6G paradigm enables massive network slicing for pervasive digitization across vertical industries, demanding scalable,sustainable, AI-driven zero-touch automation, particularly under non-IID conditions in live networks. This work introducesa cloud-native service-level agreement (SLA)-driven stochastic policy to guarantee a scalable and fast operation of constrainedfederated learning (FL)-based analytic engines (AE) that perform statistical slice-level resource provisioning at RAN-Edge domain, deploying on the kubernetes platform with real-time visualization capability. The key novelty lies in an SLA-aware adaptive learning rate policy to tune the local model’s learning rate dynamically during FL local training to enhance stability and convergence. Besides, to sustain scalability under massive slicing, we leveragesSLA-driven stochastic AE selection policy in each FL training round, maintaining computational efficiency while reducing computational overhead. Extensive experiments in both simulated and emulated environments demonstrate significant reductions in SLA violations, enhanced convergence stability, and improved scalability compared to conventional FL approaches.