» Preemption

Preemption allows Nomad to kill existing allocations in order to place allocations for a higher priority job. The evicted allocation is temporary displaced until the cluster has capacity to run it. This allows operators to run high priority jobs even under resource contention across the cluster.

» Preemption in Nomad

Every job in Nomad has a priority associated with it. Priorities impact scheduling at the evaluation and planning stages by sorting the respective queues accordingly (higher priority jobs get moved ahead in the queues).

Prior to Nomad 0.9, when a cluster is at capacity, any allocations that result from a newly scheduled or updated job remain in the pending state until sufficient resources become available - regardless of the defined priority. This leads to priority inversion, where a low priority task can prevent high priority tasks from completing.

Nomad 0.9 brings preemption capabilities to system jobs. The Nomad scheduler will evict lower priority running allocations to free up capacity for new allocations resulting from relatively higher priority jobs, sending evicted allocations back into the plan queue.

» Details

Preemption is enabled by default in Nomad 0.9. Operators can use the scheduler config API endpoint to disable preemption.

Nomad uses the job priority field to determine what running allocations can be preempted. In order to prevent a cascade of preemptions due to jobs close in priority being preempted, only allocations from jobs with a priority delta of more than 10 from the job needing placement are eligible for preemption.

For example, consider a node with the following distribution of allocations:

Job Priority Allocations Total Used capacity
cache 70 a6 2 GB Memory, 0.5 GB Disk, 1 CPU
batch-analytics 50 a4, a5 <1 GB Memory, 0.5 GB Disk, 0.5 CPU>, <1 GB Memory, 0.5 GB Disk, 0.5 CPU>
email-marketing 20 a1, a2 <0.5 GB Memory, 0.8 GB Disk>, <0.5 GB Memory, 0.2 GB Disk>

If a job webapp with priority 75 needs placement on the above node, only allocations from batch-analytics and email-marketing are considered eligible to be preempted because they are of a lower priority. Allocations from the cache job will never be preempted because its priority value 70 is lesser than the required delta of 10.

Allocations are selected starting from the lowest priority, and scored according to how closely they fit the job's required capacity. For example, if the 75 priority job needs 1GB disk and 2GB memory, Nomad will preempt allocations a1, a2 and a4 to satisfy those requirements.

» Preemption Visibility

Operators can use the allocation API or the alloc status command to get visibility into whether an allocation has been preempted. Preempted allocations will have their DesiredStatus set to “evict”. The Allocation object in the API also has two additional fields related to preemption.

  • PreemptedAllocs - This field is set on an allocation that caused preemption. It contains the allocation ids of allocations that were preempted to place this allocation. In the above example, allocations created for the job webapp will have the values a1, a2 and a4 set.
  • PreemptedByAllocID - This field is set on allocations that were preempted by the scheduler. It contains the allocation ID of the allocation that preempted it. In the above example, allocations a1, a2 and a4 will have this field set to the ID of the allocation from the job webapp.

» Integration with Nomad plan

nomad plan allows operators to dry run the scheduler. If the scheduler determines that preemption is necessary to place the job, it shows additional information in the CLI output for nomad plan as seen below.

$ nomad plan example.nomad

+ Job: "test"
+ Task Group: "test" (1 create)
  + Task: "test" (forces create)

Scheduler dry-run:
- All tasks successfully allocated.

Preemptions:

Alloc ID                              Job ID    Task Group
ddef9521                              my-batch   analytics
ae59fe45                              my-batch   analytics

Note that, the allocations shown in the nomad plan output above are not guaranteed to be the same ones picked when running the job later. They provide the operator a sample of the type of allocations that could be preempted.