Comparisons
Pairs that get conflated in real metric reviews and real design docs — batch and online, precision and recall, data drift and concept drift, validation and test. Neither column wins; what decides is the problem. Each record leads with the confusion, because the confusion is the reason the record exists.
Shadow deployment vs canary rollout
They are treated as alternatives, or the canary is run without the shadow because "we tested offline". They answer different questions in a sequence. Shadow answers "does this thing work in the serving environment" — do the features match training, is the latency within budget, does it crash on the tail of inputs the validation set never had — and it cannot answer whether users like the output, because no user sees it. A canary answers "what happens to the outcome" — and it cannot answer the shadow's questions cheaply, because by the time a skew bug shows in the canary metric, a slice of users has been served wrong predictions. The strongest form of the shadow side is that it is the only zero-risk way to observe a model on real traffic, and it makes the comparison request-by-request, which is more powerful than any aggregate. The strongest form of the canary side is that a shadow can only compare predictions to the champion, and if the champion is wrong, agreement is not a virtue; only exposure measures the outcome. The costs differ too: shadow doubles the compute for the duration; canary needs routing, a rollback path and a labelled outcome that arrives fast enough to judge the slice.
Before any user is exposed: to check that the model runs under production load, that its features arrive as expected, that its predictions agree with the champion where they should, and to catch skew and latency problems with zero product risk.
To measure the effect on real outcomes on a slice of traffic, with the ability to roll back fast, once the shadow phase has shown the model behaves. It is the first step that can move a business metric.
| Dimension | Shadow — the new model scores live traffic but its predictions are not acted on | Canary — the new model serves a small share of traffic for real |
|---|---|---|
| User exposure | None | A small share |
| Answers | Does it run, match features, meet latency, agree with champion? | Does it move the outcome? |
| Cannot answer | What users do with the output | Cheaply, whether a skew bug exists |
| Cost | Double compute for the duration | Routing, rollback, a slice of risk |
| Needs labels | No | Yes, arriving fast enough to judge |
| Comparison unit | The same request, both models | Two populations of requests |
| Order | First | Second, then A/B, then full rollout |
Model families compared
Linear, trees, boosting, neural and k-NN — compared without naming a winner, and with the block that says where the comparison stops being true.
No column is a winner. Each family is a set of assumptions about the data — linear separability, axis-aligned interactions, additive residuals, a representation that can be learned, locality in feature space — and the right one is the one whose assumptions your data happens to satisfy at the size you have. The reflex answer “XGBoost” is the red flag this table exists to catch: it is often right on tabular data and it is never right as a reflex, because it skips the baseline that would have told you whether anything more than a linear model was needed. The where this comparison misleads block on every row is the part worth reading.
Count independent entities, not rows — a million events from ten thousand users is ten-thousand-sized data for generalising to new users. And the neural column flips completely with transfer learning: a pretrained model fine-tuned on two thousand images beats every other column on that task.
The "boosting wins on tabular" claim is true for a tuned model on tens of thousands of clean rows with a validation set. It is not true for three hundred rows, for a problem whose signal is linear, for a regulator who wants coefficients, or for a team with no time to tune — and the gap to the linear model is often inside the error bar.
The columns are not competing on the same input. Once a pretrained network has produced an embedding, a linear model or k-NN on top of it is often within a few points of full fine-tuning — so "neural for images" usually means "a neural representation, then whichever head is cheapest".
Interpretability is not one property. A linear model with two hundred correlated features and L1 selection is harder to explain honestly than a depth-three tree; and every column's "explanation" is undermined equally by a leaked or proxy feature, which the explanation will present with confidence.
Training cost is dominated by the number of runs, not the run: a boosted model tuned over two hundred configurations costs more than a network fine-tuned once. And the k-NN column's zero is a loan repaid on every query.
The model is rarely the slow part. Feature retrieval, a network hop and JSON serialisation usually dwarf any of these numbers, so a latency budget is a question about the serving path before it is a question about the family.
A low burden on the modelling side does not remove the burden on the serving side: every column's features must be reproduced at prediction time with the same code, the same freshness and the same point-in-time semantics. Trees remove the need to engineer features, not the need to serve them.
Native handling is convenient and dangerous in equal measure: it lets a tree model learn that "missing" predicts the target, which is fine until serving produces missingness for a different reason — a timeout, a new form — and the model reads the outage as a signal.
Calibration only matters if a downstream decision multiplies the score by a cost or compares it to a probability threshold; a pure ranker does not need it. And calibration measured on the validation set drifts with the base rate in production, for every column alike.
Neither behaviour is "correct". A tree that predicts last year's maximum for a record-breaking day is wrong; a linear model that predicts a negative price is wrong differently. The honest answer is a monitor on inputs outside the training range and a fallback for them, whichever family serves.
Irrelevant is not the danger — leaky is. Every column will seize a feature that carries the answer, and the more capable the model the more efficiently it does so; robustness to noise says nothing about robustness to leakage, which only a point-in-time audit provides.
Every column is wrong somewhere, and the question that finds where is the same for all of them: how much data, of what modality, at what latency, with what interpretability requirement, judged by which metric, at what cost to train and serve. A model family chosen before those are answered is a guess with a library name.