Goal-Driven Resource Management for Microservices With Dual-Layer Deep Reinforcement Learning

Microservices must meet end-to-end latency SLOs while optimizing per-service resources, but heterogeneous request paths make mapping resources to latency difficult. We propose DRvalve, a dual-layer framework that decouples applicationlevel SLO feedback from service-level control via performance targets, driven by a novel objective integrating CPU utilization and throttling rates. A global deep RL controller periodically sets CPU throttling targets, and local allocators scale resources using robust rollback and self-healing mechanisms to mitigate DRL lag and suppress cascading oscillations. Moreover, median denoising and binary-tree buffers ensure stable training and O(log n) efficiency. Evaluations on three production-trace workloads show that DRvalve saves up to 29.55% CPU over the best baseline and up to 52.67% over all baselines while strictly maintaining SLO compliance.