class ClusterBuilder(BaseBuilder):
"""Builds a docker image in a Kubernetes cluster.
Args:
registry (str): Required. Registry to push image to
Example: gcr.io/kubeflow-images
base_image (str): Base image to use for the image build
preprocessor (BasePreProcessor): Preprocessor to use to modify inputs
before sending them to docker build
context_source (ContextSourceInterface): context available to the
cluster build
"""
def __init__(self,
registry=None,
context_source=gcs_context.GCSContextSource(),
preprocessor=None,
base_image=constants.DEFAULT_BASE_IMAGE,
dockerfile_path=None):
super().__init__(
registry=registry,
preprocessor=preprocessor,
base_image=base_image,
)
self.manager = KubeManager()
self.context_source = context_source
def build(self):
dockerfile_path = dockerfile.write_dockerfile(
dockerfile_path=self.dockerfile_path, base_image=self.base_image)
self.preprocessor.output_map[dockerfile_path] = 'Dockerfile'
context_path, context_hash = self.preprocessor.context_tar_gz()
self.image_tag = self.full_image_name(context_hash)
self.context_source.prepare(context_path)
labels = {'fairing-builder': 'kaniko'}
build_pod = client.V1Pod(api_version="v1",
kind="Pod",
metadata=client.V1ObjectMeta(
generate_name="fairing-builder-",
labels=labels,
),
spec=self.context_source.generate_pod_spec(
self.image_tag))
created_pod = client. \
CoreV1Api(). \
create_namespaced_pod("default", build_pod)
self.manager.log(name=created_pod.metadata.name,
namespace=created_pod.metadata.namespace,
selectors=labels)
# clean up created pod and secret
self.context_source.cleanup()
client.CoreV1Api().delete_namespaced_pod(
created_pod.metadata.name, created_pod.metadata.namespace,
client.V1DeleteOptions())
class ClusterBuilder(BaseBuilder):
"""Builds a docker image in a Kubernetes cluster.
Args:
registry (str): Required. Registry to push image to
Example: gcr.io/kubeflow-images
base_image (str): Base image to use for the image build
preprocessor (BasePreProcessor): Preprocessor to use to modify inputs
before sending them to docker build
context_source (ContextSourceInterface): context available to the
cluster build
push {bool} -- Whether or not to push the image to the registry
"""
def __init__(self,
registry=None,
image_name=constants.DEFAULT_IMAGE_NAME,
context_source=None,
preprocessor=None,
push=True,
base_image=constants.DEFAULT_BASE_IMAGE,
pod_spec_mutators=None,
namespace="kubeflow",
dockerfile_path=None):
super().__init__(
registry=registry,
image_name=image_name,
push=push,
preprocessor=preprocessor,
base_image=base_image,
)
self.manager = KubeManager()
if context_source is None:
raise RuntimeError("context_source is not specified")
self.context_source = context_source
self.pod_spec_mutators = pod_spec_mutators or []
self.namespace = namespace
def build(self):
logging.info("Building image using cluster builder.")
install_reqs_before_copy = self.preprocessor.is_requirements_txt_file_present(
)
dockerfile_path = dockerfile.write_dockerfile(
dockerfile_path=self.dockerfile_path,
path_prefix=self.preprocessor.path_prefix,
base_image=self.base_image,
install_reqs_before_copy=install_reqs_before_copy)
self.preprocessor.output_map[dockerfile_path] = 'Dockerfile'
context_path, context_hash = self.preprocessor.context_tar_gz()
self.image_tag = self.full_image_name(context_hash)
self.context_source.prepare(context_path)
labels = {'fairing-builder': 'kaniko'}
labels['fairing-build-id'] = str(uuid.uuid1())
pod_spec = self.context_source.generate_pod_spec(
self.image_tag, self.push)
for fn in self.pod_spec_mutators:
fn(self.manager, pod_spec, self.namespace)
build_pod = client.V1Pod(api_version="v1",
kind="Pod",
metadata=client.V1ObjectMeta(
generate_name="fairing-builder-",
labels=labels,
namespace=self.namespace,
),
spec=pod_spec)
created_pod = client. \
CoreV1Api(). \
create_namespaced_pod(self.namespace, build_pod)
self.manager.log(name=created_pod.metadata.name,
namespace=created_pod.metadata.namespace,
selectors=labels)
# clean up created pod and secret
self.context_source.cleanup()
client.CoreV1Api().delete_namespaced_pod(
created_pod.metadata.name,
created_pod.metadata.namespace,
body=client.V1DeleteOptions())
class Job(DeployerInterface):
"""Handle all the k8s' template building for a training
Attributes:
namespace: k8s namespace where the training's components
will be deployed.
runs: Number of training(s) to be deployed. Hyperparameter search
will generate multiple jobs.
"""
def __init__(self,
namespace=None,
runs=1,
output=None,
cleanup=True,
labels=None,
job_name=DEFAULT_JOB_NAME,
stream_log=True,
deployer_type=DEPLOPYER_TYPE,
pod_spec_mutators=None):
if namespace is None:
self.namespace = utils.get_default_target_namespace()
else:
self.namespace = namespace
# Used as pod and job name
self.job_name = job_name
self.deployment_spec = None
self.runs = runs
self.output = output
self.backend = KubeManager()
self.cleanup = cleanup
self.stream_log = stream_log
self.set_labels(labels, deployer_type)
self.pod_spec_mutators = pod_spec_mutators or []
def set_labels(self, labels, deployer_type):
self.labels = {'fairing-deployer': deployer_type}
if labels:
self.labels.update(labels)
def deploy(self, pod_spec):
self.job_id = str(uuid.uuid1())
self.labels['fairing-id'] = self.job_id
for fn in self.pod_spec_mutators:
fn(self.backend, pod_spec, self.namespace)
pod_template_spec = self.generate_pod_template_spec(pod_spec)
pod_template_spec.spec.restart_policy = 'Never'
self.deployment_spec = self.generate_deployment_spec(pod_template_spec)
if self.output:
api = k8s_client.ApiClient()
job_output = api.sanitize_for_serialization(self.deployment_spec)
print(json.dumps(job_output))
name = self.create_resource()
logger.warn("Training job {} launched.".format(name))
if self.stream_log:
self.get_logs()
def create_resource(self):
self._created_job = self.backend.create_job(self.namespace,
self.deployment_spec)
return self._created_job.metadata.name
def generate_pod_template_spec(self, pod_spec):
"""Generate a V1PodTemplateSpec initiazlied with correct metadata
and with the provided pod_spec"""
if not isinstance(pod_spec, k8s_client.V1PodSpec):
raise TypeError('pod_spec must be a V1PodSpec, but got %s' %
type(pod_spec))
return k8s_client.V1PodTemplateSpec(metadata=k8s_client.V1ObjectMeta(
name="fairing-deployer", labels=self.labels),
spec=pod_spec)
def generate_deployment_spec(self, pod_template_spec):
"""Generate a V1Job initialized with correct completion and
parallelism (for HP search) and with the provided V1PodTemplateSpec"""
if not isinstance(pod_template_spec, k8s_client.V1PodTemplateSpec):
raise TypeError("""pod_template_spec must be a V1PodTemplateSpec,
but got %s""" % type(pod_template_spec))
job_spec = k8s_client.V1JobSpec(template=pod_template_spec,
parallelism=self.runs,
completions=self.runs)
return k8s_client.V1Job(api_version="batch/v1",
kind="Job",
metadata=k8s_client.V1ObjectMeta(
generate_name=self.job_name,
labels=self.labels,
),
spec=job_spec)
def get_logs(self):
self.backend.log(self._created_job.metadata.name,
self._created_job.metadata.namespace, self.labels)
if self.cleanup:
self.do_cleanup()
def do_cleanup(self):
logger.warn("Cleaning up job {}...".format(
self._created_job.metadata.name))
k8s_client.BatchV1Api().delete_namespaced_job(
self._created_job.metadata.name,
self._created_job.metadata.namespace,
body=k8s_client.V1DeleteOptions(propagation_policy='Foreground'))
class KFServing(DeployerInterface):
"""
Serves a prediction endpoint using Kubeflow KFServing.
Attributes:
framework: The framework for the kfservice, such as Tensorflow, XGBoost and ScikitLearn etc.
default_model_uri: URI pointing to Saved Model assets for default service.
canary_model_uri: URI pointing to Saved Model assets for canary service.
canary_traffic_percent: The amount of traffic to sent to the canary, defaults to 0.
namespace: The k8s namespace where the kfservice will be deployed.
labels: Labels for the kfservice, separate with commas if have more than one.
annotations: Annotations for the kfservice, separate with commas if have more than one.
custom_default_spec: A flexible custom default specification for arbitrary customer provided containers.
custom_canary_spec: A flexible custom canary specification for arbitrary customer provided containers.
stream_log: Show log or not when kfservice started, defaults to True.
cleanup: Delete the kfserving or not, defaults to False.
"""
def __init__(self, framework, default_model_uri=None, canary_model_uri=None, canary_traffic_percent=0,
namespace=None, labels=None, annotations=None, custom_default_spec=None,
custom_canary_spec=None, stream_log=True, cleanup=False):
self.framework = framework
self.default_model_uri = default_model_uri
self.canary_model_uri = canary_model_uri
self.canary_traffic_percent = canary_traffic_percent
self.annotations = annotations
self.set_labels(labels)
self.cleanup = cleanup
self.custom_default_spec = custom_default_spec
self.custom_canary_spec = custom_canary_spec
self.stream_log = stream_log
self.backend = KubeManager()
if namespace is None:
self.namespace = utils.get_default_target_namespace()
else:
self.namespace = namespace
def set_labels(self, labels):
self.fairing_id = str(uuid.uuid1())
self.labels = {'fairing-id': self.fairing_id}
if labels:
self.labels.update(labels)
def deploy(self, template_spec):
self.kfservice = self.generate_kfservice()
self.created_kfserving = self.backend.create_kfserving(self.namespace, self.kfservice)
if self.stream_log:
self.get_logs()
kfservice_name = self.created_kfserving['metadata']['name']
logger.warn("Deployed the kfservice {} successfully.".format(kfservice_name))
if self.cleanup:
logger.warn("Cleaning up kfservice {}...".format(kfservice_name))
self.backend.delete_kfserving(kfservice_name, self.namespace)
return kfservice_name
def generate_kfservice(self):
spec = {}
spec['default'] = {}
if self.framework != 'custom':
if self.default_model_uri != None:
spec['default'][self.framework] = {}
spec['default'][self.framework]['modelUri'] = self.default_model_uri
else:
raise RuntimeError("The default_model_uri must be defined if the framework is not custom.")
else:
if self.custom_default_spec != None:
# TBD @jinchi Need to validate the custom_default_spec before executing.
spec['default'][self.framework] = self.custom_default_spec
else:
raise RuntimeError("The custom_default_spec must be defined if the framework is custom.")
if self.framework != 'custom':
if self.canary_model_uri != None:
spec['canary'] = {}
spec['canary'][self.framework] = {}
spec['canary'][self.framework]['modelUri'] = self.canary_model_uri
spec['canaryTrafficPercent'] = self.canary_traffic_percent
else:
if self.custom_default_spec != None:
spec['canary'] = {}
spec['canary'][self.framework] = self.custom_canary_spec
spec['canaryTrafficPercent'] = self.canary_traffic_percent
metadata=k8s_client.V1ObjectMeta(
generate_name = constants.KFSERVING_DEFAULT_NAME,
namespace = self.namespace,
labels = self.labels,
annotations = self.annotations
)
kfservice = {}
kfservice['kind'] = constants.KFSERVING_KIND
kfservice['apiVersion'] = constants.KFSERVING_GROUP + '/' + constants.KFSERVING_VERSION
kfservice['metadata'] = metadata
kfservice['spec'] = spec
return kfservice
def get_logs(self):
name = self.created_kfserving['metadata']['name']
namespace = self.created_kfserving['metadata']['namespace']
self.backend.log(name, namespace, self.labels,
container=constants.KFSERVING_CONTAINER_NAME, follow=False)
