Blog

AI is climbing through performance engineering and observability the way autopilots climbed through aviation: eating the mechanical work first, then the analytical work, and now reaching for the interpretive work. This essay proposes a durable reading grid — an autonomy ladder transposed from autonomous driving — to locate any tool, present or future, and to ask the only question that stays stable over time: as the machine climbs each rung, where does the human move, and what must the human still know how to do?

Most performance-testing programmes have a load generator. Few have a representative dataset. Almost none can inject failures while measuring user-perceived latency. The reasons are multi-causal — a culture that treats performance as a release-time formality, plans that under-budget the supporting work, applications whose testability was never designed in, and a tool catalogue organised by category rather than by intent. This article reframes performance testing around eight use cases — API load testing in CI/CD, full-stack validation, microservice resilience, database benchmarking, frontend optimisation, capacity planning, endurance and resource-leak detection, and pre-production data realism — and uses them as the spine of a practical campaign-setup guide: what each test is trying to prove, what testability hooks it requires, what it realistically costs, what cultural pre-requisites it has, and which combination of tools assembles it.

Tool taxonomies organise observability by metrics, traces, logs, and profiles. Practitioners organise it by intent: what am I trying to understand, debug, or prove? This article reframes the observability stack around six common intents — Golden Signals, latency propagation, high-cardinality debugging, low-overhead profiling, black-box, and cost-efficient at scale — with the workflows, the right tool combinations, the anti-patterns to avoid, and a dedicated treatment of how unified APM platforms (Datadog, New Relic, Dynatrace) fit in the intent-routing framing.

Most HTTP benchmarking tools quietly hide tail latency when the server slows down. The phenomenon is called coordinated omission, it shows up almost exclusively in p99 and beyond, and it has caused production incidents at organisations that thought their load tests were green. This post explains the mechanism, demonstrates it empirically with a reproducible benchmark of eight tools across a healthy control and four server pathologies, and shows how to fix it with a constant arrival-rate workload model.