...

To solve this kind of problem, we propose a speculative execution mechanism to Flink. The basic idea is to start mirror tasks on other nodes when a slow task is detected. The mirror task processes the same input data and produces the same data as the original task. The original task and mirror tasks are treated equally. Once any of them finishes, its output will be admitted and the remaining homogeneous tasks will be canceled.

Public Interfaces

Configuration

Speculative executions will work along with adaptive batch scheduling (FLIP-187). New configuration options will be added for the scheduler:

...

• jobmanager.adaptive-batch-scheduler.speculative.source-enabled, default to "false". It controls whether to enable speculative execution of sources. This is needed because some sources are not supported for speculative execution. 
• jobmanager.adaptive-batch-scheduler.speculative.sink-enabled, default to "false". It controls whether to enable speculative execution of sinks. This is needed because some sinks are not supported for speculative execution. 
• jobmanager.adaptive-batch-scheduler.speculative.max-concurrent-executions, default to "2". It controls how many executions (including the original one and speculative ones) of an ExecutionVertex can execute at the same time.

...

• slow-task-detector.check-interval, default to "1 s". It defines the interval to check slow tasks.
• slow-task-detector.execution-time.baseline-lower-bound, default to "1 min". It defines the lower bound of the slow task detection baseline.
• slow-task-detector.execution-time.baseline-ratio, default to "0.75". It defines the finished execution ratio threshold to compute the slow task detection baseline.
• slow-task-detector.execution-time.baseline-multiplier, default to "1.5". It defines the multiplier to compute the slow tasks detection baseline.

Metrics

We propose to add following metrics to expose job problems and show the effectiveness of speculative execution:

1. numSlowExecutionVertices: Number of slow execution vertices at the moment.
2. numEffectiveSpeculativeExecutions: Number of speculative executions which finish before their corresponding original executions finish.

Proposed Changes

When speculative execution is enabled, a SpeculativeScheduler(which extends AdaptiveBatchScheduler) will be used for task scheduling. SpeculativeScheduler will listen on slow tasks detected by SlowTaskDetector. It will create and deploy speculative executions for the slow tasks. Nodes that slow tasks located on will be treated as slow nodes and get blacklisted, so that speculative executions will not be deployed on them. Once any execution finishes, the remaining homogeneous tasks will be canceled, so that only one execution will be admitted as finished and only its output will be visible to downstream consumer tasks or in external sink services.

...

A SlowTaskBlacklistStrategy will be introduced for speculative execution. The strategy will blacklist the nodes if it is a SlowTaskException. The blacklist action will be MARK_BLACKLISTED so that future tasks will not be deployed to the node, while deployed tasks can keep running.

...

We propose to add following metrics to expose job problems and show the effectiveness of speculative execution:

1. numSlowExecutionVertices: Number of slow execution vertices at the moment.
2. numEffectiveSpeculativeExecutions: Number of speculative executions which finish before their corresponding original executions finish.

Limitations

Batch jobs only

...

Speculative execution of sources and sinks are disabled by default, this is because not all sources and sinks can work with speculative execution. 

There will be several follow up proposals to enable speculative execution of sources and sinks:

• FLIP-XXX(coming soon) is a proposal to enable speculative execution to work with most sources, without changing the source connectors
• FLIP-XXX(coming soon) is a proposal to improve Flink interfaces to support speculative execution of sinks. Sinks must implement these interfaces to work with speculative execution
• FLIP-XXX(coming soon) is a proposal to improve FileSink to work with speculative execution
• FLIP-XXX(coming soon) is a proposal to improve HiveTableSink to work with speculative execution

With these proposed improvements, the scheduler will be able to see whether the given source or sink support speculative execution. It will enable speculative execution for a JobVertex only if all the contained sources and sinks support speculative execution.

No integration with Flink's web UI

...