zaku 0.0.5

Zaku Task Queue is for distributed ML-workloads.

Zaku, a fast Task Queue for ML Workloads

To get a quick overview of what you can do with zaku, check out the following:

• take a look at the basic tutorial or the tutorial for robotics:
  • Zaku Basics
• or try to take a look at the example gallery here

Install zaku --- the latest version is {VERSION} on pypi.

pip install -U 'zaku[all]=={VERSION}'

Setting Up Zaku Server

The server script is installed as the command zaku. First, take a look at its options by running

zaku -h

This should show you the documents on all of the arguments. Now, to setup a zaku task queue server, run the following: This enables access from other than localhost.

zaku --host 0.0.0.0 --port 9000

Adding Jobs:

Supposed you have a TaskServer running at localhost:9000.

from zaku import TaskQ

queue = TaskQ(name="my-test-queue", uri="http://localhost:9000")

for i in range(100):
    queue.add_job({"job_id": i, "seed": i * 100})

Retrieving Jobs:

from zaku import TaskQ

queue = TaskQ(name="my-test-queue", uri="http://localhost:9000")

job_id, job = queue.take()

Now, after you have finished the job, you need to mark the job for completion. The way we do so is by calling

queue.mark_done(job_id)

Sometimes when you worker responsible for completeing the job encounters a failure, you need to also put the job back into the queue so that other workers can retry. You can do so by calling

queue.mark_reset()

Now, we offer a context manager TaskQ.pop, which automatically catches exceptions and resets the job (or marks it complete).

from zaku import TaskQ

queue = TaskQ(name="my-test-queue", uri="http://localhost:9000")

with queue.pop() as job:
  if job is None:
    print("No job available")
  
  print("Retrieved job:", job)

Developing Zaku (Optional)

If you want to develop zaku, you can install it in editable mode plus dependencies relevant for building the documentations:

cd zaku
pip install -e '.[dev]'

To build the documentations, run

make docs