Avoiding the Pitfalls of Autoscaling with Constant Work
summary
Claims that constant work beats reactive autoscaling for avoiding cascading failures, with concrete demos using Prometheus, Avalanche load generator, AWS ECS/Fargate and Route 53 patterns. A setup of four EC2 Avalanche targets and two Prometheus servers shows scaling one service from 8 to 20 containers overloads and stalls Prometheus; long startup/replay times (minutes) amplify cascades. Practical fixes include bounding per-instance requests, avoiding aggressive retries, keeping warm instances, or using Prometheus agent + remote_write to an external store to remove local storage startup pain.
AI-written summary · about
chapters (4)
related talks
21mAutoscaling and Progressive Delivery - A Match Made in Heaven
42mMaking Resilient, Self-Adaptive, And Autonomous Kubernetes Clusters
14mA Tale of Autoscale: Increasing Resiliency while Reducing On-Call Toil
40mThe Art of Embracing Failure at Scale
42mChaos Engineering: Why Breaking Things Should be Practiced
31mAutoscaling Apps without Supersizing Costs
18mScaling APIs to Billions: Extreme-Scale Reliability
30mAutoscaling with HashiCorp Nomad