É o comportamento esperado. O comportamento das sondas para baixo é como o fole:
New TCP connections will succeed to backend instance which is healthy and has a guest OS and application able to accept a new flow.
If a backend instance's health probe fails, established TCP connections to this backend instance continue.
If all probes for all instances in a backend pool fail, no new flows will be sent to the backend pool. Standard Load Balancer will permit established TCP flows to continue. Basic Load Balancer will terminate all existing TCP flows to the backend pool.
Because the flow is always between the client and the VM's guest OS, a pool with all probes down will cause a frontend to not respond to TCP connection open attempts as there is no healthy backend instance to receive the flow.
Referências: sondas de integridade do Load Balancer