HA is a complex topic.
If you have a cluster of nodes and one node becomes isolated, the remaining nodes holding the majority quorum can vote the isolated node out and take over services. This is normal PVE HA operation.
If a single node loses storage connectivity, for HA to act it would ideally need to determine that other nodes have a healthier access to that storage. PVE does not provide that cross-node storage-health service. It also can’t reliably determine if one node has better client network connectivity than others (assuming your cluster network is on a dedicated healthy link). Determining whether an application inside a VM is actually alive is also outside PVE’s purview.
Even if PVE could detect a single storage link issue, you then face trade-offs: would issuing an HA operation be better than riding out a 1–2 second link flap that the kernel’s IO retry would recover from? What if the failure is rolling or asymmetric: Node B looks fine now but may fail in 5 seconds, while Node A could recover faster?
The best place to handle application availability during underlying infrastructure outages is at the application layer. A load balancer can detect that an app on Node A is not responding but is OK on Node B, and redirect traffic more efficiently than anything inside PVE can.
Finally, if you have multiple switches and HBAs, deliberately reducing redundancy to a single path and then simulating a double failure by taking that path away is unusual. Your monitoring system should already notify you of a link down on the switch.
Regarding the watchdog, its normal use is to monitor whether the kernel is responsive. You can create a custom script, but be warned - there are many edge cases and potential unintended consequences.
Blockbridge : Ultra low latency all-NVME shared storage for Proxmox - https://www.blockbridge.com/proxmox
