My first experience with Lemmy was thinking that the UI was beautiful, and lemmy.ml (the first instance I looked at) was asking people not to join because they already had 1500 users and were struggling to scale.

1500 users just doesn’t seem like much, it seems like the type of load you could handle with a Raspberry Pi in a dusty corner.

Are the Lemmy servers struggling to scale because of the federation process / protocols?

Maybe I underestimate how much compute goes into hosting user generated content? Users generate very little text, but uploading pictures takes more space. Users are generating millions of bytes of content and it’s overloading computers that can handle billions of bytes with ease, what happened? Am I missing something here?

Or maybe the code is just inefficient?

Which brings me to the title’s question: Does Lemmy benefit from using Rust? None of the problems I can imagine are related to code execution speed.

If the federation process and protocols are inefficient, then everything is being built on sand. Popular protocols are hard to change. How often does the HTTP protocol change? Never. The language used for the code doesn’t matter in this case.

If the code is just inefficient, well, inefficient Rust is probably slower than efficient Python or JavaScript. Could the complexity of Rust have pushed the devs towards a simpler but less efficient solution that ends up being slower than garbage collected languages? I’m sure this has happened before, but I don’t know anything about the Lemmy code.

Or, again, maybe I’m just underestimating the amount of compute required to support 1500 users sharing a little bit of text and a few images?

  • Espi
    49•11 months ago

    I would say that it’s extremely unlikely.

    Websites in general are never limited by raw code execution, they are mostly limited by IO. Be that disk IO as files are read and written, database IO as you need to execute complex queries to gather all the data to build the user timeline, and network IO to transfer data to and from the user. For decentralized social media like Kbin or Lemmy its even more IO limited as each instance needs to go back and forth to other instances to keep up-to-date data.

    Websites usually benefit much more from caching and in-memory databases to keep frequently used data in fast storage.

    This is why simple, high level, object oriented, garbage collected languages have become so common. All the CPU performance penalties they incur don’t actually affect the website performance.

    • TortoiseWrathEnglish
      25•11 months ago

      Not relevant to lemmy (yet), but this does break down a bit at very large scales. (Source: am infra eng at YouTube.)

      System architecture (particularly storage) is certainly by far the largest contributor to web performance, but the language of choice and its execution environment can matter. It’s not so important when it’s the difference between using 51% and 50% of some server’s CPU or serving requests in 101 vs 100 ms, but when it’s the difference between running 5100 and 5000 servers or blocking threads for 101 vs 100 CPU-hours per second, you’ll feel it.

      Languages also build up cultures and ecosystems surrounding them that can lend themselves to certain architectural decisions that might not be beneficial. I think one of the major reasons they migrated the YouTube backend from Python to C++ isn’t really anything to do with the core languages themselves, but the fact that existing C++ libraries tend to be way more optimized than their Python equivalents, so we wouldn’t have to invest as much in developing more efficient libraries.

      • @terebatEnglish
        10•11 months ago

        It is fairly relevant to lemmy as is. Quite a few instances have ram constraints and are hitting swap. Consider how much worse it would be in python.

        Currently most of the issues are architectural and can be fixed with tweaking how certain things are done (i.e., image hosting on an object store instead of locally).

    • Baldur NilEnglish
      2•11 months ago

      That’s correct. I wonder if YouTube still uses Python to this day (seems like they migrated to C++?)

      Not saying there isn’t a difference in language performance, but for most world problems the architecture and algorithms matter more than the language for performance. Unless you’re in a very constrained environment such as lower end smartphones or embedded systems.

    • th3raid0rEnglish
      2•11 months ago

      In lemmy’s case, my perusal of the DB didn’t really suggest that the queries would be that complex and I suspect that moving it to a higher performance NoSQL DB might be possible, but I’d have to take a look at a few more queries to be sure.

      I wonder if this could be made to work with Aerospike Community Edition…

      Obviously it could be more effort than it’s worth though.

      • @terebatEnglish
        24•11 months ago

        The issues I’ve seen more are around images. Hosting the images on an object store (cloudflare r2, s3) and linking there would reduce a lot of federation bandwidth, as that’s probably cause higher ram/swap usage too.

        pict-rs supports storing in object stores, but when getting/serving images, it still serves through the instance as the bottleneck IIRC. That would do quite a bit to free up some resources and lower overall IO needed by the server.

      • hungrybreadEnglish
        9•11 months ago

        There’s no need to migrate the database, that shouldn’t be an issue at this size. Caching should be implemented as another comment suggested.

      • TortoiseWrathEnglish
        4•11 months ago

        Ehhhhhhh. Using a relational database for Lemmy was certainly a choice, but I don’t think it’s necessarily a bad one.

        Within Lemmy, by far the most expensive part of the database is going to be comment trees, and within the industry the consensus on the best database structure to represent these is… well, there isn’t one. The efficiency of this depends way more on how you implement it within a given database model than on the database model itself. Comment trees are actually a pretty difficult problem; you’ll notice a lot of platforms have limits on comment depth, and there’s a reason for that. Getting just one level of replies to work efficiently can be tricky, regardless of the choice of DBMS.

        Looking at the schema Lemmy uses, I see a couple opportunities to optimize it down the road. One of the first things I noticed is that comment replies don’t seem to be directly related back to the top-level post, meaning you’re restricted to a breadth-first search of the comment tree at serving time. Most comments will be at pretty shallow depths, so it sometimes makes sense to flatten the first few levels of this structure so you can get most relevant comments in a single query and rebuild the tree post-fetching. But this makes nomination (i.e. getting the “top 100” or whatever comments to show on your page) a lot more difficult, so it makes sense that it’s currently written the way it is.

        If it’s true (as another commenter said) that there’s no response caching for comment queries, that’s a much bigger opportunity for optimization than anything else in the database.