Skip to content

Latest commit

 

History

History
96 lines (70 loc) · 3.56 KB

README.md

File metadata and controls

96 lines (70 loc) · 3.56 KB
Raw

jobq: Packaging, scheduling, and monitoring Python workloads as k8s jobs

jobq is a tool to define, schedule, and observe compute jobs as Python functions in Kubernetes. It supports Docker (standard Kubernetes), Kueue, and Ray backends for workload execution.

The jobq Python package itself is split into a client and backend part. The client package is used to define jobs with the @job decorator in your Python files, and (optionally) to interact with an existing API server through a CLI. The backend contains a FastAPI-based server application that acts as an intermediary between the user and the target Kubernetes cluster, and can be easily packaged into a Docker container or a virtual machine for deployment.

Installation

Depending on what your role is, you will likely require only some parts of the jobq package.

As a data scientist, you only require the client part to define and dispatch jobs. To install the client package, simply use pip:

pip install --upgrade "aai-jobq[cli]"

This instruction includes the command-line interface to interact with an existing Kubernetes cluster. If that functionality is not needed, simply omit the bracketed extra from the install command.

Usage

Define jobs in your .py files by using the @job decorator on the function that contains your compute logic.

import time

from jobq import job

@job
def hello_world():
    """A simple job printing a log message ten times."""
    for idx in range(10):
        print(f"Hello, World!, {idx=}")
        time.sleep(2)

Within the decorator, you define the desired resources and link the environment configuration. We support Dockerfiles or a declarative .yaml file specification.

# The same example, but with configuration options for a Kueue admission.
import time
from pathlib import Path

from jobq import ImageOptions, JobOptions, ResourceOptions, SchedulingOptions, job

@job(
    options=JobOptions(
        labels={"type": "hello-world@dev", "x-jobq.io/key": "value"},
        resources=ResourceOptions(memory="1Gi", cpu="1"),
        scheduling=SchedulingOptions(
            priority_class="background", queue_name="user-queue"
        ),
    ),
    image=ImageOptions(
        spec=Path("examples/example-hello.yaml"),
        name="localhost:5000/hello-world-dev",
        tag="latest",
    ),
)
def hello_world():
    """A simple job printing a log message ten times."""
    for idx in range(10):
        print(f"Hello, World!, {idx=}")
        time.sleep(2)

Using the jobq CLI, you can execute jobs locally or submit them to one of the supported backends.

# Requires the file name in which the job is defined, and the execution mode.
jobq submit --mode=kueue example.py

Additionally, you can query for a job status and fetch logs.

# Below is a typical unique identifier of a workload inside a cluster.
jobq logs 4c0d4cbb-13a2-4d05-afde-8784e22bb940

jobq is an early stage project, so expect rough edges. We are happy for feedback and may accomodate your feature requests, so don't hesitate to get in touch.

Contributions

If you'd like to contribute to jobq, the best way to do so is to open an issue for bug reports and feature requests, or start a discussion for topics around the project, and features that might need prior feedback from the core developers.

For more information on how to contribute, refer to the contribution guide.

License

jobq is distributed under the Apache-2 license.