JobSpec

The role type. It is determined by the job type, as different job types support different roles.

• TFJob: Supports Chief, PS, Worker, Evaluator, and GraphLearn.

• PyTorchJob: Supports Master and Worker.

• XGBoostJob: Supports Master and Worker.

• OneFlowJob: Supports Master and Worker.

• ElasticBatch: Supports Master and Worker.

• RayJob: Supports Head, Worker, and Worker[-xxx].

For PyTorchJob, XGBoostJob, OneFlowJob, and ElasticBatch, the Master role is optional. If you do not specify a Master, the system designates the first Worker pod as the Master.

The Registry Address of the runtime image for this worker type. Call ListImages to get the runtime images that PAI provides. You can also specify a public third-party runtime image.

The configuration for a private runtime image.

The number of replicas.

The extra pod configuration.

The resource configuration.

Specifies whether to use spot instances.

The spot instance configuration.

The configuration for the specified scheduling node.

The auto scaling configuration.

The service configuration.

This parameter is misspelled and deprecated.

Marks whether this role is the Chief role. There can be only one Chief role.

The restart policy. Valid values are Always, Never, OnFailure, and ExitCode.

The device specification of the worker. For a detailed list of specifications, see PAI-DLC billing.

Indicates whether this role's completion is required for the job to succeed. This setting applies only when the job's success policy is Partial.