Why building a Rust LSP is hard | Hacker News Reader

Why building a Rust LSP is hard

rust-glancer.github.io

144 pointsagluszak77 comments

Reading this made me realize that I want two different things from an LSP that are sometimes at odds with each other:

1. Editing help

2. Reliable and comprehensive analysis

The editing help needs to deal with incomplete and inconsistent state and answers on a best effort basis. This is good when I'm writing code.

When I'm trying to understand code it usually is in a complete and compiling state but best effort is not enough. I expect complete and exhaustive answers.

Independently of that I'd love to read a similar analysis that compares the approaches of rust-anslyzer, rust-glancer and the JetBrains analysis engine in Rust Rover.

This is a good observation, with the observation that there are two different levels of latency requirements. Most of the time, I would be pretty happy with an untyped, unsemantic, mildly smart heuristic based ident completion while the asynchronous semantic one finishes.

The nice thing about this dual setup is that I tend to only want the semantic one if I'm thinking more, so there's naturally a larger time budget for it.

One must always think about the experience they want in UX first, rather than the tools they want to build.

I feel most LSPs I used are fairly good at both. Sometimes I get no completion suggestions, look at the sidebar and fix compile errors, then start getting suggestions. For whatever reason, it's a fine trade off for parsing errors.

A lot of things described in this article applicable not only for Rust, but for almost any language. Like it's obvious that requests should be handled asynchronously and that conversions from/to UTF-16 are needed. But it's actually not so hard.

I have written a language server for my language too. The hardest thing was to find a way allowing providing useful autocompletion for a document in edited state, when it's not syntactically-correct. This is the trickiest part how to deal with such incorrectness without missing all the context necessary.

I've not written a language server but have written a language plugin for IntelliJ.

I started with writing a correct recursive descent parser. I then extended it to detect, report, and recover from common syntax errors as I encountered them so that the parser is robust. And adding a parser test case for each of these (e.g. one test for each branch through an EBNF construction).

Some examples are:

1. missing keywords when the keyword can be detected from the current context (e.g. missing semicolon at the end of a statement);

2. using the wrong token (e.g. `:` instead of `::` in a C++ namespace qualified name);

3. detecting and ignoring whitespace in a whitespace-sensitive qualification (e.g. in XML QNames);

4. keeping in the prolog state (where functions are defined) when there are errors so that functions after the error don't get lost;

5. lexing incomplete literals like `10e` so they can be handled as integers in the parser and emitting an error for them.

> recover from common syntax errors

It's a dead-end. Sure, it can work in simple cases, but there will be always a case where such syntax recovery isn't possible. That's why relying only on syntax recovery isn't an option.

Because of that I use a different approach. I do parse on each document editing, but such parsing is guaranteed to produce valid results only up to the point with broken syntax, where editing usually takes place. Such parsing is enough to reconstruct location of the point where editing takes place (namespace/class/function) and to reconstruct local context (local variables declared prior to editing place). This allows to perform almost perfect autocompletion by suggesting global and local names available at the editing point. In order to provide proper suggestion of non-local names declared after the editing point, I do keep a structure for the most recent document state with valid syntax.

With features like "go to definition" I do the same. I store a hash-table with location to definition point mapping, but it's updated only from time to time and only if document syntax is valid. In order to be usable for cases with edits made after building such hash-table I just perform text-based position mapping using accumulated edit events.

Using a JSON TCP connection for what on Windows would be direct function calls (COM) or in Eclipse would be direct function calls between Java modules always felt a bit gross.

I agree, the only reason LSP exists as it does is for a world running on electron based applications. Plug-ins and application extensions are not new technology and they are nearly universally more efficient in the forms designed and used prior to 2005-ish. I understand why VSCode exists and why it is used so often by developers, but it is certainly a downgrade from more language specific options that could exist.

There are some arguments that do carry water in favor of using a client/server protocol transmitting JSON, in particular, the ability for a nearly complete decoupling of analysis of code and the displaying and editing of that code. Also, LSP was (to my knowledge) the first language/platform/usecase agnostic protocol intended for use in code editors.

I get why this is where a big chunk of developers have ended up, but I do bemoan the lost potential for a language to grab mindshare and popularity on the usability and performance of its tooling and developer experience via-a-vis a custom designed and hyper specific code editor. I mean, Rust and Elm received endless praise for their error messages as a massive boon to developer experience, so it is a facet of language design and implementation that can act as great advertisement. I just hate that the prevalence of LSP at this point precludes custom editors as a first choice in the current zeitgeist.

LSP helps with not having to develop the same tooling in each editor for each language.

This isn’t for Electron app only, but is also helpful for Emacs, vim or helix (especially when they lack said language plugins).

Now, some IDEs provide capabilities for a language far exceeding what can even be implemented in a LSP.

This comment makes no sense. You realize how much lsp is used in other editors like vim/neovim? The existence of lsp is exactly what lets custom editors to flourish. Any language/lsp client can add their own extensions too. Any new editor can come up with it's own way of doing things and implement it on top of lsp, as long as the lsp client supports those extra features then there's no problem. I don't know how you can get it so wrong.

They are not custom editors, they are all the same editor because they have to implement the same LSP protocol.

You don't know what an editor or LSP is, if you genuinely believe that.

It only makes sense in the context of host security and stability, as proven by plugin issues in those IDEs.

However, to come back to your point, there are much better high performance OS IPC mechanisms for inter process communication than sending JSON down through a TCP wire.

As others point out, Electron.

In order to have direct function calls, you need to load the plugin's code in your own memory. Which means you expose your own memory to the plugin. Any malicious/buggy plugin could wreak havoc on your program - even managed code doesn't solve that problem in general.

IPC is the natural solution to that - run a process in its own address space and communicate with it via I/O. TCP is not most optimal, but it is uniquous. Same thing with JSON.

In LSP, most of the processing happens in the LSP server anyway - communication overhead between client and server is negligible in comparison to it.

Fault isolation is good, but are the fault isolation benefits worth everything else though?

Windows COM does have a way to run the server out-of-process with a performance cost, and this fact is transparent to the client (if it only uses COM interfaces and not global variables or something). I assume you can even switch between the modes at runtime (with restarting the plugin).

Why not implement unix sockets too, since we want to avoid the ip/tcp stack? Now you have two platforms to support for no real gain. Also, LSP applications are shared between projects/users, how you centralize that if each system needs its own installed program and dependencies? Using networking is the path of least resistance and it's drawbacks are well understood between the ones that need to communicate, there's no point to implement IPC based communication.

I think this is a very reasonable question to ask. Language servers are certainly a significant source of slowdowns and editor unresponsiveness, and any time you're doing IPC, that's another source of latency and fragility.

My current Zed editor instance has 5 different language servers running, implemented in at least 3 different languages, some of which require a full VM runtime.

Could Zed (written in Rust) host a full .NET runtime to run the Roslyn language server for C#? Could an Electron-based editor? It feels a bit daunting. Plugins could be native shared libraries, but what happens when multiple language plugins all try to initialize huge process-wide runtimes like .NET or the JVM? Even multiple instances of the Python runtime are going to potentially be stepping on each others' toes.

IPC and separate processes is probably the pragmatic solution for now, even though I would love to live in a world where in-process was more of an option.

WASM could be the solution, but would lock language server authors into the subset of programming languages that can be compiled to WASM, and WASM still looks a lot like IPC in practice. (Zed already uses WASM components for its plugins.)

I think you can totally start a .NET Runtime for a plugin. There are windows explorer plugins written in .NET - I know because Raymond Chen analyzed how they broke things :)

If you have freedom then you have the the freedom to make mistakes. In Windows globals are tightly scoped to a DLL and cannot accidentally cross over, so multiple python interpreters are no problem if the python interpreter is statically linked in each python plugin.

It’s just much easier for an editor to manage a single interface for plugins over having to host a bunch of very large runtimes.

Also, I doubt that you would have a very great time running both .NET and JVM in the same process. For one, both of them install signal handlers, and both of them do funky things with threads.

They are not mutually exclusive.

Java has had plugin isolation libraries for decades.

Most of which run the code in a subprocess and use IPC.

Not in my experience

Care to provide a counter-example?

TCP isn't really necessary. Usually a language server just uses stdin/stdout, which are just pipes.

And overall overhead for running a language server in a separate process isn't that big. LSP is designed in such a way that only minimal amount of information is needed to be passed, like edits or short responses. Packing/unpacking JSONs isn't a bottleneck, the heaviest job like program analysis is done in the language server itself without interprocess communication overhead involved.

The particulars of Rust make this a little more difficult, I think. There’s a certain tension between making your language more concise and adding useful redundancies, and Rust has generally gone to the “concise” side, with some redundancies that can make the tooling a little more painful. Like with imports.

  impl std::fmt::Display for Blah {
  }

If your language makes you qualify your imports (like above) then your LSP can, delightfully, still reliably do certain ops like renaming, even when chunks of your project aren’t parsing. But if you glob import std::fmt, and glob import something else, you are fucked. Display could come from anywhere (maybe from a module that has a parse error at the moment). I really appreciate languages where glob imports (or their equivalent) are either disallowed entirely or where typical code doesn’t use it.

Meanwhile, if you add a new file, there’s this little dance where you say:

  mod mycoolmod;

And then you create mycoolmod.rs. Or you do it the other way around. A little redundancy (the file exists and it is declared), that seems to just create a little friction in the LSP because mycoolmod doesn’t get a working LSP until it’s declared in the parent (you have to create both, and then you get a transient diagnostic that your module is unused for a while yet). A small issue, just another little bit of friction in the tooling of Rust that has nothing to do with the type system.

Having worked with rust for nearly two years now (granted, on one team with agreed-upon standards):

- glob imports are rare in my experience, less for the LSP’s sake and more for code self-documentation

- the `mod foo` line exists because omitting it cannot fall back to a reasonable default visibility level (`pub`/`pub(crate)`/<none> (private))

Of course it could just default to private, and if you wanted something you'd have to specify that.

At this point...

- I don't think that private-default is a universally (or more importantly, widely) reasonable default

- It adds another case where Rust is unopinionated and gives the user another way to do a basic function.

I like Rust for its unopionation, generally, but I think that the returns for this kind of thing diminish as you get to really boilerplate stuff.

I'll argue the default already is private.

  mod foo
  pub mod foo
  pub(crate) mod foo

clearly when you leave the visibility modifier out, the default is private.

Wildcard imports are a horrible idea.

(And anyone pushing to have a "prelude" for their library is making it worse. Please don't.)

Clippy `wildcard_imports = "warn"` is your friend.

Tangentially, since I now write a lot of Zig and there’s no official LSP, I decided to see what life would be without an LSP for all my projects (Go, TS, Python).

Honestly, my LSP requirements are minimal: go to def, find all references, and symbol search. I’ve been wanting to try ctags for awhile and finally made the switch.

Another consideration were all the posts about how grep is better than LSPs for LLMs.

It really pushed me to just get better at grep. There have been a ton of benefits: I can symbol search outside of my editor, it’s much faster than an LSP, minimal configuration. It feels more like a universal tool I can use instead of a tool for IDEs.

I do like that vscode understands paths with line numbers in the terminal, so grep -n then ctrl+click takes you to that line.

It's faster than an lsp?? What on earth are you doing? Even just the context switch to typing the right grep command is slower than ctrl-g or whatever.

LSP load times can be brutal (or crash entirely) and I have a shortcut to grep the word under the cursor.

Ctags is a symbol indexing system so I run it once for a project and can do symbol searches in my editor or using fzf.

I have ctags index the languages stdlib as well which is really handy for quick symbol searching -> open my editor to read more

emacs has dumb-jump https://github.com/jacktasia/dumb-jump exactly for this, and is often good enough. it uses ripgrep (or some such) for faster-than-grep queries.

Enter hell: LSP assumes that it's the source of truth, but you still need to access the filesystem yourself, and do it in a synchronized way

LSP is an example of utterly horrid technical design.

Stop letting Microsoft design protocols and APIs. They are so. bad. at. it.

I've never looked into LSP, how would you design it?

Well for one I wouldn't design it so both the LSP and the editor need a synchronized view of the underlying file.

Start with synchronous function calls instead of JSON. Microsoft knows how to do that - they invented COM and OLE. Function calls enable whatever data sharing is necessary to maintain a coherent view. Imagine trying to do OLE with JSON - just wouldn't work. (Does OLE still exist?)

Stupid idea.

So the LSP crashes and/or goes into a runaway memory consumption loop. And your main application dies with it.

Or what if you want, you know, to be able to use the same LSP from TWO different applications at the same time?

Never mind issues with other managed runtimes not expecting to deal with something else in their address space.

Moreover, it's not just stupid, it also does not actually solve _anything_.

A synchronous function can also have obsolete indexing information if it races with the code updates.

But it doesn't have to, because you call AboutToUpdateCode and the LSP doesn't return until it's safe, and then you update code and call DoneUpdatingCode.

Yes, it does because the physical code update is asynchronous. Or you'd have to do a roundtrip for every little edit, resulting in unusably slow experience.

I don't know what you're talking about. What is "physical code"? A function call roundtrip costs nothing, that's the reason we are saying it should be function calls.

TCP roundtrip overhead over localhost costs approximately nothing.

But running the LSP itself is slow. It's not instant, and you can't run it in a blocking way for each keypress.

A function call costs nanoseconds. You can sprinkle them literally everywhere. A JSON-formatted TCP round trip to localhost costs, I'd guess conservatively, at least ten microseconds, 1000x slower.

OK. Here's my simple implementation of an LSP:

> func doLspStuff() { sleep(60000); }

Now sprinkle them after each keypress. Do you see the issue now?

I feel like MS actually learned their lesson with synchronously integrating the language intelligence into the IDE. Old versions of VS would hang or crash based on bugs in the language tooling trying to provide intellisense. You'd restart and it'd work fine till you hit some other weird edge case. Generally this settled to a level of rare-bugginess where you were happy enough with the advantages not to go back to Emacs/VIM, but still annoyed at the occasional restart needed.

In no way does this mean LSP is a perfect solution, but anything synchronous would be a step backwards.

That doesn't really fix anything. The filesystem is fundamentally a racy shared data structure. If you made the entrypoint API synchronous, any half-way decent editor would shove the LSP queries to the synchronous API into a different thread, because "Do Not Block the UI Thread" is a fundamental principle of good UI programming.

Once you get past that, JSON-over-TCP is just another kind of asynchronous RPC mechanism, one that has the advantage that you can build it in just about any language with out-of-the-box tools. Trying to make a plugin system or a COM or CORBA or OLE based system really cuts out the ability to build language servers in most languages, because you have to be able to build the code in just the right way.

Big assumption that the editor, which is required to make calls synchronously, would make them on a different thread. Isn't correctness more important than avoiding plugin calls in the UI thread? The good thing about having an actual function interface is that you can tailor each one to its actual needs, instead of insisting high-overhead async is okay for everything. It's like GDI vs X11.

I don't know about you, but I do not want to block on every keystroke. Program analysis can take time, asynchronous feedback is a natural fit. None of the features I use from LSP should be synchronous.

I can almost guarantee you that a synchronous API would yield a much more complicated design just because many operations can not be expected to reliably work within a few milliseconds.

You want to rename something across multiple files? Well now you add file system overhead and blow straight past reliable frame timings. Good luck waiting for that operation to finish. Of course you could make the API beginRename, and queryRename, or whatever you fancy to see if the operation was successful, but now you're back to what you wanted to avoid: an asynchronous API. Do mind that the example is actually one of the better cases, as many things you might want to do with a codebase are actually more expensive. You will feel the hiccups from waiting in the UI thread and you will loathe the program for it.

Why do you assume renaming must be completed when Rename returns? Rather, the LSP must have a consisten snapshot when AboutToSaveFile returns.

I assume that the result of the rename operation must be known. I.e. if it was successful or not, or which files were affected. That implies 2 function calls. One to start the operation, and one to query its result. That is, by definition, already an asynchronous interface. The only other (i.e. synchonous) way to achieve remote procedure calls is by blocking until it's finished.

If you insist on showing the result of a rename on the next frame, then yes, it must block the next frame until it's done.

Would you write your editor that way if it didn't use LSP but was specific to one language? If not, but you would write it that way with a language plugin, why? You still have to write a good product. You can't blame the plugin API for everything you get wrong.

Actually, COM is inspired by DCE/RPC and the initial versions had some similarities.

Parallel to that, IBM had SOM on OS/2, which was even better allowing for metaclasses and proper class inheritance, it was the key mechanism between Smalltalk and C++ on OS/2, where Smalltalk enjoyed a role similar to .NET on Windows nowadays.

OLE naturally still exists when using Office natively on Windows, other vendors seem to have forgotten about it.

COM's role on Windows has grown since Vista, and the Windows team redid many of the Longhorn ideas originally implemented in .NET into COM/C++, with WinRT being an evolution of COM.

I've never heard of anyone calling SOM 'better' at anything this century. I came across it when it was the foundation for OpenDoc at Apple, your comment brought back many bad memories

Compared with COM's design, it was much better.

Classes and metaclasses sound like architecture astronomy here. COM is fundamentally just a standard ABI and a way to look up DLL paths in the registry. There are more layers on top of that, but that's the basics and all you need for a plugin architecture.

That applies to SOM just as it does to COM.

Interesting to complain about one and then ignore exactly the same boilerplate for the other.

Leaving aside the fact that Microsoft's tooling for COM has already had multiple reboots between VB OCX, MFC, .NET Framework RCW/CCW, .NET ComWrappers, ATL, WRL, WIL, WinRT, each one with its share of astronomy.

No one uses the bare bones vtbl and nothing else.

You assume I know anything about SOM.

Everything uses the bare bones vtbl. That's how there can be so many things, they're all using the bare bones vtbl. Obviously there's application specific stuff as well because a bare bones vtbl is not useful by itself but that's like saying if TCP is so good why is there HTTP?

Because I thought this was a further discussion and did not notice another username, sorry about that.

The point being that beyond the plain vtbl, SOM was much better designed than COM, in the sense of overall architecture for application development.

> Start with synchronous function calls instead of JSON.

LSP is a function call protocol. So that is already done?

Wouldn't a normalized protocol based on AST vocabulary and tree operations make more sense ?

That would require the editor to be able to build an AST - defeating the point of LSP, because now the editor needs to know every possible language again

I meant the language server builds the normalized, lossless AST and gives it to the editor. Changes in the AST can be provided as events/diffs to the editor. The normalized, source-mapped AST is effectively the stable language protocol. (Possibly given in a columnar, database form via Apache Arrow, but thats impl specific). Was discussing alternative architectures with a collague since even the best language servers really suck on large projects.

One of the largest issues of LSP is that each language implementation needs to do everything separately, and that means each language will work differently.

Imagine if all the editing tools in microsoft word were specific to the language you used, and if you mixed German and English, each had different tooling.

Now if you mix Rust, HTML, JS, and CSS, they'll all have separate tooling, seperate "go to definition", and separate refactoring. Worst of all, none of them can see the definitions of the other ones.

So what you'd actually want to do is parse each language into your AST, with proper annotations as to what is what, and have all the "go to definition", the UI rendering, highlighting, refactoring, etc all done generically by the IDE ontop.

Which is much much closer to how Jetbrains IntelliJ does it, and why their tooling can handle "find usages" on an HTML element to find matching querySelector in .js files and matching selectors in .css

And if you subscribe to the AI stuff, you'd also want your AI to operate on this AST so it can learn skills that generalize across all languages.

> Stop letting Microsoft design protocols and APIs.

Unfortunately nobody else stepped up to do it.

I'm glad that the code editors out there didn't wait for your theoretical better designed protocol and decided to adopt LSP. Otherwise we'd still have editor that only support one language properly, and the rest is treated like text. If the price to pay is that it sucks for the handful of people who have to work with it, so be it. For every LSP developer that suffers there are tens of thousands of downstream users who benefit from better language support in their favorite editor!

we'd still have editor that only support one language properly

Emacs supported zillions of languages before LSP.

Zoom out.

As someone that got introduced to it in 1996, and was a fan of XEmacs variant, still remembers enough keybindings and Elisp, supported beyond plain syntax highlighting was very much hit and miss, even nowadays.

I have been an emacs user for over a decade now, and if you think that language “support” was as good before LSP, I have a bridge to sell you.

If the old emacs approach is so much better, why is emacs switching to using LSP?

It isn't switching. It added support for it. Because that's what emacs does; it's the kitchen sink. Its mission in life is to do absolutely everything. That includes speaking shitty protocols.

https://xkcd.com/1172/

But the effort to do so was then entirely duplicated for other editors. LSP turns an N*M effort into an N+M effort, the upshot of which being far more complete coverage for both editors and languages.

LSP is not that bad. They tried to decouple IDE features from language features.

Resolving file paths is not trivial because it's language specific.

Syncing source changes on the LSP server side is necessary to keep packet side small and patching is trivial (span, new text).

(I made an LSP-powered text editor and implemented LSP client from scratch)

> Why building a Rust LSP is hard

Not as hard as understanding what LSP means, apparently.

I was interested until I saw this project is simply an LSP server.

Reply on news.ycombinator.com