-
Notifications
You must be signed in to change notification settings - Fork 7
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
[Feature] Implement image cache max-sizing #62
Comments
You are referring to an on-disk cache, right? I'll add this to milestone v1.0.0 as I don't want users to someday find we have been eating all their data. :-) |
No. The What I would like to do -- based on the recent discussion about the performance issues you're seeing -- is:
I don't bellieve an LRU will result in fewer cache misses, because I doubt user behavior will be that stochastically predictable. I suspect users will either load several songs from an album and then move on, or be entirely random; in either case, LRU won't provide much benefit, and LRUs require a bit of code complexity and additional data storage (for time stamps). |
Thinking of that, I haven't tried running stmps on my Raspberry Pi with 512 MB RAM yet. $ ls ~/.cache/supersonic/*
artistimages covers
$ ls -1 ~/.cache/supersonic/**/*.jpg | wc -l
9687
$ du -hs ~/.cache/supersonic
78M ~/.cache/supersonic Granted, we don't need to cache 10k cover files - or should we? We could effectively get rid of a lot of network requests, which would be relevant if I were connecting to my Subsonic server remotely through a VPN. Just food for discussion, I think. |
What's your chat platform of choice? IRC? Matrix? I don't do Discord. We should have a chat room, though, even if we're not often there at the same time. Moving the cache to disk would certainly save memory, and allow larger cache sizes. On almost-embedded devices, we still probably don't want to be disk-caching large amounts of data, as they tend to also be constrained on persistent storage as well. But we could cache more, and put less load. I will point out that keeping local persistent data caches in sync with server data tends to get complex pretty quickly, with logic necessary to evaluate whether the cache is out of date. Consider when a user changes the album art on the server: AFAICT, the Subsonic API provides no I'm not saying it's a bad idea; it may even be mandatory if we want to provide support for micro-computers with highly constrained resources. An alternative would be to allow the user to disable album art loading entirely -- which could be accomplished by letting them set the max-cache-size configurable option on the as-yet-to-be-implemented cache cleaning feature to 0. If 0, never load images. I'll go on record as saying I'm bearish on disk caching. In one of my other projects, which is a mobile app and uses persistent caching heavily, the caching logic (storage, in/validation, refreshing) constitutes a large percentage of the code base, is surprisingly challenging to have function correctly, and requires good API support for timestamps -- which I believe is lacking in the Subsonic API. |
I don't really like chats for non-realtime development to be honest. It's yet another window/tab to keep open.
Ok, so the problem here is complexity. Is guess then it's a bad idea for me to say that we then should probably offer both and make it configurable. So to summarize:
|
No sweat. I haven't quite come to terms with github's discussions; I can't seem to get it integrated, and often lose track of threads. Ticket comments work fine.
I mean, I think it's fine to think we should do a thing; I'm just cautious about the complexity. If it's necessary, it's necessary. I do think we should discuss the trade-offs, though, and options like simply allowing users to disable images. It's simple to implement and maintain, completely sidesteps any network latency and memory issues that might be encountered on micro-computers, and has only the consequence that a feature of questionable value isn't available. I mean, it's not like cover art in a TUI tool is going to be anyone's mandatory requirement.
So my assumption that constrained devices are going to be constrained in storage is probably a bad one. Is it OK if we put persistent caching in a future ticket? I really think we're going to encounter some hard limitations when it comes to cache invalidation -- limitations that we have already with an in-memory cache, but which disappear with quit/load, and which I was already planing on further restricting with cache invalidation based on queue contents. We'll have to nail down assumptions and expected behaviors before implementing a persistent cache. |
That will work. Though for persistent caching I would probably use https://github.com/gokyle/filecache and a simple added-on pruning logic (maybe check |
That library is a good idea. The troublesome part will be ensuring that the cached images are in sync with the server. I wonder if Navidrome (and gonic, and all of the others) change the asset ID if the user replaces the file on disk & rescans the library? Like, the image will always be called "cover.jpg" or "album.jpg" or something like that, so the name won't change; but if the data stored in the file is changed, will Navidrome give it a new ID? If that's a reliable behavior, we could use that to detect, invalidate, and re-fetch cached assets. |
add: Makefile test task runs tests & linting add: Makefile generates changelog chore: spezifisch#62, adds a changelog docs: Adds instructions & dependency list for Makefile
The current album art cache can grow indefinitely. This feature would add cache management, such that a configurable maximum number of cache items are held at any time.
Two algorithms occur to me:
LRU
When the cache size exceeds the maximum, the algorithm should purge images for songs that firstly are not in the current queue, and secondly, are the oldest images in the cache.
Unused
Simply, when images are no longer referenced by any song in the queue, remove the image from the cache.
I think the second is more simple to implement (we would not need to consider time stamps), and would probably be sufficient. There's no reason to expect that LRU would prevent re-fetches.
The text was updated successfully, but these errors were encountered: