Deploying DUMB on Proxmox¶
The native helper runs DUMB and its managed services directly in one LXC, without a Docker layer. The helper currently uses the maintainer validation branch described below.
This keeps the normal DUMB topology simple:
Proxmox host
└── DUMB LXC
├── DUMB API and frontend
├── rclone / Decypharr / InfiniDysk
├── Arr applications
└── Plex / Jellyfin / Emby
Because every consumer is in the same LXC, DUMB-created FUSE mounts only need to exist inside that LXC. No mount propagation back to the Proxmox host is required.
Native LXC installation boundaries¶
The native helper keeps DUMB and its managed applications in one LXC. Applications in another LXC need an explicit storage-sharing plan.
flowchart TD
Host["Proxmox host: run the reviewed helper"] --> LXC["Dedicated DUMB LXC"]
LXC --> Controller["DUMB controller and dashboard"]
Controller --> Apps["Managed workflow apps and media servers"]
Controller --> Storage["Persistent config, data and media paths"]
Browser["Your browser"] --> Dashboard["LXC address on port 3005"]
Dashboard --> Controller
External["Apps in another LXC"] -. "Separate sharing plan required" .-> Storage
Install the validation build¶
While the helper is awaiting upstream acceptance, run the maintainer-fork build from the Proxmox VE host shell:
COMMUNITY_SCRIPTS_URL="https://raw.githubusercontent.com/I-am-PUID-0/ProxmoxVED/feat/dumb" \
bash -c "$(curl -fsSL https://raw.githubusercontent.com/I-am-PUID-0/ProxmoxVED/feat/dumb/ct/dumb.sh)"
COMMUNITY_SCRIPTS_URL makes the shared Community Scripts engine fetch the
matching install/dumb-install.sh from the same fork and branch. Do not omit it
while using the validation build.
The default controller source is the latest stable DUMB release. For an
advanced test installation, set DUMB Controller Branch in the script
generator's app settings or pass var_dumb_branch from the host shell. For
example, this installs the current dev branch:
var_dumb_branch=dev \
COMMUNITY_SCRIPTS_URL="https://raw.githubusercontent.com/I-am-PUID-0/ProxmoxVED/feat/dumb" \
bash -c "$(curl -fsSL https://raw.githubusercontent.com/I-am-PUID-0/ProxmoxVED/feat/dumb/ct/dumb.sh)"
Any valid branch in I-am-PUID-0/DUMB can be selected. Branch heads are
mutable development targets and may be less stable than a published release;
the same candidate validation and controller rollback protections still apply.
Take an independent backup first: controller rollback restores /opt/dumb, but
it does not reverse configuration migrations or application-state changes under
/config or /data.
Validation build
This branch is for controlled testing before the upstream pull request. Use
a new LXC, keep independent backups, and report the DUMB version plus the
relevant journalctl -u dumb output when a test fails.
After the script is validated, submitted to community-scripts/ProxmoxVED, and
promoted into community-scripts/ProxmoxVE, the production command will become:
bash -c "$(curl -fsSL https://raw.githubusercontent.com/community-scripts/ProxmoxVE/main/ct/dumb.sh)"
The default installation creates an unprivileged Debian 13 LXC with:
- 4 vCPU;
- 8 GiB RAM;
- 40 GiB disk;
- FUSE enabled for rclone and other DUMB-managed mounts;
- optional Intel/AMD GPU device passthrough; and
- DUMB plus its required runtimes installed directly on the guest OS.
Use Advanced Settings when a large stack, source builds, the install cache, or a media server needs more resources. Both amd64 and arm64 have completed the recorded native acceptance tests with DUMB 2.21.0: clean Proxmox VE installation, controller update and rollback, reboot, FUSE namespace isolation, and stopped backup/restore. That qualification applies to the tested versions; it does not establish every later application update or GPU path.
When installation finishes, open:
The frontend can take several minutes to appear on first boot while DUMB installs its initial managed services. Follow progress inside the LXC with:
Then continue with Getting Started.
.NET repository checks¶
Debian 13 has an official Microsoft .NET feed. An updater message claiming that the feed does not exist can also come from a failed network probe.
The DUMB install/update scripts wrap the existing Community Scripts .NET installer with a local repository check. On Debian, this check runs before the shared installer can remove an existing Microsoft source/keyring, retries transient failures, and reuses a successful result for the duration of that .NET setup invocation. Failed checks stop the update before repository changes. The wrapper also avoids the shared probe's reversed cached success/failure handling without modifying core or changing checks for other applications.
Use the updated DUMB helper scripts for this fix; updating only DUMB's Python
controller does not change native runtime reconciliation. No core fork or
COMMUNITY_SCRIPTS_CORE_URL override is required. Existing running services
remain on the previous controller if dependency reconciliation fails before
activation.
Native LXC paths¶
| LXC path | Purpose |
|---|---|
/opt/dumb |
DUMB controller source and Python environment |
/config |
DUMB configuration, authentication state, snapshots, and feature state |
/data |
Persistent managed-service data |
/log |
DUMB and managed-service logs |
/mnt/debrid |
FUSE mounts, generated links, and symlink libraries |
/postgres_data |
DUMB-managed PostgreSQL cluster when enabled |
/zurg/config.yml, /zurg/plex_update.sh |
Base Zurg files copied into newly configured Zurg instances |
/etc/dumb/controller-source |
Root-managed stable or branch selection used by update |
/etc/systemd/system/dumb.service |
Native DUMB systemd service |
These are normal LXC filesystem paths, not Docker bind mounts. DUMB preserves its established internal paths so the same service configuration works in the native and container-image deployments.
The helper also configures systemd-logind to retain IPC owned by DUMB's
managed-service UID. This is required for PostgreSQL POSIX shared memory after
temporary setup sessions end. Package-provided PostgreSQL, Plex, and Jellyfin
systemd units are kept inactive because DUMB, not the distribution unit, supervises
those processes inside this dedicated LXC.
GPU passthrough and Jellyfin¶
When GPU passthrough is enabled, the installer invokes the Community Scripts hardware-acceleration helper to select the vendor's drivers and required Debian repository components. The update path also reconciles drivers when updating the controller. Intel GPU support requires the appropriate VA-API driver; OpenCL tone mapping additionally depends on a compute runtime compatible with the GPU generation. Installing a package does not prove either path works.
DUMB creates and maintains its own service account. Adding only the
package-created jellyfin account to render/video does not grant access to
DUMB's Jellyfin process. After applying the installer and controller fixes,
restart DUMB and inspect the actual Jellyfin process:
pgrep -a -f '/usr/lib/jellyfin/bin/jellyfin'
ls -ln /dev/dri/renderD*
# Replace <PID> with the Jellyfin PID shown above.
grep -E '^(Uid|Gid|Groups):' /proc/<PID>/status
The command should select /usr/lib/jellyfin-ffmpeg/ffmpeg unless a custom path
was configured. The process's primary or supplementary groups must include the
render device's numeric GID. For an unprivileged LXC, unmapped ownership must be
corrected in the Proxmox device mapping; making render nodes world-writable is
not required by DUMB.
Verify playback by forcing a transcode and checking the Jellyfin FFmpeg log for the selected binary, QSV/VA-API codec, and successful device initialization. Testing only as root can conceal service-user permission failures. See the Jellyfin service guide.
Intel OpenCL tone mapping¶
QSV/VA-API video acceleration and OpenCL tone mapping use separate Intel
runtimes. QSV can keep working after a compute-runtime update drops support for
an older GPU. An OpenCL error such as CL_PLATFORM_NOT_FOUND_KHR or
Failed to get number of OpenCL platforms: -1001 means that no OpenCL platform
was available to that process.
Unreleased native helper correction
The native helper correction adds Intel's legacy compute runtime alongside the current runtime for its supported Intel GPU setup path on amd64 Debian 13 (trixie). This includes Gen9–Gen11 GPUs recognized by that path; older GPU setup paths are unchanged.
Once the corrected helper is published, run update inside the LXC.
Selecting DUMB's dev controller branch or updating Jellyfin alone does
not install this helper change. The helper checks OpenCL device enumeration
and warns if no Intel GPU is found.
Check the installed compute packages and ICD registrations inside the LXC:
Then check OpenCL as the account running DUMB's Jellyfin process, using the PID from the earlier process check:
ps -o user=,pid=,args= -p <PID>
# Replace <service-user> with the account shown above.
runuser -u <service-user> -- clinfo -l
The output should list an Intel GPU platform and device. This checks runtime
loading and device access; it does not prove that Jellyfin's FFmpeg can use
OpenCL with QSV. After applying the corrected native helper, force an HDR-to-SDR
transcode with tone mapping enabled and inspect its FFmpeg log. A listed
tonemap_opencl filter or opencl hardware-acceleration type only describes
FFmpeg's compiled features.
If playback still fails, include the package versions, clinfo output and
Jellyfin FFmpeg error in the report.
Manage and update DUMB¶
Useful commands inside the LXC:
Run the Community Scripts updater inside the LXC to update the currently selected DUMB source. Stable installations follow the latest published release, while branch installations follow the current head of their saved branch:
To switch an existing LXC to dev or another DUMB branch, pass the branch for
one update run:
The branch choice is saved only after the candidate starts and passes the API
health check, so later update runs continue following it. Return to the stable
release channel with:
Switching back to stable can be a controller downgrade. Verify that the branch did not introduce configuration or data changes that the stable release cannot read before making that switch.
The updater reconciles the required native dependencies, downloads the new
controller into a candidate directory, builds its locked Python environment,
and validates it while the current controller is still running. It then backs
up /opt/dumb, stops DUMB only for activation, starts the candidate, and waits
for the API health check. A failed build, start, or health check automatically
restores and verifies the previous controller. An interrupted update is
recovered the next time update runs.
Configuration and managed application data under /config, /data,
/postgres_data, and /mnt/debrid remain outside the replaced controller
directory. The update backup is removed after the replacement passes health
verification.
Use DUMB's own Updates controls for the individual services it manages.
Backups and FUSE¶
Back up at least /config and /data, plus any application-native backups
required by the services you enabled. /mnt/debrid normally contains live or
re-creatable mount views and should not be treated as the only copy of data.
Proxmox snapshot backups and FUSE
Proxmox strongly advises against FUSE mounts inside an LXC because freezing the container for suspend/snapshot-mode backups can fail. Prefer a stopped-mode backup window after stopping DUMB, or otherwise test the exact backup mode and mount stack before relying on it.
Existing Plex or Arr applications in other LXCs¶
An external Plex or Arr LXC cannot see /mnt/debrid merely because it uses the
same path name. It needs a separately designed storage boundary that exposes
both the symlink library and every target path those symlinks reference.
The helper deliberately does not add that boundary automatically. Reverse propagation of a mount created inside an LXC depends on the host mount namespace, LXC propagation flags, FUSE permissions, UID/GID mappings, backup behavior, and the consumer's own mount namespace. A configuration that gets one of those boundaries wrong can look healthy until a remount leaves Plex or an Arr app on a stale path.
Choose one of these approaches:
- Move the consumer into the DUMB LXC. This is the supported and simplest layout.
- Mount the remote storage on the Proxmox host, then bind-mount the same host-managed tree into DUMB and every consumer LXC. In that layout, disable the duplicate DUMB-managed rclone mount and keep the same absolute paths in all applications.
- Export the complete DUMB mount/link tree through a deliberately managed network filesystem and mount it in each consumer. Validate symlink target paths, permissions, disconnect/reconnect behavior, and startup ordering.
- Build a custom
rshared/rslaveLXC topology only after proving mount and unmount propagation in both directions on the exact Proxmox/LXC version.
Do not add a host mount over an active /mnt/debrid
A new bind mount hides the directory that was previously at that path. If DUMB is already configured, stop it, take independent backups, and migrate the mount and symlink trees deliberately. Never attach an empty host path over an active deployment and assume the old files moved into it.
For a custom propagation topology, verify all of the following before allowing an Arr import or media-server scan:
# DUMB LXC
findmnt -R /mnt/debrid
findmnt -no TARGET,PROPAGATION /mnt/debrid
# Proxmox host and every consumer LXC
findmnt -R <shared-path>
readlink <representative-symlink>
Test a provider remount while each consumer is running. A consumer should see
the replacement mount without retaining Transport endpoint is not connected.
Also confirm that consumer unmounts cannot propagate back and tear down DUMB's
producer mount.
Troubleshooting¶
The DUMB page is unavailable¶
Inside the LXC:
systemctl status dumb --no-pager
journalctl -u dumb -n 200 --no-pager
ss -lntp | grep -E ':(8000|3005)\b'
The API normally listens on port 8000; the user-facing frontend is port
3005. The LXC firewall and network must allow the frontend port from your
operator workstation.
/dev/fuse is missing¶
On the Proxmox host, inspect the generated container configuration:
The Community Scripts default enables FUSE. Stop and start the LXC after fixing its configuration; restarting only the DUMB service cannot add a missing device to the LXC.
A managed service cannot find a runtime¶
Confirm the native toolchain and controller environment:
python3.11 --version
python3.12 --version
node --version
pnpm --version
go version
dotnet --info
rclone version
/opt/dumb/venv/bin/python -m pip check
Run update to repair the DUMB controller source/environment. Service-specific
install failures remain visible in journalctl -u dumb and can then be retried
from the DUMB UI.
PostgreSQL reports a missing shared-memory segment¶
If several PostgreSQL-backed services fail together and the journal contains
could not open shared memory segment "/PostgreSQL...", verify the native LXC
safeguard:
The effective configuration must include RemoveIPC=no. Run update, then
restart DUMB, if the drop-in is absent. A healthy restart recreates the
PostgreSQL shared-memory objects and keeps them after the managed UID's setup
session closes.
Plex fails to install or cannot create its Cache directory¶
Older test builds created a direct /etc/systemd/system/plexmediaserver.service
mask before Plex was installed. Plex's Debian package rejects that path during
its pre-install check. If the package was then repaired manually, the generated
/plex/Plex Media Server tree could also remain owned by the package account.
Current LXC helpers use a supported systemd drop-in with an unmet condition. That lets the Debian package install successfully without starting its own Plex process; DUMB remains the only process supervisor and repairs the writable tree.
test ! -e /etc/systemd/system/plexmediaserver.service
systemctl cat plexmediaserver.service | grep -F 'ConditionPathExists=/run/dumb-allow-package-services'
systemctl is-active plexmediaserver.service
The first command should produce no output, the second should show the condition,
and the package-owned unit should report inactive. For an LXC created with an
older test build, reconcile the helper and restart the DUMB controller:
If dpkg still reports an interrupted installation, complete its pending work
after update has removed the legacy mask: