Skip to content

Executing batch jobs in Ray Cluster

In this example we will: - Create a Ray cluster - Submit a job python file - Terminate the cluster after job is completed.

Before you begin

  • Create "ray" under your "~/my" folder
  • And copy job.py under this folder
import practicuscore as prt

job_dir = "~/my/ray"

distributed_config = prt.DistJobConfig(
    job_type = prt.DistJobType.ray,
    job_dir = job_dir,
    py_file = "job.py",
    worker_count = 2,
)

worker_config = prt.WorkerConfig(
    # Please note that Ray requires a specific worker image
    worker_image="practicus-ray",
    worker_size="Medium",
    distributed_config=distributed_config,
    log_level="DEBUG",
)

coordinator_worker = prt.create_worker(
    worker_config=worker_config,
)
prt.distributed.live_view(
    job_dir=job_dir,
    job_id=coordinator_worker.job_id,
)
# You can view the logs during or after the job is completed
# To view coordinator (master) set rank = 0
rank = 0
# To view other workers set rank = 1,2, ..

prt.distributed.view_log(
    job_dir=job_dir,
    job_id=coordinator_worker.job_id,
    rank=rank
)

Wrapping up

  • Once the job is completed, you can view the results in ~/my/ray/result.csv/
  • Please note that result.csv is a folder that can contain parts of the processed file by each worker (Ray executors)
  • Also note that you do not need to terminate the cluster since it has a 'py_file' to execute, which defaults terminate_on_completion parameter to True.
  • You can change terminate_on_completion to False to keep the cluster running after the job is completed to troubleshoot issues.
  • You can view other prt.DistJobConfig properties to customize the cluster

Supplementary Files

job.py

import practicuscore as prt

ray = prt.distributed.get_client()


@ray.remote
def square(x):
    return x * x


def calculate():
    numbers = [i for i in range(10)]
    futures = [square.remote(i) for i in numbers]
    results = ray.get(futures)
    print("Distributed square results of", numbers, "is", results)


if __name__ == "__main__":
    calculate()
    ray.shutdown()

Previous: Use Cluster | Next: Modin > Start Cluster