Make Descheduler Smarter and Safer - Applying Reinforcement Learning in Descheduling Strategies

Explore how reinforcement learning can be applied to enhance descheduling strategies in Kubernetes clusters. Dive into a presentation that examines the limitations of current rule-based descheduler approaches and introduces a novel solution using reinforcement learning. Learn about defining reward functions and training agents to address resource fragmentation and workload hotspots. Discover the process of utilizing historical scheduling data and a scheduler simulator to train and validate the RL-based descheduler without disrupting production services. Gain insights into making descheduling smarter and safer for complex environments with diverse workload categories and multiple influencing factors.