Kubeflow
zenml.integrations.kubeflow
special
Initialization of the Kubeflow integration for ZenML.
The Kubeflow integration sub-module powers an alternative to the local orchestrator. You can enable it by registering the Kubeflow orchestrator with the CLI tool.
KubeflowIntegration (Integration)
Definition of Kubeflow Integration for ZenML.
Source code in zenml/integrations/kubeflow/__init__.py
class KubeflowIntegration(Integration):
"""Definition of Kubeflow Integration for ZenML."""
NAME = KUBEFLOW
REQUIREMENTS = ["kfp==1.8.16"]
@classmethod
def flavors(cls) -> List[Type[Flavor]]:
"""Declare the stack component flavors for the Kubeflow integration.
Returns:
List of stack component flavors for this integration.
"""
from zenml.integrations.kubeflow.flavors import (
KubeflowOrchestratorFlavor,
)
return [KubeflowOrchestratorFlavor]
flavors()
classmethod
Declare the stack component flavors for the Kubeflow integration.
Returns:
| Type | Description |
|---|---|
List[Type[zenml.stack.flavor.Flavor]] |
List of stack component flavors for this integration. |
Source code in zenml/integrations/kubeflow/__init__.py
@classmethod
def flavors(cls) -> List[Type[Flavor]]:
"""Declare the stack component flavors for the Kubeflow integration.
Returns:
List of stack component flavors for this integration.
"""
from zenml.integrations.kubeflow.flavors import (
KubeflowOrchestratorFlavor,
)
return [KubeflowOrchestratorFlavor]
flavors
special
Kubeflow integration flavors.
kubeflow_orchestrator_flavor
Kubeflow orchestrator flavor.
KubeflowOrchestratorConfig (BaseOrchestratorConfig, KubeflowOrchestratorSettings)
pydantic-model
Configuration for the Kubeflow orchestrator.
Attributes:
| Name | Type | Description |
|---|---|---|
kubeflow_hostname |
Optional[str] |
The hostname to use to talk to the Kubeflow Pipelines API. If not set, the hostname will be derived from the Kubernetes API proxy. |
kubeflow_namespace |
str |
The Kubernetes namespace in which Kubeflow
Pipelines is deployed. Defaults to |
kubernetes_context |
str |
Optional name of a kubernetes context to run pipelines in. If not set, will try to spin up a local K3d cluster. |
local |
bool |
If |
skip_local_validations |
bool |
If |
Source code in zenml/integrations/kubeflow/flavors/kubeflow_orchestrator_flavor.py
class KubeflowOrchestratorConfig( # type: ignore[misc] # https://github.com/pydantic/pydantic/issues/4173
BaseOrchestratorConfig, KubeflowOrchestratorSettings
):
"""Configuration for the Kubeflow orchestrator.
Attributes:
kubeflow_hostname: The hostname to use to talk to the Kubeflow Pipelines
API. If not set, the hostname will be derived from the Kubernetes
API proxy.
kubeflow_namespace: The Kubernetes namespace in which Kubeflow
Pipelines is deployed. Defaults to `kubeflow`.
kubernetes_context: Optional name of a kubernetes context to run
pipelines in. If not set, will try to spin up a local K3d cluster.
local: If `True`, the orchestrator will assume it is connected to a
local kubernetes cluster and will perform additional validations and
operations to allow using the orchestrator in combination with other
local stack components that store data in the local filesystem
(i.e. it will mount the local stores directory into the pipeline
containers).
skip_local_validations: If `True`, the local validations will be
skipped.
"""
kubeflow_hostname: Optional[str] = None
kubeflow_namespace: str = "kubeflow"
kubernetes_context: str # TODO: Potential setting
local: bool = False
skip_local_validations: bool = False
@root_validator(pre=True)
def _validate_deprecated_attrs(
cls, values: Dict[str, Any]
) -> Dict[str, Any]:
"""Pydantic root_validator for deprecated attributes.
This root validator is used for backwards compatibility purposes. E.g.
it handles attributes that are no longer available or that have become
mandatory in the meantime.
Args:
values: Values passed to the object constructor
Returns:
Values passed to the object constructor
Raises:
ValueError: If the attributes or their values are not valid.
"""
provisioning_attrs = [
"skip_cluster_provisioning",
"skip_ui_daemon_provisioning",
"kubeflow_pipelines_ui_port",
]
provisioning_attrs_used = [
attr for attr in provisioning_attrs if attr in values
]
msg_header = (
"The ability to automatically provision and manage a Kubeflow "
"instance with `zenml stack up` on top of a local K3D cluster "
"is no longer available in the current version of ZenML "
"client. Please use the `k3d-modular` ZenML stack recipe to "
"achieve the same results (and more). Automatically exposing the "
"Kubeflow UI TCP port locally as part of the stack provisioning "
"has also been removed in favor of methods better suited for this "
"purpose, such as using an Ingress controller in the remote "
"cluster. \n"
"As a result, the `kubernetes_context` attribute is no longer "
"optional and the following Kubeflow orchestrator configuration "
"attributes have been deprecated: "
f"{provisioning_attrs}.\n"
)
if provisioning_attrs_used:
logger.warning(
msg_header
+ "To get rid of this warning, you should remove the deprecated "
"attributes from your orchestrator configuration (e.g. by "
"using the `zenml orchestrator remove-attribute <attr-name>` "
"CLI command)."
)
# remove deprecated attributes from values dict
for attr in provisioning_attrs_used:
del values[attr]
context = values.get("kubernetes_context")
if not context:
raise ValueError(
msg_header
+ "Please set the `kubernetes_context` attribute to the name "
"of the Kubernetes config context pointing to the cluster "
"where Kubeflow is installed (e.g. the K3D cluster provisioned "
"by the `k3d-modular` ZenML stack recipe) and also set the "
"`local` configuration flag."
)
# TODO: remove this in a future release. kept here for backwards
# compatibility with old stack configs
elif (
isinstance(context, str)
and context.startswith("k3d-zenml-kubeflow-")
and "local" not in values
):
values["local"] = True
return values
@property
def is_remote(self) -> bool:
"""Checks if this stack component is running remotely.
This designation is used to determine if the stack component can be
used with a local ZenML database or if it requires a remote ZenML
server.
Returns:
True if this config is for a remote component, False otherwise.
"""
return not self.local
@property
def is_local(self) -> bool:
"""Checks if this stack component is running locally.
This designation is used to determine if the stack component can be
shared with other users or if it is only usable on the local host.
Returns:
True if this config is for a local component, False otherwise.
"""
return self.local
is_local: bool
property
readonly
Checks if this stack component is running locally.
This designation is used to determine if the stack component can be shared with other users or if it is only usable on the local host.
Returns:
| Type | Description |
|---|---|
bool |
True if this config is for a local component, False otherwise. |
is_remote: bool
property
readonly
Checks if this stack component is running remotely.
This designation is used to determine if the stack component can be used with a local ZenML database or if it requires a remote ZenML server.
Returns:
| Type | Description |
|---|---|
bool |
True if this config is for a remote component, False otherwise. |
KubeflowOrchestratorFlavor (BaseOrchestratorFlavor)
Kubeflow orchestrator flavor.
Source code in zenml/integrations/kubeflow/flavors/kubeflow_orchestrator_flavor.py
class KubeflowOrchestratorFlavor(BaseOrchestratorFlavor):
"""Kubeflow orchestrator flavor."""
@property
def name(self) -> str:
"""Name of the flavor.
Returns:
The name of the flavor.
"""
return KUBEFLOW_ORCHESTRATOR_FLAVOR
@property
def docs_url(self) -> Optional[str]:
"""A url to point at docs explaining this flavor.
Returns:
A flavor docs url.
"""
return self.generate_default_docs_url()
@property
def sdk_docs_url(self) -> Optional[str]:
"""A url to point at SDK docs explaining this flavor.
Returns:
A flavor SDK docs url.
"""
return self.generate_default_sdk_docs_url()
@property
def logo_url(self) -> str:
"""A url to represent the flavor in the dashboard.
Returns:
The flavor logo.
"""
return "https://public-flavor-logos.s3.eu-central-1.amazonaws.com/orchestrator/kubeflow.png"
@property
def config_class(self) -> Type[KubeflowOrchestratorConfig]:
"""Returns `KubeflowOrchestratorConfig` config class.
Returns:
The config class.
"""
return KubeflowOrchestratorConfig
@property
def implementation_class(self) -> Type["KubeflowOrchestrator"]:
"""Implementation class for this flavor.
Returns:
The implementation class.
"""
from zenml.integrations.kubeflow.orchestrators import (
KubeflowOrchestrator,
)
return KubeflowOrchestrator
config_class: Type[zenml.integrations.kubeflow.flavors.kubeflow_orchestrator_flavor.KubeflowOrchestratorConfig]
property
readonly
Returns KubeflowOrchestratorConfig config class.
Returns:
| Type | Description |
|---|---|
Type[zenml.integrations.kubeflow.flavors.kubeflow_orchestrator_flavor.KubeflowOrchestratorConfig] |
The config class. |
docs_url: Optional[str]
property
readonly
A url to point at docs explaining this flavor.
Returns:
| Type | Description |
|---|---|
Optional[str] |
A flavor docs url. |
implementation_class: Type[KubeflowOrchestrator]
property
readonly
Implementation class for this flavor.
Returns:
| Type | Description |
|---|---|
Type[KubeflowOrchestrator] |
The implementation class. |
logo_url: str
property
readonly
A url to represent the flavor in the dashboard.
Returns:
| Type | Description |
|---|---|
str |
The flavor logo. |
name: str
property
readonly
Name of the flavor.
Returns:
| Type | Description |
|---|---|
str |
The name of the flavor. |
sdk_docs_url: Optional[str]
property
readonly
A url to point at SDK docs explaining this flavor.
Returns:
| Type | Description |
|---|---|
Optional[str] |
A flavor SDK docs url. |
KubeflowOrchestratorSettings (BaseSettings)
pydantic-model
Settings for the Kubeflow orchestrator.
Attributes:
| Name | Type | Description |
|---|---|---|
synchronous |
bool |
If |
timeout |
int |
How many seconds to wait for synchronous runs. |
client_args |
Dict[str, Any] |
Arguments to pass when initializing the KFP client. |
client_username |
Optional[str] |
Username to generate a session cookie for the kubeflow client. Both |
client_password |
Optional[str] |
Password to generate a session cookie for the kubeflow client. Both |
user_namespace |
Optional[str] |
The user namespace to use when creating experiments and runs. |
node_selectors |
Dict[str, str] |
Deprecated: Node selectors to apply to KFP pods. |
node_affinity |
Dict[str, List[str]] |
Deprecated: Node affinities to apply to KFP pods. |
pod_settings |
Optional[zenml.integrations.kubernetes.pod_settings.KubernetesPodSettings] |
Pod settings to apply. |
Source code in zenml/integrations/kubeflow/flavors/kubeflow_orchestrator_flavor.py
class KubeflowOrchestratorSettings(BaseSettings):
"""Settings for the Kubeflow orchestrator.
Attributes:
synchronous: If `True`, running a pipeline using this orchestrator will
block until all steps finished running on KFP. This setting only
has an effect when specified on the pipeline and will be ignored if
specified on steps.
timeout: How many seconds to wait for synchronous runs.
client_args: Arguments to pass when initializing the KFP client.
client_username: Username to generate a session cookie for the kubeflow client. Both `client_username`
and `client_password` need to be set together.
client_password: Password to generate a session cookie for the kubeflow client. Both `client_username`
and `client_password` need to be set together.
user_namespace: The user namespace to use when creating experiments
and runs.
node_selectors: Deprecated: Node selectors to apply to KFP pods.
node_affinity: Deprecated: Node affinities to apply to KFP pods.
pod_settings: Pod settings to apply.
"""
synchronous: bool = False
timeout: int = 1200
client_args: Dict[str, Any] = {}
client_username: Optional[str] = SecretField()
client_password: Optional[str] = SecretField()
user_namespace: Optional[str] = None
node_selectors: Dict[str, str] = {}
node_affinity: Dict[str, List[str]] = {}
pod_settings: Optional[KubernetesPodSettings] = None
@root_validator
def _validate_and_migrate_pod_settings(
cls, values: Dict[str, Any]
) -> Dict[str, Any]:
"""Validates settings and migrates pod settings from older version.
Args:
values: Dict representing user-specified runtime settings.
Returns:
Validated settings.
Raises:
AssertionError: If old and new settings are used together.
ValueError: If username and password are not specified together.
"""
has_pod_settings = bool(values.get("pod_settings"))
node_selectors = cast(
Dict[str, str], values.get("node_selectors") or {}
)
node_affinity = cast(
Dict[str, List[str]], values.get("node_affinity") or {}
)
has_old_settings = any([node_selectors, node_affinity])
if has_old_settings:
logger.warning(
"The attributes `node_selectors` and `node_affinity` of the "
"Kubeflow settings will be deprecated soon. Use the "
"attribute `pod_settings` instead.",
)
if has_pod_settings and has_old_settings:
raise AssertionError(
"Got Kubeflow pod settings using both the deprecated "
"attributes `node_selectors` and `node_affinity` as well as "
"the new attribute `pod_settings`. Please specify Kubeflow "
"pod settings only using the new `pod_settings` attribute."
)
elif has_old_settings:
from kubernetes import client as k8s_client
affinity = {}
if node_affinity:
match_expressions = [
k8s_client.V1NodeSelectorRequirement(
key=key,
operator="In",
values=values,
)
for key, values in node_affinity.items()
]
affinity = k8s_client.V1Affinity(
node_affinity=k8s_client.V1NodeAffinity(
required_during_scheduling_ignored_during_execution=k8s_client.V1NodeSelector(
node_selector_terms=[
k8s_client.V1NodeSelectorTerm(
match_expressions=match_expressions
)
]
)
)
)
pod_settings = KubernetesPodSettings(
node_selectors=node_selectors, affinity=affinity
)
values["pod_settings"] = pod_settings
values["node_affinity"] = {}
values["node_selectors"] = {}
# Validate username and password for auth cookie logic
username = values.get("client_username")
password = values.get("client_password")
client_creds_error = "`client_username` and `client_password` both need to be set together."
if username and password is None:
raise ValueError(client_creds_error)
if password and username is None:
raise ValueError(client_creds_error)
return values
orchestrators
special
Initialization of the Kubeflow ZenML orchestrator.
kubeflow_orchestrator
Implementation of the Kubeflow orchestrator.
KubeflowOrchestrator (ContainerizedOrchestrator)
Orchestrator responsible for running pipelines using Kubeflow.
Source code in zenml/integrations/kubeflow/orchestrators/kubeflow_orchestrator.py
class KubeflowOrchestrator(ContainerizedOrchestrator):
"""Orchestrator responsible for running pipelines using Kubeflow."""
@property
def config(self) -> KubeflowOrchestratorConfig:
"""Returns the `KubeflowOrchestratorConfig` config.
Returns:
The configuration.
"""
return cast(KubeflowOrchestratorConfig, self._config)
@property
def kubernetes_context(self) -> str:
"""Gets the kubernetes context associated with the orchestrator.
Returns:
The kubernetes context associated with the orchestrator.
"""
return self.config.kubernetes_context
def get_kubernetes_contexts(self) -> Tuple[List[str], Optional[str]]:
"""Get the list of configured Kubernetes contexts and the active context.
Returns:
A tuple containing the list of configured Kubernetes contexts and
the active context.
"""
try:
contexts, active_context = k8s_config.list_kube_config_contexts()
except k8s_config.config_exception.ConfigException:
return [], None
context_names = [c["name"] for c in contexts]
active_context_name = active_context["name"]
return context_names, active_context_name
@property
def settings_class(self) -> Type[KubeflowOrchestratorSettings]:
"""Settings class for the Kubeflow orchestrator.
Returns:
The settings class.
"""
return KubeflowOrchestratorSettings
@property
def validator(self) -> Optional[StackValidator]:
"""Validates that the stack contains a container registry.
Also check that requirements are met for local components.
Returns:
A `StackValidator` instance.
"""
def _validate_local_requirements(stack: "Stack") -> Tuple[bool, str]:
container_registry = stack.container_registry
# should not happen, because the stack validation takes care of
# this, but just in case
assert container_registry is not None
contexts, active_context = self.get_kubernetes_contexts()
if self.kubernetes_context not in contexts:
if not self.config.is_local:
return False, (
f"Could not find a Kubernetes context named "
f"'{self.kubernetes_context}' in the local Kubernetes "
f"configuration. Please make sure that the Kubernetes "
f"cluster is running and that the kubeconfig file is "
f"configured correctly. To list all configured "
f"contexts, run:\n\n"
f" `kubectl config get-contexts`\n"
)
elif active_context and self.kubernetes_context != active_context:
logger.warning(
f"The Kubernetes context '{self.kubernetes_context}' "
f"configured for the Kubeflow orchestrator is not the "
f"same as the active context in the local Kubernetes "
f"configuration. If this is not deliberate, you should "
f"update the orchestrator's `kubernetes_context` field by "
f"running:\n\n"
f" `zenml orchestrator update {self.name} "
f"--kubernetes_context={active_context}`\n"
f"To list all configured contexts, run:\n\n"
f" `kubectl config get-contexts`\n"
f"To set the active context to be the same as the one "
f"configured in the Kubeflow orchestrator and silence "
f"this warning, run:\n\n"
f" `kubectl config use-context "
f"{self.kubernetes_context}`\n"
)
silence_local_validations_msg = (
f"To silence this warning, set the "
f"`skip_local_validations` attribute to True in the "
f"orchestrator configuration by running:\n\n"
f" 'zenml orchestrator update {self.name} "
f"--skip_local_validations=True'\n"
)
if (
not self.config.skip_local_validations
and not self.config.is_local
):
# if the orchestrator is not running in a local k3d cluster,
# we cannot have any other local components in our stack,
# because we cannot mount the local path into the container.
# This may result in problems when running the pipeline,
# because the local components will not be available inside the
# Kubeflow containers.
# go through all stack components and identify those that
# advertise a local path where they persist information that
# they need to be available when running pipelines.
for stack_comp in stack.components.values():
local_path = stack_comp.local_path
if not local_path:
continue
return False, (
f"The Kubeflow orchestrator is configured to run "
f"pipelines in a remote Kubernetes cluster designated "
f"by the '{self.kubernetes_context}' configuration "
f"context, but the '{stack_comp.name}' "
f"{stack_comp.type.value} is a local stack component "
f"and will not be available in the Kubeflow pipeline "
f"step.\nPlease ensure that you always use non-local "
f"stack components with a remote Kubeflow "
f"orchestrator, otherwise you may run into pipeline "
f"execution problems. You should use a flavor of "
f"{stack_comp.type.value} other than "
f"'{stack_comp.flavor}'.\n"
+ silence_local_validations_msg
)
# if the orchestrator is remote, the container registry must
# also be remote.
if container_registry.config.is_local:
return False, (
f"The Kubeflow orchestrator is configured to run "
f"pipelines in a remote Kubernetes cluster designated "
f"by the '{self.kubernetes_context}' configuration "
f"context, but the '{container_registry.name}' "
f"container registry URI "
f"'{container_registry.config.uri}' "
f"points to a local container registry. Please ensure "
f"that you always use non-local stack components with "
f"a remote Kubeflow orchestrator, otherwise you will "
f"run into problems. You should use a flavor of "
f"container registry other than "
f"'{container_registry.flavor}'.\n"
+ silence_local_validations_msg
)
return True, ""
return StackValidator(
required_components={
StackComponentType.CONTAINER_REGISTRY,
StackComponentType.IMAGE_BUILDER,
},
custom_validation_function=_validate_local_requirements,
)
@property
def root_directory(self) -> str:
"""Path to the root directory for all files concerning this orchestrator.
Returns:
Path to the root directory.
"""
return os.path.join(
io_utils.get_global_config_directory(),
"kubeflow",
str(self.id),
)
@property
def pipeline_directory(self) -> str:
"""Returns path to a directory in which the kubeflow pipeline files are stored.
Returns:
Path to the pipeline directory.
"""
return os.path.join(self.root_directory, "pipelines")
def _configure_container_op(
self,
container_op: dsl.ContainerOp,
settings: KubeflowOrchestratorSettings,
) -> None:
"""Makes changes in place to the configuration of the container op.
Configures persistent mounted volumes for each stack component that
writes to a local path. Adds some labels to the container_op and applies
some functions to ir.
Args:
container_op: The kubeflow container operation to configure.
settings: Orchestrator settings for this step.
"""
volumes: Dict[str, k8s_client.V1Volume] = {}
stack = Client().active_stack
if self.config.is_local:
stack.check_local_paths()
local_stores_path = GlobalConfiguration().local_stores_path
host_path = k8s_client.V1HostPathVolumeSource(
path=local_stores_path, type="Directory"
)
volumes[local_stores_path] = k8s_client.V1Volume(
name="local-stores",
host_path=host_path,
)
logger.debug(
"Adding host path volume for the local ZenML stores (path: %s) "
"in kubeflow pipelines container.",
local_stores_path,
)
if sys.platform == "win32":
# File permissions are not checked on Windows. This if clause
# prevents mypy from complaining about unused 'type: ignore'
# statements
pass
else:
# Run KFP containers in the context of the local UID/GID
# to ensure that the artifact and metadata stores can be shared
# with the local pipeline runs.
container_op.container.security_context = (
k8s_client.V1SecurityContext(
run_as_user=os.getuid(),
run_as_group=os.getgid(),
)
)
logger.debug(
"Setting security context UID and GID to local user/group "
"in kubeflow pipelines container."
)
container_op.container.add_env_variable(
k8s_client.V1EnvVar(
name=ENV_ZENML_LOCAL_STORES_PATH,
value=local_stores_path,
)
)
container_op.add_pvolumes(volumes)
# Add some pod labels to the container_op
for k, v in KFP_POD_LABELS.items():
container_op.add_pod_label(k, v)
if settings.pod_settings:
apply_pod_settings(
container_op=container_op, settings=settings.pod_settings
)
# Disable caching in KFP v1 only works like this, replace by the second
# line in the future
container_op.execution_options.caching_strategy.max_cache_staleness = (
"P0D"
)
# container_op.set_caching_options(enable_caching=False)
@staticmethod
def _configure_container_resources(
container_op: dsl.ContainerOp,
resource_settings: "ResourceSettings",
) -> None:
"""Adds resource requirements to the container.
Args:
container_op: The kubeflow container operation to configure.
resource_settings: The resource settings to use for this
container.
"""
if resource_settings.cpu_count is not None:
container_op = container_op.set_cpu_limit(
str(resource_settings.cpu_count)
)
if resource_settings.gpu_count is not None:
container_op = container_op.set_gpu_limit(
resource_settings.gpu_count
)
if resource_settings.memory is not None:
memory_limit = resource_settings.memory[:-1]
container_op = container_op.set_memory_limit(memory_limit)
def prepare_or_run_pipeline(
self,
deployment: "PipelineDeploymentResponseModel",
stack: "Stack",
) -> Any:
"""Creates a kfp yaml file.
This functions as an intermediary representation of the pipeline which
is then deployed to the kubeflow pipelines instance.
How it works:
-------------
Before this method is called the `prepare_pipeline_deployment()`
method builds a docker image that contains the code for the
pipeline, all steps the context around these files.
Based on this docker image a callable is created which builds
container_ops for each step (`_construct_kfp_pipeline`).
To do this the entrypoint of the docker image is configured to
run the correct step within the docker image. The dependencies
between these container_ops are then also configured onto each
container_op by pointing at the downstream steps.
This callable is then compiled into a kfp yaml file that is used as
the intermediary representation of the kubeflow pipeline.
This file, together with some metadata, runtime configurations is
then uploaded into the kubeflow pipelines cluster for execution.
Args:
deployment: The pipeline deployment to prepare or run.
stack: The stack the pipeline will run on.
Raises:
RuntimeError: If trying to run a pipeline in a notebook
environment.
"""
# First check whether the code running in a notebook
if Environment.in_notebook():
raise RuntimeError(
"The Kubeflow orchestrator cannot run pipelines in a notebook "
"environment. The reason is that it is non-trivial to create "
"a Docker image of a notebook. Please consider refactoring "
"your notebook cells into separate scripts in a Python module "
"and run the code outside of a notebook when using this "
"orchestrator."
)
assert stack.container_registry
# Create a callable for future compilation into a dsl.Pipeline.
def _construct_kfp_pipeline() -> None:
"""Create a container_op for each step.
This should contain the name of the docker image and configures the
entrypoint of the docker image to run the step.
Additionally, this gives each container_op information about its
direct downstream steps.
If this callable is passed to the `_create_and_write_workflow()`
method of a KFPCompiler all dsl.ContainerOp instances will be
automatically added to a singular dsl.Pipeline instance.
"""
# Dictionary of container_ops index by the associated step name
step_name_to_container_op: Dict[str, dsl.ContainerOp] = {}
for step_name, step in deployment.step_configurations.items():
image = self.get_image(
deployment=deployment, step_name=step_name
)
# The command will be needed to eventually call the python step
# within the docker container
command = StepEntrypointConfiguration.get_entrypoint_command()
# The arguments are passed to configure the entrypoint of the
# docker container when the step is called.
arguments = (
StepEntrypointConfiguration.get_entrypoint_arguments(
step_name=step_name, deployment_id=deployment.id
)
)
# Create a container_op - the kubeflow equivalent of a step. It
# contains the name of the step, the name of the docker image,
# the command to use to run the step entrypoint
# (e.g. `python -m zenml.entrypoints.step_entrypoint`)
# and the arguments to be passed along with the command. Find
# out more about how these arguments are parsed and used
# in the base entrypoint `run()` method.
container_op = dsl.ContainerOp(
name=step.config.name,
image=image,
command=command,
arguments=arguments,
)
settings = cast(
KubeflowOrchestratorSettings, self.get_settings(step)
)
self._configure_container_op(
container_op=container_op,
settings=settings,
)
if self.requires_resources_in_orchestration_environment(step):
self._configure_container_resources(
container_op=container_op,
resource_settings=step.config.resource_settings,
)
# Find the upstream container ops of the current step and
# configure the current container op to run after them
for upstream_step_name in step.spec.upstream_steps:
upstream_container_op = step_name_to_container_op[
upstream_step_name
]
container_op.after(upstream_container_op)
# Update dictionary of container ops with the current one
step_name_to_container_op[step.config.name] = container_op
orchestrator_run_name = get_orchestrator_run_name(
pipeline_name=deployment.pipeline_configuration.name
)
# Get a filepath to use to save the finished yaml to
fileio.makedirs(self.pipeline_directory)
pipeline_file_path = os.path.join(
self.pipeline_directory, f"{orchestrator_run_name}.yaml"
)
# write the argo pipeline yaml
KFPCompiler()._create_and_write_workflow(
pipeline_func=_construct_kfp_pipeline,
pipeline_name=deployment.pipeline_configuration.name,
package_path=pipeline_file_path,
)
logger.info(
"Writing Kubeflow workflow definition to `%s`.", pipeline_file_path
)
# using the kfp client uploads the pipeline to kubeflow pipelines and
# runs it there
self._upload_and_run_pipeline(
deployment=deployment,
pipeline_file_path=pipeline_file_path,
run_name=orchestrator_run_name,
)
def _upload_and_run_pipeline(
self,
deployment: "PipelineDeploymentResponseModel",
pipeline_file_path: str,
run_name: str,
) -> None:
"""Tries to upload and run a KFP pipeline.
Args:
deployment: The pipeline deployment.
pipeline_file_path: Path to the pipeline definition file.
run_name: The Kubeflow run name.
Raises:
RuntimeError: If Kubeflow API returns an error.
"""
pipeline_name = deployment.pipeline_configuration.name
settings = cast(
KubeflowOrchestratorSettings, self.get_settings(deployment)
)
user_namespace = settings.user_namespace
try:
logger.info(
"Running in kubernetes context '%s'.",
self.kubernetes_context,
)
# upload the pipeline to Kubeflow and start it
client = self._get_kfp_client(settings=settings)
if deployment.schedule:
try:
experiment = client.get_experiment(
pipeline_name, namespace=user_namespace
)
logger.info(
"A recurring run has already been created with this "
"pipeline. Creating new recurring run now.."
)
except (ValueError, ApiException):
experiment = client.create_experiment(
pipeline_name, namespace=user_namespace
)
logger.info(
"Creating a new recurring run for pipeline '%s'.. ",
pipeline_name,
)
logger.info(
"You can see all recurring runs under the '%s' experiment.",
pipeline_name,
)
interval_seconds = (
deployment.schedule.interval_second.seconds
if deployment.schedule.interval_second
else None
)
result = client.create_recurring_run(
experiment_id=experiment.id,
job_name=run_name,
pipeline_package_path=pipeline_file_path,
enable_caching=False,
cron_expression=deployment.schedule.cron_expression,
start_time=deployment.schedule.utc_start_time,
end_time=deployment.schedule.utc_end_time,
interval_second=interval_seconds,
no_catchup=not deployment.schedule.catchup,
)
logger.info("Started recurring run with ID '%s'.", result.id)
else:
logger.info(
"No schedule detected. Creating a one-off pipeline run.."
)
try:
result = client.create_run_from_pipeline_package(
pipeline_file_path,
arguments={},
run_name=run_name,
enable_caching=False,
namespace=user_namespace,
)
except ApiException:
raise RuntimeError(
f"Failed to create {run_name} on kubeflow! "
"Please check stack component settings and configuration!"
)
logger.info(
"Started one-off pipeline run with ID '%s'.", result.run_id
)
if settings.synchronous:
client.wait_for_run_completion(
run_id=result.run_id, timeout=settings.timeout
)
except urllib3.exceptions.HTTPError as error:
logger.warning(
f"Failed to upload Kubeflow pipeline: %s. "
f"Please make sure your kubernetes config is present and the "
f"{self.kubernetes_context} kubernetes context is configured "
f"correctly.",
error,
)
def get_orchestrator_run_id(self) -> str:
"""Returns the active orchestrator run id.
Raises:
RuntimeError: If the environment variable specifying the run id
is not set.
Returns:
The orchestrator run id.
"""
try:
return os.environ[ENV_KFP_RUN_ID]
except KeyError:
raise RuntimeError(
"Unable to read run id from environment variable "
f"{ENV_KFP_RUN_ID}."
)
def _get_kfp_client(
self,
settings: KubeflowOrchestratorSettings,
) -> kfp.Client:
"""Creates a KFP client instance.
Args:
settings: Settings which can be used to
configure the client instance.
Returns:
A KFP client instance.
"""
client_args = {
"kube_context": self.config.kubernetes_context,
}
client_args.update(settings.client_args)
# The host and namespace are stack component configurations that refer
# to the Kubeflow deployment. We don't want these overwritten on a
# run by run basis by user settings
if self.config.kubeflow_hostname:
client_args["host"] = self.config.kubeflow_hostname
client_args["namespace"] = self.config.kubeflow_namespace
# Handle username and password, ignore the case if one is passed and not the other
# Also do not attempt to get cookie if cookie is already passed in client_args
if settings.client_username and settings.client_password:
# If cookie is already set, then ignore
if "cookie" in client_args:
logger.warning(
"Cookie already set in `client_args`, ignoring `client_username` and `client_password`..."
)
else:
session_cookie = self._get_session_cookie(
username=settings.client_username,
password=settings.client_password,
)
client_args["cookies"] = session_cookie
return kfp.Client(**client_args)
def _get_session_cookie(self, username: str, password: str) -> str:
"""Gets session cookie from username and password.
Args:
username: Username for kubeflow host.
password: Password for kubeflow host.
Raises:
RuntimeError: If the cookie fetching failed.
Returns:
Cookie with the prefix `authsession=`.
"""
if self.config.kubeflow_hostname is None:
raise RuntimeError(
"You must configure the Kubeflow orchestrator "
"with the `kubeflow_hostname` parameter which usually ends "
"with `/pipeline` (e.g. `https://mykubeflow.com/pipeline`). "
"Please update the current kubeflow orchestrator with: "
f"`zenml orchestrator update {self.name} "
"--kubeflow_hostname=<MY_KUBEFLOW_HOST>`"
)
# Get cookie
logger.info(
f"Attempting to fetch session cookie from {self.config.kubeflow_hostname} "
"with supplied username and password..."
)
session = requests.Session()
try:
response = session.get(self.config.kubeflow_hostname)
response.raise_for_status()
except (
requests.exceptions.HTTPError,
requests.exceptions.ConnectionError,
requests.exceptions.Timeout,
requests.exceptions.RequestException,
) as e:
raise RuntimeError(
f"Error while trying to fetch kubeflow cookie: {e}"
)
headers = {
"Content-Type": "application/x-www-form-urlencoded",
}
data = {"login": username, "password": password}
try:
response = session.post(response.url, headers=headers, data=data)
response.raise_for_status()
except requests.exceptions.HTTPError as errh:
raise RuntimeError(
f"Error while trying to fetch kubeflow cookie: {errh}"
)
cookie_dict = session.cookies.get_dict() # type: ignore[no-untyped-call]
if "authservice_session" not in cookie_dict:
raise RuntimeError("Invalid username and/or password!")
logger.info("Session cookie fetched successfully!")
return "authservice_session=" + str(cookie_dict["authservice_session"])
def get_pipeline_run_metadata(
self, run_id: UUID
) -> Dict[str, "MetadataType"]:
"""Get general component-specific metadata for a pipeline run.
Args:
run_id: The ID of the pipeline run.
Returns:
A dictionary of metadata.
"""
hostname = self.config.kubeflow_hostname
if not hostname:
return {}
hostname = hostname.rstrip("/")
pipeline_suffix = "/pipeline"
if hostname.endswith(pipeline_suffix):
hostname = hostname[: -len(pipeline_suffix)]
run = Client().get_pipeline_run(run_id)
settings_key = settings_utils.get_stack_component_setting_key(self)
run_settings = self.settings_class.parse_obj(
run.pipeline_configuration.get(settings_key, self.config)
)
user_namespace = run_settings.user_namespace
if user_namespace:
run_url = (
f"{hostname}/_/pipeline/?ns={user_namespace}#"
f"/runs/details/{self.get_orchestrator_run_id()}"
)
return {
METADATA_ORCHESTRATOR_URL: Uri(run_url),
}
else:
return {
METADATA_ORCHESTRATOR_URL: Uri(f"{hostname}"),
}
config: KubeflowOrchestratorConfig
property
readonly
Returns the KubeflowOrchestratorConfig config.
Returns:
| Type | Description |
|---|---|
KubeflowOrchestratorConfig |
The configuration. |
kubernetes_context: str
property
readonly
Gets the kubernetes context associated with the orchestrator.
Returns:
| Type | Description |
|---|---|
str |
The kubernetes context associated with the orchestrator. |
pipeline_directory: str
property
readonly
Returns path to a directory in which the kubeflow pipeline files are stored.
Returns:
| Type | Description |
|---|---|
str |
Path to the pipeline directory. |
root_directory: str
property
readonly
Path to the root directory for all files concerning this orchestrator.
Returns:
| Type | Description |
|---|---|
str |
Path to the root directory. |
settings_class: Type[zenml.integrations.kubeflow.flavors.kubeflow_orchestrator_flavor.KubeflowOrchestratorSettings]
property
readonly
Settings class for the Kubeflow orchestrator.
Returns:
| Type | Description |
|---|---|
Type[zenml.integrations.kubeflow.flavors.kubeflow_orchestrator_flavor.KubeflowOrchestratorSettings] |
The settings class. |
validator: Optional[zenml.stack.stack_validator.StackValidator]
property
readonly
Validates that the stack contains a container registry.
Also check that requirements are met for local components.
Returns:
| Type | Description |
|---|---|
Optional[zenml.stack.stack_validator.StackValidator] |
A |
get_kubernetes_contexts(self)
Get the list of configured Kubernetes contexts and the active context.
Returns:
| Type | Description |
|---|---|
Tuple[List[str], Optional[str]] |
A tuple containing the list of configured Kubernetes contexts and the active context. |
Source code in zenml/integrations/kubeflow/orchestrators/kubeflow_orchestrator.py
def get_kubernetes_contexts(self) -> Tuple[List[str], Optional[str]]:
"""Get the list of configured Kubernetes contexts and the active context.
Returns:
A tuple containing the list of configured Kubernetes contexts and
the active context.
"""
try:
contexts, active_context = k8s_config.list_kube_config_contexts()
except k8s_config.config_exception.ConfigException:
return [], None
context_names = [c["name"] for c in contexts]
active_context_name = active_context["name"]
return context_names, active_context_name
get_orchestrator_run_id(self)
Returns the active orchestrator run id.
Exceptions:
| Type | Description |
|---|---|
RuntimeError |
If the environment variable specifying the run id is not set. |
Returns:
| Type | Description |
|---|---|
str |
The orchestrator run id. |
Source code in zenml/integrations/kubeflow/orchestrators/kubeflow_orchestrator.py
def get_orchestrator_run_id(self) -> str:
"""Returns the active orchestrator run id.
Raises:
RuntimeError: If the environment variable specifying the run id
is not set.
Returns:
The orchestrator run id.
"""
try:
return os.environ[ENV_KFP_RUN_ID]
except KeyError:
raise RuntimeError(
"Unable to read run id from environment variable "
f"{ENV_KFP_RUN_ID}."
)
get_pipeline_run_metadata(self, run_id)
Get general component-specific metadata for a pipeline run.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
run_id |
UUID |
The ID of the pipeline run. |
required |
Returns:
| Type | Description |
|---|---|
Dict[str, MetadataType] |
A dictionary of metadata. |
Source code in zenml/integrations/kubeflow/orchestrators/kubeflow_orchestrator.py
def get_pipeline_run_metadata(
self, run_id: UUID
) -> Dict[str, "MetadataType"]:
"""Get general component-specific metadata for a pipeline run.
Args:
run_id: The ID of the pipeline run.
Returns:
A dictionary of metadata.
"""
hostname = self.config.kubeflow_hostname
if not hostname:
return {}
hostname = hostname.rstrip("/")
pipeline_suffix = "/pipeline"
if hostname.endswith(pipeline_suffix):
hostname = hostname[: -len(pipeline_suffix)]
run = Client().get_pipeline_run(run_id)
settings_key = settings_utils.get_stack_component_setting_key(self)
run_settings = self.settings_class.parse_obj(
run.pipeline_configuration.get(settings_key, self.config)
)
user_namespace = run_settings.user_namespace
if user_namespace:
run_url = (
f"{hostname}/_/pipeline/?ns={user_namespace}#"
f"/runs/details/{self.get_orchestrator_run_id()}"
)
return {
METADATA_ORCHESTRATOR_URL: Uri(run_url),
}
else:
return {
METADATA_ORCHESTRATOR_URL: Uri(f"{hostname}"),
}
prepare_or_run_pipeline(self, deployment, stack)
Creates a kfp yaml file.
This functions as an intermediary representation of the pipeline which is then deployed to the kubeflow pipelines instance.
How it works:
Before this method is called the prepare_pipeline_deployment()
method builds a docker image that contains the code for the
pipeline, all steps the context around these files.
Based on this docker image a callable is created which builds
container_ops for each step (_construct_kfp_pipeline).
To do this the entrypoint of the docker image is configured to
run the correct step within the docker image. The dependencies
between these container_ops are then also configured onto each
container_op by pointing at the downstream steps.
This callable is then compiled into a kfp yaml file that is used as the intermediary representation of the kubeflow pipeline.
This file, together with some metadata, runtime configurations is then uploaded into the kubeflow pipelines cluster for execution.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
deployment |
PipelineDeploymentResponseModel |
The pipeline deployment to prepare or run. |
required |
stack |
Stack |
The stack the pipeline will run on. |
required |
Exceptions:
| Type | Description |
|---|---|
RuntimeError |
If trying to run a pipeline in a notebook environment. |
Source code in zenml/integrations/kubeflow/orchestrators/kubeflow_orchestrator.py
def prepare_or_run_pipeline(
self,
deployment: "PipelineDeploymentResponseModel",
stack: "Stack",
) -> Any:
"""Creates a kfp yaml file.
This functions as an intermediary representation of the pipeline which
is then deployed to the kubeflow pipelines instance.
How it works:
-------------
Before this method is called the `prepare_pipeline_deployment()`
method builds a docker image that contains the code for the
pipeline, all steps the context around these files.
Based on this docker image a callable is created which builds
container_ops for each step (`_construct_kfp_pipeline`).
To do this the entrypoint of the docker image is configured to
run the correct step within the docker image. The dependencies
between these container_ops are then also configured onto each
container_op by pointing at the downstream steps.
This callable is then compiled into a kfp yaml file that is used as
the intermediary representation of the kubeflow pipeline.
This file, together with some metadata, runtime configurations is
then uploaded into the kubeflow pipelines cluster for execution.
Args:
deployment: The pipeline deployment to prepare or run.
stack: The stack the pipeline will run on.
Raises:
RuntimeError: If trying to run a pipeline in a notebook
environment.
"""
# First check whether the code running in a notebook
if Environment.in_notebook():
raise RuntimeError(
"The Kubeflow orchestrator cannot run pipelines in a notebook "
"environment. The reason is that it is non-trivial to create "
"a Docker image of a notebook. Please consider refactoring "
"your notebook cells into separate scripts in a Python module "
"and run the code outside of a notebook when using this "
"orchestrator."
)
assert stack.container_registry
# Create a callable for future compilation into a dsl.Pipeline.
def _construct_kfp_pipeline() -> None:
"""Create a container_op for each step.
This should contain the name of the docker image and configures the
entrypoint of the docker image to run the step.
Additionally, this gives each container_op information about its
direct downstream steps.
If this callable is passed to the `_create_and_write_workflow()`
method of a KFPCompiler all dsl.ContainerOp instances will be
automatically added to a singular dsl.Pipeline instance.
"""
# Dictionary of container_ops index by the associated step name
step_name_to_container_op: Dict[str, dsl.ContainerOp] = {}
for step_name, step in deployment.step_configurations.items():
image = self.get_image(
deployment=deployment, step_name=step_name
)
# The command will be needed to eventually call the python step
# within the docker container
command = StepEntrypointConfiguration.get_entrypoint_command()
# The arguments are passed to configure the entrypoint of the
# docker container when the step is called.
arguments = (
StepEntrypointConfiguration.get_entrypoint_arguments(
step_name=step_name, deployment_id=deployment.id
)
)
# Create a container_op - the kubeflow equivalent of a step. It
# contains the name of the step, the name of the docker image,
# the command to use to run the step entrypoint
# (e.g. `python -m zenml.entrypoints.step_entrypoint`)
# and the arguments to be passed along with the command. Find
# out more about how these arguments are parsed and used
# in the base entrypoint `run()` method.
container_op = dsl.ContainerOp(
name=step.config.name,
image=image,
command=command,
arguments=arguments,
)
settings = cast(
KubeflowOrchestratorSettings, self.get_settings(step)
)
self._configure_container_op(
container_op=container_op,
settings=settings,
)
if self.requires_resources_in_orchestration_environment(step):
self._configure_container_resources(
container_op=container_op,
resource_settings=step.config.resource_settings,
)
# Find the upstream container ops of the current step and
# configure the current container op to run after them
for upstream_step_name in step.spec.upstream_steps:
upstream_container_op = step_name_to_container_op[
upstream_step_name
]
container_op.after(upstream_container_op)
# Update dictionary of container ops with the current one
step_name_to_container_op[step.config.name] = container_op
orchestrator_run_name = get_orchestrator_run_name(
pipeline_name=deployment.pipeline_configuration.name
)
# Get a filepath to use to save the finished yaml to
fileio.makedirs(self.pipeline_directory)
pipeline_file_path = os.path.join(
self.pipeline_directory, f"{orchestrator_run_name}.yaml"
)
# write the argo pipeline yaml
KFPCompiler()._create_and_write_workflow(
pipeline_func=_construct_kfp_pipeline,
pipeline_name=deployment.pipeline_configuration.name,
package_path=pipeline_file_path,
)
logger.info(
"Writing Kubeflow workflow definition to `%s`.", pipeline_file_path
)
# using the kfp client uploads the pipeline to kubeflow pipelines and
# runs it there
self._upload_and_run_pipeline(
deployment=deployment,
pipeline_file_path=pipeline_file_path,
run_name=orchestrator_run_name,
)
local_deployment_utils
Utils for the local Kubeflow deployment behaviors.
add_hostpath_to_kubeflow_pipelines(kubernetes_context, local_path)
Patches the Kubeflow Pipelines deployment to mount a local folder.
This folder serves as a hostpath for visualization purposes.
This function reconfigures the Kubeflow pipelines deployment to use a shared local folder to support loading the TensorBoard viewer and other pipeline visualization results from a local artifact store, as described here:
https://github.com/kubeflow/pipelines/blob/master/docs/config/volume-support.md
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
kubernetes_context |
str |
The kubernetes context on which Kubeflow Pipelines should be patched. |
required |
local_path |
str |
The path to the local folder to mount as a hostpath. |
required |
Source code in zenml/integrations/kubeflow/orchestrators/local_deployment_utils.py
def add_hostpath_to_kubeflow_pipelines(
kubernetes_context: str, local_path: str
) -> None:
"""Patches the Kubeflow Pipelines deployment to mount a local folder.
This folder serves as a hostpath for visualization purposes.
This function reconfigures the Kubeflow pipelines deployment to use a
shared local folder to support loading the TensorBoard viewer and other
pipeline visualization results from a local artifact store, as described
here:
https://github.com/kubeflow/pipelines/blob/master/docs/config/volume-support.md
Args:
kubernetes_context: The kubernetes context on which Kubeflow Pipelines
should be patched.
local_path: The path to the local folder to mount as a hostpath.
"""
logger.info("Patching Kubeflow Pipelines to mount a local folder.")
pod_template = {
"spec": {
"serviceAccountName": "kubeflow-pipelines-viewer",
"containers": [
{
"volumeMounts": [
{
"mountPath": local_path,
"name": "local-artifact-store",
}
]
}
],
"volumes": [
{
"hostPath": {
"path": local_path,
"type": "Directory",
},
"name": "local-artifact-store",
}
],
}
}
pod_template_json = json.dumps(pod_template, indent=2)
config_map_data = {"data": {"viewer-pod-template.json": pod_template_json}}
config_map_data_json = json.dumps(config_map_data, indent=2)
logger.debug(
"Adding host path volume for local path `%s` to kubeflow pipeline"
"viewer pod template configuration.",
local_path,
)
subprocess.check_call(
[
"kubectl",
"--context",
kubernetes_context,
"-n",
"kubeflow",
"patch",
"configmap/ml-pipeline-ui-configmap",
"--type",
"merge",
"-p",
config_map_data_json,
]
)
deployment_patch = {
"spec": {
"template": {
"spec": {
"containers": [
{
"name": "ml-pipeline-ui",
"volumeMounts": [
{
"mountPath": local_path,
"name": "local-artifact-store",
}
],
}
],
"volumes": [
{
"hostPath": {
"path": local_path,
"type": "Directory",
},
"name": "local-artifact-store",
}
],
}
}
}
}
deployment_patch_json = json.dumps(deployment_patch, indent=2)
logger.debug(
"Adding host path volume for local path `%s` to the kubeflow UI",
local_path,
)
subprocess.check_call(
[
"kubectl",
"--context",
kubernetes_context,
"-n",
"kubeflow",
"patch",
"deployment/ml-pipeline-ui",
"--type",
"strategic",
"-p",
deployment_patch_json,
]
)
wait_until_kubeflow_pipelines_ready(kubernetes_context=kubernetes_context)
logger.info("Finished patching Kubeflow Pipelines setup.")
check_prerequisites(skip_k3d=False, skip_kubectl=False)
Checks prerequisites for a local kubeflow pipelines deployment.
It makes sure they are installed.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
skip_k3d |
bool |
Whether to skip the check for the k3d command. |
False |
skip_kubectl |
bool |
Whether to skip the check for the kubectl command. |
False |
Returns:
| Type | Description |
|---|---|
bool |
Whether all prerequisites are installed. |
Source code in zenml/integrations/kubeflow/orchestrators/local_deployment_utils.py
def check_prerequisites(
skip_k3d: bool = False, skip_kubectl: bool = False
) -> bool:
"""Checks prerequisites for a local kubeflow pipelines deployment.
It makes sure they are installed.
Args:
skip_k3d: Whether to skip the check for the k3d command.
skip_kubectl: Whether to skip the check for the kubectl command.
Returns:
Whether all prerequisites are installed.
"""
k3d_installed = skip_k3d or shutil.which("k3d") is not None
kubectl_installed = skip_kubectl or shutil.which("kubectl") is not None
logger.debug(
"Local kubeflow deployment prerequisites: K3D - %s, Kubectl - %s",
k3d_installed,
kubectl_installed,
)
return k3d_installed and kubectl_installed
create_k3d_cluster(cluster_name, registry_name, registry_config_path)
Creates a K3D cluster.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
cluster_name |
str |
Name of the cluster to create. |
required |
registry_name |
str |
Name of the registry to create for this cluster. |
required |
registry_config_path |
str |
Path to the registry config file. |
required |
Source code in zenml/integrations/kubeflow/orchestrators/local_deployment_utils.py
def create_k3d_cluster(
cluster_name: str, registry_name: str, registry_config_path: str
) -> None:
"""Creates a K3D cluster.
Args:
cluster_name: Name of the cluster to create.
registry_name: Name of the registry to create for this cluster.
registry_config_path: Path to the registry config file.
"""
logger.info("Creating local K3D cluster '%s'.", cluster_name)
local_stores_path = GlobalConfiguration().local_stores_path
subprocess.check_call(
[
"k3d",
"cluster",
"create",
cluster_name,
"--image",
K3S_IMAGE_NAME,
"--registry-create",
registry_name,
"--registry-config",
registry_config_path,
"--volume",
f"{local_stores_path}:{local_stores_path}",
]
)
logger.info("Finished K3D cluster creation.")
delete_k3d_cluster(cluster_name)
Deletes a K3D cluster with the given name.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
cluster_name |
str |
Name of the cluster to delete. |
required |
Source code in zenml/integrations/kubeflow/orchestrators/local_deployment_utils.py
def delete_k3d_cluster(cluster_name: str) -> None:
"""Deletes a K3D cluster with the given name.
Args:
cluster_name: Name of the cluster to delete.
"""
subprocess.check_call(["k3d", "cluster", "delete", cluster_name])
logger.info("Deleted local k3d cluster '%s'.", cluster_name)
deploy_kubeflow_pipelines(kubernetes_context)
Deploys Kubeflow Pipelines.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
kubernetes_context |
str |
The kubernetes context on which Kubeflow Pipelines should be deployed. |
required |
Source code in zenml/integrations/kubeflow/orchestrators/local_deployment_utils.py
def deploy_kubeflow_pipelines(kubernetes_context: str) -> None:
"""Deploys Kubeflow Pipelines.
Args:
kubernetes_context: The kubernetes context on which Kubeflow Pipelines
should be deployed.
"""
logger.info("Deploying Kubeflow Pipelines.")
subprocess.check_call(
[
"kubectl",
"--context",
kubernetes_context,
"apply",
"-k",
f"github.com/kubeflow/pipelines/manifests/kustomize/cluster-scoped-resources?ref={KFP_VERSION}&timeout=5m",
]
)
subprocess.check_call(
[
"kubectl",
"--context",
kubernetes_context,
"wait",
"--timeout=60s",
"--for",
"condition=established",
"crd/applications.app.k8s.io",
]
)
subprocess.check_call(
[
"kubectl",
"--context",
kubernetes_context,
"apply",
"-k",
f"github.com/kubeflow/pipelines/manifests/kustomize/env/platform-agnostic-pns?ref={KFP_VERSION}&timeout=5m",
]
)
wait_until_kubeflow_pipelines_ready(kubernetes_context=kubernetes_context)
logger.info("Finished Kubeflow Pipelines setup.")
k3d_cluster_exists(cluster_name)
Checks whether there exists a K3D cluster with the given name.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
cluster_name |
str |
Name of the cluster to check. |
required |
Returns:
| Type | Description |
|---|---|
bool |
Whether the cluster exists. |
Source code in zenml/integrations/kubeflow/orchestrators/local_deployment_utils.py
def k3d_cluster_exists(cluster_name: str) -> bool:
"""Checks whether there exists a K3D cluster with the given name.
Args:
cluster_name: Name of the cluster to check.
Returns:
Whether the cluster exists.
"""
output = subprocess.check_output(
["k3d", "cluster", "list", "--output", "json"]
)
clusters = json.loads(output)
for cluster in clusters:
if cluster["name"] == cluster_name:
return True
return False
k3d_cluster_running(cluster_name)
Checks whether the K3D cluster with the given name is running.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
cluster_name |
str |
Name of the cluster to check. |
required |
Returns:
| Type | Description |
|---|---|
bool |
Whether the cluster is running. |
Source code in zenml/integrations/kubeflow/orchestrators/local_deployment_utils.py
def k3d_cluster_running(cluster_name: str) -> bool:
"""Checks whether the K3D cluster with the given name is running.
Args:
cluster_name: Name of the cluster to check.
Returns:
Whether the cluster is running.
"""
output = subprocess.check_output(
["k3d", "cluster", "list", "--output", "json"]
)
clusters = json.loads(output)
for cluster in clusters:
if cluster["name"] == cluster_name:
server_count: int = cluster["serversCount"]
servers_running: int = cluster["serversRunning"]
return servers_running == server_count
return False
kubeflow_pipelines_ready(kubernetes_context)
Returns whether all Kubeflow Pipelines pods are ready.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
kubernetes_context |
str |
The kubernetes context in which the pods should be checked. |
required |
Returns:
| Type | Description |
|---|---|
bool |
Whether all Kubeflow Pipelines pods are ready. |
Source code in zenml/integrations/kubeflow/orchestrators/local_deployment_utils.py
def kubeflow_pipelines_ready(kubernetes_context: str) -> bool:
"""Returns whether all Kubeflow Pipelines pods are ready.
Args:
kubernetes_context: The kubernetes context in which the pods
should be checked.
Returns:
Whether all Kubeflow Pipelines pods are ready.
"""
try:
subprocess.check_call(
[
"kubectl",
"--context",
kubernetes_context,
"--namespace",
"kubeflow",
"wait",
"--for",
"condition=ready",
"--timeout=0s",
"pods",
"-l",
"application-crd-id=kubeflow-pipelines",
],
stdout=subprocess.DEVNULL,
stderr=subprocess.DEVNULL,
)
return True
except subprocess.CalledProcessError:
return False
start_k3d_cluster(cluster_name)
Starts a K3D cluster with the given name.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
cluster_name |
str |
Name of the cluster to start. |
required |
Source code in zenml/integrations/kubeflow/orchestrators/local_deployment_utils.py
def start_k3d_cluster(cluster_name: str) -> None:
"""Starts a K3D cluster with the given name.
Args:
cluster_name: Name of the cluster to start.
"""
subprocess.check_call(["k3d", "cluster", "start", cluster_name])
logger.info("Started local k3d cluster '%s'.", cluster_name)
start_kfp_ui_daemon(pid_file_path, log_file_path, port, kubernetes_context)
Starts a daemon process that forwards ports.
This is so the Kubeflow Pipelines UI is accessible in the browser.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
pid_file_path |
str |
Path where the file with the daemons process ID should be written. |
required |
log_file_path |
str |
Path to a file where the daemon logs should be written. |
required |
port |
int |
Port on which the UI should be accessible. |
required |
kubernetes_context |
str |
The kubernetes context for the cluster where Kubeflow Pipelines is running. |
required |
Source code in zenml/integrations/kubeflow/orchestrators/local_deployment_utils.py
def start_kfp_ui_daemon(
pid_file_path: str,
log_file_path: str,
port: int,
kubernetes_context: str,
) -> None:
"""Starts a daemon process that forwards ports.
This is so the Kubeflow Pipelines UI is accessible in the browser.
Args:
pid_file_path: Path where the file with the daemons process ID should
be written.
log_file_path: Path to a file where the daemon logs should be written.
port: Port on which the UI should be accessible.
kubernetes_context: The kubernetes context for the cluster where
Kubeflow Pipelines is running.
"""
command = [
"kubectl",
"--context",
kubernetes_context,
"--namespace",
"kubeflow",
"port-forward",
"svc/ml-pipeline-ui",
f"{port}:80",
]
if not networking_utils.port_available(port):
modified_command = command.copy()
modified_command[-1] = "PORT:80"
logger.warning(
"Unable to port-forward Kubeflow Pipelines UI to local port %d "
"because the port is occupied. In order to access the Kubeflow "
"Pipelines UI at http://localhost:PORT/, please run '%s' in a "
"separate command line shell (replace PORT with a free port of "
"your choice).",
port,
" ".join(modified_command),
)
elif sys.platform == "win32":
logger.warning(
"Daemon functionality not supported on Windows. "
"In order to access the Kubeflow Pipelines UI at "
"http://localhost:%d/, please run '%s' in a separate command "
"line shell.",
port,
" ".join(command),
)
else:
from zenml.utils import daemon
def _daemon_function() -> None:
"""Port-forwards the Kubeflow Pipelines UI pod."""
subprocess.check_call(command)
daemon.run_as_daemon(
_daemon_function, pid_file=pid_file_path, log_file=log_file_path
)
logger.info(
"Started Kubeflow Pipelines UI daemon (check the daemon logs at %s "
"in case you're not able to view the UI). The Kubeflow Pipelines "
"UI should now be accessible at http://localhost:%d/.",
log_file_path,
port,
)
stop_k3d_cluster(cluster_name)
Stops a K3D cluster with the given name.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
cluster_name |
str |
Name of the cluster to stop. |
required |
Source code in zenml/integrations/kubeflow/orchestrators/local_deployment_utils.py
def stop_k3d_cluster(cluster_name: str) -> None:
"""Stops a K3D cluster with the given name.
Args:
cluster_name: Name of the cluster to stop.
"""
subprocess.check_call(["k3d", "cluster", "stop", cluster_name])
logger.info("Stopped local k3d cluster '%s'.", cluster_name)
stop_kfp_ui_daemon(pid_file_path)
Stops the KFP UI daemon process if it is running.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
pid_file_path |
str |
Path to the file with the daemons process ID. |
required |
Source code in zenml/integrations/kubeflow/orchestrators/local_deployment_utils.py
def stop_kfp_ui_daemon(pid_file_path: str) -> None:
"""Stops the KFP UI daemon process if it is running.
Args:
pid_file_path: Path to the file with the daemons process ID.
"""
if fileio.exists(pid_file_path):
if sys.platform == "win32":
# Daemon functionality is not supported on Windows, so the PID
# file won't exist. This if clause exists just for mypy to not
# complain about missing functions
pass
else:
from zenml.utils import daemon
daemon.stop_daemon(pid_file_path)
fileio.remove(pid_file_path)
logger.info("Stopped Kubeflow Pipelines UI daemon.")
wait_until_kubeflow_pipelines_ready(kubernetes_context)
Waits until all Kubeflow Pipelines pods are ready.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
kubernetes_context |
str |
The kubernetes context in which the pods should be checked. |
required |
Source code in zenml/integrations/kubeflow/orchestrators/local_deployment_utils.py
def wait_until_kubeflow_pipelines_ready(kubernetes_context: str) -> None:
"""Waits until all Kubeflow Pipelines pods are ready.
Args:
kubernetes_context: The kubernetes context in which the pods
should be checked.
"""
logger.info(
"Waiting for all Kubeflow Pipelines pods to be ready (this might "
"take a few minutes)."
)
while True:
logger.info("Current pod status:")
subprocess.check_call(
[
"kubectl",
"--context",
kubernetes_context,
"--namespace",
"kubeflow",
"get",
"pods",
]
)
if kubeflow_pipelines_ready(kubernetes_context=kubernetes_context):
break
logger.info(
"One or more pods not ready yet, waiting for 30 seconds..."
)
time.sleep(30)
write_local_registry_yaml(yaml_path, registry_name, registry_uri)
Writes a K3D registry config file.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
yaml_path |
str |
Path where the config file should be written to. |
required |
registry_name |
str |
Name of the registry. |
required |
registry_uri |
str |
URI of the registry. |
required |
Source code in zenml/integrations/kubeflow/orchestrators/local_deployment_utils.py
def write_local_registry_yaml(
yaml_path: str, registry_name: str, registry_uri: str
) -> None:
"""Writes a K3D registry config file.
Args:
yaml_path: Path where the config file should be written to.
registry_name: Name of the registry.
registry_uri: URI of the registry.
"""
yaml_content = {
"mirrors": {registry_uri: {"endpoint": [f"http://{registry_name}"]}}
}
yaml_utils.write_yaml(yaml_path, yaml_content)
utils
KFP utilities.
apply_pod_settings(container_op, settings)
Applies Kubernetes Pod settings to a KFP container.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
container_op |
ContainerOp |
The container to which to apply the settings. |
required |
settings |
KubernetesPodSettings |
The settings to apply. |
required |
Source code in zenml/integrations/kubeflow/utils.py
def apply_pod_settings(
container_op: "ContainerOp",
settings: KubernetesPodSettings,
) -> None:
"""Applies Kubernetes Pod settings to a KFP container.
Args:
container_op: The container to which to apply the settings.
settings: The settings to apply.
"""
from kubernetes.client.models import V1Affinity, V1Toleration
for key, value in settings.node_selectors.items():
container_op.add_node_selector_constraint(label_name=key, value=value)
if settings.affinity:
affinity: V1Affinity = (
serialization_utils.deserialize_kubernetes_model(
settings.affinity, "V1Affinity"
)
)
container_op.add_affinity(affinity)
for toleration_dict in settings.tolerations:
toleration: V1Toleration = (
serialization_utils.deserialize_kubernetes_model(
toleration_dict, "V1Toleration"
)
)
container_op.add_toleration(toleration)
resource_requests = settings.resources.get("requests") or {}
for name, value in resource_requests.items():
container_op.add_resource_request(name, value)
resource_limits = settings.resources.get("limits") or {}
for name, value in resource_limits.items():
container_op.add_resource_limit(name, value)