shepherd.sheep

Classes

• BaseSheep: A base class for container adapters - classes that allow launching different kinds of containers.

• DockerSheep: Sheep running its jobs in docker containers.

• BareSheep: An adapter that running models on bare metal with shepherd-runner.

class shepherd.sheep.BaseSheep(socket, sheep_data_root)

Bases: object

A base class for container adapters - classes that allow launching different kinds of containers.

class Config(raw_data=None, trusted_data=None, deserialize_mapping=None, init=True, partial=True, strict=True, validate=False, app_data=None, lazy=False, **kwargs)

Bases: schematics.deprecated.patch_models.<locals>.Model

__init__(socket, sheep_data_root)

Create new BaseSheep.

Parameters

• socket (Socket) – socket for feeding sheep’s runner with InputMessages

• sheep_data_root (str) – sheep data root with job working directories

_load_model(model_name, model_version)

Tell the sheep to prepare a new model (without restarting).

Return type

None

abstract property running: bool

Is the sheep running, i.e. capable of accepting computation requests?

Return type

bool

start(model_name, model_version)

(Re)start the sheep with the given model name and version. Any unfinished jobs will be lost, socket connection will be reset.

Parameters

• model_name (str) – model name

• model_version (str) – model version

Return type

None

class shepherd.sheep.DockerSheep(config, registry_config, command=None, **kwargs)

Bases: shepherd.sheep.base_sheep.BaseSheep

Sheep running its jobs in docker containers. To enable GPU computation, specify the gpu devices in the configuration and sheep will attempt to use nvidia docker 2.

class Config(raw_data=None, trusted_data=None, deserialize_mapping=None, init=True, partial=True, strict=True, validate=False, app_data=None, lazy=False, **kwargs)

Bases: shepherd.sheep.base_sheep.BaseSheep.Config

_CONTAINER_POINT = 9999

Container port to bind the socket to.

__init__(config, registry_config, command=None, **kwargs)

Create new DockerSheep.

Parameters

• config (Dict[str, Any]) – docker sheep configuration

• registry_config (RegistryConfig) – docker registry configuration

• command (Optional[List[str]]) – optional docker container run command

• kwargs – BaseSheep’s kwargs

_load_model(model_name, model_version)

Pull docker image of the given name and version from the previously configured docker registry.

Parameters

• model_name (str) – docker image name

• model_version (str) – docker image version

Return type

None

property running: bool

Check if the underlying docker container is running.

Return type

bool

slaughter()

Kill the underlying docker container.

Return type

None

start(model_name, model_version)

Run a docker command starting the docker runner.

Parameters

• model_name (str) – docker image name

• model_version (str) – docker image version

Return type

None

class shepherd.sheep.BareSheep(config, **kwargs)

Bases: shepherd.sheep.base_sheep.BaseSheep

An adapter that running models on bare metal with shepherd-runner. This might be useful when Docker isolation is impossible or not necessary, for example in deployments with just a few models.

class Config(raw_data=None, trusted_data=None, deserialize_mapping=None, init=True, partial=True, strict=True, validate=False, app_data=None, lazy=False, **kwargs)

Bases: shepherd.sheep.base_sheep.BaseSheep.Config

__init__(config, **kwargs)

Create new BareSheep.

Parameters

• config (Dict[str, Any]) – bare sheep configuration (BareSheep.Config)

• kwargs – parent’s kwargs

_load_model(model_name, model_version)

Set up runner config path to working_directory / model_name / model_version / config.yaml.

Parameters

• model_name (str) – model name

• model_version (str) – model version

Raises

SheepConfigurationError – if the runner config path does not exist

Return type

None

property running: bool

Check if the underlying runner (subprocess) is running.

Return type

bool

slaughter()

Kill the underlying runner (subprocess).

Return type

None

start(model_name, model_version)

Start a subprocess with the sheep runner.

Parameters

• model_name (str) – model name

• model_version (str) – model version

Return type

None