This project turns a weekly request-rate pattern into a controlled experiment. A Go gateway and worker service expose latency metrics; reactive KEDA-style scaling is compared with linear regression and gradient-boosted tree forecasts, with guardrails that keep observed load as a safety floor.
The point is not that prediction wins everywhere. The local results show the benefit depends on forecast horizon, workload shape, calibration, and drift — exactly the conditions that make a latency/SLO experiment worth publishing.