THIS IS A TEST INSTANCE. ALL YOUR CHANGES WILL BE LOST!!!!
Status
...
| Page properties | |
|---|---|
|
...
...
|
...
|
...
Please keep the discussion on the mailing list rather than commenting on the wiki (wiki discussions get unwieldy fast).
...
- 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.
- jobmanager.adaptive-batch-scheduler.speculative.block-slow-node-duration, default to "1 min". It controls how long an identified slow node should be blocked for.
New configuration options will be added in SlowTaskDetectorOptions for slow task detection:
...
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 blacklistedblocked, 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.
...
- SpeculativeScheduler needs to be able to directly deploy an Execution, while AdaptiveBatchScheduler can only perform ExecutionVertex level deployment.
- SpeculativeScheduler does not restart the ExecutionVertex if an execution fails when any other current execution is still making progress
- SpeculativeScheduler listens on slow tasks. Once there are slow tasks, it will blacklist the block the slow nodes and deploy speculative executions of the slow tasks on other nodes.
- Once any execution finishes, SpeculativeScheduler will cancel all the remaining executions of the same execution vertex.
...
Once notified about slow tasks, the SpeculativeScheduler will handle them as below:
- Blacklist Block nodes that the slow tasks locate on. To achieve this, the scheduler will notify locations of slow tasks with a SlowTaskException add the slow nodes to the BlacklistHandlerblocklist. The BlacklistHandler will use a SlowTaskBlacklistStrategy (see section below) to make decisions.block action will be MARK_BLOCKED so that future tasks will not be deployed to the slow node, while deployed tasks can keep running. (See FLIP-224 Blocklist Mechanism for more details)
- Create speculative executions for slow tasks until the current executions of each execution vertex reach the concurrency limit (defined via config jobmanager.adaptive-batch-scheduler.speculative.max-concurrent-executions)
- Deploy the newly created speculative executions
SlowTaskBlacklistStrategy
...
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.
Limitations
Batch jobs only
...
Currently, AdaptiveBatchScheduler does not support jobs with PIPELINED data exchanges. As a result, speculative execution does not support PIPELINED data exchanges either. Requiring all data exchanges to be BLOCKING also simplifies things, because each ExecutionVertex is an individual pipelined region in this case and can have individual speculations. Otherwise multiple ExecutionVertex from one pipelined region need to do speculative execution together.
This also means that
Speculative execution of sources and sinks are disabled by default
...
The web UI does not show all the concurrent executions of each ExecutionVertex/subtask. It only shows the one with the fastest progress.
User defined functions must not be affected by its speculative instances
When a user defined function and its speculative instances run concurrently, they must not affect each other. For example,
- access to the same exclusive resources
- overriding the output to external services which happens as a side effect, i.e. not via Flink sinks
- competition for data ingestion. Note that it includes cases that
- user defined source function competition
- data ingestion happens as a side effect, i.e. not via Flink sources.
- ...
Once the concurrent instances can affect each other, it may result in task failures, or even worse, data inconsistency. So that speculative executions should not be enabled in this case.
Compatibility, Deprecation, and Migration Plan
...