The Apache Spark scheduler in Databricks automatically preempts tasks to enforce fair sharing. This guarantees interactive response times on clusters with many concurrently running jobs.
When tasks are preempted by the scheduler, their kill reason will be set to
preempted by scheduler. This reason is visible in the Spark UI and can be used to debug preemption behavior.
By default, preemption is conservative: jobs can be starved of resources for up to 30 seconds before the scheduler intervenes. You can tune preemption by setting the following Spark configuration properties at cluster launch time:
Whether preemption should be enabled.
The fair share fraction to guarantee per job. Setting this to 1.0 means the scheduler will aggressively attempt to guarantee perfect fair sharing. Setting this to 0.0 effectively disables preemption. The default setting is 0.5, which means at worst a jobs will get half of its fair share.
How long a job must remain starved before preemption kicks in. Setting this to lower values will provide more interactive response times, at the cost of cluster efficiency. Recommended values are from 1-100 seconds.
How often the scheduler will check for task preemption. This should be set to less than the preemption timeout.
For further information on job scheduling, see Scheduling Within an Application.