dougiteixeira/proxmoxve 5.3.2 on GitHub

Fixed

A node that is switched off is no longer mistaken for the host that is gone. Every path names a node, and the configured host forwards what is not its own — so the reads for a switched-off node's guests, storages, tasks and updates fail on a host that is perfectly well, with 595 No route to host after up to 25 seconds or a dropped connection. Taken as "this host stopped answering", the integration moved to another node, repeated the same impossible read there and moved again; on a cluster of two it moved to the node that was actually down, which is what 5.3.1 did not fix. The host is now asked version before the integration moves, and kept when it answers — a refusal included, because a host that says 401 is a host that is there. Thank you, @mundschenk-at, for the report and for testing (#692).
Reads for a node the cluster itself reports as offline are not made at all, so they no longer wait for a timeout once a minute per guest and storage. The log says Node X is offline instead.
Error fetching … data: said nothing: the refused-resource path raised without a message. It carries the reason now, which is what makes a report like #595 answerable in the first place.

When the configured host is down in the docs describes what the integration asks before it moves, with a row in the comparison against core.