Azure Service Fabric carries something most orchestrators don't: stateful services with data living on the nodes themselves. That changes what a node failure costs. Losing a stateless replica is an inconvenience. Losing a stateful primary before its state has replicated is a data problem.

Replication and failover are built in and they work, but their guarantees depend on replica counts and partition placement configured early and rarely revisited. A service running with fewer replicas than intended, or with replicas that landed on the same underlying node, has less protection than its configuration suggests.

Install the Gremlin Agent on your VM scale set nodes and run Shutdown experiments against a cluster carrying real state. You'll learn how long the failover transition takes, whether in-flight operations survive it, and whether clients reconnect without intervention. Process Killer experiments isolate a single service process so you can test failover without removing the whole node.

Those answers matter more here than in a stateless environment, and they're much harder to infer from documentation than to measure directly.