stevezau/media_preview_generator 5.2.0

Remote nodes, a Windows Plex node and GPU hand-off

on GitHub

5.2.0 lets the app use other machines. A worker node lends its GPU or CPU to your jobs, and a Plex database node runs beside Plex, now on Windows as well as Docker. It also stops previews being remade after a remount or move, and a file one GPU can't decode now tries your other GPUs before the CPU.

Before you upgrade

• Plex marker helper users: the helper is now called the Plex database node. Update it to ghcr.io/stevezau/plex-marker-agent:1.2.0 together with the app, and add a /config volume (the new compose file has one) or it writes no loudness. Your helper moves to Settings → Nodes by itself and keeps its address and key. AGENT_TOKEN and AGENT_PORT still work; the new names are NODE_KEY and NODE_PORT.
• Going back to 5.1.0 after upgrading: 5.1.0 can't read the new settings. Restore settings.json.pre-v23.bak from your config folder first.
• Nothing else changes in your settings.

New

Remote nodes (Settings → Nodes)

• Run the app's image on another machine with MODE=node, add it under Settings → Nodes, and put its GPU or CPU in a worker group. It then takes previews, Intro & Credits detection and loudness like the app's own devices, with the same hours and job types.
• The app stays in charge: it keeps the queue, your server tokens and every write. A node only needs to read your media, and never gets a token.
• Test access checks the node sees real files from each library at the path you mapped. A file the node can't see, or a node that goes offline mid-file, goes back to another worker instead of failing.
• Each node has its own status, Setup Health and log. Worker nodes run the same image tag as the app; a node on another version says Update needed and gets no work until you update it.
• One container on the Plex machine can be both a worker and the Plex database node.
• Guide: Remote nodes.

Plex database node, now on Windows

• Plex installed on Windows (no Docker) can get markers, chapter thumbnails, loudness and Library Health counts. Download PlexDbNode-windows.zip from the Plex database node 1.2.0 release and run install.ps1 as the user Plex runs as. Use -AtStartup if Plex runs as a service or starts before anyone signs in.
• The Docker node now writes loudness and counts Library Health too, so a Plex on another machine gets everything a local one does.

Other GPUs before the CPU

• When one GPU can't decode a file, your other GPUs try it before it falls back to the CPU. The job shows "waiting for another GPU" meanwhile.

Schedules and jobs

• One schedule can run several job types: previews, Intro & Credits and Plex loudness, over whole libraries or recently added.
• Processing order and filters from Start new job now apply to loudness and Intro & Credits jobs as well.
• Loudness analysis runs at a lower CPU priority, so other work on the machine comes first (Settings → Processing → Job execution; default 10, set 0 for the old behaviour).

Library health

• Counts chapter thumbnails, and shows files it can't read as "can't" instead of to-do.
• After the nightly check, a Plex library with up to 500 files still missing previews or chapter thumbnails gets one low-priority job for just those files.
• Setup Health suggests turning off Plex's own loudness analysis, with a button to do it.

Improved

• Library scans, Library health and Intro & Credits checks are much faster on large Plex libraries.
• GPU, Vulkan and NVIDIA driver errors say what's actually wrong and which screen fixes it.
• The container stops in a couple of seconds instead of waiting up to 30.

Fixes

Previews

• Files that didn't change no longer get new previews after a remount, rename, move or unRAID mover run.
• Thumbnails line up with their time in the video. Clips shorter than the interval get a thumbnail, and 3D MKVs get previews.
• Rotated and mirrored videos look the same from the GPU as from the CPU. A GPU decode that comes out green or striped is redone on the CPU.
• A GPU missing a codec, profile or 4:2:2 support hands the file to the CPU quietly instead of logging driver errors.
• A full or read-only working folder fails the file with the folder named, and never publishes a short preview.
• The app no longer makes Plex run its scheduled jobs early. It waits for Plex's maintenance window when those jobs are on.
• Emby: a new preview size removes the old size's files.

Webhooks and workers

• A webhook only makes the jobs its file needs. Libraries with Previews off get no preview job, and recently-added scans skip them.
• Idle GPU workers get work before idle CPU workers.
• A cancelled job stays cancelled, and retrying a Regenerate job regenerates.

Intro & Credits

• Files moved to another folder keep their intros and credits.
• Every episode in a multi-episode file gets the file's markers.
• Credit detection works on Intel GPUs that don't run GuC firmware.
• TheIntroDB lookups use its new v4 API.
• Jellyfin collections list their movies instead of the collection itself.

Chapter thumbnails and loudness

• A file shared by two Plex libraries gets chapter thumbnails on both items.
• A final chapter just before the end of the video gets its thumbnail.
• Tracks stored as -inf are measured again once. Near-silent tracks get Plex's silence values instead of failing every run, and trailers and extras are skipped instead of failed.

Settings and backups

• Restoring a backup applies quiet hours, pause and schedules at once, refuses a backup from a newer version, and never reopens setup.
• A settings file that can't be read is left alone at start-up instead of being overwritten.