Primitives

Jobs, Services, and VMs are the three Workload types. Kubernetes Deployments provide reusable applications that accept Jobs, and Buckets provide storage that persists independently of compute. Choose the resources that match your workload's lifecycle and access needs.

Primitive	Choose it when	Lifetime	Primary access
Job	Work should run to completion	Until a terminal outcome	Status; logs and exec for VM-backed Jobs
Service	An HTTP workload should stay online	Until upgrade or termination	Stable public URL
VM	You need direct, interactive control	Until termination	SSH to the container or host
Kubernetes Deployment	Jobs should share a persistent application	Until deletion	Targeted Jobs and Deployment status
Bucket	Data must outlive its compute	Until the bucket is deleted	Mounted files in supported workloads

Jobs are the default for finite work such as training, evaluation, and batch inference. Services manage long-running HTTP workloads and give them a stable public URL. VMs provide an interactive shell and direct host control without a managed HTTP endpoint. Buckets complement supported Jobs and Services with durable object storage; they are not a compute environment.

Jobs run on dedicated VMs or target Kubernetes Deployments on shared node pool capacity. A targeted Job runs inside the Deployment's existing application container; the application remains available after that Job ends. Services run on VMs or on node pools. Workload remains the shared record for Jobs, Services, and VMs; each Deployment has its own record and lifecycle.