Tech & AI

Split, shed, then scale — and why the intuitive order takes the database down.

A restore drill that corrected its own premise and made RPO explicit for the first time.

What changed, and what still needs an engineer

"Models are getting better, so the orchestration layer will disappear" is the same claim as "CPUs got faster, so the operating system will disappear." Faster CPUs did kill memory overlays and hand-rolled DMA — but process management, permissions and drivers got more important, not less. Orchestration isn't dying either; it's redistributing into three places. With a testable criterion for which is which.

A working harness that refactors SRE oncall triage from a human-executed procedure into an agent that investigates on its own while every irreversible action stays behind a deterministic gate. Built on four control primitives — Spec / Loop / Hook / Fork — with the reliability discipline of a Kubernetes control plane wrapped around a non-deterministic reasoning core.

Integer tail latencies are human constants: a periodic spike traced to a cache timeout hiding inside a rate-limiter plugin.

The radar's dashed target line is a commitment, not decoration — three prioritized gaps, each with a testable definition of closed.

Why the radar has nine axes — a capability model built on problem classes, not tool inventories, and where the moat actually is.

A Kafka consumer-lag alert that was really a terabyte of upgrade residue inside ClickHouse, detonated by a one-second background merge

"Familiar with Jenkins" is in every SRE job description, but writing pipeline scripts is not the skill. The skill is treating the delivery pipeline as a production system: idempotent procedures, observable job state, configuration in code — and knowing when automation itself becomes the new source of entropy. Grounded in a real Jenkins migration and its diagnose/fix script pairs.