Every Android team has lived this story. A screen feels sluggish, someone attaches a profiler on a well-charged flagship, the numbers look fine, and the ticket is closed as “cannot reproduce”. Two weeks later the p95 for the same screen is still ugly in your analytics, and nobody can say why.

The gap is not a lack of tools. It is that each tool answers a different question, and most teams use only the first one. This post walks through a four-stage loop that connects them, using a fictional notes app called Marigold as the running example, and spends real time on the places where each stage breaks.

Why a single profiler is never enough

A local profile is deep but narrow: one device, one build, one data set, one person deciding what to click. A CI benchmark is repeatable but synthetic. A field capture is honest but expensive and sparse. The mistake is treating any one of them as ground truth.

The more useful framing is a staircase.…