Architecture Component Breakdown

Created: 2023-03-02 11:10
#quicknote

High level architecture:

1. Evaluation layer: it checks the equivalence of two models and it can be used to monitor production models;
2. Scoring layer: it transforms features into predictions (scikit learn is industry standard);
3. Feature layer: it is responsible for generating feature data in a transparent, reusable, and scalable manner;
4. Data layer: it provides access to all of our data sources which simplifies the challenge of data reproducibility.

Offline training phase (from the bottom):

• training data -> functions or application that load and/or transform data;
• feature extraction -> scripts that generates features;
• model builder -> it transforms model in a form that can be deployed;
• output: trained model

Prediction phase:

• feature extractor as similar as possible to the one used in the offline phase;
• trained model obtained from the offline phase

Resources

1. FBLearner Flow
2. Scaling ML as a Service: Uber
3. System architectures for personalization and recommendation: Netflix
4. TFX: Google

Tags

#mlops #deployment #course