It is 2015 in a Mountain View war room, and the postmortem is embarrassing: the machine-learned model at the heart of the product is a few hundred lines, but it is wrapped in tens of thousands of lines of glue — feature pipelines nobody owns, data dependencies nobody declared, a training set that quietly diverged from what the servers see in production. Accuracy fell for a month before anyone noticed, because no test covers "the world changed." Engineer ML as a system: detect training-serving skew, version data like code, monitor the input distribution, and make every consumed signal a declared, tested dependency. Get it wrong and the organization ships a model that was excellent at training time into a production reality that has already moved on.
The professor puts up a slide titled 'MLOps Best Practices' with a stock photo of gears, and by the time he reaches bullet point two the room has already diagnosed the training-serving skew from the symptom description alone. He once shipped a demo where the feature pipeline silently used yesterday's exchange rate for three weeks; nobody noticed until a student pointed out the numbers looked 'confidently wrong,' which is now his favorite phrase to describe his own code. He teaches this exact 2015 Google postmortem every semester and still, this morning, deployed a notebook that trained on one column ordering and served on another. Dean built the infrastructure that makes this problem tractable; the professor built the cautionary anecdote that makes it memorable. Score: zero, earned honestly, cited in his own lecture notes.
Bellman's dynamic programming provides a general framework for structuring multistage decision problems, and his curse-of-dimensionality insight is relevant, in the abstract, to why complex production ML pipelines with many interacting components become difficult to reason about and easy to break silently as they scale. But Bellman's actual career-defining contribution was algorithmic and combinatorial optimization theory, developed decades before production machine-learning systems, data pipelines, or feature-engineering infrastructure existed as engineering categories; he never worked on anything resembling training-serving skew or dependency management. His relevance to this specific 2015 problem is generic mathematical adjacency rather than a direct historical engineering precedent, since nothing in his multistage decision framework addresses who owns a broken feature pipeline.
Battle #22 · 8/9/2026, 5:14:52 PM · this result is deterministic: the same two personas on this problem always resolve the same way.