Preemption
Quota management is the only AWS Batch scheduling algorithm to incorporate
preemption, where AWS Batch stops a SCHEDULED,
STARTING, or RUNNING job to create capacity for a
RUNNABLE job.
Cross-share preemption
Quota management uses cross-share preemption to restore borrowed capacity back to a quota share when jobs arrive.
Administrators lowering a quota share's capacity limits can also make SCHEDULED,
STARTING, or RUNNING jobs within that quota share eligible
for preemption, if the capacity consumed by that quota share is now above its configured
capacity limits.
In-share preemption
Quota shares can be configured to enable in-share
preemption, which allows higher priority RUNNABLE jobs to
trigger preemptions of lower priority jobs within the same quota share that have entered
SCHEDULED, STARTING, or RUNNING.
Preemption selection algorithm
When choosing which jobs to preempt, AWS Batch considers the number and type of
instances a job is borrowing, relative priorities of the jobs and the duration of the
jobs, and applies a custom heuristic. You can use the UpdateServiceJob API
call to update the schedulingPriority of a job after submission. This can
be useful to either lower the priority of a RUNNING job (increasing
likelihood of preemption) or raise the priority of a RUNNABLE job in a
quota share with in-share preemption enabled, making it possible for the job to preempt
an already-running job.
Preemption retries
The default is for preempted jobs to be re-queued as RUNNABLE without
limit. To limit the number of preemptions a job experiences, set
preemptionRetriesBeforeTermination on job submission. When
preemptionRetriesBeforeTermination is set to 0, jobs go to
FAILED on their first preemption.
A sliding window of recent preemption attempts is stored on the job, and visible via DescribeServiceJob.
