What measurement forbids

Six claims no measurement supports

A number without its configuration.
A comparison without a noise floor.
A percentile without a composition rule.
A capacity claim without a latency constraint.
A conclusion without a control.
An extrapolation without a named boundary.

Each article's own list names the rest. Every stage can pass while the next fails, and a failure at any stage makes everything after it decorative.

Case, resolvedConstructed case · Ingrid at Bellwether Tickets

Staging is the optimistic case

Before the first real on-sale, Mateus asked what happens if a Tessera node dies mid-sale and whether it is still better on the invoice. Where it started: Closing the map →

Production. For the new-season on-sale, Ingrid sets up shadow traffic: every real hold request is duplicated to a Tessera cluster whose responses are discarded. This gives her the real fan retries and the real seat hot spots. She adds a canary of 1% of requests for the two minutes after the on-sale opens. She also notes the interference risk: both arms share the seat database, so she makes sure the Tessera arm uses a separate connection pool and she states the remaining shared parts in her report.

The dead node. She writes the hypothesis: "If one Tessera node is killed during peak load, hold-write p99 stays under 100 ms and the error rate under 0.1% after 30 seconds." The abort condition is a 2% error rate. The result: it takes 95 seconds, because a client library retries against the dead node before trying another. The failover that was a claim is now a measured number, and a fixable one.

The invoice. She reports dollars per million holds, and a frontier of p99 against monthly cost, for Engine A and for tuned Tessera, at the 1×, 3×, 10× and 30× scale.

The one-page decision. Mateus receives a single page. It says: with group commit enabled, Tessera improves on-sale p99 by 11% at the same cost, in the range 1 to 30 TB. This is below the 15% threshold that Ingrid wrote before testing, so the rule says "do not ship for this on-sale." It gives the mechanism (the flush path), the prediction that confirmed it, the limits (cold start, one instance type, compaction and failover under load still untested) and the next experiment. Mateus wanted yes or no. He gets "no, for now, and here is exactly what would change the answer". He finds that he can defend it to his own manager, because every sentence on the page is either a measurement with its configuration or an argument that is labelled as one.