batch

This module wraps a subset of the GCP Batch functionalities that are useful and adequate for typical applications.

User code runs within a Docker container. Running a standalone script is not supported by this wrapper module.

class cloudly.gcp.batch.Job

Bases: object

classmethod create(name: str, config: JobConfig | dict) → Job

There are some restrictions on the form of name; see GCP doc for details or cloudly.gcp.compute. In addition, the batch name must be unique in the project and the region.

classmethod list(*, region: str) → list[Job]

__init__(name_or_obj: str | Job, /)

name is like ‘projects/<project-id>/locations/<location>/jobs/<name>’. This can also be considered the job “ID” or “URI”. This is available from the object returned by create().

job is not necessary because it can be created if needed. It is accepted in case you already have it. See create(), list().

property create_time

property update_time

property definition: dict

status() → JobStatus

The returned JobStatus object has some useful attributes you can access; see state() for an example.

state() → Literal['STATE_UNSPECIFIED', 'QUEUED', 'SCHEDULED', 'RUNNING', 'SUCCEEDED', 'FAILED', 'DELETION_IN_PROGRESS']

delete() → None

class cloudly.gcp.batch.JobConfig

Bases: object

class BootDisk

Bases: object

__init__(*, size_gb: int, disk_type: Literal['pd-balanced', 'pd-extreme', 'pd-ssd', 'pd-standard'] = 'pd-balanced', image: str | None = None)

image: ‘batch-debian’ seems to be a good value for GPUs; otherwise batch-cos may

also work well. Leave it at None until needed. See https://cloud.google.com/batch/docs/vm-os-environment-overview

property disk: Disk

class LocalSSD

Bases: object

Local SSDs are attached to each worker node for use during the lifetime of the tasks. They are not “persistent” storage that lives beyond the batch job.

__init__(*, size_gb: int, device_name: str = 'local-ssd', mount_path: str = '/mnt', mode: Literal['ro', 'rw'] = 'rw')

size_gb should be a multiple of 375. If not, the next greater multiple of 375 will be used.

property disk: AttachedDisk

property volume: Volume

class GPU

Bases: object

__init__(*, gpu_type: str, gpu_count: int)

Use gcloud compute accelerator-types list to see valid values of gpu_type. Some examples: ‘nvidia-tesla-t4’, ‘nvidia-l4’, ‘nvidia-tesla-a100’, ‘nvidia-tesla-v100’

property accelerator: Accelerator

classmethod task_group(*, task_spec: dict, task_count: int = 1, task_count_per_node: int = 1, parallelism: int | None = None, permissive_ssh: bool = True, **kwargs) → TaskGroup

Parameters:

task_count

Number of tasks to be created.

task_count_per_node

Number of tasks that can be running on a work node at any time.

parallelism

Number of tasks that can be running across all nodes at any time.

classmethod allocation_policy(*, region: str, labels: dict, network_uri: str | None = None, subnet_uri: str | None = None, machine_type: str, no_external_ip_address: bool = False, provisioning_model: Literal['standard', 'spot', 'preemptible'] = 'standard', boot_disk: dict | None = None, gpu: GPU | None = None, local_ssd_disk: LocalSSD | None = None, install_gpu_drivers: bool | None = None, **kwargs) → batch_v1.AllocationPolicy

Parameters:

region

Like ‘us-central1’, ‘us-west1’, etc.

network_uri, subnet_uri

If missing and no_external_ip_address is False, the default for the project in the specified region will be used.

If no_external_ip_address is True, then both must be provided.

See https://cloud.google.com/compute/docs/networking/network-overview and google.cloud.batch_v1.types.job.AllocationPolicy.NetworkInterface.

labels, gpu, local_ssd_disk

These are provided by JobConfig.__init__. User should not provide them directly.

classmethod labels(**kwargs) → dict[str, str]

__init__(*, task_group: dict, allocation_policy: dict, labels: dict[str, str] | None = None, logs_policy: LogsPolicy | None = None, gpu: dict | None = None, local_ssd: dict | None = None, **kwargs)

property job: Job

property definition: dict

property region: str