August 2022 Summaries

Kubernetes autoscaling adjusts infrastructure resources to meet demand, using vertical scaling to increase resources assigned to nodes or pods and horizontal scaling to add nodes or pod replicas. The walkthrough explains how to verify a Horizontal Pod Autoscaler (HPA) by preparing a namespace, confirming that the Kubernetes metrics server is running, deploying a sample PHP-Apache service with CPU requests and limits, and configuring it to scale from one to 15 replicas when CPU use exceeds 50 percent. Autoscaling can be tested directly with Kubernetes by running a BusyBox-based load generator and monitoring CPU targets and replica counts, although appropriate resource settings and metrics-server configuration are necessary for results. It also describes using Speedscale to capture generated or production traffic, create a snapshot, and replay that traffic against a deployment under configurable load conditions while reviewing request success rates, latency, and resource consumption. Speedscale’s replay approach can use realistic historical workloads, such as peak-event traffic, and can mock outbound dependencies to isolate a service during testing, though full-infrastructure testing remains optional.

Running ARM-based nodes alongside existing architectures in an Amazon EKS Kubernetes cluster can reduce production compute costs while preserving support for legacy services and utilities that cannot run on ARM. The approach involves creating an ARM-compatible node group, applying an `arch=arm` taint to prevent unintended scheduling, and updating compatible workload manifests with a matching toleration and an optional `kubernetes.io/arch: arm64` node selector to target ARM nodes explicitly. Kubernetes scheduling and autoscaling account for these constraints automatically, while workloads that support multiple architectures can omit the selector and use available capacity across node types. Although managing multiple node groups and building multi-architecture images adds operational overhead, the reported savings can be substantial, with Speedscale citing a roughly 32% reduction in EC2 costs and potential savings of up to 40% depending on workloads and instance types. The same Kubernetes concepts apply to other cloud providers, including Google Cloud and Azure, and can also support specialized node groups for GPUs or other workload requirements.