Ceph Umbrella is already in Proxmox, but not for your cluster

For a few days now, Proxmox's Ceph installation wizard has had a new entry in its dropdown: Umbrella, version 21.2. If you administer a cluster, you will see it this week. And the decision in front of you is not set by any end-of-life date, however much that is the only thing anyone talks about: it is set by a version number, 20.2.4, which appears twice, arriving from two directions that never meet.

We run Proxmox VE clusters with Ceph in production, and from staring at this calendar long enough we have ended up with a rule of thumb: the date is the first thing everyone looks at and the last thing that decides anything. What decides sits two layers below, in version numbers that make no headlines. This week both moved, and neither was announced.

Twelve days

The commit that moved two dates and announced neither

The file doc/releases/releases.yml in the Ceph repository is what feeds the active-releases table and the lifecycle chart. On 4 November two lines changed:

Read that again. A release published seven days earlier, one that everybody is advised to upgrade to, lost almost five months of planned life in one go. And Squid, officially already dead, was handed 32 more days. The commit subject is "doc/releases: update EOL dates", and there is no body underneath it: just the sign-off. No release note, no blog post, no thread.

What we measured in the repository

Release notes describe intentions. The repository describes what is actually there. This is what download.proxmox.com returned on the morning of 9 November 2026, looking at the trixie indexes:

Seventy-seven packages against four hundred and eighty-six. And, above all, a single version. What matters is not that 21.2.0-pve2 does not exist yet — it will — but that there is nothing behind it: ceph-squid carries eleven versions and ceph-tentacle nine, so if one goes badly you can pin the previous one. In ceph-umbrella there is no previous one. All three repositories, moreover, expose exactly the same two components: no-subscription and test. The word "preview" appears in the wizard, in the wiki and in the forum thread title; where it does not appear is in the repository metadata, which is what your apt actually reads.

You can repeat the whole measurement without touching your cluster — two HTTP requests:

And a detail that falls out of comparing two clocks: that ceph-umbrella index was built at 00:44:31 UTC on the Thursday, and the forum post announcing it is stamped 01:48:52+01:00, i.e. 00:48:52 UTC. Four minutes. The repository was available before anyone said so, which is exactly the order in which your apt finds out about things.

The hinge is called 20.2.4

Here is the part no headline covers. The same number closes both doors, and it does so from two sources that are not talking to each other.

It is worth not turning that pve5 into a metaphor: it does not mean "five rebuilds is what maturity costs". It means something specific, and the wiki says so in its known issues: OSDs that ran Tentacle up to 20.2.4-pve4 can see their META usage grow, which is why the floor sits at pve5. In fact 20.2.5-pve1 also meets the requirement, so the rebuild number is not the criterion: the criterion is the point at which one concrete thing got fixed.

Incidentally, 20.2.4 is the 19 August release, the one that closed four authentication and authorisation CVEs: two signature-verification flaws in RGW (STS tokens and the SigV4 verifier), an authentication bypass in CephX caused by AES-CBC misuse, and an unauthorised read of the monitor's config-key store. If the key-type mess rings a bell, it is the same thread we pulled on when we covered the cephx key migration: when Ceph touches authentication cryptography, the bill does not arrive on announcement day — it arrives in your next maintenance window.

An honesty that forces us to soften our own rule: that 9.2.25 is not carved in stone either. The raw wikitext of that same page carries two comments the browser never shows you. One reads "TODO: set the first pve-manager version that knows about the Umbrella release and its repository" — the number is a placeholder awaiting a decision. The other: "TODO: drop the next sentence once the required pve-manager version above is 9.2.23 or newer". And the forum says the wizard has offered Umbrella since 9.2.23, two revisions earlier. So requirements move too; what does not move is that there is always one underneath, and that nobody is watching it.

The announcement is more cautious than its headline

An abbreviated version of the announcement is doing the rounds: that Proxmox recommends moving Squid clusters "sooner rather than later", to Tentacle or to Umbrella. The full sentence says something else: "You should upgrade any existing Ceph Squid setups sooner rather than later, either to Ceph Tentacle or, once it is fully supported, to Ceph Umbrella." And it then links a single guide: Squid to Tentacle. The caveat was there from the start; it is the summary that loses it.

The same thread declares some mileage, and it is only fair to quote it: "We have been running and testing the Umbrella release candidates and this release internally for a few weeks and found no major issues." A few weeks of internal testing is not the same as a branch with nine published versions behind it, but it is not nothing either. Does that mean you cannot go from Squid to Umbrella directly? You can, and the wiki covers it explicitly — it is worth reading in full before deciding: "While it is possible to upgrade from the older Ceph Squid (19.2+) to Umbrella (21.2+) release directly, we primarily test and recommend upgrading to Ceph Tentacle first." And it adds: "If you want to skip one upgrade, we recommend testing this first on a non-production setup." The path is not missing: it is handed to you labelled as the one they do not test. There is even a different operational detail, because coming from Squid the cluster must report require_osd_release squid rather than tentacle, "otherwise, the Umbrella OSDs will refuse to start".

When not to — and that means now

The very Proxmox wiki page that explains how to jump to Umbrella opens with the sentence that settles it: "Mature tech preview, not yet available in the enterprise repository." In practice: if you pay for a subscription and feed from the enterprise repository, installing Umbrella means pointing your cluster's storage layer at no-subscription or test. For the system holding the disks of every virtual machine you run, that is not an upgrade: it is a change of support regime.

And there are things Umbrella changes underneath you, whether or not you touch anything:

What we would do this month

If you are still on Squid, the date that matters is neither 31 October nor 2 December: it is that the Proxmox forum says Squid has "only one last update currently in progress". Once that ships, the safety net is gone. Move to Tentacle and land on 20.2.4 or later — the repository is serving 20.2.5-pve1 today — which conveniently leaves you on the step from which you will one day jump to Umbrella without touching anything again.

If you are already on Tentacle, our recommendation is boring: do not touch Umbrella yet. Tentacle is supported until June 2027 "for the time being" — that qualifier is theirs, not ours — and that is a good seven months in which to let the new branch accumulate rebuilds. Meanwhile, the three checks we run that change nothing: ceph versions to see what each daemon is running, apt-cache policy ceph-common to see which repository you are feeding from, and pveversion --verbose to know whether you reach 9.2.25 and where Ceph stands.

And when the day comes, the wiki's order is not decorative: monitors, managers, OSDs and finally the MDS, with noout set for the window and restarting OSDs one node at a time — they put it in bold, "to avoid loss of data redundancy" — and ceph osd require-osd-release umbrella at the end, not halfway through. One step is missing from every summary and follows directly from those PG numbers: you must pause the autoscaler with ceph osd pool set noautoscale before upgrading the packages, or the cluster will start splitting PGs while daemons of different versions coexist. The wiki explains why this falls to you, and settles the CPU question along the way: "Ceph pauses the autoscaler automatically only for upgrades managed by cephadm, which Proxmox VE does not use." Note too that, unlike noout, this flag "does not show up as a health warning": if you leave it set, nobody will remind you.

What we would not do

An installation wizard with a dropdown is an interface without memory: it offers you all three branches looking the same, and none of them tells you how many rebuilds it carries. That part is on you, and it lives in two places: in ceph versions, and in the IT maintenance that makes sure somebody looks at that number before an incident does.

Sources (verified 9 November 2026): dates and versions in doc/releases/releases.yml and commit 18c8b132 (4 Nov 2026) — github.com/ceph/ceph; 21.2.0 Umbrella release notes (28 Oct 2026) — ceph.io; the upgrade warning from Squid and the Tentacle releases — docs.ceph.com; jump requirements and window order — Proxmox VE wiki; package counts, versions and index dates, measured by us against download.proxmox.com.