#13: Multidimensional Pod Autoscaling & Machine Learning for Cloud Optimization

Dr. Haoran Qiu, a fresh PhD from the University of Illinois Urbana-Champaign, joins our host James Wilson, VP of Engineering at nOps. They’re diving into multidimensional autoscaling, an area in which Haoran’s pioneering research is making waves in the Kubernetes community. 

Some workloads work better with Horizontal Pod Autoscaler (HPA), others with Vertical Pod Autoscaler (VPA).  Running them together can create conflicts, but using only one limits efficiency gains. A Multidimensional Pod Autoscaler solves this dilemma by combining the benefits of both VPA and HPA to dynamically adjust both the number and size of pods.

But is MPA poised to redefine resource optimization? What problems does it solve, and what fresh complexities are involved in its implementation?

Haoran and James dig into these questions while debating traditional heuristic versus Machine Learning approaches, industry versus academia, and other hot topics in Kubernetes.

Listen now to discover if MPA is the holy grail of cloud optimization as we discuss the evolution of autoscaling technologies and their impact on cost, sustainability, and developer experience.

Chapters:

0:00 - 2:20: Haoran Chu and the state of cloud resource management

2:20-6:00: Historical evolution of autoscaling

6:01 - 10:45: HPA, VPA and Multidimensional Autoscaling

10:46 - 18:50: Challenges of MPA: heuristics versus machine learning

18:51 - 24:20: How to quantify excess capacity?

24:21 - 32:16: The state of ML in autoscaling

32:16 - 37:37: Operationalizing ML in production environments
37:37 - 42:01: The near-term future of autoscaling