Choosing the EC2 launch type means you wanted control over the instances behind your tasks. That control comes with the responsibility for what happens when one of those instances fails, and ECS won't make that decision correctly on its own.

Task placement strategies, health check grace periods, and capacity provider settings interact in ways that look reasonable in the console and behave differently under pressure. A grace period tuned for a fast-starting service will mark a slow-starting one unhealthy and cycle it repeatedly. You typically discover this during a deploy that coincides with real traffic.

Regular testing surfaces these interactions while the stakes are low. It also gives you something more durable than confidence: a record of how your cluster behaved the last time an instance disappeared, and whether that behavior improved after you changed the configuration.