MLOps Conference

Audience Level: Intermediate
Nature: Conceptual

https://drive.google.com/file/d/1WpKUXeC0i93f72vC8yXtUR6unPEaB4ma/view?usp=sharing

Scribble has built and operate feature stores for companies for the
past few years. This is a perspective talk on why feature stores came
about, what is being built today, and what we foresee over the next
few years.

1. Feature store introduction and history

2. Understanding existing feature stores
   (a) Architecture: Integrated/standalone
   (b) Scale: Peta/Tera
   (c) Core abstraction: SQL-like/program
   (d) Application scope: ML/Non-ML
   (e) Programming Iterface: Tight/Open
   (f) Data classes: Streaming/Timeseries, document, transactions

3. Classes of decisions
   (a) What, Shallow why, Deep why, Why not
   (b) How these are addressed today & gaps

4. Feature Stores 1.0: Passive, robust, scalable
   (a) Focused on ML usecases
   (b) Focus on scale & abstractions
   (c) Passive but robust

5. Feature Stores 2.0: Intelligent, trusted, end-to-end
   (a) Context-aware - Integrates with upstream and downstream
   About data, nature of processing, risks involved
   Changes operations, resources, observation levels
   (b) Knowledge management - Help ip creation
   Better and efficient processes
   (c) Risk management - Trust and safety as first class goal
   Reduce risks from insecure, poor/changing code & data
   Change handling, impact assessment
   (d) Proactive - Actively observes and recommends
   Suggests features, impact assessment
   (e) Scope - Expanded classes of decisions and users
   All classes of advanced data needs (shallow why...)
   (f) Distributed - Handle constraints (time, volume etc)
   Data cannot/should not flow to centralized
   Distributed discovery

6. Some niche contexts where new classes of
   stores might emerge:
   (a) Constrained devices (handhelds)
   (b) Classes of data (geospatial)
   (c) Computational complexity (1000s of models)

7. Key Takeaways
   (a) Feature stores are now a standard component
   (b) Understanding the journey will help future-proof your implementation
   (c) Feature stores 2.0 will be different from 1.0