Arkitekt hub on a FRAME
This is a working state from November 2026, not a released feature. It runs on one test FRAME and needs manual steps. Branches and open points are listed at the end.
The idea
Arkitekt connects microscopes, analysis tools and people. A hub stores the images (mikro) and passes on remote actions (rekuest). A coordination server (go.arkitekt.live) only holds logins, organizations, and which hubs exist and where to reach them.
Each FRAME runs its own hub, installed with the rest of the system by the openUC2 OS pallet:
- ImSwitch registers with the hub as an Arkitekt app. It offers actions such as Tile Scan and Acquire Frame, and after a scan it uploads the canvas (one OME-Zarr image with its pyramid) to mikro.
- Orkestrator (the Arkitekt desktop app) logs in at go.arkitekt.live, picks the FRAME's hub, and then gets images and actions straight from the FRAME, over the local network or the tailnet. Image data never leaves the lab.
- The hub's database, objects and container images live on an NVMe, not on the SD card.
What runs where
| Part | Where | Deployed by |
|---|---|---|
| ImSwitch (with the Arkitekt client, arkitekt 16) | container imswitch | pallet deployment imswitch, feature arkitekt keeps its login in /home/pi/arkitekt/imswitch |
| Hub: gateway (Caddy, ports 2754/2755), rekuest, mikro, Postgres, Redis, RustFS | compose project arkitekt-hub | pallet deployment arkitekt-hub |
| Hub configs, certificate, keys | /home/pi/arkitekt/arkitekt-hub | konstruktor hub create, once, from arkitekt-hub-setup.service at boot |
| Container images, hub data | NVMe at /mnt/nvme | systemd mount units (not yet in the pallet) |
| Logins, organizations | go.arkitekt.live | Arkitekt |
Setting up a FRAME
- NVMe. Format it with the label
frame-nvme, then mount it and bind-mount/var/lib/containerd,/var/lib/dockerand/home/pi/arkitektonto it with systemd mount units. Docker and containerd getRequiresMountsFor=for those paths. An/etc/fstabline does not work on openUC2 OS:/etcis an overlay mounted after systemd reads fstab. - Pallet. Switch to the os-rpi branch with the hub, then enable it:
forklift plt switch github.com/openUC2/os-rpi@feat/arkitekt-hubforklift plt enable-depl arkitekt-hubforklift plt enable-depl-feat imswitch arkitektforklift plt stage && sudo reboot
- Connect the hub once. Open the printed link and accept the hub into your organization:
konstruktor hub connect --in /home/pi/arkitekt/arkitekt-hub --server go.arkitekt.live --no-open
- ImSwitch. Add an
arkitektblock to the setup file, then press Connect in the Arkitekt panel and approve this FRAME's hub:"arkitekt": {"enabled": true, "appName": "frame-<name>", "url": "https://go.arkitekt.live","autoConnect": true, "useMikro": true,"uploadExperiments": true, "deleteAfterUpload": false} - Orkestrator. Log in at go.arkitekt.live and pick the hub named
openuc2-<name>. On macOS, allow Orkestrator under Privacy → Local Network. Check thatcurl -k https://<frame address>:2755/mikro/htprints200on the Mac.
First measurements
Synthetic scan, 10 × 10 RGB tiles of 3072 × 2048 pixels, 12 bit (3.8 GB raw), written with ImSwitch's own OME-Zarr writer, then its pyramid. Same ImSwitch image, Raspberry Pi 5 with 8 GB.
| Step | NVMe | SD card |
|---|---|---|
| Write the 100 tiles | 83 s | 418 s |
| Build the pyramid (5 levels) | 35 s | 546 s |
| Total, including flushing to disk | 127 s | 977 s |
| CPU time spent waiting for the disk | 1 % | 69 % |
The NVMe is about 7.7 times faster for the scan itself. The CPU now spends its time on the work, not on waiting. Peak memory was 1.1 GB in both runs, and the Pi did not throttle.
SD cards differ a lot. On FRAME ID10 the same pyramid took 128 s on its card, and the whole way from the end of the scan to the image in mikro took 6.5 minutes (259 s of it for the upload). The upload to mikro on the NVMe has not been measured yet.
Known limits
- Every scan is stored twice, once as ImSwitch's canvas and once as mikro's re-encoded copy,
unless
deleteAfterUploadis on. mikro does not downscale on the server, so ImSwitch uploads its pyramid. - Orkestrator is macOS arm64 only so far. It draws at most 64 layers, so a scan has to be one canvas, not one image per tile.
- The hub's addresses are fixed when it connects. After an IP change run
konstruktor hub connect … --reach this-network, or reserve the DHCP lease. - Forklift removes compose projects it does not know. Start the hub only through the pallet
deployment, never with
konstruktor up. - Logins and token refresh need go.arkitekt.live, so the FRAME needs internet. A local coordination
server (
konstruktor coord create) would work offline but has not been tried. - Docker publishes ports 2754 and 2755 on every interface, whatever the firewall zones say.
Open work
- Bring the hub branch up to date with
edge, and merge it. - Put the NVMe setup into the pallet, as a deployment that can be enabled.
- Merge ImSwitch PR #333 (
feat/arkitekt-hub-mikro) and pin the image built from it. - Show the hub's connect link in the machine portal, so no terminal is needed.
- Test the boot path end to end: setup unit, connect, upload, reboot.
| ImSwitch | branch feat/arkitekt-hub-mikro, PR #333 |
| Pallet | os-rpi branch feat/arkitekt-hub, package deployments/arkitekt-hub.pkg |