Scaling your object store is complex and payloads vary in size - objects can be as large as virtual machine images or as small as emails. In behaviour - some are mostly reading, writing, and listing objects. Other payloads delete objects, and some keep them forever. Using CPU and RAM to autoscale the pods horizontally or vertically is limited and may adversely affect the system. Treating our object store as a queueing system: converting HTTP requests to actions on disks may be the solution! Please note that this session was originally scheduled for 18:30.
In this talk, we would show how we use KEDA and Prometheus to track the backlog of work in an object store system orchestrated by Rook and to autoscale the object store pods based on different metrics. We show how autoscaling is impacting the performance and resource utilization in the environment by taking Ceph’s object store frontend (the RADOS Gateway daemon) as a use case. We will demonstrate how to use KEDA and Prometheus to solve a scaling problem that cannot be quickly resolved by the traditional HPA (horizontal pod autoscaling). We will show how the Rook operator orchestrates the different parts of the system. Attendees will learn how to use cloud-native tools to perform autoscaling, better utilize resources like CPU and I/O and increase the performance of their solutions.
Speakers: Jiffin Tony Thottan