Early Backpressure Detection for High-Throughput Distributed Systems: A Latency Gradient Based Approach Using Dual EWMA and Queue Occupancy

This paper presents a lightweight, zero-allocation backpressure detection algorithm for high-throughput distributed message pipelines. It combines two signals: a latency gradient computed via dual EWMA (short-term and long-term exponentially weighted moving averages), and ring buffer occupancy. Unlike traditional queue-occupancy-only detectors, which react late, the latency gradient flags degradation before the queue fills — as validated experimentally, detecting congestion ~200ms before full queue saturation. The design draws from TCP congestion control (Vegas, gradient-based load management) adapted for application-level queues. It runs in constant O(1) time per event, with ~500ns overhead measured via JMH microbenchmarks, and includes hysteresis to prevent state flapping plus an AIMD-based advisory rate control. Full reference implementation and evaluation harness available on GitHub.