Auto Scaling Groups — How They Work

Auto Scaling Groups maintain a fleet of EC2 instances or ECS tasks between a minimum and maximum count, replacing unhealthy instances and adjusting capacity in response to load — which means the fleet self-heals and grows or shrinks automatically.

How scaling works

• An ASG scales out (adds instances) when a CloudWatch metric crosses a threshold — CPU > 70%, request count per instance above a target — and scales in (removes instances) when load drops
• Target tracking scaling is the simplest policy: set a target metric value like CPUUtilization = 50% and the ASG adjusts the instance count to keep the metric near the target — you don't write scale-in/scale-out rules separately
• Scheduled scaling pre-warms capacity for known traffic patterns — add 20 instances at 8am on weekdays, remove them at 8pm — which is more reliable than reactive scaling for predictable load

Health checks and self-healing

• ASG performs EC2 health checks (is the instance running?) and optionally ELB health checks (is the application responding to HTTP health checks?)
• An instance that fails the health check is terminated and replaced automatically — which means a JVM crash or OOM that kills the process results in a new instance starting with no manual intervention
• Configure the ELB health check to call /actuator/health with a timeout that gives Spring Boot enough time to initialize before failing

Warm-up period — the Spring Boot cold start problem

• New EC2 instances need time to bootstrap the OS, start the JVM, and initialize Spring's ApplicationContext before serving traffic — configure the instance warm-up period to match your application's startup time
• Without a warm-up period, the ASG counts the new instance's metrics immediately, causing premature scale-in or misleading averages during startup

ECS Auto Scaling — ECS services have their own Application Auto Scaling separate from ASGs; the pattern is the same (target tracking on CPU or a custom metric) but operates at the task level, not the instance level.