FunPilot: Runtime Performance Diagnosis and Remediation for Serverless Applications with LLMs

Abstract

In serverless computing, optimizing performance and cost is challenging due to the large configuration space in the presence of time-varying workloads and rapidly scaling infrastructure. Human-driven optimization is fundamentally limited in this setting: it is too slow for the short control timescales of serverless platforms, prohibitively expensive at scale, and unable to continuously explore high-dimensional configuration spaces in real time. Existing diagnostic agents are designed for long-running services and do not adapt well to the unique characteristics of serverless, such as short timescales, cost efficiency, and ephemeral runtimes. We present FunPilot, a system that enables rapid LLM-assisted diagnosis and remediation for serverless applications. FunPilot uses an event-driven control loop to diagnose active symptoms, derive control knob updates, and validate remediation decisions while coordinating with the underlying autoscaler. FunPilot completes diagnosis in 6.04 s at a cost of $0.0082 per diagnosis on average, fitting within a single metrics collection interval while incurring modest cost relative to the applications it manages. Starting from the initial application configuration, FunPilot reduces average alert active time by 70.9% and resource consumption by 38.3%. Across real-world workloads and synthetic workloads with injected anomalies, firing alerts resolve in 80.57 s on average after the FunPilot diagnosis engine is triggered.