Speculative Execution

This page describes the background of speculative execution, how to use it, and how to check the effectiveness of it.

Background

Speculative execution is a mechanism to mitigate job slowness which is caused by problematic nodes. A problematic node may have hardware problems, accident I/O busy, or high CPU load. These problems may make the hosted tasks run much slower than tasks on other nodes, and affect the overall execution time of a batch job.

In such cases, speculative execution will start new attempts of the slow task on nodes that are not detected as problematic. The new attempts process the same input data and produces the same data as the old one. The old attempt will not be affected and will keep running. The first finished attempt will be admitted, its output will be seen and consumed by the downstream tasks, and the remaining attempts will be canceled.

To achieve this, Flink uses the slow task detector to detect slow tasks. The nodes that the slow tasks locate in will be identified as problematic nodes and get blocked via the blocklist mechanism. The scheduler will create new attempts for the slow tasks and deploy them on nodes that are not blocked.

Usage

This section describes how to use speculative execution, including how to enable it, how to tune it, and how to develop/improve custom sources to work with speculative execution.

Note: Flink does not support speculative execution of DataSet jobs because DataSet will be deprecated in near future. DataStream API is now the recommended low level API to develop Flink batch jobs.

Enable Speculative Execution

You can enable speculative execution through the following configuration items：

• execution.batch.speculative.enabled: true

Note that currently only Adaptive Batch Scheduler supports speculative execution. And Flink batch jobs will use this scheduler by default unless another scheduler is explicitly configured.

Tuning Configuration

To make speculative execution work better for different jobs, you can tune below configuration options of the scheduler:

• execution.batch.speculative.max-concurrent-executions
• execution.batch.speculative.block-slow-node-duration

You can also tune below configuration options of the slow task detector:

• slow-task-detector.check-interval
• slow-task-detector.execution-time.baseline-lower-bound
• slow-task-detector.execution-time.baseline-multiplier
• slow-task-detector.execution-time.baseline-ratio

Currently, speculative execution uses the slow task detector based on execution time to detect slow tasks. The detector will periodically count all finished executions, if the finished execution ratio reaches the configured ratio(slow-task-detector.execution-time.baseline-ratio), the baseline will be defined as the execution time median multiplied by the configured multiplier(slow-task-detector.execution-time.baseline-multiplier), then the running task whose execution time exceeds the baseline will be detected as a slow task. It is worth mentioning that the execution time will be weighted with the input data volume of the execution vertex, so the executions with large data volume differences but close computing power will not be detected as a slow task, when data skew occurs. This helps to avoid starting unnecessary speculative attempts.

Note: If the node is Source or the Hybrid Shuffle mode is used, the optimization that execution time weighted with input data volume will not take effect, because the input data volume cannot be judged.

Enable Sources for Speculative Execution

If your job uses a custom Source , and the source uses custom SourceEvent , you need to change the SplitEnumerator of that source to implement SupportsHandleExecutionAttemptSourceEvent interface.

public interface SupportsHandleExecutionAttemptSourceEvent {
    void handleSourceEvent(int subtaskId, int attemptNumber, SourceEvent sourceEvent);
}

This means the SplitEnumerator should be aware of the attempt which sends the event. Otherwise, exceptions will happen when the job manager receives a source event from the tasks and lead to job failures.

No extra change is required for other sources to work with speculative execution, including SourceFunction sources , InputFormat sources , and new sources . All the source connectors offered by Apache Flink can work with speculative execution.

Enable Sinks for Speculative Execution

Speculative execution is disabled by default for a sink unless it implements SupportsConcurrentExecutionAttempts interface. This is due to compatibility considerations.

public interface SupportsConcurrentExecutionAttempts {}

The SupportsConcurrentExecutionAttempts works for Sink , SinkFunction and OutputFormat .

If any operator in a task does not support speculative execution, the entire task would be marked as not supporting speculative execution. That means if the Sink does not support speculative execution, the task containing the Sink operator cannot be speculatively executed.

For the Sink implementation, Flink disables speculative execution for Committer ( including the operators extended by WithPreCommitTopology and WithPostCommitTopology ). Because the concurrent committing may cause some unexpected problems if the user is not experienced with it. And the committer is very unlikely to be the bottleneck of the batch job.

Checking the Effectiveness of Speculative Execution

After enabling speculative execution, when there are slow tasks that trigger speculative execution, the web UI will show the speculative attempts on the SubTasks tab of vertices on the job page. The web UI also shows the blocked taskmanagers on the Flink cluster Overview and Task Managers pages.

You can also check these metrics to see the effectiveness of speculative execution.