What are Challenges and Key Principles for ML Systems

---

Challenges and Key Principles for ML Systems

Specific Challenges of ML Systems

• The need for reproducibility (versioning everywhere)
• Entanglement
• Data Dependencies(데이터 의존성은 이미 수행된 데이터의 변화가 뒤의 수행 결과에 영향을 끼치는 것을 의미)
• Configuration issues
• Data and feature preparation
• Model errors can be hard to detect with traditional tests
• Separation of expertise

ML System COntributors

The architecure of a production machine learning system needs to take into account business requirements, as well as the unique challenges at the intersection of data science, software engineering and devops.

Research vs. Production Environments

• Research : Seperate from customer facing software, Can be taken offline
• Production : Reproduciblity matter, Infrastructure planning required

Key Principles for ML Sysytem ARchitecture

• Reproducibility : Have the ability to replicate a given ML prediction
• Automation : Retrain, update and deploy models as part of an automated pipeline
• Extensibility : Have the ability to easily add and update model
• Modularity : Preprocessing/feature engineering code used in training should be organized into clear pipeline
• Scalability : Ability to serve moel predictions to large numbers of customer(within time constraints)
• Testing : Test variation between model versions

Unknown Word

• Entanglement : a situation or relationship that you are involved in and that is difficult to escape from
  (The book describes the complex emotional and sexual entanglements between the members of the group.)