202305081026-data_science_questions_in_enterprise_context
A short path to failure is to develop ML models without thinking about the deployment context. We need to ask ourselves several questions on the broader data pipeline: data management, quality assurance, operational context, the target platform, etc.:
- What data pipeline do we need to support an optimized ML pipeline?
- Does the ML run in the cloud or on edge?
- How will data be acquired and secured in production environments and does this constrain your ML design?
- Does ML model require data preparation and a certain data quality?
- How will data preparation be done in an operational context?
- Does this constraint the target environment or ML model?
- Do you actually own and have access to the data for training and running your models in production?
- Are the features used during development also available in production at the required speed?
- E.g. if you need to do federated learning in production, perhaps we need local dummy data for developing our initial model locally.
- How is data quality approached?
- Which data quality requirements to the Data Architect or Data Engineer need to know from the get go?
- How do you bring data quality into production?
- How to avoid unwanted bias in your training set.
- How will you monitor in production?
- This may impact choices you make and priorities you set now
- How will you retrain your models in production?