DevOps Daily with Fexingo · 2026-07-21 · 10 min
Kubernetes Vertical Pod Autoscaler (VPA) is supposed to right-size container resource requests automatically. But in practice, VPA's recommender often suggests CPU and memory values that lead to pod evictions or wasted capacity. Drawing on real incidents from a mid-2026 production cluster running 50 microservices, Lucas and Luna unpack why VPA's default OOM-aware policy can misinterpret short-lived memory spikes, how its recommendation window of 8 days lags behind traffic patterns, and why the 'lower bound' mode sometimes recommends requests below actual usage. They walk through a specific case where VPA recommended 512 MiB of memory for a Go service that actually needed 1.2 GiB under peak load, causing repeated OOMKills. The episode closes with practical mitigations: setting custom OOM-scoring thresholds, using VPA in 'initial' mode for batch workloads, and combining VPA with Horizontal Pod Autoscaler using resource metrics. Listeners come away with a clear mental model of VPA's blind spots and how to compensate for them.
How Kubernetes Namespace Quotas Trigger Unexpected Failures58 / 100Why Your Kubernetes Cost Reports Are Lying To YouHow Kubernetes Node Pools Can Cost You More Than You ThinkHow Kubernetes Pod Overhead Changes Node Capacity CalculationsHow Kubernetes Pod Security Admission Blocks Risky Workloads