flink-tests/src/test/java/org/apache/flink/test/checkpointing/ApproximateLocalRecoveryDownstreamITCase.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.test.checkpointing;

import org.apache.flink.api.common.functions.RichMapFunction;
import org.apache.flink.api.common.restartstrategy.RestartStrategies;
import org.apache.flink.configuration.Configuration;
import org.apache.flink.configuration.MemorySize;
import org.apache.flink.configuration.TaskManagerOptions;
import org.apache.flink.runtime.state.KeyGroupRangeAssignment;
import org.apache.flink.runtime.testutils.MiniClusterResourceConfiguration;
import org.apache.flink.streaming.api.environment.StreamExecutionEnvironment;
import org.apache.flink.streaming.api.functions.sink.RichSinkFunction;
import org.apache.flink.streaming.api.functions.source.RichParallelSourceFunction;
import org.apache.flink.test.util.MiniClusterWithClientResource;
import org.apache.flink.test.util.SuccessException;
import org.apache.flink.util.TestLogger;

import org.junit.Ignore;
import org.junit.Rule;
import org.junit.Test;
import org.junit.rules.Timeout;

import java.util.Arrays;
import java.util.Collections;

import static org.apache.flink.test.util.TestUtils.tryExecute;

/**
 * To test approximate downstream failover.
 *
 * <p>If a task fails, all its downstream tasks restart, including itself. - Test the failed task
 * can reconnect successfully. - Test partial records are cleaned up correctly. - Test only
 * downstream are tasks restart.
 */
@Ignore("Approximate local recovery has currently no scheduler support")
public class ApproximateLocalRecoveryDownstreamITCase extends TestLogger {
    private static final int BUFFER_SIZE = 4096;

    @Rule
    public MiniClusterWithClientResource cluster =
            new MiniClusterWithClientResource(
                    new MiniClusterResourceConfiguration.Builder()
                            .setConfiguration(createConfig())
                            .setNumberTaskManagers(4)
                            .setNumberSlotsPerTaskManager(1)
                            .build());

    @Rule public final Timeout timeout = Timeout.millis(300000L);

    /**
     * Test the following topology.
     *
     * <pre>
     *     (source1/1) -----> (map1/1) -----> (sink1/1)
     * </pre>
     *
     * <p>(map1/1) fails, (map1/1) and (sink1/1) restart
     */
    @Test
    public void localTaskFailureRecoveryThreeTasks() throws Exception {
        final int failAfterElements = 150;
        final StreamExecutionEnvironment env = StreamExecutionEnvironment.getExecutionEnvironment();
        env.setParallelism(1)
                .setBufferTimeout(0)
                .setMaxParallelism(128)
                .disableOperatorChaining()
                .setRestartStrategy(RestartStrategies.fixedDelayRestart(1, 0));
        env.getCheckpointConfig().enableApproximateLocalRecovery(true);

        env.addSource(new AppSourceFunction())
                .slotSharingGroup("source")
                .map(new FailingMapper<>(failAfterElements))
                .slotSharingGroup("map")
                .addSink(new ValidatingAtMostOnceSink(300))
                .slotSharingGroup("sink");

        FailingMapper.failedBefore = false;
        tryExecute(env, "testThreeTasks");
    }

    /**
     * Test the following topology.
     *
     * <pre>
     *     (source1/1) -----> (map1/2) -----> (sink1/1)
     *         |                                ^
     *         -------------> (map2/2) ---------|
     * </pre>
     *
     * <p>(map1/2) fails, (map1/2) and (sink1/1) restart
     */
    @Test
    public void localTaskFailureRecoveryTwoMapTasks() throws Exception {
        final int failAfterElements = 20;
        final int keyByChannelNumber = 2;
        final StreamExecutionEnvironment env = StreamExecutionEnvironment.getExecutionEnvironment();
        env.setParallelism(1)
                .setBufferTimeout(0)
                .disableOperatorChaining()
                .setMaxParallelism(128)
                .setRestartStrategy(RestartStrategies.fixedDelayRestart(1, 0));
        env.getCheckpointConfig().enableApproximateLocalRecovery(true);

        env.addSource(
                        new AppSourceFunction(
                                BUFFER_SIZE, env.getMaxParallelism(), keyByChannelNumber))
                .slotSharingGroup("source")
                .keyBy(InvertedKeyTuple::key)
                .map(new FailingMapper<>(failAfterElements))
                .setParallelism(keyByChannelNumber)
                .slotSharingGroup("map")
                .addSink(new ValidatingAtMostOnceSink(200, keyByChannelNumber))
                .slotSharingGroup("sink");

        FailingMapper.failedBefore = false;
        tryExecute(env, "testTwoMapTasks");
    }

    /**
     * source function used to generate InvertedKeyTuple.
     *
     * <p>InvertedKeyTuple includes four fields: - index: increased upon each tuple is generated. -
     * key: based on which the tuple to be partitioned for downstream consumption (a map, as shown
     * in the testcase for example). - selectedChannel: the selected channel (the selected instance
     * of the downstream operator) the tuple to be sent to. Usually, selectedChannel is decided by
     * the hash-partition function applied on the key. However in this case, we invert the process
     * by selecting the channel to be sent first, and then decide a possible key the tuple may
     * contain. In this way, we can control exactly how many tuples are sent to each instance of the
     * downstream operator. - longOrShortString: either long or short string used to test network
     * transition upon reconnection.
     *
     * <p>AppSourceFunction makes sure tuples generated are evenly distributed amongst downstream
     * operator instances.
     */
    private static class AppSourceFunction extends RichParallelSourceFunction<InvertedKeyTuple> {
        private final String shortString =
                "I am a very long string to test partial records hohoho hahaha ";
        private final String longOrShortString;
        private final int maxParallelism;
        private final int numberOfChannels;
        private final int[] keys;
        private int index = 0;
        private volatile boolean running = true;

        // short-length string
        AppSourceFunction() {
            this.longOrShortString = shortString;
            this.maxParallelism = 128;
            this.numberOfChannels = 1;
            this.keys = initKeys(numberOfChannels);
        }

        // long-length string
        AppSourceFunction(int bufferSize, int maxParallelism, int numberOfChannels) {
            this.maxParallelism = maxParallelism;
            this.numberOfChannels = numberOfChannels;
            this.keys = initKeys(numberOfChannels);

            StringBuilder builder = new StringBuilder(shortString);
            for (int i = 0; i <= 2 * bufferSize / shortString.length() + 1; i++) {
                builder.append(shortString);
            }
            this.longOrShortString = builder.toString();
        }

        @Override
        public void run(SourceContext<InvertedKeyTuple> ctx) throws Exception {
            while (running) {
                synchronized (ctx.getCheckpointLock()) {
                    if (index % 100 == 0) {
                        Thread.sleep(50);
                    }
                    int selectedChannel = index % numberOfChannels;
                    int key = keys[selectedChannel];
                    ctx.collect(
                            new InvertedKeyTuple(index, key, selectedChannel, longOrShortString));
                }
                index++;
            }
        }

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

        // Find the first key which falls into the key group of each downstream channel
        private int[] initKeys(int numberOfChannels) {
            int[] keys = new int[numberOfChannels];

            for (int i = 0; i < numberOfChannels; i++) {
                int key = 0;
                while (key < 1000 && selectedChannel(key) != i) {
                    key++;
                }
                assert key < 1000 : "Can not find a key within number 1000";
                keys[i] = key;
            }

            return keys;
        }

        private int selectedChannel(int key) {
            return KeyGroupRangeAssignment.assignKeyToParallelOperator(
                    key, maxParallelism, numberOfChannels);
        }
    }

    /** Fails once the first map instance reaches failCount. */
    private static class FailingMapper<T> extends RichMapFunction<T, T> {
        private static final long serialVersionUID = 6334389850158703L;

        private static volatile boolean failedBefore;

        private final int failCount;
        private int numElementsTotal;

        private boolean failer;

        FailingMapper(int failCount) {
            this.failCount = failCount;
        }

        @Override
        public void open(Configuration parameters) {
            failer = getRuntimeContext().getIndexOfThisSubtask() == 0;
        }

        @Override
        public T map(T value) throws Exception {
            numElementsTotal++;

            if (!failedBefore) {
                Thread.sleep(10);

                if (failer && numElementsTotal >= failCount) {
                    failedBefore = true;
                    throw new Exception("Artificial Test Failure");
                }
            }

            return value;
        }
    }

    /**
     * Validating sink to make sure each selectedChannel (of map) gets at least numElementsTotal
     * elements.
     *
     * <p>Notice that the source generates tuples evenly distributed amongst downstream operator
     * (map) instances. Besides, the index generated is continuous and monotonic increasing as long
     * as the source is not restarted. Hence, - In the case of approximate local recovery is NOT
     * enabled, where the entire job, including source, is restarted after map fails, source
     * regenerates tuples indexed from 0. Upon reach numElementsTotal, the maximal possible index is
     * numElementsTotal * numberOfInputChannels - 1 - In the case of approximate local recovery is
     * enabled, where only the downstream of the failed task restart, source does not regenerating
     * tuples, the maximal possible index > numElementsTotal * numberOfInputChannels - 1
     */
    private static class ValidatingAtMostOnceSink extends RichSinkFunction<InvertedKeyTuple> {
        private static final long serialVersionUID = 1748426382527469932L;
        private final int numElementsTotal;
        private final int[] numElements;
        private final Integer[] indexReachingNumElements;
        private final int numberOfInputChannels;

        ValidatingAtMostOnceSink(int numElementsTotal, int numberOfInputChannels) {
            this.numElementsTotal = numElementsTotal;
            this.numberOfInputChannels = numberOfInputChannels;
            this.numElements = new int[numberOfInputChannels];
            this.indexReachingNumElements = new Integer[numberOfInputChannels];
        }

        ValidatingAtMostOnceSink(int numElementsTotal) {
            this.numElementsTotal = numElementsTotal;
            this.numberOfInputChannels = 1;
            this.numElements = new int[numberOfInputChannels];
            this.indexReachingNumElements = new Integer[numberOfInputChannels];
        }

        @Override
        public void invoke(InvertedKeyTuple tuple, Context ctx) throws Exception {
            assert tuple.selectedChannel < numberOfInputChannels;
            numElements[tuple.selectedChannel]++;

            boolean allReachNumElementsTotal = true;
            for (int i = 0; i < numberOfInputChannels; i++) {
                if (numElements[i] == numElementsTotal) {
                    indexReachingNumElements[i] = tuple.index;
                } else if (numElements[i] < numElementsTotal) {
                    allReachNumElementsTotal = false;
                }
            }
            if (allReachNumElementsTotal) {
                assert Collections.max(Arrays.asList(indexReachingNumElements))
                        >= numElementsTotal * numberOfInputChannels;
                throw new SuccessException();
            }
        }
    }

    private static Configuration createConfig() {
        Configuration config = new Configuration();
        config.set(
                TaskManagerOptions.MEMORY_SEGMENT_SIZE,
                MemorySize.parse(Integer.toString(BUFFER_SIZE)));

        return config;
    }

    private static class InvertedKeyTuple {
        int index;
        /** Key based on which the tuple is partitioned. */
        int key;
        /** The selected channel of this tuple after key-partitioned. */
        int selectedChannel;

        String longOrShortString;

        InvertedKeyTuple(int index, int key, int selectedChannel, String longOrShortString) {
            this.index = index;
            this.key = key;
            this.selectedChannel = selectedChannel;
            this.longOrShortString = longOrShortString;
        }

        int key() {
            return key;
        }
    }
}