Autoscaling Stateful Workloads in Kubernetes (DoK Day EU 2022) // Mohammad Fahim Abrar & Md. Kamol Hasan

Data on Kubernetes Community

https://go.dok.community/slack
https://dok.community/
From the DoK Day EU 2022 (https://youtu.be/Xi-h4XNd5tE)
Managing stateful workloads in a containerized environment has always been a concern. However, as Kubernetes developed, the whole community worked hard to bring stateful workloads to meet the needs of their enterprise users. As a result, Kubernetes introduced StatefulSets which supports stateful workloads since Kubernetes version 1.9. Users of Kubernetes now can use stateful applications like databases, AI workloads, and big data. Kubernetes support for stateful workloads comes in the form of StatefulSets. And as we all know, Kubernetes lets us automate many administration tasks along with provisioning and scaling. Rather than manually allocating resources, we can generate automated procedures that save time, it lets us respond faster when peaks in demand, and reduce costs by scaling this down when resources are not required. So, it’s really important to capture autoscaling in terms of stateful workloads in Kubernetes for better fault tolerance, high availability, and cost management. There are still a few challenges regarding Autoscaling Stateful Workloads in Kubernetes. They are related to horizontal/vertical scaling and automating the scaling process. In Horizontal Scaling when we are scaling up the workloads, we need to make sure that the infant workloads join the existing workloads in terms of collaboration, integration, load-sharing, etc. And make sure that no data is lost, also the ongoing tasks have to be completed/transferred/aborted while scaling down the workloads. If the workloads are in primary-standby architecture, we need to make sure that scale-up or scale-down happens on standby workloads first, so that the failovers are minimized. While scaling down some workloads, we also need to ensure that the targeted workloads are excluded from the voting to prevent quorum loss. Similarly, while scaling up some workloads, we need to ensure that new workloads join the voting. When new resources are required, we have to make the tradeoff between vertical scaling and horizontal scaling. And when it comes to Automation, we have to determine how to generate resource (CPU/memory) recommendations for the workloads. Also, when to trigger the autoscaling? Let’s say, a group of workloads may need to be autoscaled together. For example, In sharded databases, each shard is represented by one StatefulSet. But, all the shards are treated similarly by the database operator. Each shard may have its own recommendations. So, we have to find a way to scale them with the same recommendations. Also, we need to determine what happens when an autoscaling operation fails and what will happen to the future recommendations after the failure? There can be some workloads that may need a managed restart. For example, in a database, secondary nodes may need to be restarted before the primary. In this case, how to do a managed restart while autoscaling? Also, we need to figure out what happens when the workloads are going through maintenance? We will try to answer some of those questions throughout our session. ----- Fahim is a Software Engineer, working at AppsCode Inc. He has been involved with Kubernetes project since 2018 and is very enthusiastic about Kubernetes and open source in general. ----- MD Kamol Hasan is a Professional Software Developer with expertise in Kubernetes and backend development in Go. One of the lead engineers of KubeDB and KubeVault projects. Competitive contest programmer participated in different national and international programming contests including ACM ICPC, NCPC, etc