This is an unfinished paper. I am writing. After finish I will send e-mail to dev@flink.apache.org.

Status

Current state: Under Discussion

Discussion thread: here (<- link to https://mail-archives.apache.org/mod_mbox/flink-dev/)

JIRA:  Unable to render Jira issues macro, execution error.

Released: <Flink Version>

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

Motivation

In Flink batch job, the job is usually divided into multiple parallel tasks executed cross many nodes in the cluster. It is common to encounter the performance degradation on some nodes due to hardware problems, or accident I/O busy, or high CPU load. This kind of degradation can probably cause the running tasks on the node to be quite slow, that is so called long tail tasks. Although the long tail tasks will not fail, they can severely affect the total job running time. Flink task scheduling does not take this long tail problem into account currently.

Here we propose the speculative execution strategy FLINK-10644 to handle the problem. The basic idea is to run a copy of task on another node when the original task is identified to be long tail. The speculative task is executed in parallel with the original one and share the same failure retry mechanism. Once either task complete, the scheduler admits its output as the final result and cancel the other running one. I will introduce a blacklist module to schedule the long tail task on different machine from the original task. And modify FileOutputFormat.java to adapter speculative execution mechanism.

The preliminary experiments have demonstrated the effectiveness in our product cluster.

Proposed Changes

General design

Detection of Long Tail Tasks

A task will be classified as a long tail task when it meets the following three rules.

Finished Tasks Percentage

In one ExecutionJobVertex, when a configurable percentage of executions has been finished, the speculative execution thread begin really work.

Long Running Time

In speculative execution thread all executions' interval between the current time and it first create/deploying time before it failover in one ExecutionJobVertex are calculated. when the running time of a execution is greater than a configurable multiple of the median of the running time of other finished executions, this execution is defined as long tail execution.

Slow Processing Throughput

In our product cluster of Alibaba in China, we only considered running time could solve the long tail task problem sufficiently. So first version I will not take throughput into consider. After release this feature if Flink community if users have this requirement I will delevop next version. 

Scheduling of Speculative Executions

Refactoring of ExecutionVertex

Because of introduction speculative execution, the ExecutionVertex can't default to only one execution is running at the same time. Some member-variables in ExecutionVertex need bigger refactor.

There are two ways of code refactoring:

1. Add a member-variable named speculativeExecution that similar to currentExecution. But this way will lead to lots of if judgments in the scheduler and failover code. Also this way will reduce the flexible of code if there are more than two executions existed at the same time. 
2. Change the currentExecution in ExecutionVertex to an ArrayList named executionList which means there can be multiple executions in an ExecutionVertex at the same time. For each execution in executionList there is no difference of the behavior such as failover, request slot etc between them.

private List<Execution> executionList = new ArrayList<>();

Introduction SpeculativeScheduler thread

In each ExecutionJobVertex there is a SpeculativeScheduler thread used for detect the long tail execution in this ExecutionJobVertex periodically based on above rules. Its member-variables included SchedulerNG and ExecutionVertex[] etc. SchedulerNG used for scheduling the speculative executions and the ExecutionVertex[] used for get execution state timestamp in this ExecutionJobVertex.

Modify of scheduler logical

The scheduler should scheduling an execution according to the index of this execution in executionList instead of default to scheduling the currentExecution. So we need change ExecutionVertexID to ExecutionVertexIDAndExecutionIndex which represent which execution in ExecutionVertex should be scheduling in scheduler logical. Also when task failover, executionIndex also should be calculated by fail task's ExecutionAttemptID so that scheduler knows which execution in executionList should be restart.

public class ExecutionVertexIDWithExecutionIndex {
    private ExecutionVertexID executionVertexID;
    private Integer executionIndex;
}

In order to reuse the code of scheduler, we need to extend the interface with an additional method then SchedulerBase should implements it.

public interface SchedulerNG extends AutoCloseableAsync {
	default void schedulingSpeculativeExecutions(List<ExecutionVertexIDWithExecutionIndex> verticesToSchedule) {
        throw new UnsupportedOperationException();
    }
}

Process failover situation

Just like the normal tasks, the speculative task is executed in parallel with the original one and share the same failover and restart strategy.

The original long tail task and speculative tasks can still retry with failure on their own track.

But I think it should not be restarted globally when the speculation execution failover count reach the max-retry-counts. 

When a task fail, we could calculate its index(executionIndex) in executionList by executionAttemptID. Then the scheduler takes a series of processing for the corresponding execution according to the executionIndex as shown below.

In order to better failover logic, I will extend the calss FailureHandlingResult with an additional member-variable.

public class FailureHandlingResult {
	@Nullable private final Integer executionIndex;
}

Black list

Most long tail tasks are caused by machine problems, so the speculative execution must runs on a different machine from origin execution.

I will introduce blacklist module into Flink used for filter node when the speculative executions are request slots.

basic plan

Blacklist is a kind of scheduling constraint. According to FLINK-11000 description this is a bigger feature.

There are several levels of blacklist, including (JobVertex, TaskManager) blacklist, (JobVertex Host) blacklist, (Job, TaskManager) blacklist, (Job, Host) blacklist, (Session, TaskManager) blacklist and (Session Host) blacklist.

I must implement (Job, Host) blacklist for speculative execution feature. In order to implement FLINK-11000  friendly in the future, my interface also suit other blacklist descripted above.

The blacklist module is a thread that maintains the black machines of this job and removes expired elements periodically. Each element in blacklist contains IP and timestamp. The timestamp is used to decide whether the elements of the blacklist is expired or not. Since request resource logical both existed in JobMaster and ResourceManager side, so both JobMaster and ResourceManager need to get the blacklist information for filtering black node. But beacuse I only implement the (Job, Host) level blacklist in this feature, so I only consider maintaining the blacklist on the JobMaster side.

On the JobMaster side, the blacklist is encapsulated as an independent system component, which updates the blacklist by receiving the information of speculative execution happened, and filters resources through the blacklist when request resource.

On the ResourceManager side, it does not maintain the blacklist temporarily. When JobMaster request resources from ResourceManager, it pass the blacklist information to ResourceManager. ResourceManager needs to consider blacklist when request or filters resources and blacklist information will be encapsulated in the form of filtering resources required by external resource management system(such as yarn) when request new container.

Classes and Interfaces of (Job, Host) blacklist

（1）Abstract Class BlackList, each type of blacklist will extends it.

public abstract class BlackList implements Serializable {
	/** Black list configuration. */
	protected final BlackListConfig blackListConfig;

	public BlackList(BlackListConfig blackListConfig) {
        this.blackListConfig = blackListConfig;
    }

    /**
     * Remove time out black list records.
     * @param timeout Time out time.
     * @return Minimum timestamp of black list records.
     */
    public abstract long removeTimeoutBlackList(long timeout);

    /** Clear black list. */
    public abstract void clear();

	/** Is black list empty. */
    public abstract boolean isEmpty();
}

（2）JobBlackList, the job-level black list.

public class JobBlackList extends BlackList {

	/** The list of the black ip. This list is mainly for time out checking. */
    private final Queue<BlackListRecord> jobToIpBlackListRecords;

	/** The set of the black ip. This set is mainly for black host filter. */
    private final Set<String> jobBlackIpSet;

	public JobBlackList(BlackListConfig blackListConfig) {
        super(blackListConfig);
        this.jobToIpBlackListRecords = new ConcurrentLinkedQueue<>();
        this.jobBlackIpSet = Collections.newSetFromMap(new ConcurrentHashMap<>());
    }
	
	/** Add a ip to this black list. */
	public boolean addIpToBlackList(String ip) {}

	public Set<String> getAllBlackListIpSet() {}

	/** clear (job, ip) blacklist. */
	@Override
    public void clear() {}

	@Override
    public long removeTimeoutBlackList(long timeout) {}

	@Override
    public boolean isEmpty() {}

	/** Whether the given host has been added to black list. */
	public boolean containsIp(String ip) {}
}

（3）BlackListRecord， the item of blackList

public class BlackListRecord implements Serializable {
	/**
 	 * Black list record which stores the black host ip and
 	 * the time stamp when this record is added.
 	 */
	public class BlackListRecord implements Serializable {}

	/** The black host ip. */
    private final String ip;

    /** The time stamp when this black list record is added. */
    private final long timeStamp;
}

（4）A abstract Class called BlackListTracker, a thread that maintain blacklist info

public abstract class BlackListTracker implements Runnable {

	/** The executor to run the time out checking task. */
    private final ScheduledExecutor executor;
	
	/** The black list configuration. */
    protected final BlackListConfig blackListConfig;

	/** The black list timeout check future, will be canceled when black black list destroyed. */
    private AtomicReference<ScheduledFuture> timeoutCheckFuture;

	public BlackListTracker(ScheduledExecutor executor, BlackListConfig blackListConfig) {
        Preconditions.checkNotNull(blackListConfig);
        this.executor = executor;
        this.blackListConfig = blackListConfig;
        this.timeoutCheckFuture = new AtomicReference<>();
    }

	/**
     * Given the minimum time stamp of black list record. The function schedules a task to remove the black list
     * record when it got timeout.
     * @param minTimeStamp The minimum time stamp of black list record.
     */
    public void scheduleTimeOutCheck(long minTimeStamp) {}

	public boolean isBlackListEnabled() {
        return blackListConfig.isBlackListEnabled();
    }

	/** Clear the black list. */
    public abstract void clearBlackList();

    /** Get all black list ip. */
    public abstract Set<String> getAllBlackListIp();

	/** Clear the black list and cancel the timeout check task. */
    public void destroy() {}
}

（5）JobBlackListSpeculativeListener, event listener

/** Listener to be notified when speculative execution happened. */
public interface JobBlackListSpeculativeListener {
    /**
     * When a speculative execution happened, the listener will be notified.
     * @param ip the ip
     */
    void onSpeculativeExecutionHappened(String ip);
}

（5）JobBlackListTracker, per-job blackList tracker

public class JobBlackListTracker extends BlackListTracker implements JobBlackListSpeculativeListener {

	/** The black list of this job. */
    private final JobBlackList jobBlackList;

	public JobBlackListTracker(ScheduledExecutor executor, BlackListConfig blackListConfig) {
        super(executor, blackListConfig);
        this.jobBlackList = new JobBlackList(blackListConfig);
        scheduleTimeOutCheck(blackListConfig.getBlackListTimeOutInMillis());
    }

	@Override
    public void clearBlackList() {
        jobBlackList.clear();
    }

	@Override
    public Set<String> getAllBlackListIp() {
        if (blackListConfig.isBlackListEnabled()) {
            return jobBlackList.getAllBlackListIpSet();
        }
        return Collections.emptySet();
    }

	/** The time out checking task to be scheduled. /
    @Override
    public void run() {
        long minTimeStamp = jobBlackList.removeTimeoutBlackList(blackListConfig.getBlackListTimeOutInMillis());
        scheduleTimeOutCheck(minTimeStamp);
    }

    @Override
    public void onSpeculativeExecutionHappened(String ip) {
        if (blackListConfig.isBlackListEnabled()) {
            jobBlackList.addIpToBlackList(ip);
        }
    }
}

Init black list 

In DefaultExecutionGraph I will add a member-variable jobBlackListSpeculativeListeners and when after create ExecutionGraph jobBlackListTracker will be add in jobBlackListSpeculativeListeners. JobBlackListSpeculativeListener.onSpeculativeExecutionHappened() will be called when the SpeculativeExecution detected a long tail task and start notify scheduler to scheduling a speculative execution.

public interface ExecutionGraph extends AccessExecutionGraph {
	void registerJobBlackListSpeculativeListener(JobBlackListTracker jobBlackListTracker);
}

public class DefaultExecutionGraph implements ExecutionGraph, InternalExecutionGraphAccessor {

	private final List<JobBlackListSpeculativeListener> jobBlackListSpeculativeListeners;
	
	public DefaultExecutionGraph() {
		this.jobBlackListSpeculativeListeners = new ArrayList<>();
	}
	
	@Override
    public void registerJobBlackListSpeculativeListener(JobBlackListTracker listener) {
        if (listener != null) {
            jobBlackListSpeculativeListeners.add(listener);
        }
    }

	@Override
    public List<JobBlackListSpeculativeListener> getJobBlackListSpeculativeListeners() {
        return jobBlackListSpeculativeListeners;
    }
}

Add element to black list

First JobBlackListSpeculativeListener.onSpeculativeExecutionHappened() will be called when the SpeculativeExecution detected a long tail task and start notify scheduler to scheduling a speculative execution.

Second ExecutionGraph will notify listener(JobBlackListTracker) it. 

Third JobBlackList in JobBlackListTracker will be added ip of original execution location.

Remove element in black list

The BlackListTracker has implement Runnable that responsibility is maintains the black machines of this job and removes expired elements periodically. Each element in blacklist contains IP and timestamp. The timestamp is used to decide whether each element in the blacklist is expired or not. 


Pass the blacklist information to cluster ResourceManager

Yarn

First, nodes' attributes should include machine ip attribute, then we can control containers do not on some mechines by yarn PlacementConstraints.

Now Flink use hadoop-2.x and request container by ResourceRequest api, it don't support PlacementConstraints.

So for use PlacementConstraints, I will introduce hadoop-3.x SchedulingRequest api by java reflect mechanism.

When the executions are scheduled, I will add blacklist information to yarn PlacementConstraint. In this way, I can ensure that the yarn container is not on the machines in the blacklist.

Kubernetes

We could achieve the same goal with Yarn PlacementConstraint in K8s integration by node-affinity if nodes had attached ip label.

The same as yarn, when the executions are scheduled, I will add blacklist information to k8s PodSpec.

Mesos

According to FLINK-22352, the community has decided to deprecate Mesos support in Apache Flink.

So we don’t need to think about how to pass the blacklist information to mesos.

Manage input and output of ExecutionVertex

Manage InputSplit

After I read FLINK-10205, pr-6684 and code, I found that Flink now can't ensure the different attempt of a same ExecutionVertex to have the same inputs. Because when a task failover, now simply returning the input splits to the assigner and letting the next idling task take it should work. This is no problem because it should not matter which tasks processes which input split. If a failure occurs and some other task takes over the failed input splits, it would as if this task had processed these input splits from the very beginning.

But because of introduce speculative execution, we must ensure that the InputSplits processed by speculative execution is the same as the original execution. So I will add a Map in ExecutionVertex that key indicates ExecutionAttemptId and value indicates the index of InputSplit that this execution currently consumes.

public class ExecutionVertex
        implements AccessExecutionVertex, Archiveable<ArchivedExecutionVertex> {
	private final ArrayList<InputSplit> inputSplits;
    private final Map<ExecutionAttemptID, Integer> inputSplitIndexMap;
}

When execution failover, in resetForNewExecutionInternal(), we clear up the information related to failover execution in inputSplitIndexMap instead of call returnInputSplit() now.

For example, as shown below.

(a) In a ExecutionVertex, after execution_1 has consumed inputSplit_0, it go on to consume inputSplit_1.

      Now inputSplitIndexMap data is { (execution_1, 1) }.

(b) Specutative execution_2 has inited, it will consume inputSplit_0 in inputSplits first.

      Now inputSplitIndexMap data is { (execution_1, 1), (execution_2, 0) }.

(c) A failure occurs of execution_1, then resetForNewExecutionInternal() in ExecutionVertex will be called.

      Now inputSplitIndexMap data is { (execution_2, 0) }.

(d) execution_2 consumed inputSplit_0 finished, it go on to consume inputSplit_1. And after execution_1 failover, execution_1_new occurs.

      Now inputSplitIndexMap data is { (execution_1_new, 0), (execution_2, 1) }.

Manage middle ResultPartition 

I think that speculative execution means that multiple executions in a ExecutionVertex running at the same time, and failover means that two executions running at two different time.

Multiple executions of a upstream ExecutionVertex will produce Multiple ResultPartitions. When all upstream ExecutionVertexs run finished, the inputChannel of down stream executions will be updated to consume the fastest finished execution of upstream. So add a member-variable named fastestFinishedExecution in ExecutionVertex used for create PartitionInfo which used for update down stream executions' inputChannels. 

public class ExecutionVertex
        implements AccessExecutionVertex, Archiveable<ArchivedExecutionVertex> {
	private Execution fastestFinishedExecution = null;
	private volatile boolean vertexFinished = false;
}

Some code should be modifyed to avoid cause problems when multiple executions in one ExecutionVertex finished at the same time. 

A member-variable named vertexFinished was added in ExecutionVertex used for indicates whether this ExecutionVertex run finished and double checked locking pattern in ExecutionVertex.executionFinished().  

finishPartitionsAndUpdateConsumers() in Execution.markFinished() should be moved to ExecutionVertex.executionFinished() and be called after double checked locking pattern in ExecutionVertex.executionFinished()

when the down stream execution meet DataConsumptionException. It will restart with the upstream execution that has been finished.

If a task read from a blocking result partition, when its input is not available, we can ‘revoke’ the produce task, set the task fail and rerun the upstream task to regenerate data.

In certain scenarios producer data was transferred in blocking mode or data was saved in persistent store. If the partition was missing, we need to revoke/rerun the produce task to regenerate the data.

这里要画一张图。把failover的情况加进去，特别是下游带着上游重启的情况。

Manage sink files

这个就是对于batch的sink file，到每个文件后面加全局唯一的后缀，然后最后在作业全部结束的时候，在主节点对于各种情况，去rename或者delete.

Metrics

We can use six metrics to measure and evaluate the efficiency of speculative execution that can be summarized in the job status and Web page.

// SpeculationRatio is defined as the proportion of speculative tasks in the total tasks of a Job, and SuccessfulSpeculationRatio is the percent of speculative task which finish before the original long tail ones in the total launched speculative tasks. 

(1)
// 必须要多少execution完成了，才开始预测执行的metrics
private AtomicInteger minFinishedForSpeculationThresholdMetrics;

(2)
// 已经完成的execution数的metrics
private AtomicInteger finishedExecutionCountMetrics;

(3)
// 有多少execution发生了预测执行的metrics
private Counter speculativeExecutionCountMetrics;

(4)
// execution运行时间维度，发生预测执行的阈值
private AtomicDouble speculativeThresholdOfTime;

(5)
// 每个ExecutionVertex中原execution的运行时间的metrics
private Map<String, AtomicDouble> executionIndex2RunningTimespan;

(6)

// 如果最快结束的execution是预测执行的execution，那么对相应的metrics进行汇报

private Counter speculativeExecutionFastestFinishedCountMetrics;

Web UI

todo...

Limitations

（1）JobType is Batch.

（2）Cluster ResourceManager is Yarn or K8s.

（3）input split的限制？

（4）yarn要对ip打标

（5）k8s的限制？

（6）The JobVertex with all input edges and all output edges are blocking could enable speculative execution feature.

Configuration

All configurations related to this feature are added in JobManagerOptions.class.

public static final ConfigOption<Boolean> SPECULATIVE_EXECUTION_ENABLED =
        key("flink.batch.speculative.enabled")
                .booleanType()
                .defaultValue(false)
                .withDescription("Whether to enable speculation of batch job.");

public static final ConfigOption<Long> SPECULATIVE_EXECUTION_INTERVAL_IN_MILLIS =
        key("flink.batch.speculative.interval.in.millis")
                .longType()
                .defaultValue(100L)
                .withDescription("How often to check for speculative tasks.");

public static final ConfigOption<Double> SPECULATIVE_EXECUTION_MULTIPLIER =
        key("flink.batch.speculative.multiplier")
                .doubleType()
                .defaultValue(1.5)
                .withDescription("When the running time of a unfinished executionVertex is several times of" +
                        " the median of all completed executionVertexs, it will be speculated.");

public static final ConfigOption<Double> SPECULATIVE_EXECUTION_QUANTILE =
        key("flink.batch.speculative.quantile")
                .doubleType()
                .defaultValue(0.75)
                .withDescription("When the percentage of ExecutionVertex in an executionJobVertex are completed," +
                        " the speculative execution mechanism will be started. 0.9 means that 90% of the" +
                        " executionVertex are completed, and the speculative execution mechanism will be started.");

public static final ConfigOption<Long> SPECULATIVE_EXECUTION_LOG_INTERVAL_IN_MILLIS =
        key("flink.batch.speculative.log.interval.in.millis")
                .longType()
                .defaultValue(1000 * 60 * 5L)
                .withDescription("Interval in millis for speculative related log.");

public static final ConfigOption<Boolean> BLINK_BLACKLIST_ENABLE =
        key("flink.blacklist.enable")
                .booleanType()
                .defaultValue(false)
                .withDescription("Whether to enable blacklist job.");

public static final ConfigOption<Long> BLINK_BLACKLIST_TIMEOUT_IN_MILLIS =
        key("flink.blacklist.timeout.in.millis")
                .longType()
                .defaultValue(60 * 1000L)
                .withDescription("indicates how long a black list record will be removed after added to black list.");

Compatibility, Deprecation, and Migration Plan

• What impact (if any) will there be on existing users? 
• If we are changing behavior how will we phase out the older behavior? 
• If we need special migration tools, describe them here.
• When will we remove the existing behavior?

Test Plan

Describe in few sentences how the FLIP will be tested. We are mostly interested in system tests (since unit-tests are specific to implementation details). How will we know that the implementation works as expected? How will we know nothing broke?

Rejected Alternatives

If there are alternative ways of accomplishing the same thing, what were they? The purpose of this section is to motivate why the design is the way it is and not some other way.