Repository navigation
Notes from running Wholphin behind a ranking/augmentation proxy #1848
Replies: 2 comments 5 replies
Do you know what version(s) of Wholphin? In the past few versions, I'd expect only child count (which will be removed by #1852) and user data to be included on grid pages. I think maybe older versions might have also used recursive querying which would have been slower as well.
I'm not sure I follow this
I think this change was done because Jellyfin's default search ordering isn't great. But I guess this also clobbers search plugins that a better. This definitely needs to be reviewed again, especially using Jellyfin 12. |
|
Quick update, and an answer to your version question. Back then I was running this against Jellyfin 10.11.11, where the collections grid was slow but fine. I have since upgraded the server to 12.1, and the collections story changed enough that it is worth writing down here. On 12.1 the grid does not just get slow, it OOM-kills the server. The 12.0 migration normalized BoxSet membership into a The relevant part for Wholphin: the field-stripping that fixed the 10.11 slowness does not help here, and I do not think #1852 does either. I tried trimming the grid to a minimal field set and it still OOMs. The cost is not the ChildCount / MediaStreams / Trickplay fields, it is the BoxSet-to-LinkedChildren join multiplied by the image include, and a poster grid still has to enable images to get the tags. So this is a server bug, not something a client field change can dodge. I filed it upstream with a repro and a build-tested one-line fix ( Until that lands, my proxy no longer strips fields for collections. It bypasses Jellyfin's BoxSet query entirely: it synthesizes the grid from my own database, serves the boxset posters/backdrops from there, and answers "open a collection" by resolving the members to movie/show ids and fetching those (which do not cartesian). Not something most people would want to build, but it is the only way I found to keep the collections page usable on 12.x with a large collection count. Mostly flagging this as a heads-up: any Wholphin user on 12.x with a lot of collections will hit a hard server crash on the collections page, and there is nothing obvious to do client-side about a server OOM. If Wholphin ever wanted to be defensive about it, the only lever I can think of is not issuing the whole-library BoxSet enumeration at all on 12.x until the server fix ships, but that is your call and it may not be worth special-casing. |
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
I run Wholphin as the main TV client against a Jellyfin server for myself, friends, and relatives - currently fronted by Caddy plus a small shim service, and I've been able to fix a surprising number of client-visible things entirely server-side; no client changes, works for every client we run. Sharing what worked and where I hit the limits of the approach, partly as roadmap signal and partly because some of it may interest you.
Fixed transparently at the proxy
Collections grid, 30 seconds to 100ms. I have 150 very large custom collections. This is the reason I first installed Caddy. My users were avoiding these powerful collections for slowness, and I was trying to standardize on Wholphin.
Jellyfin itself isn't the bottleneck — jellyfin-web renders the same collections instantly. The difference is query shape: Wholphin's BoxSet grid request asks for ChildCount, MediaStreams, MediaSources, Trickplay, Chapters plus per-item user data, and that combination is what costs 20–30s. Stripping those fields and appending EnableUserData=false at the proxy made the grid as fast as web. Client-side equivalent: a slim field set for grids, hydrate on focus.
Recently Added / Latest Released rows. The rows are built from raw DateCreated / MaxPremiereDate sort keys, so one season drop eats an entire row. I recognize those queries by shape and serve curated, series-grain orderings instead.
Fuzzy search. /Search/Hints and /Items?searchTerm= get translated into a trigram-ranked Ids fetch (Postgres pg_trgm over my own catalog), and Jellyfin materializes the authentic items — any client gets ranked results this way.
Where the proxy runs out of road — needs the client
Search ordering. SearchPage re-sorts every result set with SearchRelevance.score() + alphabetical tiebreak, which discards the server's ranking. Concrete example: searching "Station 11" — my server returns "Station Eleven" first, but "11" isn't a substring of "eleven" and Levenshtein lands at 0.571, so it scores 400 alongside every unrelated match and alphabetizes to ~15 cards deep, while the Jellyseerr Discover row (alias-aware) shows it immediately. Typing station eleven scores 0 and it's first. A response-index tiebreak instead of sortName would fix this without giving up the scorer's wins on stock servers. Happy to file it as a [FEA] with a PR if you'd like it tracked.
People in search (I see this is already #1750 / #646). The missing piece is only the query + row: Jellyfin's /Items search can't return people, but /Persons?searchTerm= can. My shim serves fuzzy people results on that endpoint and it works in jellyfin-web today. Since PersonPage, people-in-favorites, and people home rows all already exist, a People row in search plugs straight into navigation you've already built. Glad to add these details on #1750 if useful.
There are a few more in both categories — happy to share Caddyfile snippets, shim details, and anything else - if any of this is interesting. Love your client and willing to help in any way I can.
All reactions