Scheduling in Nomad
There are four primary "nouns" in Nomad; jobs, nodes, allocations, and evaluations. Jobs are submitted by users and represent a desired state. A job is a declarative description of tasks to run which are bounded by constraints and require resources. Tasks can be scheduled on nodes in the cluster running the Nomad client. The mapping of tasks in a job to clients is done using allocations. An allocation is used to declare that a set of tasks in a job should be run on a particular node. Scheduling is the process of determining the appropriate allocations and is done as part of an evaluation.
An evaluation is created any time the external state, either desired or emergent, changes. The desired state is based on jobs, meaning the desired state changes if a new job is submitted, an existing job is updated, or a job is deregistered. The emergent state is based on the client nodes, and so we must handle the failure of any clients in the system. These events trigger the creation of a new evaluation, as Nomad must evaluate the state of the world and reconcile it with the desired state.
This diagram shows the flow of an evaluation through Nomad:
The lifecycle of an evaluation begins with an event causing the evaluation to
be created. Evaluations are created in the
pending state and are enqueued
into the evaluation broker. There is a single evaluation broker which runs on
the leader server. The evaluation broker is used to manage the queue of pending
evaluations, provide priority ordering, and ensure at least once delivery.
Nomad servers run scheduling workers, defaulting to one per CPU core, which are
used to process evaluations. The workers dequeue evaluations from the broker,
and then invoke the appropriate scheduler as specified by the job. Nomad ships
service scheduler that optimizes for long-lived services, a
scheduler that is used for fast placement of batch jobs,
sysbatch schedulers that are used to run jobs on every node, and a
scheduler which is used for internal maintenance.
Schedulers are responsible for processing an evaluation and generating an allocation plan. The plan is the set of allocations to evict, update, or create. The specific logic used to generate a plan may vary by scheduler, but generally the scheduler needs to first reconcile the desired state with the real state to determine what must be done. New allocations need to be placed and existing allocations may need to be updated, migrated, or stopped.
Placing allocations is split into two distinct phases, feasibility checking and ranking. In the first phase the scheduler finds nodes that are feasible by filtering nodes in datacenters and node pools not used by the job, unhealthy nodes, those missing necessary drivers, and those failing the specified constraints.
The second phase is ranking, where the scheduler scores feasible nodes to find the best fit. Scoring is primarily based on bin packing, which is used to optimize the resource utilization and density of applications, but is also augmented by affinity and anti-affinity rules. Nomad automatically applies a job anti-affinity rule which discourages colocating multiple instances of a task group. The combination of this anti-affinity and bin packing optimizes for density while reducing the probability of correlated failures.
Once the scheduler has ranked enough nodes, the highest ranking node is selected and added to the allocation plan.
When planning is complete, the scheduler submits the plan to the leader which adds the plan to the plan queue. The plan queue manages pending plans, provides priority ordering, and allows Nomad to handle concurrency races. Multiple schedulers are running in parallel without locking or reservations, making Nomad optimistically concurrent. As a result, schedulers might overlap work on the same node and cause resource over-subscription. The plan queue allows the leader node to protect against this and do partial or complete rejections of a plan.
As the leader processes plans, it creates allocations when there is no conflict and otherwise informs the scheduler of a failure in the plan result. The plan result provides feedback to the scheduler, allowing it to terminate or explore alternate plans if the previous plan was partially or completely rejected.
Once the scheduler has finished processing an evaluation, it updates the status of the evaluation and acknowledges delivery with the evaluation broker. This completes the lifecycle of an evaluation. Allocations that were created, modified or deleted as a result will be picked up by client nodes and will begin execution.