worker.py

"""

Module containing the Kubernetes worker used for executing flow runs as Kubernetes jobs.

To start a Kubernetes worker, run the following command:

```bash
prefect worker start --pool 'my-work-pool' --type kubernetes
```

Replace `my-work-pool` with the name of the work pool you want the worker
to poll for flow runs.

### Securing your Prefect Cloud API key
If you are using Prefect Cloud and would like to pass your Prefect Cloud API key to
created jobs via a Kubernetes secret, set the
`PREFECT_INTEGRATIONS_KUBERNETES_WORKER_CREATE_SECRET_FOR_API_KEY` environment variable before
starting your worker:

```bash
export PREFECT_INTEGRATIONS_KUBERNETES_WORKER_CREATE_SECRET_FOR_API_KEY="true"
prefect worker start --pool 'my-work-pool' --type kubernetes
```

Note that your work will need permission to create secrets in the same namespace(s)
that Kubernetes jobs are created in to execute flow runs.

### Using a custom Kubernetes job manifest template

The default template used for Kubernetes job manifests looks like this:
```yaml
---
apiVersion: batch/v1
kind: Job
metadata:
labels: "{{ labels }}"
namespace: "{{ namespace }}"
generateName: "{{ name }}-"
spec:
ttlSecondsAfterFinished: "{{ finished_job_ttl }}"
template:
    spec:
    parallelism: 1
    completions: 1
    restartPolicy: Never
    serviceAccountName: "{{ service_account_name }}"
    containers:
    - name: prefect-job
        env: "{{ env }}"
        image: "{{ image }}"
        imagePullPolicy: "{{ image_pull_policy }}"
        args: "{{ command }}"
```

Each values enclosed in `{{ }}` is a placeholder that will be replaced with
a value at runtime. The values that can be used a placeholders are defined
by the `variables` schema defined in the base job template.

The default job manifest and available variables can be customized on a work pool
by work pool basis. These customizations can be made via the Prefect UI when
creating or editing a work pool.

For example, if you wanted to allow custom memory requests for a Kubernetes work
pool you could update the job manifest template to look like this:

```yaml
---
apiVersion: batch/v1
kind: Job
metadata:
labels: "{{ labels }}"
namespace: "{{ namespace }}"
generateName: "{{ name }}-"
spec:
ttlSecondsAfterFinished: "{{ finished_job_ttl }}"
template:
    spec:
    parallelism: 1
    completions: 1
    restartPolicy: Never
    serviceAccountName: "{{ service_account_name }}"
    containers:
    - name: prefect-job
        env: "{{ env }}"
        image: "{{ image }}"
        imagePullPolicy: "{{ image_pull_policy }}"
        args: "{{ command }}"
        resources:
            requests:
                memory: "{{ memory }}Mi"
            limits:
                memory: 128Mi
```

In this new template, the `memory` placeholder allows customization of the memory
allocated to Kubernetes jobs created by workers in this work pool, but the limit
is hard-coded and cannot be changed by deployments.

For more information about work pools and workers,
checkout out the [Prefect docs](https://docs.prefect.io/concepts/work-pools/).
"""

from __future__ import annotations

import asyncio
import base64
import enum
import json
import logging
import shlex
import tempfile
import warnings
from contextlib import asynccontextmanager
from datetime import datetime
from functools import partial
from typing import (
    TYPE_CHECKING,
    Any,
    AsyncGenerator,
    Dict,
    List,
    Optional,
    Tuple,
    TypeVar,
    Union,
)

import anyio
import anyio.abc
import kubernetes_asyncio
from jsonpatch import JsonPatch
from kubernetes_asyncio import config
from kubernetes_asyncio.client import (
    ApiClient,
    BatchV1Api,
    CoreV1Api,
    V1Job,
    V1Pod,
)
from kubernetes_asyncio.client.exceptions import ApiException
from kubernetes_asyncio.client.models import (
    CoreV1Event,
    CoreV1EventList,
    V1ObjectMeta,
    V1Secret,
)
from pydantic import Field, field_validator, model_validator
from tenacity import retry, stop_after_attempt, wait_fixed, wait_random
from typing_extensions import Literal, Self

import prefect
from prefect.client.schemas.objects import Flow as APIFlow
from prefect.exceptions import (
    InfrastructureError,
)
from prefect.futures import PrefectFlowRunFuture
from prefect.states import Pending
from prefect.utilities.collections import get_from_dict
from prefect.utilities.dockerutils import get_prefect_image_name
from prefect.utilities.templating import find_placeholders
from prefect.utilities.timeout import timeout_async
from prefect.workers.base import (
    BaseJobConfiguration,
    BaseVariables,
    BaseWorker,
    BaseWorkerResult,
)
from prefect_kubernetes.credentials import KubernetesClusterConfig
from prefect_kubernetes.events import KubernetesEventsReplicator
from prefect_kubernetes.settings import KubernetesSettings
from prefect_kubernetes.utilities import (
    KeepAliveClientRequest,
    _slugify_label_key,
    _slugify_label_value,
    _slugify_name,
)

if TYPE_CHECKING:
    from prefect.client.schemas.objects import FlowRun
    from prefect.client.schemas.responses import DeploymentResponse
    from prefect.flows import Flow

# Captures flow return type
R = TypeVar("R")


MAX_ATTEMPTS = 3
RETRY_MIN_DELAY_SECONDS = 1
RETRY_MIN_DELAY_JITTER_SECONDS = 0
RETRY_MAX_DELAY_JITTER_SECONDS = 3


def _get_default_job_manifest_template() -> Dict[str, Any]:
    """Returns the default job manifest template used by the Kubernetes worker."""
    return {
        "apiVersion": "batch/v1",
        "kind": "Job",
        "metadata": {
            "labels": "{{ labels }}",
            "namespace": "{{ namespace }}",
            "generateName": "{{ name }}-",
        },
        "spec": {
            "backoffLimit": 0,
            "ttlSecondsAfterFinished": "{{ finished_job_ttl }}",
            "template": {
                "spec": {
                    "parallelism": 1,
                    "completions": 1,
                    "restartPolicy": "Never",
                    "serviceAccountName": "{{ service_account_name }}",
                    "containers": [
                        {
                            "name": "prefect-job",
                            "env": "{{ env }}",
                            "image": "{{ image }}",
                            "imagePullPolicy": "{{ image_pull_policy }}",
                            "args": "{{ command }}",
                        }
                    ],
                }
            },
        },
    }


def _get_base_job_manifest():
    """Returns a base job manifest to use for manifest validation."""
    return {
        "apiVersion": "batch/v1",
        "kind": "Job",
        "metadata": {"labels": {}},
        "spec": {
            "template": {
                "spec": {
                    "parallelism": 1,
                    "completions": 1,
                    "restartPolicy": "Never",
                    "containers": [
                        {
                            "name": "prefect-job",
                        }
                    ],
                }
            }
        },
    }


class KubernetesImagePullPolicy(enum.Enum):
    """Enum representing the image pull policy options for a Kubernetes job."""

    IF_NOT_PRESENT = "IfNotPresent"
    ALWAYS = "Always"
    NEVER = "Never"


class KubernetesWorkerJobConfiguration(BaseJobConfiguration):
    """
    Configuration class used by the Kubernetes worker.

    An instance of this class is passed to the Kubernetes worker's `run` method
    for each flow run. It contains all of the information necessary to execute
    the flow run as a Kubernetes job.

    Attributes:
        name: The name to give to created Kubernetes job.
        command: The command executed in created Kubernetes jobs to kick off
            flow run execution.
        env: The environment variables to set in created Kubernetes jobs.
        labels: The labels to set on created Kubernetes jobs.
        namespace: The Kubernetes namespace to create Kubernetes jobs in.
        job_manifest: The Kubernetes job manifest to use to create Kubernetes jobs.
        cluster_config: The Kubernetes cluster configuration to use for authentication
            to a Kubernetes cluster.
        job_watch_timeout_seconds: The number of seconds to wait for the job to
            complete before timing out. If `None`, the worker will wait indefinitely.
        pod_watch_timeout_seconds: The number of seconds to wait for the pod to
            complete before timing out.
        stream_output: Whether or not to stream the job's output.
    """

    namespace: str = Field(default="default")
    job_manifest: Dict[str, Any] = Field(
        json_schema_extra=dict(template=_get_default_job_manifest_template())
    )
    cluster_config: Optional[KubernetesClusterConfig] = Field(default=None)
    job_watch_timeout_seconds: Optional[int] = Field(default=None)
    pod_watch_timeout_seconds: int = Field(default=60)
    stream_output: bool = Field(default=True)

    env: Union[Dict[str, Optional[str]], List[Dict[str, Any]]] = Field(
        default_factory=dict
    )

    # internal-use only
    _api_dns_name: Optional[str] = None  # Replaces 'localhost' in API URL

    @model_validator(mode="after")
    def _validate_job_manifest(self) -> Self:
        """
        Validates the job manifest by ensuring the presence of required fields
        and checking for compatible values.
        """
        job_manifest = self.job_manifest
        # Ensure metadata is present
        if "metadata" not in job_manifest:
            job_manifest["metadata"] = {}

        # Ensure labels is present in metadata
        if "labels" not in job_manifest["metadata"]:
            job_manifest["metadata"]["labels"] = {}

        # Ensure namespace is present in metadata
        if "namespace" not in job_manifest["metadata"]:
            job_manifest["metadata"]["namespace"] = self.namespace

        # Check if job includes all required components
        patch = JsonPatch.from_diff(job_manifest, _get_base_job_manifest())
        missing_paths = sorted([op["path"] for op in patch if op["op"] == "add"])
        if missing_paths:
            raise ValueError(
                "Job is missing required attributes at the following paths: "
                f"{', '.join(missing_paths)}"
            )

        # Check if job has compatible values
        incompatible = sorted(
            [
                f"{op['path']} must have value {op['value']!r}"
                for op in patch
                if op["op"] == "replace"
            ]
        )
        if incompatible:
            raise ValueError(
                "Job has incompatible values for the following attributes: "
                f"{', '.join(incompatible)}"
            )

        return self

    @field_validator("env", mode="before")
    @classmethod
    def _coerce_env(cls, v):
        if isinstance(v, list):
            return v
        return {k: str(v) if v is not None else None for k, v in v.items()}

    @staticmethod
    def _base_flow_run_labels(flow_run: "FlowRun") -> Dict[str, str]:
        """
        Generate a dictionary of labels for a flow run job.
        """
        return {
            "prefect.io/flow-run-id": str(flow_run.id),
            "prefect.io/flow-run-name": flow_run.name,
            "prefect.io/version": _slugify_label_value(
                prefect.__version__.split("+")[0]
            ),
        }

    def prepare_for_flow_run(
        self,
        flow_run: "FlowRun",
        deployment: Optional["DeploymentResponse"] = None,
        flow: Optional["APIFlow"] = None,
    ):
        """
        Prepares the job configuration for a flow run.

        Ensures that necessary values are present in the job manifest and that the
        job manifest is valid.

        Args:
            flow_run: The flow run to prepare the job configuration for
            deployment: The deployment associated with the flow run used for
                preparation.
            flow: The flow associated with the flow run used for preparation.
        """

        super().prepare_for_flow_run(flow_run, deployment, flow)
        # Update configuration env and job manifest env
        self._update_prefect_api_url_if_local_server()
        self._populate_env_in_manifest()
        # Update labels in job manifest
        self._slugify_labels()
        # Add defaults to job manifest if necessary
        self._populate_image_if_not_present()
        self._populate_command_if_not_present()
        self._populate_generate_name_if_not_present()

    def _populate_env_in_manifest(self):
        """
        Populates environment variables in the job manifest.

        When `env` is templated as a variable in the job manifest it comes in as a
        dictionary. We need to convert it to a list of dictionaries to conform to the
        Kubernetes job manifest schema.

        This function also handles the case where the user has removed the `{{ env }}`
        placeholder and hard coded a value for `env`. In this case, we need to prepend
        our environment variables to the list to ensure Prefect setting propagation.
        An example reason the a user would remove the `{{ env }}` placeholder to
        hardcode Kubernetes secrets in the base job template.
        """
        transformed_env = [{"name": k, "value": v} for k, v in self.env.items()]

        template_env = self.job_manifest["spec"]["template"]["spec"]["containers"][
            0
        ].get("env")

        # If user has removed `{{ env }}` placeholder and hard coded a value for `env`,
        # we need to prepend our environment variables to the list to ensure Prefect
        # setting propagation.
        if isinstance(template_env, list):
            self.job_manifest["spec"]["template"]["spec"]["containers"][0]["env"] = [
                *transformed_env,
                *template_env,
            ]
        # Current templating adds `env` as a dict when the kubernetes manifest requires
        # a list of dicts. Might be able to improve this in the future with a better
        # default `env` value and better typing.
        else:
            self.job_manifest["spec"]["template"]["spec"]["containers"][0]["env"] = (
                transformed_env
            )

    def _update_prefect_api_url_if_local_server(self):
        """If the API URL has been set by the base environment rather than the by the
        user, update the value to ensure connectivity when using a bridge network by
        updating local connections to use the internal host
        """
        if self.env.get("PREFECT_API_URL") and self._api_dns_name:
            self.env["PREFECT_API_URL"] = (
                self.env["PREFECT_API_URL"]
                .replace("localhost", self._api_dns_name)
                .replace("127.0.0.1", self._api_dns_name)
            )

    def _slugify_labels(self):
        """Slugifies the labels in the job manifest."""
        all_labels = {**self.job_manifest["metadata"].get("labels", {}), **self.labels}
        self.job_manifest["metadata"]["labels"] = {
            _slugify_label_key(k): _slugify_label_value(v)
            for k, v in all_labels.items()
        }

    def _populate_image_if_not_present(self):
        """Ensures that the image is present in the job manifest. Populates the image
        with the default Prefect image if it is not present."""
        try:
            if (
                "image"
                not in self.job_manifest["spec"]["template"]["spec"]["containers"][0]
            ):
                self.job_manifest["spec"]["template"]["spec"]["containers"][0][
                    "image"
                ] = get_prefect_image_name()
        except KeyError:
            raise ValueError(
                "Unable to verify image due to invalid job manifest template."
            )

    def _populate_command_if_not_present(self):
        """
        Ensures that the command is present in the job manifest. Populates the command
        with the `prefect -m prefect.engine` if a command is not present.
        """
        try:
            command = self.job_manifest["spec"]["template"]["spec"]["containers"][
                0
            ].get("args")
            if command is None:
                self.job_manifest["spec"]["template"]["spec"]["containers"][0][
                    "args"
                ] = shlex.split(self._base_flow_run_command())
            elif isinstance(command, str):
                self.job_manifest["spec"]["template"]["spec"]["containers"][0][
                    "args"
                ] = shlex.split(command)
            elif not isinstance(command, list):
                raise ValueError(
                    "Invalid job manifest template: 'command' must be a string or list."
                )
        except KeyError:
            raise ValueError(
                "Unable to verify command due to invalid job manifest template."
            )

    def _populate_generate_name_if_not_present(self):
        """Ensures that the generateName is present in the job manifest."""
        manifest_generate_name = self.job_manifest["metadata"].get("generateName", "")
        has_placeholder = len(find_placeholders(manifest_generate_name)) > 0
        # if name wasn't present during template rendering, generateName will be
        # just a hyphen

        manifest_generate_name_templated_with_empty_string = (
            manifest_generate_name == "-"
        )
        if (
            not manifest_generate_name
            or has_placeholder
            or manifest_generate_name_templated_with_empty_string
            or manifest_generate_name == "None-"
        ):
            generate_name = None
            if self.name:
                generate_name = _slugify_name(self.name)
            # _slugify_name will return None if the slugified name in an exception
            if not generate_name:
                generate_name = "prefect-job"
            self.job_manifest["metadata"]["generateName"] = f"{generate_name}-"


class KubernetesWorkerVariables(BaseVariables):
    """
    Default variables for the Kubernetes worker.

    The schema for this class is used to populate the `variables` section of the default
    base job template.
    """

    namespace: str = Field(
        default="default", description="The Kubernetes namespace to create jobs within."
    )
    image: Optional[str] = Field(
        default=None,
        description="The image reference of a container image to use for created jobs. "
        "If not set, the latest Prefect image will be used.",
        examples=["docker.io/prefecthq/prefect:3-latest"],
    )
    service_account_name: Optional[str] = Field(
        default=None,
        description="The Kubernetes service account to use for job creation.",
    )
    image_pull_policy: Literal["IfNotPresent", "Always", "Never"] = Field(
        default=KubernetesImagePullPolicy.IF_NOT_PRESENT,
        description="The Kubernetes image pull policy to use for job containers.",
    )
    finished_job_ttl: Optional[int] = Field(
        default=None,
        title="Finished Job TTL",
        description="The number of seconds to retain jobs after completion. If set, "
        "finished jobs will be cleaned up by Kubernetes after the given delay. If not "
        "set, jobs will be retained indefinitely.",
    )
    job_watch_timeout_seconds: Optional[int] = Field(
        default=None,
        description=(
            "Number of seconds to wait for each event emitted by a job before "
            "timing out. If not set, the worker will wait for each event indefinitely."
        ),
    )
    pod_watch_timeout_seconds: int = Field(
        default=60,
        description="Number of seconds to watch for pod creation before timing out.",
    )
    stream_output: bool = Field(
        default=True,
        description=(
            "If set, output will be streamed from the job to local standard output."
        ),
    )
    cluster_config: Optional[KubernetesClusterConfig] = Field(
        default=None,
        description="The Kubernetes cluster config to use for job creation.",
    )


class KubernetesWorkerResult(BaseWorkerResult):
    """Contains information about the final state of a completed process"""


class KubernetesWorker(
    BaseWorker[
        "KubernetesWorkerJobConfiguration",
        "KubernetesWorkerVariables",
        "KubernetesWorkerResult",
    ]
):
    """Prefect worker that executes flow runs within Kubernetes Jobs."""

    type: str = "kubernetes"
    job_configuration = KubernetesWorkerJobConfiguration
    job_configuration_variables = KubernetesWorkerVariables
    _description = (
        "Execute flow runs within jobs scheduled on a Kubernetes cluster. Requires a "
        "Kubernetes cluster."
    )
    _display_name = "Kubernetes"
    _documentation_url = "https://docs.prefect.io/integrations/prefect-kubernetes"
    _logo_url = "https://cdn.sanity.io/images/3ugk85nk/production/2d0b896006ad463b49c28aaac14f31e00e32cfab-250x250.png"  # noqa

    def __init__(self, *args: Any, **kwargs: Any):
        super().__init__(*args, **kwargs)
        self._created_secrets: dict[
            tuple[str, str], KubernetesWorkerJobConfiguration
        ] = {}

    async def run(
        self,
        flow_run: "FlowRun",
        configuration: KubernetesWorkerJobConfiguration,
        task_status: Optional[anyio.abc.TaskStatus[int]] = None,
    ) -> KubernetesWorkerResult:
        """
        Executes a flow run within a Kubernetes Job and waits for the flow run
        to complete.

        Args:
            flow_run: The flow run to execute
            configuration: The configuration to use when executing the flow run.
            task_status: The task status object for the current flow run. If provided,
                the task will be marked as started.

        Returns:
            KubernetesWorkerResult: A result object containing information about the
                final state of the flow run
        """
        logger = self.get_flow_run_logger(flow_run)
        async with self._get_configured_kubernetes_client(configuration) as client:
            logger.info("Creating Kubernetes job...")

            job = await self._create_job(configuration, client)

            pid = await self._get_infrastructure_pid(job, client)
            # Indicate that the job has started
            if task_status is not None:
                task_status.started(pid)

            # Monitor the job until completion
            events_replicator = KubernetesEventsReplicator(
                client=client,
                job_name=job.metadata.name,
                namespace=configuration.namespace,
                worker_resource=self._event_resource(),
                related_resources=self._event_related_resources(
                    configuration=configuration
                ),
                timeout_seconds=configuration.pod_watch_timeout_seconds,
            )
            async with events_replicator:
                status_code = await self._watch_job(
                    logger, job.metadata.name, configuration, client
                )

            return KubernetesWorkerResult(identifier=pid, status_code=status_code)

    async def submit(
        self,
        flow: "Flow[..., R]",
        parameters: dict[str, Any] | None = None,
        job_variables: dict[str, Any] | None = None,
    ) -> "PrefectFlowRunFuture[R]":
        """
        EXPERIMENTAL: The interface for this method is subject to change.

        Submits a flow to run in a Kubernetes job.

        Args:
            flow: The flow to submit
            parameters: The parameters to pass to the flow

        Returns:
            A flow run object
        """
        warnings.warn(
            "The `submit` method on the Kubernetes worker is experimental. The interface "
            "and behavior of this method are subject to change.",
            category=FutureWarning,
        )
        if self._runs_task_group is None:
            raise RuntimeError("Worker not properly initialized")
        flow_run = await self._runs_task_group.start(
            partial(
                self._submit_adhoc_run,
                flow=flow,
                parameters=parameters,
                job_variables=job_variables,
            ),
        )
        return PrefectFlowRunFuture(flow_run_id=flow_run.id)

    async def _submit_adhoc_run(
        self,
        flow: "Flow[..., R]",
        parameters: dict[str, Any] | None = None,
        job_variables: dict[str, Any] | None = None,
        task_status: anyio.abc.TaskStatus["FlowRun"] | None = None,
    ):
        """
        Submits a flow run to the Kubernetes worker.
        """
        from prefect._experimental.bundles import (
            convert_step_to_command,
            create_bundle_for_flow_run,
        )

        if TYPE_CHECKING:
            assert self._client is not None
            assert self._work_pool is not None
        flow_run = await self._client.create_flow_run(
            flow, parameters=parameters, state=Pending()
        )
        if task_status is not None:
            # Emit the flow run object to .submit to allow it to return a future as soon as possible
            task_status.started(flow_run)
        # Avoid an API call to get the flow
        api_flow = APIFlow(id=flow_run.flow_id, name=flow.name, labels={})
        logger = self.get_flow_run_logger(flow_run)

        # TODO: Migrate steps to their own attributes on work pool when hardening this design
        upload_step = json.loads(
            get_from_dict(
                self._work_pool.base_job_template,
                "variables.properties.env.default.PREFECT__BUNDLE_UPLOAD_STEP",
                "{}",
            )
        )
        execute_step = json.loads(
            get_from_dict(
                self._work_pool.base_job_template,
                "variables.properties.env.default.PREFECT__BUNDLE_EXECUTE_STEP",
                "{}",
            )
        )

        upload_command = convert_step_to_command(upload_step, str(flow_run.id))
        execute_command = convert_step_to_command(execute_step, str(flow_run.id))

        job_variables = (job_variables or {}) | {"command": " ".join(execute_command)}

        configuration = await self.job_configuration.from_template_and_values(
            base_job_template=self._work_pool.base_job_template,
            values=job_variables,
            client=self._client,
        )
        configuration.prepare_for_flow_run(flow_run=flow_run, flow=api_flow)

        bundle = create_bundle_for_flow_run(flow=flow, flow_run=flow_run)

        logger.debug("Uploading execution bundle")
        with tempfile.TemporaryDirectory() as temp_dir:
            await (
                anyio.Path(temp_dir)
                .joinpath(str(flow_run.id))
                .write_bytes(json.dumps(bundle).encode("utf-8"))
            )

            try:
                await anyio.run_process(
                    upload_command + [str(flow_run.id)],
                    cwd=temp_dir,
                )
            except Exception as e:
                self._logger.error(
                    "Failed to upload bundle: %s", e.stderr.decode("utf-8")
                )
                raise e

        logger.debug("Successfully uploaded execution bundle")

        try:
            result = await self.run(flow_run, configuration)

            if result.status_code != 0:
                await self._propose_crashed_state(
                    flow_run,
                    (
                        "Flow run infrastructure exited with non-zero status code"
                        f" {result.status_code}."
                    ),
                )
        except Exception as exc:
            # This flow run was being submitted and did not start successfully
            logger.exception(
                f"Failed to submit flow run '{flow_run.id}' to infrastructure."
            )
            message = f"Flow run could not be submitted to infrastructure:\n{exc!r}"
            await self._propose_crashed_state(flow_run, message)

    async def teardown(self, *exc_info: Any):
        await super().teardown(*exc_info)

        await self._clean_up_created_secrets()

    async def _clean_up_created_secrets(self):
        """Deletes any secrets created during the worker's operation."""
        for key, configuration in self._created_secrets.items():
            async with self._get_configured_kubernetes_client(configuration) as client:
                v1 = CoreV1Api(client)
                result = await v1.delete_namespaced_secret(
                    name=key[0],
                    namespace=key[1],
                )

                if isinstance(result, Exception):
                    self._logger.warning(
                        "Failed to delete created secret with exception: %s", result
                    )

    @asynccontextmanager
    async def _get_configured_kubernetes_client(
        self, configuration: KubernetesWorkerJobConfiguration
    ) -> AsyncGenerator["ApiClient", None]:
        """
        Returns a configured Kubernetes client.
        """
        client = None
        settings = KubernetesSettings()

        if configuration.cluster_config:
            config_dict = configuration.cluster_config.config
            context = configuration.cluster_config.context_name
            client = await config.new_client_from_config_dict(
                config_dict=config_dict,
                context=context,
            )
        else:
            # Try to load in-cluster configuration
            try:
                config.load_incluster_config()
                client = ApiClient()
            except config.ConfigException:
                # If in-cluster config fails, load the local kubeconfig
                client = await config.new_client_from_config()

        if settings.worker.add_tcp_keepalive:
            client.rest_client.pool_manager._request_class = KeepAliveClientRequest

        try:
            yield client
        finally:
            await client.close()

    async def _replace_api_key_with_secret(
        self,
        configuration: KubernetesWorkerJobConfiguration,
        client: "ApiClient",
        secret_name: Optional[str] = None,
    ):
        """Replaces the PREFECT_API_KEY environment variable with a Kubernetes secret"""
        manifest_env = configuration.job_manifest["spec"]["template"]["spec"][
            "containers"
        ][0].get("env")
        manifest_api_key_env = next(
            (
                env_entry
                for env_entry in manifest_env
                if env_entry.get("name") == "PREFECT_API_KEY"
            ),
            {},
        )
        api_key = manifest_api_key_env.get("value")
        if api_key and not secret_name:
            secret_name = f"prefect-{_slugify_name(self.name)}-api-key"
            secret = await self._upsert_secret(
                name=secret_name,
                value=api_key,
                namespace=configuration.namespace,
                client=client,
            )
            # Store configuration so that we can delete the secret when the worker shuts
            # down
            self._created_secrets[(secret.metadata.name, secret.metadata.namespace)] = (
                configuration
            )
        if secret_name:
            new_api_env_entry = {
                "name": "PREFECT_API_KEY",
                "valueFrom": {"secretKeyRef": {"name": secret_name, "key": "value"}},
            }
            manifest_env = [
                entry if entry.get("name") != "PREFECT_API_KEY" else new_api_env_entry
                for entry in manifest_env
            ]
            configuration.job_manifest["spec"]["template"]["spec"]["containers"][0][
                "env"
            ] = manifest_env

    @retry(
        stop=stop_after_attempt(MAX_ATTEMPTS),
        wait=wait_fixed(RETRY_MIN_DELAY_SECONDS)
        + wait_random(
            RETRY_MIN_DELAY_JITTER_SECONDS,
            RETRY_MAX_DELAY_JITTER_SECONDS,
        ),
        reraise=True,
    )
    async def _create_job(
        self, configuration: KubernetesWorkerJobConfiguration, client: "ApiClient"
    ) -> "V1Job":
        """
        Creates a Kubernetes job from a job manifest.
        """
        settings = KubernetesSettings()
        if settings.worker.api_key_secret_name:
            await self._replace_api_key_with_secret(
                configuration=configuration,
                client=client,
                secret_name=settings.worker.api_key_secret_name,
            )
        elif settings.worker.create_secret_for_api_key:
            await self._replace_api_key_with_secret(
                configuration=configuration, client=client
            )

        try:
            batch_client = BatchV1Api(client)
            job = await batch_client.create_namespaced_job(
                configuration.namespace,
                configuration.job_manifest,
            )
        except kubernetes_asyncio.client.exceptions.ApiException as exc:
            # Parse the reason and message from the response if feasible
            message = ""
            if exc.reason:
                message += ": " + exc.reason
            if exc.body and "message" in (body := json.loads(exc.body)):
                message += ": " + body["message"]

            raise InfrastructureError(
                f"Unable to create Kubernetes job{message}"
            ) from exc

        return job

    async def _upsert_secret(
        self, name: str, value: str, namespace: str, client: "ApiClient"
    ):
        encoded_value = base64.b64encode(value.encode("utf-8")).decode("utf-8")
        core_client = CoreV1Api(client)
        try:
            # Get the current version of the Secret and update it with the
            # new value
            current_secret = await core_client.read_namespaced_secret(
                name=name, namespace=namespace
            )
            current_secret.data = {"value": encoded_value}
            secret = await core_client.replace_namespaced_secret(
                name=name, namespace=namespace, body=current_secret
            )
        except ApiException as exc:
            if exc.status != 404:
                raise
            # Create the secret if it doesn't already exist
            metadata = V1ObjectMeta(name=name, namespace=namespace)
            secret = V1Secret(
                api_version="v1",
                kind="Secret",
                metadata=metadata,
                data={"value": encoded_value},
            )
            secret = await core_client.create_namespaced_secret(
                namespace=namespace, body=secret
            )
        return secret

    @asynccontextmanager
    async def _get_batch_client(
        self, client: "ApiClient"
    ) -> AsyncGenerator["BatchV1Api", None]:
        """
        Context manager for retrieving a Kubernetes batch client.
        """
        try:
            yield BatchV1Api(api_client=client)
        finally:
            await client.close()

    async def _get_infrastructure_pid(self, job: "V1Job", client: "ApiClient") -> str:
        """
        Generates a Kubernetes infrastructure PID.

        The PID is in the format: "<cluster uid>:<namespace>:<job name>".
        """
        cluster_uid = await self._get_cluster_uid(client)
        pid = f"{cluster_uid}:{job.metadata.namespace}:{job.metadata.name}"
        return pid

    def _parse_infrastructure_pid(
        self, infrastructure_pid: str
    ) -> Tuple[str, str, str]:
        """
        Parse a Kubernetes infrastructure PID into its component parts.

        Returns a cluster UID, namespace, and job name.
        """
        cluster_uid, namespace, job_name = infrastructure_pid.split(":", 2)
        return cluster_uid, namespace, job_name

    async def _get_cluster_uid(self, client: "ApiClient") -> str:
        """
        Gets a unique id for the current cluster being used.

        There is no real unique identifier for a cluster. However, the `kube-system`
        namespace is immutable and has a persistence UID that we use instead.

        PREFECT_KUBERNETES_CLUSTER_UID can be set in cases where the `kube-system`
        namespace cannot be read e.g. when a cluster role cannot be created. If set,
        this variable will be used and we will not attempt to read the `kube-system`
        namespace.

        See https://github.com/kubernetes/kubernetes/issues/44954
        """
        settings = KubernetesSettings()
        # Default to an environment variable
        env_cluster_uid = settings.cluster_uid
        if env_cluster_uid:
            return env_cluster_uid

        # Read the UID from the cluster namespace
        v1 = CoreV1Api(client)
        namespace = await v1.read_namespace("kube-system")
        cluster_uid = namespace.metadata.uid
        return cluster_uid

    async def _stream_job_logs(
        self,
        logger: logging.Logger,
        pod_name: str,
        job_name: str,
        configuration: KubernetesWorkerJobConfiguration,
        client,
    ):
        core_client = CoreV1Api(client)

        logs = await core_client.read_namespaced_pod_log(
            pod_name,
            configuration.namespace,
            follow=True,
            _preload_content=False,
            container="prefect-job",
        )
        try:
            async for line in logs.content:
                if not line:
                    break
                print(line.decode().rstrip())
        except Exception:
            logger.warning(
                (
                    "Error occurred while streaming logs - "
                    "Job will continue to run but logs will "
                    "no longer be streamed to stdout."
                ),
                exc_info=True,
            )

    async def _job_events(
        self,
        batch_client: BatchV1Api,
        job_name: str,
        namespace: str,
        watch_kwargs: dict,
    ):
        """
        Stream job events.

        Pick up from the current resource version returned by the API
        in the case of a 410.

        See https://kubernetes.io/docs/reference/using-api/api-concepts/#efficient-detection-of-changes  # noqa
        """
        watch = kubernetes_asyncio.watch.Watch()
        resource_version = None
        async with watch:
            while True:
                try:
                    async for event in watch.stream(
                        func=batch_client.list_namespaced_job,
                        namespace=namespace,
                        field_selector=f"metadata.name={job_name}",
                        **watch_kwargs,
                    ):
                        yield event
                except ApiException as e:
                    if e.status == 410:
                        job_list = await batch_client.list_namespaced_job(
                            namespace=namespace,
                            field_selector=f"metadata.name={job_name}",
                        )

                        resource_version = job_list.metadata.resource_version
                        watch_kwargs["resource_version"] = resource_version
                    else:
                        raise

    async def _monitor_job_events(self, batch_client, job_name, logger, configuration):
        job = await batch_client.read_namespaced_job(
            name=job_name, namespace=configuration.namespace
        )
        completed = job.status.completion_time is not None
        watch_kwargs = (
            {"timeout_seconds": configuration.job_watch_timeout_seconds}
            if configuration.job_watch_timeout_seconds
            else {}
        )

        while not completed:
            async for event in self._job_events(
                batch_client,
                job_name,
                configuration.namespace,
                watch_kwargs,
            ):
                if event["type"] == "DELETED":
                    logger.error(f"Job {job_name!r}: Job has been deleted.")
                    completed = True
                elif event["object"].status.completion_time:
                    if not event["object"].status.succeeded:
                        # Job failed, exit while loop and return pod exit code
                        logger.error(f"Job {job_name!r}: Job failed.")
                    completed = True
                # Check if the job has reached its backoff limit
                # and stop watching if it has
                elif (
                    event["object"].spec.backoff_limit is not None
                    and event["object"].status.failed is not None
                    and event["object"].status.failed
                    > event["object"].spec.backoff_limit
                ):
                    logger.error(f"Job {job_name!r}: Job reached backoff limit.")
                    completed = True
                # If the job has no backoff limit, check if it has failed
                # and stop watching if it has
                elif (
                    not event["object"].spec.backoff_limit
                    and event["object"].status.failed
                ):
                    completed = True
                if completed:
                    break

    async def _watch_job(
        self,
        logger: logging.Logger,
        job_name: str,
        configuration: KubernetesWorkerJobConfiguration,
        client: "ApiClient",
    ) -> int:
        """
        Watch a job.

        Return the final status code of the first container.
        """

        logger.debug(f"Job {job_name!r}: Monitoring job...")

        job = await self._get_job(logger, job_name, configuration, client)
        if not job:
            return -1

        pod = await self._get_job_pod(logger, job_name, configuration, client)
        if not pod:
            return -1

        # Create a list of tasks to run concurrently
        async with self._get_batch_client(client) as batch_client:
            tasks = [
                self._monitor_job_events(
                    batch_client,
                    job_name,
                    logger,
                    configuration,
                )
            ]
            try:
                with timeout_async(seconds=configuration.job_watch_timeout_seconds):
                    if configuration.stream_output:
                        tasks.append(
                            self._stream_job_logs(
                                logger,
                                pod.metadata.name,
                                job_name,
                                configuration,
                                client,
                            )
                        )

                    results = await asyncio.gather(*tasks, return_exceptions=True)
                    if any(isinstance(result, Exception) for result in results):
                        for result in results:
                            if isinstance(result, Exception):
                                logger.error(
                                    f"Error during task execution: {result}",
                                    exc_info=True,
                                )
            except TimeoutError:
                logger.error(
                    f"Job {job_name!r}: Job did not complete within "
                    f"timeout of {configuration.job_watch_timeout_seconds}s."
                )
                return -1

            core_client = CoreV1Api(client)
            # Get all pods for the job
            pods = await core_client.list_namespaced_pod(
                namespace=configuration.namespace, label_selector=f"job-name={job_name}"
            )
            # Get the status for only the most recently used pod
            pods.items.sort(
                key=lambda pod: pod.metadata.creation_timestamp, reverse=True
            )
            most_recent_pod = pods.items[0] if pods.items else None
            first_container_status = (
                getattr(most_recent_pod, "status", None)
                and getattr(most_recent_pod.status, "container_statuses", None)
                and most_recent_pod.status.container_statuses[0]
            )

        if not first_container_status:
            logger.error(f"Job {job_name!r}: No pods found for job.")
            return -1

        # In some cases, the pod will still be running at this point.
        # We can assume that the job is still running and return 0 to prevent marking the flow run as crashed
        elif first_container_status.state and (
            first_container_status.state.running is not None
            or first_container_status.state.waiting is not None
        ):
            logger.warning(
                f"The worker's watch for job {job_name!r} has exited early. Check the logs for more information."
                " The job is still running, but the worker will not wait for it to complete."
            )
            # Return 0 to prevent marking the flow run as crashed
            return 0

        # In some cases, such as spot instance evictions, the pod will be forcibly
        # terminated and not report a status correctly.
        elif (
            first_container_status.state is None
            or first_container_status.state.terminated is None
            or first_container_status.state.terminated.exit_code is None
        ):
            logger.error(
                f"Could not determine exit code for {job_name!r}."
                "Exit code will be reported as -1."
                f"First container status info did not report an exit code."
                f"First container info: {first_container_status}."
            )
            return -1

        return first_container_status.state.terminated.exit_code

    async def _get_job(
        self,
        logger: logging.Logger,
        job_id: str,
        configuration: KubernetesWorkerJobConfiguration,
        client: "ApiClient",
    ) -> Optional["V1Job"]:
        """Get a Kubernetes job by id."""

        try:
            batch_client = BatchV1Api(client)
            job = await batch_client.read_namespaced_job(
                name=job_id, namespace=configuration.namespace
            )
        except kubernetes_asyncio.client.exceptions.ApiException:
            logger.error(f"Job {job_id!r} was removed.", exc_info=True)
            return None
        return job

    async def _get_job_pod(
        self,
        logger: logging.Logger,
        job_name: str,
        configuration: KubernetesWorkerJobConfiguration,
        client: "ApiClient",
    ) -> Optional["V1Pod"]:
        """Get the first running pod for a job."""

        watch = kubernetes_asyncio.watch.Watch()
        logger.info(f"Job {job_name!r}: Starting watch for pod start...")
        last_phase = None
        last_pod_name: Optional[str] = None
        core_client = CoreV1Api(client)
        async with watch:
            async for event in watch.stream(
                func=core_client.list_namespaced_pod,
                namespace=configuration.namespace,
                label_selector=f"job-name={job_name}",
                timeout_seconds=configuration.pod_watch_timeout_seconds,
            ):
                pod: V1Pod = event["object"]
                last_pod_name = pod.metadata.name
                phase = pod.status.phase
                if phase != last_phase:
                    logger.info(f"Job {job_name!r}: Pod has status {phase!r}.")

                if phase != "Pending":
                    return pod

                last_phase = phase
            # If we've gotten here, we never found the Pod that was created for the flow run
            # Job, so let's inspect the situation and log what we can find.  It's possible
            # that the Job ran into scheduling constraints it couldn't satisfy, like
            # memory/CPU requests, or a volume that wasn't available, or a node with an
            # available GPU.
            logger.error(f"Job {job_name!r}: Pod never started.")
            await self._log_recent_events(
                logger, job_name, last_pod_name, configuration, client
            )

    async def _log_recent_events(
        self,
        logger: logging.Logger,
        job_name: str,
        pod_name: Optional[str],
        configuration: KubernetesWorkerJobConfiguration,
        client: "ApiClient",
    ) -> None:
        """Look for reasons why a Job may not have been able to schedule a Pod, or why
        a Pod may not have been able to start and log them to the provided logger."""

        def best_event_time(event: CoreV1Event) -> datetime:
            """Choose the best timestamp from a Kubernetes event"""
            return event.event_time or event.last_timestamp

        def log_event(event: CoreV1Event):
            """Log an event in one of a few formats to the provided logger"""
            if event.count and event.count > 1:
                logger.info(
                    "%s event %r (%s times) at %s: %s",
                    event.involved_object.kind,
                    event.reason,
                    event.count,
                    best_event_time(event),
                    event.message,
                )
            else:
                logger.info(
                    "%s event %r at %s: %s",
                    event.involved_object.kind,
                    event.reason,
                    best_event_time(event),
                    event.message,
                )

        core_client = CoreV1Api(client)

        events: CoreV1EventList = await core_client.list_namespaced_event(
            configuration.namespace
        )

        event: CoreV1Event
        for event in sorted(events.items, key=best_event_time):
            if (
                event.involved_object.api_version == "batch/v1"
                and event.involved_object.kind == "Job"
                and event.involved_object.namespace == configuration.namespace
                and event.involved_object.name == job_name
            ):
                log_event(event)

            if (
                pod_name
                and event.involved_object.api_version == "v1"
                and event.involved_object.kind == "Pod"
                and event.involved_object.namespace == configuration.namespace
                and event.involved_object.name == pod_name
            ):
                log_event(event)