Executing batch jobs in Dask Cluster
In this example we will: - Create a Dask cluster - Submit a job python file - Terminate the cluster after job is completed.
Before you begin
- Create "dask" under your "~/my" folder
- And copy job.py under this folder
import practicuscore as prt
job_dir = "~/my/dask"
distributed_config = prt.distributed.JobConfig(
job_type = prt.distributed.JobType.dask,
job_dir = job_dir,
py_file = "job.py",
worker_count = 2,
)
worker_config = prt.WorkerConfig(
worker_size="X-Small",
distributed_config=distributed_config,
log_level="DEBUG",
)
coordinator_worker = prt.create_worker(
worker_config=worker_config,
)
# 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/dask/result.csv/
- Please note that result.csv is a folder that can contain
parts of the processed file
by each worker (Dask 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.distributed.JobConfig
properties to customize the cluster
Supplementary Files
job.py
import practicuscore as prt
import dask.dataframe as dd
# Let's get a Dask session
print("Getting Dask session")
dask = prt.distributed.get_client()
print("Reading diamond data")
df = dd.read_csv('/home/ubuntu/samples/diamond.csv')
print("Calculating")
df["New Price"] = df["Price"] * 0.8
print("Since Dask is a lazy execution engine,")
print(" actual calculations will happen when you call compute() or save.")
print("Saving")
df.to_csv('/home/ubuntu/my/dask/result.csv')
# Note: the save location must be accessible by all workers
# A good place to save for distributed processing is object storage
Previous: Use Cluster | Next: Distributed Training > XGBoost