when to use a service mesh versus direct client libraries for service communication?

Hey folks, I'm trying to figure out the best way for our services to talk to each other on this next microservices project. We could build our own client libraries with retry logic and service discovery built-in, or use a service mesh like Istio or Linkerd. A mesh gives us observability and traffic management stuff right away, but wow, it seems like a lot of extra complexity. When did you guys find a service mesh was actually needed, instead of just putting that logic in the clients? What made you choose one way or the other?