Targets that fail the right way, not the wrong way.

The defaults that came with the target group had been "what the wizard suggested when we set this up." They had not been revisited as the application matured. Unhealthy-threshold of 5 meant 50 seconds before a misbehaving target got pulled; healthy-threshold of 5 meant 50 more seconds before a recovered target rejoined.

We tuned: unhealthy-threshold to 2 with a 5-second interval (10-second detection), healthy-threshold to 2 (10-second rejoin), connection draining to 15 seconds. Slow-target detection added a CloudWatch alarm on response-time p99 that flagged slow-but-healthy targets for restart via ECS deployment circuit breaker.

Deploy-window incidents dropped 92% year-over-year. The remaining 8% are now caught faster because slow targets get detected before they accumulate enough customer impact to be noticed. Drain time fell to 15 seconds; healthy-rejoin to 10 seconds. The customer experience during deploys is finally what the team always claimed it was.