Back to :latest + a watchtower service checking daily at 05:00, cleanup
on, ntfy notification to arrsstack-alerts on every update. Rollback net
= Plex's own scheduled DB backups + the NAS media_pc backup. shell-mcp
is a local build, watchtower ignores it.
Same PMS build the database was born on. SQLite schema upgrades are
one-way; a surprise :latest bump no longer gets a vote. Bump the pin
deliberately when choosing to upgrade.
subprocess.run inside the async handler serialized every request behind
the slowest command (found 2026-08-01 when a du wedged the server mid-
migration). Run it in the default thread pool via run_in_executor so
concurrent calls actually run concurrently. Timeout default 30->60,
cap 300->600, schema text matched.
No Windows-side robocopy. Everything on D: (PMS dir, Calibre, Nikola
iPad) is copied to the NAS media_pc share from Linux, Media/ staged via
the SSD, D: formatted ext4 BEFORE Plex first starts, thumbnails live on
/srv/data. Fixes the M3 restore path that pointed at the unmounted
Share. Per Mike 2026-07-31.
Bind mount is now active, not a commented Phase-2 option — Mike's call
2026-07-31 (337 GB is too large for the SSD). Quick Sync confirmed as
the transcode device; NAS account is read-only on the library shares.
NAS account is read-only on the libraries at the Synology side; the ro
mount option is belt-and-braces. media_pc is this box's one write share
(PMS backup, Calibre, Nikola iPad). Per Mike 2026-07-31.
The forward-only decision was reversed: the server identity is now
migrated, not minted. Claiming would destroy the exact thing the
migration exists to preserve.
- remove every PLEX_CLAIM path
- check the six real mount points by exact share name
- refuse to start against an un-remapped database
- refuse to start without ProcessedMachineIdentifier in Preferences.xml
- assert machineIdentifier after start
Reverses the forward-only decision. That call predated measuring the library;
at 27 TB / 25,148 video items / 337 GB of generated preview cache, rebuilding
costs 1-3 weeks of saturated gigabit and permanently loses every manual match
fix across 23,493 episodes. Migration preserves watch history, resume points,
collections, playlists, manual matches, artwork choices, added-at dates, the
337 GB cache, and server identity - so shared users stay invited, clients do
not re-add, and there is no claim step at all.
migration/remap.sql
Four path columns remapped: media_parts.file (80,126), section_locations
.root_path (11 -> 9), media_streams.url (14,837) and metadata_items.guid (9).
Two formats, not one: backslash/UNC for the first two, file:// with %20
encoding for the last two - decoding those %20s would break every subtitle
reference containing a space, which given share names like Radio Shows is
most of them.
A scan of all 80 text columns across 82 tables found SIX columns matching
korval. Only four are paths. taggings.text is one row reading Dr. Korval,
and metadata_items.summary is four Liaden Universe blurbs about Clan Korval.
A bare REPLACE on the word would have corrupted a cast credit and four book
summaries, so every statement is anchored to a path prefix. Proven against a
synthetic database built from the real path shapes: 12/12, including both
false positives surviving byte-identical.
Also drops the Audio Books and Music Organized roots - configured as library
roots but holding 0 files / 0 bytes. The consolidation had already happened;
only the dead roots remained.
migration/migrate-db.sh
Runs the remap through Plex own SQLite build borrowed from the container
image, keeps a .pre-remap rollback, asserts the counts moved 1:1 and the
false positives did not, then stats 200 random remapped paths against the
real filesystem. That last check is the one that matters - SQL running
without error proves nothing.
migration/gen-preferences.ps1
Plex live settings store on Windows is the REGISTRY, not Preferences.xml,
and the two disagree here. Folds ~50 values into one Linux file, dropping
Windows-only keys including the per-GPU limit keyed by 10de:1b81, the GTX
1070 PCI ID. Preserves MachineIdentifier and the online token, which is what
keeps the server identity. The existing Preferences.xml is malformed anyway
(duplicate allowedNetworks) and will not parse strictly.
scripts/20-cifs.sh
Eight shares down to six. Mount points now mirror the share names verbatim -
/mnt/nas/Home Movies, space and all - because that reduces the remap to one
uniform prefix substitution instead of eleven special cases. New requirement
this creates: \040 escaping in BOTH fstab fields, not just the share name.
Verified, plus a round-trip check that a remapped DB path lands under the
generated mount point.
plex/docker-compose.yml
Six read-only NAS binds whose source path equals target path, so database,
host and container agree with no translation. No PLEX_CLAIM. Phase 2 Media
relocation present but commented.
scripts/60-media-relocate.sh
Post-soak, optional: moves the 337 GB cache to the 3TB ext4 disk so future
growth (~28 MB per content-hour) stops eating the SSD. Plex does not support
this, so the script forces generation on one title afterwards to prove writes
survive the EXDEV boundary, and rollback is deleting a compose override.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Autoinstall lays down a thin base (sshd, key, DHCP, Docker CE, /srv) and hands
off to this repo on first boot. Everything interesting stays in git so it is
reviewable and re-runnable, rather than frozen onto a USB nobody can diff.
Stages, all idempotent:
00-preflight asserts hardware/BIOS state, changes nothing. Catches a BIOS
update having silently re-enabled Secure Boot, which would stop
the NVIDIA DKMS module loading on a box with no keyboard.
10-secrets ADD-ONLY seeder for /srv/secrets/stacks.env. Never overwrites an
existing key. Verified against a pre-populated file: existing
values, unrelated keys, the operator tier and existing manifest
lines all survive byte-for-byte; a second run is a no-op.
20-cifs the 8 shares Plex actually uses (Share is excluded, it is not a
library root). \040 escaping, nofail + x-systemd.automount +
_netdev. Managed-block rewrite verified not to duplicate or to
drop the root fstab entry.
30-nvidia nvidia-driver-580 explicitly: 580 is the LAST branch supporting
Pascal, and the -open modules need Turing+. Pins against newer
branches. Not in late-commands because DKMS needs the installed
kernel, not the installer's.
40-shell-mcp builds the native MCP locally for amd64; refuses to finish
unless /sse returns 401 without a token.
50-plex run by hand: PLEX_CLAIM expires in 4 minutes. Refuses to start
against missing mounts and disables autoEmptyTrash, which with
read-write NAS credentials is the most dangerous default here.
shell-mcp was built on arm64 originally. It builds clean on amd64 (whole dep
tree resolves to prebuilt manylinux x86_64 wheels, no compiler needed), but
dependencies are now pinned - the original installed mcp/starlette/uvicorn
unpinned and starlette has since gone 1.x. Port moved to 8103 so NPM host 42
can simply be repointed, and the tool description now says media box rather
than arrsstack.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Runs the stages in order and fails soft. A stage failure must never wedge
boot: there is no keyboard attached to this box, so a machine that comes up
with sshd and a broken GPU driver is recoverable and one that hangs is not.