Release state
Replaced #
This was the Latest release until 2026.11.4 replaced it. Before it became Latest, we updated our own Moorhold server to it from 2026.11.2, and it passed a restore drill there. Its files stay on get.moorhold.dev and still work. If your server runs it, update to the Latest release, 2026.11.4.
We wrote this release's notes, including their Still to be tested line, before we ran it on our own server.
This release lets Moorhold run on smaller servers and moves updates onto the service that already does what an admin approves. The installer now accepts a 12 GB server for the full profile and a 6 GB one for core, and 4 GB for light in local mode, figures taken from measuring what each profile uses. Updates approved in the portal, and automatic updates, are installed by moorhold-requests.service once a server runs this release. Servers tell search engines not to index the portal or any published tool unless you turn indexing on. Installs and updates wait and try again when the download server is busy. Starting Moorhold, for example after a restart, no longer stops part-way on two jobs that collided. Adding the lake to a server, and the sample dashboard, the first-run page and the Lineage menu, get fixes.
https://moorhold.dev/releases/2026.11.3Before you update #
- What changes
- The installer accepts smaller servers, sized from measuring what each profile uses: 12 GB for full and 6 GB for core, and 4 GB for light in local mode. Updates approved in the portal, and automatic updates, are now installed by the same service as the other things an admin approves. Servers tell search engines not to index the portal or published tools. Installs and updates wait and try again when the download server is busy, starting Moorhold after a restart no longer stops part-way, and adding the lake to a server gets two fixes.
- Downtime
- Moorhold is unavailable from when the update stops the running release until the new one has started.
- Backup
- Yes, on a server with backups on, which is the default: before it changes your Stack or the running release, the update takes a rollback point, a backup of every database on this server apart from the copies an earlier rollback kept. If the new release doesn't start, it goes back to that point by itself. If you've turned backups off, the update takes no rollback point.
- Manual steps
- None for a server running 2026.11.2. If you're updating from an earlier release, read 2026.11.2's Updating first, and the Updating of any release between.
- Still to be tested
- Owed: the first update installed by moorhold-requests.service, which can only be the update from this release to the next one; automatic updates on a server, and a moorhold-update.timer masked before the update; many restarts in a row of an updated full server; an install of this release on a 12 GB server with the full profile or a 6 GB one with core, and the light profile in 4 GB; adding the lake to an installed server; the wait for a busy download server, on a server and in the Moorhold app for Mac; the Lineage menu on a Stack without DataHub; and the sample dashboard's Month column in the Moorhold app for Mac. Also still owed from 2026.11.2: the Updates page offering a release once it's the latest, the update window on a whole update, the message behind a proxy that refuses get.moorhold.dev, the updates from releases before 2026.10.16, a fresh install with the sample data, a daily backup on a server that holds kept copies, and a restore drill with a kept copy of a long name.
Updating #
On a server running 2026.11.2, update from the server's command line with:
sudo moorhold upgrade 2026.11.3
It checks the release and the server before it changes anything, and takes the same rollback point as the installer. Once this release is the latest one, an admin can also press "Check for updates" on the Updates page in the portal and ask for the update there. On a server running 2026.11.2, that update is installed by 2026.11.2's moorhold-update.service, as before. Once your server runs this release, updates approved in the portal are installed by moorhold-requests.service (see New).
You can instead download this release's installer and run it on the server as you did to install, but without --hostname:
curl -fsSLo install.sh https://get.moorhold.dev/2026.11.3/install.sh sudo bash install.sh
The installer sees the release that's running and updates it in place, with the same checks and the same rollback point. On a server that already runs this release, it says there's nothing to update and changes nothing.
If you're updating from a release before 2026.11.2, read 2026.11.2's Updating as well, and the Updating of any release between.
New #
- The installer now accepts smaller servers, sized from measuring what each profile really uses: 12 GB of memory for the full profile, and 6 GB for core (8 GB recommended). Before, it asked for 32 GB for full and 16 GB for core. Local mode's light profile, for the Moorhold app for Mac, now needs 3.5 GiB, which its 4 GB virtual machine has.
- The portal and every published tool now tell search engines not to index them, with an X-Robots-Tag header on every response. Their robots.txt is meant to say the same, but in this release it comes out as one line: see Still known. You can turn indexing on in the Stack if you want the site listed.
- Once a server runs this release, updates approved in the portal, and automatic updates, are installed by moorhold-requests.service, the service that already does the other things an admin approves, such as giving a tool more memory. The Updates page now reaches the server through the same folder as those requests, /var/lib/moorhold/host-requests. While an update installs, other approved requests wait until it has finished. The timer for automatic updates keeps its name, moorhold-update.timer, and now starts moorhold-requests.service. A timer you masked stays masked. You can still install an approved update by hand with
sudo moorhold upgrade --approved. On a server updated from an earlier release, moorhold-update.path and moorhold-update.service are still in place, because a rollback to that release needs them. Until the first update run on this release, moorhold-update.timer can still start the old service once. That runs this release'ssudo moorhold upgrade --approved, which installs as normal and then points the timer at moorhold-requests.service. The update after this one removes the old path and service, and says so. A server installed with this release never has them.
Changed #
- A comment in your Stack about search engines. The update adds two comment lines under the
publishline of your Stack'sspec.public, saying how to let search engines index the server. It adds nothing if your Stack already setsindex, or has nopublishline. - The cloud-init files name smaller servers. This release's cloud-init files suggest smaller servers for the full profile: on IONOS, 4 cores and 12 GB of memory, and on Hetzner, a CX43 with 16 GB. For the core profile they suggest an 8 GB server.
Fixed #
- An install or update no longer stops when the download server is busy. It says so, waits and tries again, up to four more times over about three minutes, and if the server is still busy it tells you plainly that trying again later should work. The Mac app's window shows the wait and the final message too.
- Just after you add the sample data from the first-run page on a server, the page now says it's waiting for Apache Airflow to pick up the load job and checks again by itself, for up to five minutes, rather than saying straight away that Airflow hasn't picked it up.
- If you add SeaweedFS and Polaris to a Moorhold Stack, the installer's first start no longer fails on the lake backup job. That job now only runs when you ask for it with
moorhold backup --lake. - On a Stack with no DataHub, the portal no longer shows a Lineage item in its menu. If you open the Lineage page by its address, it says lineage isn't shown for this Stack yet and points you to Assets, instead of telling you to add a component you can't add from the portal.
- If your Moorhold server was installed before the lake, adding the lake no longer leaves Polaris without a database. A short job now creates the Polaris login and database when they're missing, and leaves them alone when they're there.
- On the Mac app, the sample dashboard's "Top Hacker News stories" chart now shows its Month column, instead of failing with a "function date_format does not exist" error. Its introduction text now says the data is queried through Postgres there, not Trino.
- Starting Moorhold on a server, for example after a restart, could stop part-way because warehouse-grants or warehouse-semantic-login failed with "tuple concurrently updated", and the services that wait for them, such as Airflow's scheduler, didn't start. Each of the two jobs gives a login access to the warehouse database as Moorhold starts, and they could run at the same moment. Now warehouse-semantic-login waits for warehouse-grants to finish, and each gives that access only when the login doesn't have it yet.
Verified #
- Unit and golden tests: pass, on the check run of the commit this release's images were built from.
- Images: each of Moorhold's 12 images was built for x86_64 and for ARM and published as one image that holds both. Every image passed the licence check on both platforms, and every image our test server runs is one the manifest names.
- On our test server: a full server on 2026.11.2 took a rollback point first and updated to 2026.11.3 with all 26 services healthy, and the restore drill passed after it. The server kept its data and settings. After the update, moorhold-requests.path was enabled and active, and the portal had one folder for requests, with no separate one for updates. The old moorhold-update.path was still enabled and moorhold-update.service still in place, and that service didn't start after the update. The portal and Superset sent X-Robots-Tag: noindex, nofollow.
- Updates on our test server: on 2026.11.3, pressing "Check for updates" on the Updates page started moorhold-requests.service, whose update step said "No approved update is waiting.", and moorhold-update.service didn't start. After a rollback to 2026.11.2, an update to 2026.11.3 approved on the Updates page was installed by 2026.11.2's moorhold-update.service, which said "Updated to 2026.11.3." The restore drill passed after that update too, and all 26 services were healthy.
- A restart on our test server: after a reboot on 2026.11.3, Moorhold started at boot and ran to "Moorhold is running" without trying again. warehouse-grants finished without an error before warehouse-semantic-login started, and all 26 services were healthy.
- In the code and its tests: the installer's memory figures for each profile and the checks before install, updates installed by the requests service, the doorbell the Updates page and the memory page share, the automatic updates timer and a masked timer, the old update units left in place on an updated server, a rollback waking the old update service, the noindex header and the setting that turns indexing on, the comment the update adds to a Stack, the retries when the download server is busy, the first-run page's wait for Airflow, the lake backup job left out of the first start, the Polaris database job, the Lineage menu, the sample dashboard's Month column on Postgres, and the order of the two start jobs and their grants.
- Owed: the first update installed by moorhold-requests.service, which can only be the update from this release to the next one. Automatic updates on a server, and a moorhold-update.timer that was masked before the update. Many restarts in a row of an updated full server. An install of this release on a 12 GB server with the full profile or a 6 GB one with core, and the light profile in 4 GB. Adding the lake to an installed server. The wait for a busy download server, on a server and in the Moorhold app for Mac. The Lineage menu on a Stack without DataHub, and the sample dashboard's Month column in the Moorhold app for Mac. Also still owed from 2026.11.2: the Updates page offering a release once it's the latest, the update window on a whole update, and the message behind a proxy that refuses get.moorhold.dev. So are the updates from releases before 2026.10.16, a fresh install with the sample data, a daily backup on a server that holds kept copies, and a restore drill on a server that holds a kept copy with a name of more than 47 bytes.
Images #
Every one of Moorhold's own images is rebuilt from this release's source, for x86_64 and ARM. No tool changes version, and the images Moorhold takes from upstream keep the digests they had in 2026.11.2.
Still known #
- A second admin for updates and Proposals. One admin is enough for the sample data, for more memory and for the mail password set in the portal. On a server with two or more admins, an update asked for on the Updates page still waits for another admin to approve it, and a Proposal is still approved by someone other than the person who made it.
- An approval while an update runs by hand. If an admin approves an update in the portal while someone runs
sudo moorhold upgradeon the server, the approved update can wait until the next daily update check. Runningsudo moorhold upgrade --approvedinstalls it sooner. - Many people behind one address. The server limits how often one address can load the first-run pages. If many people reach Moorhold from one address, or someone keeps mistyping a password, an open first-run page can still be refused and show a blank page. Wait a minute, then reload it.
- The audit page and one admin's request. The audit log stores "Asked for and approved by" and the admin's name for a sample data or memory request, but the portal's audit page doesn't show those words. On the page, an ask for the sample data is a
setup.sample-askrow, and the server's work on it is two rows namedrequests.add-sample-data, one when it starts and one when it's done, with nothing on the page to tell them apart. - A source that runs the sample's own code. Removing the sample also removes its code, so while the sample is removed, a source that runs the sample's code is refused until the sample is added back. We've seen this only with a test source on our test server.
- robots.txt comes out as one line. The portal and published tools serve robots.txt as a single line,
User-agent: *\nDisallow: /\n, with each\nsent as the two characters, soDisallow: /isn't on a line of its own. The X-Robots-Tag header that tells search engines not to index them is right. 2026.11.4 fixes robots.txt. - ARM servers. Moorhold servers still need an x86_64 (amd64) processor. ARM is for the Moorhold app for Mac only. An install on a real ARM machine is owed.
- The Moorhold app for Mac is early access. A first setup from its disk image and an update of the app and the stack over an install are owed on a Mac.
- A Mac with 8 GB of memory. The Moorhold app for Mac needs 16 GB of memory. An 8 GB Mac hasn't been measured.
- Removing Moorhold from a Mac. The removal has been tested on throwaway folders only, not on a Mac with Moorhold installed. Lima's cached Linux image stays after a removal, and the app names it.
- The sample's metrics aren't saved or run. The sample data declares seven metrics, but its marts list no readers and have no row-level rule, so Superset doesn't save them and no one can run them through an agent. On a Stack of your own, Superset saves a mart's metrics once the mart lists readers, under
readersor under its row-level rule. An agent can run them only on a mart with a row-level rule. - Metrics on the native Assets page. The page says it lists metrics, but no metric appears on it.
- Lineage on a Stack without DataHub. The portal reads lineage from DataHub, so on a Stack without DataHub no page shows lineage. The lineage service stores Airflow's events, and no page shows them.
- Changing the catalog after install. A Stack that still declares the DataHub component can't read the native catalog, and the render says so. Changing the catalog after install isn't available in this release.
- Two lines on the native Assets page ask you to reload. Where the catalog is still being built, or is empty, the page says to reload it to check again. It doesn't check by itself.
- Signing back in to DataHub returns to its home page. After you sign out of DataHub and sign in again from the signed-out page, you land on DataHub's home page, not the page you left.
- Signing back in after Superset's sign-out. After you sign out of Superset and sign in again from the signed-out page, you land on the portal's Proposals page. Open Superset again from the launchpad.
- More memory on a small server. On a full server with less than about 24 GB of memory (the 12 GB and 16 GB sizes included), or a core server with 8 GB, more memory for a tool is refused: the portal says the server hasn't room, because Moorhold keeps 4.5 GiB spare above its apps' limits.
- More memory for a sleeping DataHub. If a request for more memory for DataHub's services comes due while DataHub sleeps, the resize can fail and put the old size back. If it does, open DataHub to wake it and ask again.
- The launchpad and a sleeping tool. The launchpad can't tell that a tool is asleep, so a sleeping tool's tile can show it as not answering. Opening it wakes it.
- DataHub on a shared computer. If the last person to use the browser didn't sign out, DataHub can still open as them, with their access, until its session runs out, even after you sign in to Moorhold as yourself. Sign out before someone else uses the browser.
- Users and Groups in DataHub. If your role can't manage users, DataHub's Users and Groups page shows "Failed to load users!". That's DataHub's message when your role isn't allowed to see the list of users.
- Working only in Airflow or Superset. Time spent in Airflow or Superset alone doesn't keep your Moorhold session going, so if you use only one of them until the session times out, you see the sign-in page.
- API tokens made in Airflow. They keep working after you sign out, until they expire or are deleted in Airflow.
- An agent working alone. An agent's calls to DataHub's MCP server don't count as use, so DataHub can go to sleep while an agent is working with it. The agent's next call wakes it.
- System tables for agents. Agents that search through DataHub's own MCP server still see the system tables, because DataHub applies its default view only on its own pages.
- A reboot after a rollback. 2026.10.7 and earlier releases can't stop Moorhold at shutdown. So after a rollback to one of them, a reboot shows services failing and restarting until Moorhold has started, as before. They recover by themselves.
- Airflow's Triggerer. Airflow shows its Triggerer as unhealthy because Moorhold doesn't run one. Nothing in Moorhold uses deferrable tasks, and Moorhold's own health check ignores it.
- The words reconciler and reconciled. The pass summary after "Update the tools now", such as "3 resources reconciled", and its refusals, a Waiting reason that ends "on the next reconcile", an agent's reply to a new proposal, the audit log and Moorhold's own logs still use those words.
Upgrading #
On a server running 2026.10.7 or later, update from the server's command line with sudo moorhold upgrade <version>, giving the release you're moving to. On a server running 2026.10.6 or later, you can instead download the new release's installer and run it on the server as you did to install, but without --hostname.
This release's notes give its steps under Updating, and say which releases it updates from. If they don't cover the release your server runs, email hello@moorhold.dev and we'll help. More about upgrading
Recovery steps #
There's no automatic rollback yet, so there are no recovery steps to follow on your own. If something has gone wrong with your Moorhold, email hello@moorhold.dev with the version you're on and what you saw, and we'll help. Please don't send passwords or keys.
Download and verify #
The signed release is on get.moorhold.dev. Check the signature on release.lock first, then each file's SHA-256.
| File | SHA-256 |
|---|---|
| moorhold-2026.11.3.tar.gz | 6e4ad75d3c4a0058e51974757a05ff0dbbd827069829d238e999c20164bcfca7 |
| manifest.yaml | 4fbb38adc644d366e4568702b79b0f82f2180f87c48f3b87e6e4fb0eeb5ea5f6 |
| release.lock | 30c7fec552efdf9d57a1dd3fed1803319126e45ba03c1fecc58c0926bbad5497 |
| install.sh | 5efd01031c3fa2a629c70d05726c34ebc2f4411b700daea230ffba520e8eba1d |
| cloud-init-ionos.yaml | 677d8e45b11c364ddde5c3c6dda6fbd2fd794b265017ba4f7f0165df8d816f10 |
| cloud-init-hetzner.yaml | 239413416d492fa5c02bb000a5d5533cce2511387fee1db98ea3049b5d7a7b03 |
| Moorhold-2026.11.3.dmg | c0a04a3733b02c7c0044af55019eafc04cb59254cc404e2a499bbcf12685b387 |
| release.lock.minisig | Signature, key 7F6D12F3F6217F9C |
| SHA256SUMS | The checksums above |
minisign -Vm release.lock -p moorhold-release.pub sha256sum -c SHA256SUMS
The Mac disk image is signed and notarised. It isn't in release.lock or SHA256SUMS: check it against the SHA-256 in its row.
Public key: get.moorhold.dev/moorhold-release.pub