Datum Batching

By default, HPE Machine Learning Data Management processes each datum independently. This means that your user code is called once for each datum. This can be inefficient and costly if you have a large number of small datums or if your user code is slow to start.

When you have a large number of datums, you can batch them to optimize performance. HPE Machine Learning Data Management provides a next datum command that you can use to batch datums.

Flow Diagram #

flowchart LR
    user_code(User Code)
    ndsuccess(NextDatum)
    nderror("NextDatum(error)")
    response(NextDatumResponse)
    process_datum{process datum}

    cmd_err(Run cmd_err)
    kill[Kill User Code]  
    datum?{datum exists?}
    retry?{retry?}
    cmd_err?{cmd_err defined}

    user_code ==>ndsuccess
    ndsuccess =====> datum?
    datum? ==>|yes| process_datum
    process_datum ==>|success| response
    response ==> user_code

    datum? -->|no| kill
    process_datum -->|fail| nderror
    nderror --> cmd_err?
    cmd_err? -->|yes| cmd_err
    cmd_err? -->|no|kill
    cmd_err --> retry?
    retry? -->|yes| response
    retry? -->|no| kill

How to Batch Datums #

Define your user code and build a docker image. Your user code must call pachctl next datum to get the next datum to process.

transformation() {
  # Your transformation code goes here
  echo "Transformation function executed"
}

echo "Starting while loop"
while true; do
  pachctl next datum
  echo "Next datum called"
  transformation
done

Your user code can apply the @batch_all_datums convenience decorator to iterate through all datums. This will perform the NextDatum calls for you as well as prepare the environment for each datum.

import os
from pachyderm_sdk import batch_all_datums

@batch_all_datums
def main():
   # Processing code goes here.
   # This function will be run for each datum until all are processed.
   # Once all datums are processed, the process is terminated.
   print(f'datum processed: {os.environ["PACH_DATUM_ID"]}')

def init():
   # Initializing code goes here.
   # When this function is called, no input data is present.
   print('Preparing for datum batching job')

if __name__ == '__main__':
    init()
    print('Starting datum processing')
    main()

Create a repo (e.g., pachctl create repo repoName).
Define a pipeline spec in YAML or JSON that references your Docker image and repo.

Add the following to the transform section of your pipeline spec:

datum_batching: true

pipeline:
  name: p_datum_batching_example
input:
  pfs:
    repo: repoName
    glob: "/*"
transform:
  datum_batching: true
  image: user/docker-image:tag

Create the pipeline (e.g., pachctl update pipeline -f pipeline.yaml).
Monitor the pipeline’s state either via Console or via pachctl list pipeline.

Tip

You can view the printed confirmation of “Next datum called” in the logs your pipeline’s job.

FAQ #

Q: My pipeline started but no files from my input repo are present. Where are they?

A: Files from the first datum are mounted following the first call to NextDatum or, when using the Python client, when code execution enters the decorated function.

Q: How can I set environment variables when the datum runs?

A: You can use the .env file accessible from the /pfs directory. To easily locate your .env file, you can do the following:

def find_files(pattern):
    return [f for f in glob.glob(os.path.join("/pfs", "**", pattern), recursive=True)]

env_file = find_files(".env")

Article Summarization

Datum Batching

Flow Diagram #

How to Batch Datums #

FAQ #