Managed Instance Groups on Google Compute Engine have two separate health concepts, and conflating them is a common and consequential mistake. The load balancer health check decides where traffic goes. The autohealing health check decides whether an instance gets destroyed and recreated.

Both misconfigurations look identical in the console: a green health check. Install the Gremlin Agent on your instance template and you can tell them apart. With Gremlin, you can:

• Catch an autohealing check that's too shallow. A process listening on a port is enough to look healthy. Blackhole experiments break the application without stopping the process, revealing whether a broken instance stays in rotation indefinitely.
• Catch one that's too aggressive. CPU and memory experiments simulate an ordinary load spike so you can confirm your group doesn't destroy instances precisely when it needs them.
• Validate autoscaling thresholds. Testing shows where your scaling triggers sit relative to the load that actually degrades your response times.

It's a narrow question with an outsized effect, and far easier to answer deliberately than to infer from an incident where instances were cycling for reasons nobody could explain.