Questions people actually ask
Every one of these has been asked more than once. The answer is a sentence; the link is where the reasoning lives.
Can I put the board on the internet?
Section titled “Can I put the board on the internet?”No. It can reflash four computers, its one account is a full administrator with no authorization layer under it, and its login deliberately does not depend on anything outside the board — because the board is what you reach for when everything outside it is broken. Expose one page over every board instead, which holds no credential of its own.
Does upgrading the firmware reboot my nodes?
Section titled “Does upgrading the firmware reboot my nodes?”On a v2.5, no. Upstream power-cycles all four compute modules to patch the management computer beside them; this fork leaves their rails alone. Measured on a nine-node cluster during a live upgrade: 0 modules cycled, 0 cluster nodes lost, and 48 seconds for the board itself to come back.
The board's own network drops for those 48 seconds, because the board drives the switch. The modules keep running throughout.
That was measured on a v2.5 board, whose nodes have latches on their power rails. A v2.4 board drives the rails from the BMC, so there any BMC reboot, including the one that boots a new image, power-cycles all four nodes: shut them down first. Until v2.46.0 the web interface's reboot dialog said the opposite on a v2.4 (the fault). I have no v2.4 board to measure it on.
Can I use my own certificate?
Section titled “Can I use my own certificate?”Yes, and you should. If you run a private authority, import its root into your browser once and every certificate it issues is trusted — which is also the only fix for the serial console, because a click-through exception is not applied to a WebSocket.
Getting your certificate onto the board is a control now — the Security page,
tpi tls install, or the API — and it takes effect without a restart.
Automatic renewal over ACME is still on the roadmap. Let's
Encrypt is possible and is the wrong tool here, for three reasons.
The password form will not let me type. Is my board broken?
Section titled “The password form will not let me type. Is my board broken?”No, and it was not you. Every release from v2.28.0 to v2.36.0 had a fault in the shared text field that made password boxes read-only: the keystroke arrived and the interface put the old value straight back, silently. v2.37.0 fixes it.
Until you can install that, change the password over SSH — it is the same account the interface means:
$ ssh root@BMC$ passwdOn a board still using the password it shipped with, that also clears the page which blocks the rest of the interface. Nothing needs undoing afterwards. The full account is in what is and isn't fixed.
Why did my password and certificate reset after I upgraded?
Section titled “Why did my password and certificate reset after I upgraded?”If you upgraded from an SD card, that is why. The card is an installer, and
through v2.41.0 it erased everything on the board: the password went back to
turing, a new self-signed certificate was made, and the network, NTP, fan and
node settings went with them. v2.42.0 fixes it: the card now keeps them
when it can verify them, and says on the serial console when it cannot.
If you used the .tpu (the web interface's Firmware page, tpi firmware), the
settings are not touched by the upgrade, so the other cause is more likely: the
settings volume is damaged and the board falls back to a factory state in RAM at
every reboot. An upgrade is only when you notice, because that is when the
BMC reboots. From v2.46.0 the board tells you: every page of the web
interface, including the factory-password page, shows a banner with the
kernel's explanation. Install v2.46.1 rather than v2.46.0, because only v2.46.1
offers the repair button on a real board. To tell the two causes apart:
- After a card install from v2.41.0 or earlier, the reset happens once, at install time, and the new password then survives reboots. The board showed the installer's prompt on the serial console, or waited on the front panel buttons, before it came back.
- With a damaged volume, the reset repeats at every reboot. On v2.46.0 and
later the banner says so. Press its plain Reboot first, because a mount
failure can be a one-off. If the banner comes back, Repair settings
storage (offered only where it is safe) removes the volume that could not
be mounted and reboots; the next boot makes a new one. It erases the saved
settings for good, which were not loading anyway. On an earlier release, or
without the web interface, SSH shows it:
mount | grep ' / 'givesupperdir=/tmp/safemode/upper, anddmesg | grep -i ubifsa UBIFS error. This is a known fault, with the manual repair.
Holding KEY1 at power-on is a deliberate safe mode and also gives the factory
password; that is expected, and its banner, from v2.46.0, never offers the
repair. To reset on purpose with the new card, put factory-reset.txt next to
install.txt.
The banner and the repair have been tried on a board with healthy settings only (nothing to show, and the repair refused); I have not seen them on a board that is really in this state.
Grafana's Save & test fails with "failed to get Prometheus heuristics". What is wrong?
Section titled “Grafana's Save & test fails with "failed to get Prometheus heuristics". What is wrong?”The data source URL points at the board. The board serves a page of numbers on
port 9110 and cannot answer queries; Grafana needs a Prometheus server (or
VictoriaMetrics, or Mimir) that scrapes the board, and its URL is that
server's — http://<prometheus host>:9090 — never the board's. That curl
against :9110 returns data is correct and is not what Grafana is asking
for. The monitoring guide
has the one-container Prometheus for people who have Grafana and nothing
scraping yet.
Flashing a module in slot 1 or 2 fails with "USB interface not found". Why?
Section titled “Flashing a module in slot 1 or 2 fails with "USB interface not found". Why?”Look at the mini-PCIe slot of that node. The slots of nodes 1 and 2 carry USB 2.0, and a USB device in one of them (an LTE modem, a Zigbee stick, a serial card) appears to share the USB pair the board uses to flash the module. With two devices on one pair the module cannot enumerate, and the flash stops with "Compute module's USB interface not found or supported". The board's kernel log shows the module failing with error -71 and then "unable to enumerate".
Power the node off, take the mini-PCIe card out, flash, and put it back. Or move the module to slot 3 or 4 for the flash; those slots have no mini-PCIe USB. From v2.44.0 the error says what it saw instead: it names the other USB devices on the node's lines and says that a device in the node's mini-PCIe slot may be in the way.
This comes from one user report on Discord (2026-10-15) and from reading the firmware source and the factory tests, not from a schematic, so I call it likely, not proven. The entry in known faults has the details.
I split the board's network into two bridges by hand. Will the web interface undo it?
Section titled “I split the board's network into two bridges by hand. Will the web interface undo it?”No. From v2.45.0 the Network page reads the layout you made, shows the
address from whichever interface carries it, and lists every interface the
board has. It manages br0 only, so on a board without br0 it says so and
leaves the card read-only; change the file over SSH, and the page notices.
If you do apply an address on a board whose /etc/network/interfaces was
written by hand, it asks first. Only if you agree is the file replaced, and the
old one is kept as /etc/network/interfaces.hand-written.<time>, the newest
five. On v2.35.0 to v2.44.0 the same board showed no address and every change
failed with "status code 500", and confirming would have overwritten your file
without a copy.
The board's .local name needed one more release. mDNS is bound to exactly one
interface, and until v2.46.0 that was always br0, which a split board does not
have. From v2.46.0 it picks br0 if there is one, otherwise the bridge that
carries the default route, and re-checks every five minutes. I tried it on the
stock layout of my own boards, where it still binds br0; a hand-split board is
not something I have here.
Which board revisions does it run on?
Section titled “Which board revisions does it run on?”v2.5.2, on the two boards this fork is developed against, and v2.4, from two readers' reports — every one dated and recorded on the install guide. The other revisions are upstream's compatibility list, not this fork's proof, and the install guide says so rather than implying a test matrix that does not exist.
The clock says "not synchronised". What now?
Section titled “The clock says "not synchronised". What now?”Since v2.35.0 the Time card on the Maintenance page lists every source chrony knows and what it thinks of it, in chrony's own terms. Two states cover nearly every report:
- unresolved — the board could not look the name up. It has no working resolver, which is what a static address set by hand over SSH leaves behind on this image. Give it one on the Network page (the address card takes resolvers), or use the server's address instead of its name.
- unreachable — the server never answered: wrong address, a firewall, or a router that does not serve NTP at all.
A server that reports itself unsynchronised (stratum 16) is refused, and the card says so; chrony will not take time from it. The pool the image ships with is a name too, so a board with no resolver has no source at all — which is how "not synchronised" looked before the card could say why.
If the firmware check also says nothing, the two have one cause. Until v2.36.0 a source the board could not reach was reported as offering nothing rather than as unreachable, so a board with no resolver showed an empty catalogue and no error. It says which source failed and why now.
If it was fine until a reboot, and the address is static: on this image
/etc/resolv.conf is a link into a memory filesystem and starts empty at
every boot; the address card's stanza rebuilds it as the bridge comes up.
Firmware v2.35.0 wrote that stanza with a # in it, which the board's
ifup reads as a comment, so the rebuild failed silently at every boot and
the resolvers were gone by the time you looked — reported from a 2.4 board
on 2026-09-22. v2.36.0 fixes the stanza and repairs a board that
v2.35.0 already wrote, the first time its daemon starts. If you wrote a
resolv.conf into the overlay by hand to get past it, it does no harm and
is no longer needed.
How much of this is written by a machine?
Section titled “How much of this is written by a machine?”Some of it, and the parts that matter are read line by line — the API, the build scripts, the TLS handling, the update gate. Every release is run against a live board and the patch releases that follow minors are the visible evidence of that. What CI can prove without hardware it proves; what only a board can prove is listed as such.
What is checked, and by what →
Where do I report something, or ask for a feature?
Section titled “Where do I report something, or ask for a feature?”A bug goes to Issues, a question to Q&A, and a feature to Ideas, where 7 are waiting and 23 votes have been cast. The roadmap is ordered by those votes.
What changed recently?
Section titled “What changed recently?”Firmware v2.52.0 is the newest of 167 releases across four components. The changelog is each repository's own, not a summary, and there is a feed.
The news → · Every component's changelog → · Follow by feed →