1 Year of Autoscaling Postgres: How it’s going, and what’s next

Neon has been operating its autoscaling feature for Postgres serverless instances for over a year, and it has seen significant adoption with almost a third of customers using it. The feature performs zero-downtime vertical scaling, provisioning extra CPU and memory as needed to handle peak demand without incuring peak infrastructure costs. Neon's architecture uses virtual machines in Kubernetes clusters, which provides strong isolation boundaries, dynamic resource adjustment, and seamless live migration. Despite some challenges, such as issues with cgroup signals and memory scaling, Neon has continued to prioritize stability and improve its autoscaling algorithm. Future improvements include expanding the set of workloads that autoscaling is a good fit for, smarter autoscaling using local file cache metrics, accommodating rapid memory allocation with swap, and enhancing observability into swap usage.