ML Platform Evolution — Week 3 CheckWeek 3 of Data Engineering for ML & AI: decade-long platform arcs — Michelangelo, Metaflow, Griffin, LyftLearn — and what each rewrite fixed.

Uber's Michelangelo, Netflix's Metaflow, Instacart's Griffin — the recurring first problem every ML platform solved:

Instacart's Griffin tripled ML applications in a year. Platform leverage of that kind comes from:

ML monitoring differs from service monitoring because models fail:

Training jobs are data-engineering jobs wearing lab coats. A skewed join OOM-ing a training pipeline (this week's Spark scenario) is diagnosed:

Model retraining pipelines need REPRODUCIBILITY. The data-side requirement:

The platform decision 'build vs adopt' (Stripe adapting Chronon vs building fresh) should weigh: