deployment

A canary release sends a slice of real traffic to a new version and asks whether it is healthy. This lesson covers what to measure, why comparing the canary against the current version beats comparing against history, the statistics problem that makes small canaries weak evidence, and how automated promotion and rollback turn a judgement call into a rule.

Deploying is the moment a working system is replaced by a different one while people are using it. This lesson covers what actually goes wrong at that moment, the research finding that shipping fast and shipping safely are not opposites, and the four deployment strategies as answers to one question: how many users meet a bad version before you find out.

Every deployment strategy assumes you can go back, and that assumption is the one most often false when it matters. This lesson covers what actually makes a rollback work, the database migration pattern that keeps schema changes reversible, the one-way doors that no amount of tooling can undo, and how to tell which kind of change you are about to ship.

A feature flag turns shipping code and exposing behaviour into two independent decisions, which is what lets a team deploy continuously while releasing on someone else's schedule. This lesson covers the four kinds of flag and their very different lifespans, the debt they accumulate, and the discipline that keeps a flag system from becoming untestable.

Quantization damage does not show up where people look for it. Perplexity barely moves while hard tasks degrade, and long reasoning suffers most because error compounds. This lesson covers building an evaluation that detects real loss, where the published cliffs are, KV cache quantization as a separate lever, end-to-end memory sizing, and the rollout that catches what evals miss.

Decoding is memory-bandwidth-bound, so fewer bits per weight means more tokens per second, not fewer. This lesson builds the mechanism from the arithmetic up: the affine mapping, a worked example done by hand, why group size costs fractional bits, the difference between W4A16 and W8A8, and which formats the hardware actually accelerates.

What happens when a prediction model leaves the benchmark: adversarial fragility, distribution shift, the feedback loop where your own robot changes the behaviour it predicts, and how planners consume uncertainty safely.

Self-hosting is usually justified on privacy grounds, and the argument is often weaker than assumed. This lesson examines what actually stays private, which regulatory requirements genuinely demand it, the reasons that hold up better than privacy, and the honest total cost against an API.