Often Google are right, but sometimes not.
Recently I was looking at why GCP TCP Load Balancer healthchecks were failing, despite I have configured strictly my local OS firewall to allow the documented health check IP ranges :
(https://cloud.google.com/compute/docs/load-balancing/network/ – Section Firewall rules and Network load balancing –
126.96.36.199/16 at the time of writing).
Network Services–>Load Balancing–>Load Balancers tab was stating my instances were unhealthy:
Instance <instance name> is unhealthy for <LBIP>
It turned out Google is sending healthcheck from their metadata IP address instead – metadata.google.internal (169.254.169.254)
Allowing this in the OS firewall made the HCs succeed!