<![CDATA[Remote Procedure Call (RPC) frameworks form the connective substrate of modern distributed systems, enabling coordination across microservices, storage engines, recommendation pipelines, search infrastructure, and machine learning inference services deployed at hyperscale. As distributed systems grow in scale and heterogeneity, average latency increasingly fails as a performance adequacy metric. User-facing reliability is governed by tail latency, particularly at the 95th, 99th, and 99.9th percentiles. A single slow RPC in a fan-out request can dominate end-to-end response time, making tail latency behavior a first-class engineering concern for systems operating at the scale of modern cloud infrastructure. This paper presents Congestion-Aware RPC Scheduling (CARS), a distributed scheduling framework that dynamically routes, prioritizes, delays, retries, and hedges RPCs based on real-time multi-layer congestion telemetry. Unlike conventional RPC load balancing, which relies primarily on endpoint health metrics and recent latency observations, CARS integrates congestion signals from endpoints, application queues, transport layers, and network paths into a unified congestion scoring function. The scheduling objective is not merely to identify healthy endpoints but to identify execution paths least likely to violate request-level tail-latency objectives under current congestion conditions. CARS introduces three main contributions: a unified congestion scoring model combining five categories of multi-layer telemetry signals into a single replica scoring function; a deadline-aware scheduling algorithm routing requests based on estimated tail-latency risk relative to remaining deadline budgets; and adaptive hedging and admission control mechanisms that suppress duplicate work when congestion cost exceeds expected latency benefit. Experimental evaluation using trace-driven simulation across microservice fan-out workloads demonstrates that CARS achieves approximately 43 percent improvement in 99th-percentile latency and 49 percent improvement in 99.9th-percentile latency relative to exponentially weighted moving average latency-based routing]]>
