Storage Plugins

Nomad has built-in support for scheduling compute resources such as CPU, memory, and networking. Nomad's storage plugin support extends this to allow scheduling tasks with externally created storage volumes. Storage plugins are third-party plugins that conform to the Container Storage Interface (CSI) specification.

Storage plugins are created dynamically as Nomad jobs, unlike device and task driver plugins that need to be installed and configured on each client. Each dynamic plugin type has its own type-specific job spec block; currently there is only the csi_plugin type. Nomad tracks which clients have instances of a given plugin, and communicates with plugins over a Unix domain socket that it creates inside the plugin's tasks.

CSI Plugins

Every storage vendor has its own APIs and workflows, and the industry-standard Container Storage Interface specification unifies these APIs in a way that's agnostic to both the storage vendor and the container orchestrator. Each storage provider can build its own CSI plugin. Jobs can claim storage volumes from AWS Elastic Block Storage (EBS) volumes, GCP persistent disks, Ceph, Portworx, vSphere, etc. The Nomad scheduler will be aware of volumes created by CSI plugins and schedule workloads based on the availability of volumes on a given Nomad client node. A list of available CSI plugins can be found in the Kubernetes CSI documentation. Any of these plugins should work with Nomad out of the box.

A CSI plugin task requires the csi_plugin block:

csi_plugin {
  id        = "csi-hostpath"
  type      = "monolith"
  mount_dir = "/csi"
}

There are three types of CSI plugins. Controller Plugins communicate with the storage provider's APIs. For example, for a job that needs an AWS EBS volume, Nomad will tell the controller plugin that it needs a volume to be "published" to the client node, and the controller will make the API calls to AWS to attach the EBS volume to the right EC2 instance. Node Plugins do the work on each client node, like creating mount points. Monolith Plugins are plugins that perform both the controller and node roles in the same instance. Not every plugin provider has or needs a controller; that's specific to the provider implementation.

You should almost always run node plugins as Nomad system jobs to ensure volume claims are released when a Nomad client is drained. Use constraints for the node plugin jobs based on the availability of volumes. For example, AWS EBS volumes are specific to particular availability zones with a region. Controller plugins can be run as service jobs.

Nomad exposes a Unix domain socket named csi.sock inside each CSI plugin task, and communicates over the gRPC protocol expected by the CSI specification. The mount_dir field tells Nomad where the plugin expects to find the socket file.

Plugin Lifecycle and State

CSI plugins report their health like other Nomad jobs. If the plugin crashes or otherwise terminates, Nomad will launch it again using the same restart and reschedule logic used for other jobs. If plugins are unhealthy, Nomad will mark the volumes they manage as "unscheduable".

Storage plugins don't have any responsibility (or ability) to monitor the state of tasks that claim their volumes. Nomad sends mount and publish requests to storage plugins when a task claims a volume, and unmount and unpublish requests when a task stops.

The dynamic plugin registry persists state to the Nomad client so that it can restore volume managers for plugin jobs after client restarts without disrupting storage.

Volume Lifecycle

The Nomad scheduler decides whether a given client can run an allocation based on whether it has a node plugin present for the volume. But before a task can use a volume the client needs to "claim" the volume for the allocation. The client makes an RPC call to the server and waits for a response; the allocation's tasks won't start until the volume has been claimed and is ready.

If the volume's plugin requires a controller, the server will send an RPC to the Nomad client where that controller is running. The Nomad client will forward this request over the controller plugin's gRPC socket. The controller plugin will make the request volume available to the node that needs it.

Once the controller is done (or if there's no controller required), the server will increment the count of claims on the volume and return to the client. This count passes through Nomad's state store so that Nomad has a consistent view of which volumes are available for scheduling.

The client then makes RPC calls to the node plugin running on that client, and the node plugin mounts the volume to a staging area in the Nomad data directory. Nomad will bind-mount this staged directory into each task that mounts the volume.

This cycle is reversed when a task that claims a volume becomes terminal. The client updates the server frequently about changes to allocations, including terminal state. When the server receives a terminal state for a job with volume claims, it creates a volume claim garbage collection (GC) evaluation to to handled by the core job scheduler. The GC job will send "detach" RPCs to the node plugin. The node plugin unmounts the bind-mount from the allocation and unmounts the volume from the plugin (if it's not in use by another task). The GC job will then send "unpublish" RPCs to the controller plugin (if any), and decrement the claim count for the volume. At this point the volume’s claim capacity has been freed up for scheduling.