How would you split HTTP-layer coverage across handler-only, slice and full-stack tests for a service with hundreds of routes?

## Start from the cost shape, not from a ratio The three tiers do not differ only in speed; they differ in **how their cost scales**, and that is what decides where each kind of case belongs. | Tier | Fixed cost | Scales with | Evidence it uniquely produces | |---|---|---|---| | Direct handler call | None | Number of cases | Branch and error behaviour inside the body | | Router-and-hooks slice | One framework start per configuration | Number of distinct configurations | Route matching, binding, hooks, status, headers, serialisation | | Full stack on real wiring | A real boot, plus real resources | Number of environments and wiring variants | The deployed graph builds, configures and answers | A ratio like "seventy, twenty, ten" is a symptom of a good split, never a rule that produces one. What produces one is asking, per case, **which tier is the cheapest place this fault can be seen**. ## A workable split at hundreds of routes 1. **Volume goes to the bottom.** Every validation branch, every domain error, every awkward edge in the handler body is a direct call. At this scale that is thousands of cases costing seconds in total. 2. **One contract assertion per route in a slice.** For each route, a single test proves it matches under the right method, binds its inputs, and answers with the agreed status, the headers clients act on, and a payload of the promised shape. Hundreds of such tests on a handful of shared configurations remain cheap, because the start-up is amortised. 3. **Configuration discipline is the slice budget.** Treat a new slice configuration as an expensive purchase; consolidate aggressively, because suite runtime tracks configuration count far more than test count. 4. **The full stack is narrow and deliberate.** One proof that the application starts on real wiring, plus one real request per route family whose wiring is genuinely distinct. Assert coarsely there — status and a minimal shape — and leave detail to the slice. 5. **True externals stay replaced even at the top.** The point of the full stack is your graph, not someone else's availability. ## Steer by what escapes, not by a target The honest feedback signal is the defect that reached production, classified by which tier could have caught it: - Escaping **logic** faults mean the bottom tier is thin or its cases are shallow. - Escaping **binding, status, header or serialisation** faults mean routes lack a slice-level contract test, or those tests assert the wrong thing. - Escaping **construction, registration or configuration** faults mean the full-stack tier is too narrow. - Faults escaping in more than one class at once usually mean the suite is concentrated in a single tier and everything outside it is untested by construction. This classification is worth keeping as a lightweight running list. It converts an argument about testing philosophy into an argument about observed escapes, which is far easier to settle. ## The failure modes of a single-tier suite - **All handler calls.** Fast, green, and unable to prove any endpoint exists. Contract drift ships freely. - **All slices.** The HTTP surface is well covered, but nothing ever builds the real graph, so every deployment carries construction risk. - **All full stack.** Every fault is catchable in principle, but the suite is slow, the feedback loop collapses, flakiness arrives with real resources, and people start skipping it — which removes more coverage than any deliberate choice would have. ## The organisational half of the answer A split only holds if the cheapest tier is also the **easiest to write**. If adding a direct handler call requires assembling an elaborate imitation of a request, engineers will write a slice test instead, and the suite drifts upward one commit at a time until it is slow. So the leverage is in the code shape: handlers that take plain values and return plain results, a small number of shared slice configurations, and one obvious place to add a full-stack case. Where that shape is not available — because handlers are written against framework-owned request and response objects — accept it, put the volume in the slice, and spend the saved argument on keeping configurations few. Finally, state the boundary you are buying. At any split there is a region only production exercises; naming it, rather than pretending it is closed, is what makes the plan defensible when something escapes anyway.