• Overview
• Site Worker
• Compute Worker
• Submission Container

Overview

When you make a submission to Codabench, the information and file are saved to the database. Afterwards, a Celery task gets sent to the compute worker (Default queue or a compute worker attached to a custom queue). From there the compute worker spins up another docker container with either the default docker image for submissions, or a custom one supplied by the organizer.

Site Worker

The site worker can be thought of exactly how it sounds. It's a local celery worker for the site to handle different Celery tasks and other miscellaneous functions. This is what is responsible for creating competition dumps, unpacking competitions, firing off the tasks for re-running submissions, etc.

Compute Worker

Is a remote and/or local celery worker that listens to the main RabbitMQ server for tasks. A compute worker can be given a special queue to listen to, or listen to the default queue. For more information on setting up a custom queue and compute worker, see:

• Queue Management
• Compute Worker Management and Setup

Submission Container

The submission container is the container that participant's submissions get ran in. The image used to create this container is either the Codabench default if the organizer's haven't specified an image, or the custom image they specified in the competition. The default docker image is codalab/codalab-legacy:py37. Organizers can specify their own image using either the YAML file or the editor.

This container is also created with some specific directories, some of which are only available at specific steps of the run. For example, generally for a submission there are 2 steps. The prediction step (If the submission needs to make predictions) and the scoring step, where the predictions are scored against the truth reference data.

Here are the following directories:

• /app/input_data Where input data will be (Only exists if input data is supplied for the task)
• /app/output Where all submission output should go (Should always exist)
• /app/ingestion_program Where the ingestion program files should exist (Only available in ingestion)
• /app/program Where the scoring program and/or the ingestion program should be located (Should always exist)
• /app/input Where any input for this step should be. (I.E: Previous predictions from ingestion for scoring. Only exists on scoring step)
• /app/input/ref Where the reference data should live (Not available to submissions. Only available on scoring step.)
• /app/input/res Where predictions/output from the prediction step should live. (Only available on scoring step)
• /app/shared Where any data that needs to be shared between the ingestion program and submission should exist. (Only available on scoring steps.)
• /app/ingested_program Where submission code should live if it is a code submission. (Only available in ingestion)