def get_builder(self,
preprocessor,
base_image,
registry,
needs_deps_installation=True,
pod_spec_mutators=None):
"""Creates a builder instance with right config for GKE
:param preprocessor: Preprocessor to use to modify inputs
:param base_image: Base image to use for this job
:param registry: Registry to push image to. Example: gcr.io/kubeflow-images
:param needs_deps_installation: need depends on installation(Default value = True)
:param pod_spec_mutators: list of functions that is used to mutate the podsspec.
e.g. fairing.cloud.gcp.add_gcp_credentials_if_exists
This can used to set things like volumes and security context.
(Default value =None)
"""
pod_spec_mutators = pod_spec_mutators or []
pod_spec_mutators.append(gcp.add_gcp_credentials_if_exists)
if not needs_deps_installation:
return AppendBuilder(preprocessor=preprocessor,
base_image=base_image,
registry=registry)
elif (utils.is_running_in_k8s()
or not ml_tasks_utils.is_docker_daemon_exists()):
return ClusterBuilder(preprocessor=preprocessor,
base_image=base_image,
registry=registry,
pod_spec_mutators=pod_spec_mutators,
namespace=self._namespace,
context_source=self._build_context_source)
elif ml_tasks_utils.is_docker_daemon_exists():
return DockerBuilder(preprocessor=preprocessor,
base_image=base_image,
registry=registry)
else:
msg = ["Not able to guess the right builder for this job!"]
if not utils.is_running_in_k8s():
msg.append(
" Also If you are using 'sudo' to access docker in your system you can"
" solve this problem by adding your username to the docker group. "
"Reference: https://docs.docker.com/install/linux/linux-postinstall/"
"#manage-docker-as-a-non-root-user You need to logout and login to "
"get change activated.")
message = " ".join(msg)
raise RuntimeError(message)
def __init__(self,
config_file=None,
context=None,
client_configuration=None,
persist_config=True):
"""
:param config_file: kubeconfig file, defaults to ~/.kube/config. Note that for the case
that the SDK is running in cluster and you want to operate in another remote
cluster, user must set config_file to load kube-config file explicitly.
:param context: kubernetes context
:param client_configuration: The kubernetes.client.Configuration to set configs to.
:param persist_config: If True, config file will be updated when changed
"""
self.config_file = config_file
self.context = context
self.client_configuration = client_configuration
self.persist_config = persist_config
if config_file or not is_running_in_k8s():
config.load_kube_config(
config_file=self.config_file,
context=self.context,
client_configuration=self.client_configuration,
persist_config=self.persist_config)
else:
config.load_incluster_config()
def get_builder(self, preprocessor, base_image, registry, needs_deps_installation=True, # pylint:disable=arguments-differ
pod_spec_mutators=None):
"""Creates a builder instance with right config for the given Kubernetes
:param preprocessor: Preprocessor to use to modify inputs
:param base_image: Base image to use for this job
:param registry: Registry to push image to. Example: gcr.io/kubeflow-images
:param needs_deps_installation: need depends on installation(Default value = True)
:param pod_spec_mutators: list of functions that is used to mutate the podsspec.
e.g. fairing.cloud.gcp.add_gcp_credentials_if_exists
This can used to set things like volumes and security context.
(Default value =None)
"""
if not needs_deps_installation:
return AppendBuilder(preprocessor=preprocessor,
base_image=base_image,
registry=registry)
elif utils.is_running_in_k8s():
return ClusterBuilder(preprocessor=preprocessor,
base_image=base_image,
registry=registry,
pod_spec_mutators=pod_spec_mutators,
namespace=self._namespace,
context_source=self._build_context_source)
elif ml_tasks_utils.is_docker_daemon_exists():
return DockerBuilder(preprocessor=preprocessor,
base_image=base_image,
registry=registry)
else:
# TODO (karthikv2k): Add more info on how to reolve this issue
raise RuntimeError(
"Not able to guess the right builder for this job!")
def __init__(self, namespace=None, build_context_source=None):
if not namespace and not utils.is_running_in_k8s():
logger.warning("Can't determine namespace automatically. "
"Using 'default' namespace but recomend to provide namespace explicitly"
". Using 'default' namespace might result in unable to mount some "
"required secrets in cloud backends.")
self._namespace = namespace or utils.get_default_target_namespace()
self._build_context_source = build_context_source
def __init__(self):
if is_running_in_k8s():
config.load_incluster_config()
else:
config.load_kube_config()
def __init__(self, namespace=None, build_context_source=None):
if not namespace and not utils.is_running_in_k8s():
namespace = "kubeflow"
super(KubeflowGKEBackend, self).__init__(
namespace, build_context_source)
def execute(config,
docker_registry,
base_image="gcr.io/kubeflow-fairing/lightgbm:latest",
namespace=None,
stream_log=True,
cores_per_worker=None,
memory_per_worker=None,
pod_spec_mutators=None):
"""Runs the LightGBM CLI in a single pod in user's Kubeflow cluster.
Users can configure it to be a train, predict, and other supported tasks
by using the right config.
Please refere https://github.com/microsoft/LightGBM/blob/master/docs/Parameters.rst
for more information on config options.
:param config: config entries
:param docker_registry: docker registry name
:param base_image: base image (Default value = "gcr.io/kubeflow-fairing/lightgbm:latest")
:param namespace: k8s namespace (Default value = None)
:param stream_log: should that stream log? (Default value = True)
:param cores_per_worker: number of cores per worker (Default value = None)
:param memory_per_worker: memory value per worker (Default value = None)
:param pod_spec_mutators: pod spec mutators (Default value = None)
"""
if not namespace and not fairing_utils.is_running_in_k8s():
namespace = "kubeflow"
namespace = namespace or fairing_utils.get_default_target_namespace()
config_file_name = None
if isinstance(config, str):
config_file_name = config
config = utils.load_properties_config_file(config)
elif isinstance(config, dict):
config_file_name = utils.save_properties_config_file(config)
else:
raise RuntimeError("config should be of type dict or string(filepath) "
"but got {}".format(type(dict)))
utils.scrub_fields(config, BLACKLISTED_FIELDS)
_, num_machines = utils.get_config_value(config, NUM_MACHINES_FILEDS)
num_machines = num_machines or 1
if num_machines:
try:
num_machines = int(num_machines)
except ValueError:
raise ValueError(
"num_machines value in config should be an int >= 1 "
"but got {}".format(config.get('num_machines')))
if num_machines < 1:
raise ValueError(
"num_machines value in config should >= 1 but got {}".format(
num_machines))
if num_machines > 1:
config['machine_list_file'] = "mlist.txt"
output_map = generate_context_files(config, config_file_name, num_machines)
preprocessor = BasePreProcessor(command=[ENTRYPOINT],
output_map=output_map)
builder = AppendBuilder(registry=docker_registry,
base_image=base_image,
preprocessor=preprocessor)
builder.build()
pod_spec = builder.generate_pod_spec()
pod_spec_mutators = pod_spec_mutators or []
pod_spec_mutators.append(gcp.add_gcp_credentials_if_exists)
pod_spec_mutators.append(
k8s_utils.get_resource_mutator(cores_per_worker, memory_per_worker))
if num_machines == 1:
# non-distributed mode
deployer = Job(namespace=namespace,
pod_spec_mutators=pod_spec_mutators,
stream_log=stream_log)
else:
# distributed mode
deployer = TfJob(namespace=namespace,
pod_spec_mutators=pod_spec_mutators,
chief_count=1,
worker_count=num_machines - 1,
stream_log=stream_log)
deployer.deploy(pod_spec)
return deployer
