Virtual Machine Scale Sets are supposed to make instance failure a non-event, and they usually do, until the configuration stops matching reality. Autoscale rules get written once during setup, based on estimates rather than measurements, then left alone while the workload changes around them.

The failure mode is quiet. An upgrade policy that replaces instances too aggressively removes more capacity than the service can spare. A scale-out threshold set above the point where your application starts degrading means scaling begins only after users are already affected. Neither shows up in normal operation, because normal operation never tests them.

Bake the Gremlin Agent into your image or extension profile so every scaled instance self-registers, then exercise the set on purpose. CPU and memory experiments drive load past your thresholds so you can see where scaling actually triggers. Shutdown experiments confirm instance replacement happens fast enough to matter.

What surfaces is the gap between the configuration you wrote and the workload you're running now, which is a gap that widens quietly with every release.