Option to stop compute worker from sending logs and/or prediction files to Codabench

[ObadaS]

A brief description of the purpose of the changes contained in this PR.

• This PR adds the ability to stop the compute worker from sending the competition container logs and/or prediction files to Codabench. Instead, the logs will be saved locally on disk.
• Rename the folders that are created to be more descriptive and less random.
• Updated the documentation to show the new options in the compute worker .env

Checklist

• Code review by me
• Hand tested by me
• I'm proud of my work
• Code review by reviewer
• Hand tested by reviewer
• CircleCi tests are passing
• Ready to merge

[Didayolo]

Hi @ObadaS,

Some points looks problematic to me:

1. Readability issue in local log file

In push_logs, data is bytes, so f.write(str(data)) writes b'...\n...' on a single line instead of the actual text.

This is not critical but the logs produced will be annoying, with no line breaks.

Possible solution: open the file in "wb" mode and write the bytes directly (ideally with a header per program/stream).

2. Logs are deleted right after being written

With only COMPUTE_WORKER_DISABLE_LOG_UPLOAD=True, push_logs() writes the file inside root_dir, then clean_up() deletes root_dir. Maybe we could write them outside root_dir (e.g. /codabench/logs/) or force NO_CLEANUP like for predictions.

3. Local prediction transfer fails silently with multiple workers

I think in the doc it is said that this work only with 1 worker in the queue. But that is not explicitly reported in case of problem.

With COMPUTE_WORKER_DISABLE_PREDICTION_UPLOAD, scoring must run on the same worker as prediction, but nothing enforces or checks this. If the folder is missing, copytree raises a raw FileNotFoundError.

I think we need to check that the folder exists and raise a clear SubmissionException.

4. Predictions copied scoring_program folder

Predictions are copied into scoring_program/submission/. From what I get, the folder contains the submission data sent during scoring, so code and predictions end up mixed in both /app/ingested_program and /app/input/res, no? Not sure about this point. But maybe the predictions could be copied directly into input/res, as usual.