION_TIME.equals(timeType)) {
            env.setStreamTimeCharacteristic(TimeCharacteristic.IngestionTime);
        }

        // create input stream of an single integer
        DataStream<Integer> inputStream = env.addSource(new SimpleSource(maxCount));

        // create async function, which will *wait* for a while to simulate the process of async i/o
        AsyncFunction<Integer, String> function =
                new SampleAsyncFunction(sleepFactor, failRatio, shutdownWaitTS);

        // add async operator to streaming job
        DataStream<String> result;
        if (ORDERED.equals(mode)) {
            result = AsyncDataStream.orderedWait(
                inputStream,
                function,
                timeout,
                TimeUnit.MILLISECONDS,
                20).setParallelism(taskNum);
        }
        else {
            result = AsyncDataStream.unorderedWait(
                inputStream,
                function,
                timeout,
                TimeUnit.MILLISECONDS,
                20).setParallelism(taskNum);
        }

        // add a reduce to get the sum of each keys.
        result.flatMap(new FlatMapFunction<String, Tuple2<String, Integer>>() {
            private static final long serialVersionUID = -938116068682344455L;

            @Override
            public void flatMap(String value, Collector<Tuple2<String, Integer>> out) throws Exception {
                out.collect(new Tuple2<>(value, 1));
            }
        }).keyBy(0).sum(1).print();

        // execute the program
        env.execute("Async IO Example");
    }
}