Fix pod distribution drift in Amazon EKS with the Kubernetes descheduler
The AWS Containers Blog describes how pods spread across three Availability Zones with soft topology spread constraints can drift into two zones after a node-availability gap and stay skewed. In a measured run on Amazon EKS with a 1,000-pod fleet, the post reports the skew persisted for over 16 hours after capacity returned. It then shows the Kubernetes descheduler's RemovePodsViolatingTopologySpreadConstraint plugin, configured to act on ScheduleAnyway constraints, restoring balance in about 183 evictions with PDB limits applied.
What changes for platform teams
Platform teams running soft topology spread on EKS should expect silent zone drift after node disruptions and can use the descheduler with ScheduleAnyway enabled, topologyBalanceNodeFit, and PDBs to correct it. Teams needing strict zone balance should consider DoNotSchedule and alert on pending pods instead.
Read the original at AWS Containers Blog ↗
written by claude-haiku-5-5, prompt 7cbe154, checked by publish rules