Hello, world! · Rust Glancer

I want to present a project that I've been working on for the past 4 months: an alternative Rust LSP implementation that is built with a focus on low memory usage. It has two main features: It can use very little memory (target <100mb for reasonable projects). There are caveats, these are described below. It allows immediate indexing after restart: if your project was indexed, restarting the editor will not require re-indexing. Your browser does not support embedded videos. You can download the recording instead. Note: throughout this video, the used RAM remained under 100mb These features make Rust Glancer suitable for the older computers: I have tested it on my old MacBook Pro M1 2020 with 8GB RAM, and it was pretty good. MachineLSPBase indexing (engine usable)Full indexing MacBook Pro M4 Max, 36GB (2025)Rust Glancer5 seconds8 seconds MacBook Pro M4 Max, 36GB (2025)rust-analyzer6 seconds13 seconds MacBook Pro M1, 8GB (2020)Rust Glancer6 seconds9 seconds MacBook Pro M1, 8GB (2020)rust-analyzer7 seconds14 seconds As you can imagine, 4 months is not a lot of time for a project as big as a Rust LSP. Rust Glancer is not a complete LSP yet, it has a lot of missing functionality, it has some known bugs, and it has a lot of things I want to improve. At the same time, it is already pretty capable: it has a full indexing pipeline with type inference and a trait solver (chalk), most of the "normal" Rust syntax is supported, and most of the "normal" LSP actions do work as well: goto definition, hover, inlay hints, completions, you name it. If you are interested, you can already try it out: just install the VS Code extension here, or, if you prefer, build and install the vsix from the repository. The rest of the post contains the history of the project: motivation, LLM use, plans and roadmap. If you're not interested, you might want to check out the project documentation instead. Difference with rust-analyzer There are several reasons why rust-analyzer consumes a lot of memory: Rust workspaces genuinely have a lot of information that must be indexed: thousands of functions, structures, traits, relationships between these, function bodies and statements in them, etc. Each of these needs to be analyzed and remembered, and you can't really cheat if you want to have things like "find all references to this structure". rust-analyzer uses salsa as its database. It's an incremental query-based database, which lazily computes all the data you need without having to explicitly "record" everything. It is a very cool approach, but it's inherently tied to memory, which makes it hard to move parts of data from memory elsewhere. rust-analyzer uses rowan for syntax tree representation. The cool property here is that it allows partial invalidation: if only a part of the file changed, only the relevant bits have to be reparsed, which makes it faster than having to re-parse the whole file on each keystroke. However, the tree-like representation inside of it can cause heavy memory fragmentation (meaning that the amount of RAM taken from the OS is higher than the amount of "actually used" RAM). (1) is something we have to live with (though there are a few optimizations we can do there which Rust Glancer does), but (2) and (3) are the consequences of the rust-analyzer architecture. rust-analyzer chose them to make the LSP faster, and it does work for that purpose. The idea I had when I started the project: what if we don't try to make an incremental LSP? What if all we have is a frozen analysis result that gets invalidated on save? It obviously will not be as fast as rust-analyzer, but it will give us the properties we seek: analysis results can be offloaded to the filesystem and loaded to memory only when they are actually needed. saved analysis is reusable, and since it's already offloaded to the filesystem, it can be reused after the editor restart. This is the core idea of Rust Glancer. It indexes the workspace once and preserves results in the filesystem, and then whenever queries need something, they can load the required information for the duration of the query. It doesn't come for free though: frozen workspace analysis is slower than lazy incremental by definition, since loading and deserializing data from filesystem is slower than loading from memory. To mitigate that, Rust Glancer has to use some tricks: for example, when you type, it doesn't perform full blown analysis on each keystroke, it instead attempts shallow analysis of the current body and reuses the previous complete index. This makes completions reasonably fast, but it also means that new items (imports, structures, traits) are not "indexed" until you save the document. Which, hopefully, should not be a problem: you really get used to it fast, and at least in my case it does not feel overly wrong after a while. If that sounds scary, I suggest to just try it, it really is not. For people who rely on agentic workflows, Rust Glancer is also optimized for large amount of out-of-editor changes. I'm not sure why, but in rust-analyzer I've observed that when agents edit the code, inlay hints can get out of place, and I had the same problem in Rust Glancer initially, but it was resolved by implementing a custom file watcher and tweaking it somewhat. The server also has lower priority for out-of-editor changes, so agentic changes do not cause rapid re-indexing. Still, it's important to understand that Rust Glancer has some benefits, but also has some drawbacks (besides being incomplete, obviously) compared to rust-analyzer. Maybe I will manage to solve some of them eventually, but it's highly unlikely that Rust Glancer will ever become "just like rust-analyzer, but better". I imagine that rust-analyzer will remain the default choice for projects that care about completeness and keystroke accuracy, while Rust Glancer will work for people with weaker machines or people who are ready for some sacrifices to reduce RAM usage. How and why it happened I have been writing Rust professionally for ~7 years, and since pretty early on I started observing how the compiler and its tooling are developed. I've made some contributions to rustc, clippy, and rust-analyzer, and I've spent dozens of hours reading its source code just to teach myself. So I was pretty much aware how big of a project a Rust LSP is. At the same time, I have a love-hate relationship with rust-analyzer. It is absolutely beautiful except for two things: memory usage and initial indexing (especially with build scripts / proc macros enabled). These problems seem to be brought up quite a lot, but in my case they are even more drastic: I have a rather stupid workflow where I have two identical IDEs open on two displays with a bunch of projects inside a workspace. So the memory consumption is roughly 2N, and with my last set of the projects I had to work on, rust analyzer was consuming 16GB of memory that I, ugh, would prefer to have available for other uses; not to mention that each time I opened VS Code, my PC fans would go brr because of a ton of parallel indexing jobs. At some point I thought that I am fairly confident in my Rust knowledge, so I probably don't need a full-blown LSP, and can use something simpler and more memory efficient. I decided to try building a "smart ctags for Rust". I very explicitly did not want to build an alternative LSP, because of how insane of a task it is. Little did I know... The initial progress was going pretty smoothly: I made use of rust-analyzer's syntax library, lowered items to internal representations, then built definition maps and module structure, got all the declarations indexed. It was so surprisingly straightforward that I decided to do some primitive body lowering. Then I decided to add very very simple type propagation. Then it turned out that naive type propagation doesn't give me much -- but I already had these nice inlay hints, so I wanted more. Overall, I don't care about complex cases and nightly features, right? (Right?...). So then came naive trait resolving via impl header matching. It's quite addictive, you get it. The illusion, however, broke when I decided that it is pretty reasonable to expect the following code to be supported as well: fn mul_by_two(vals: &[u8]) -> Vec<u8> { vals.iter().copied().map(|v| v * 2).collect() } The code is pretty simple, but in order to support it we need: Slice type support Closures / Fn traits Trait solving Associated type projection A bunch of nightly stuff the last item is funny: I wanted to avoid nightly, but I somehow didn't think that std (or sysroot in general) breathes nightly. Welp. So all in all, one feature after another, I slowly was getting from "smart ctags" to a "real LSP". Probably, the three biggest milestones were: Declarative macro expansion (I hate declarative macros now). Thankfully, I was able to reuse most of rust-analyzer's infrastructure for that. Proper type inference engine. It was a big "oh wow" moment when I truly realized how type inference works (in short: we "link" all related type bindings in a big inference table, and then we try to get evidence from all possible places, where providing evidence can solve types for multiple places). It was the moment that probably brought me the most joy during the work on this project so far. Proper trait solving engine. I initially wrote "it's highly unlikely that we will have a trait solver in this project", but then I really wanted to get the abovementioned iterator example to work properly. I resisted integrating trait solver for a while, trying to have naive hacks like naive trait impl matching + specialized handlers for std traits, but it was getting more and more complex while working pretty poorly. Then I gave up and integrated Chalk, which turned out to be significantly simpler than the whole hierarchy I have built. Making Chalk fast was another challenge, though.