Kubernetes has become the de facto platform for deploying machine learning (ML) workloads, but its default scheduler was designed for long-running web services—not batch jobs, distributed training, or GPU-heavy tasks. As a result, teams running ML workloads often face poor cluster utilization, job starvation, and costly inefficiencies—all because the scheduler isn’t optimized for their needs.

In this talk, we’ll explore:
1. Why the default kube-scheduler falls short for ML workloads (e.g., no gang scheduling, poor spot instance handling).
2. How specialized schedulers (like Apache Yunikorn, Kube-batch, or Volcano) solve these problems with features like topology-aware placement, fairness queues, and elastic scaling.
3. Real-world case studies of teams that improved job completion times and reduced costs by switching schedulers.
4. A decision framework for choosing the right scheduler based on workload requirements.