How to Elm-ify Your ML

on YouTube ↗41m

" summary

Stripe builds a small, principled feature system: a free applicative feature DSL that describes feature logic as an AST, compiles to batch (MapReduce/Scalding/Summingbird) or a Kafka-backed push backend, and enforces time-aware lookups so temporal leakage is impossible. Unit tests run as lazy streams (example: dog bark/jump/howl test). Real-time scoring hits ~10–50 ms for typical graphs. The claim is that separating business logic from execution yields consistent training vs. scoring and simpler, optimizable runtimes.

AI-written summary · about

" chapters (14)

1. 0:00Intro
2. 0:52What is machine learning
3. 3:06Key challenges
4. 6:41Functional programming and time
5. 10:35Magickal matrices
6. 14:21Scala type classes
7. 15:59combinators
8. 18:47free implementation
9. 22:13map
10. 24:01lookup
11. 28:30how to run
12. 30:55pushbackend
13. 31:52conclusion
14. 34:20windowing

" related talks

• 10mFast Sub-ML Usecase Development using Feature StoresAchint Thomas · Feature Store Summit 2022
• 17mCreating & operating ML models from event-based data using feature stor...Charna Parkey · Global Artificial Intelligence Virtual Conference 2022
• 13mDesigning Feature DSLs: Principles and TradeoffsGreg Kuhlmann · Feature Store Summit 2022
• 31mLeveraging GraphQL for Continual Learning in Real-Time ML SystemsAryan Magoon · GraphQL Summit 2023
• 53mAdventures in Elm: Events, Reproducibility, and KindnessJessica Kerr · Philly ETE 2016
• 36mBuild prod ML systems with feature /training /inference pipelinesJim Dowling · Conf42 Python 2023
• 20mSemblance: Feature Generation In Real-Time and Batch Without Time-TravelTianshi Zhu · TWIMLcon: AI Platforms 2019
• 01h 28mFunctional Programming for the Web with ElmAustin Bingham · ACCU 2017