MLOps Best Practices

At Task-TS we aim to set broadly applicable standards to versioning, tracking, deploying, monitoring, and re-training models. Moreover, we want to show how software engineering and DevOps best practices can be synthesized with cutting edge ML-Research.

Reproducible Results

In order for experiments to be reproducible three major things need to be met:

1. Data Versioning

   a. Store all data on a daily basis to GCS and Dataverse.

   b. Maintain historical versioned snapshots of data.

   c. Allow direct loading into flow from a historical snapshot.

2. Experiment tracking

3. Code/Config versioning a. Store all experiment code to GitHub and tag each experiment with commit hash. b. Store notebooks full notebooks to Github

** Extendibility Best Practices**