Fixed
- Two nodes that report the same hardware address both keep a device of their own. Proxmox names a port after its permanent address, and a board whose address lives in flash that was cloned hands out the same one twice — one node then overrides it, while the listing still shows the original. Home Assistant gives an address to one device per entry, so the second claim was refused and the refusal failed the whole setup: every version since 5.1.0 was unusable on such a cluster. Thank you, @vdaluz, for a report that had the diagnosis in it (#700).
- A read the proxy could not finish is asked once more before the entity goes unavailable.
596meanspveproxypassed the request on and no answer came back in time — a busy moment, not a refusal — and one of them used to take a sensor out until the next poll a minute later.595is left alone: that is a node that is switched off. Thank you, @litecross91, for chasing the reason down to an iSCSI timeout on your own side (#595). - A start with a node switched off no longer waits out that node's timeouts before the entry loads.
- The repair for a refused SMART read or ZFS pool names the path Proxmox really checks,
['perm','/',['Sys.Audit']]rather than the node. Granting it on the node listed the disks and was refused for everything behind them. - Retired cluster peers no longer stay in the fallback addresses for good — a membership answer that names nodes replaces them, while an empty or failed read leaves them alone. Thank you, @triantos (#705).
Changed
- The
updates_listattribute of the Total updates sensor names the packages (pve-manager - 9.0.6) instead of carrying apt's short description. It had been the description since 4.0.2, so a template matching on that text needs adjusting; the descriptions are still in the update entity's release notes. Thank you, @diegocjorge (#702).