Cadey is coffee
<Cadey> Hello! Thank you for visiting my website. You seem to be using an ad-blocker. I understand why you do this, but I'd really appreciate if it you would turn it off for my website. These ads help pay for running the website and are done by Ethical Ads. I do not receive detailed analytics on the ads and from what I understand neither does Ethical Ads. If you don't want to disable your ad blocker, please consider donating on Patreon or sending some extra cash to xeiaso.eth or 0xeA223Ca8968Ca59e0Bc79Ba331c2F6f636A3fB82. It helps fund the website's hosting bills and pay for the expensive technical editor that I use for my longer articles. Thanks and be well!

This isn't the way to speed up Rust compile times

Read time in minutes: 35

Cadey is coffee
<Cadey> By the way, this problem was fixed with the release of serde v1.0.185. Please enjoy the technical overview of the problem and problem space regardless.

Mara is happy
<Mara> You could have read this a week ago if you subscribed to Patreon.

Recently serde, one of the most popular Rust libraries made a decision that supposedly sped up compile times by using a precompiled version of a procedural macro instead of compiling it on the fly. Like any technical decision, there are tradeoffs and advantages to everything. I don't think the inherent ecosystem risks in slinging around precompiled binaries are worth the build speed advantages, and in this article I'm going to cover all of the moving parts for this space.

hero image waifu-perch
Anything v3 -- 1girl, green hair, green eyes, smile, hoodie, skirt, river, bridge in the distance, long hair, wearing uggs, summer, beach, space needle, crabs, -amputee


serde is one of the biggest libraries in the Rust ecosystem. It provides the tooling for serializing and deserializing (ser/de) arbitrary data structures into arbitrary formats. The main difference between serde and other approaches is that serde doesn't prefer an individual encoding format. Compare this struct in Rust vs the equivalent struct in Go:

#[derive(Debug, Deserialize, Eq, PartialEq, Clone, Serialize)]
pub struct WebMention {
    pub source: String,
    pub title: Option<String>,
type WebMention struct {
    Source string  `json:"source"`
    Title  *string `json:"title"`

Besides syntax, the main difference is in how the serialization/deserialization works. In Go the encoding/json package uses runtime reflection to parse the structure metadata. This does work, but it's expensive compared to having all that information already there.

The way serde works is by having an implementation of Deserialize or Serialize on the data types you want to encode or decode. This effectively pushes all of the data that is normally inspected at runtime with reflection into compile-time data. In the process, this makes the code run a little bit faster and more deterministically, but at the cost of adding some time at compile time to determine that reflection data up front.

I think this is a fair tradeoff due to the fundamental improvements in developer experience. In Go, you have to declare the encoding/decoding rules for every codec individually. This can lead to stuctures that look like this:

type WebMention struct {
    Source string  `json:"source" yaml:"source" toml:"source"`
    Title  *string `json:"title" yaml:"title" toml:"source"`

Aoi is wut
<Aoi> Hey, in your code you have the struct tag toml:"source" defined on the Title field, didn't you mean to say toml:"title"?
Cadey is coffee
<Cadey> Good catch! The fact that you have to declare the same thing over and over again makes it ripe for messing things up in annoyingly trivial ways. It would be so much better if this was all declared once. Here's the correct way to tag this struct:

type WebMention struct {
    Source string  `json:"source" yaml:"source" toml:"source"`
    Title  *string `json:"title" yaml:"title" toml:"title"`

This becomes unwieldy and can make your code harder to read. Some codecs get around this by reading and using the same tag rules that encoding/json does, but the Rust equivalent works for any codec that can be serialized into or deserialized from. That same WebMention struct works with JSON, YAML, TOML, msgpack, or anything else you can imagine. serde is one of the most used packages for a reason: it's so convenient and widespread that it's widely seen as being effectively in the standard library.

If you need to add additional behavior such as parsing a string to markdown, you can do that with your own implementation of the Deserialize trait. I do this with the VODs pages in order to define my stream VOD information in configuration. The markdown inside strings compiles to the HTML you see on the VOD page, including the embedded video on XeDN. This is incredibly valuable to me and something I really want to keep doing until I figure out how to switch my site to using something like contentlayer and MDX.

Mara is hacker
<Mara> All of the VOD information is stored in Dhall (read: JSON with imports, functions, and static types) files. There's Dhall support for serde with serde_dhall, which this website's code uses heavily for all of the static data. This includes signalboost entries, salary transparency data, blog series metadata, and the character information sheets.

The downsides

It's not all sunshine, puppies and roses though. The main downside to the serde approach is the fact that it relies on a procedural macro. Procedural macros are effectively lisp-style "syntax hygenic" macros. Effectively you can view them as a function that takes in some syntax, does stuff to it, and then returns the result to be compiled in the program.

This is how it can derive the serialization/deserialization code, it takes the tokens that make up the struct type, walks through the fields, and inserts the correct serialization or deserialization code so that you can construct values correctly. If it doesn't know how to deal with a given type, it will blow up at compile-time, meaning that you may need to resort to increasingly annoying hacks to get things working.

Cadey is coffee
<Cadey> Pedantically, this whole support works at the language token level, not at the type level. You need to write wrappers around remote types in order to add serde support because proc macros don't have access to the tokens that make up other type definitions. You could do all of this at compile time in theory with a perfectly spherical compiler that supports type-level metaprogramming, but the Rust compiler of today can't do that.

When you write your own procedural macro, you create a separate crate for this. This separate crate is compiled against a special set of libraries that allow it to take tokens from the Rust compiler and emit tokens back to the rust compiler. These compiled proc macros are run as dynamic libraries inside invocations of the Rust compiler. This means that proc macros can do anything as the permissions of the Rust compiler, including crashing the compiler, stealing your SSH key and uploading it to a remote server, running arbitrary commands with sudo power, and much more.

Mara is hacker
<Mara> To be fair, most people do use this power for good. The library sqlx will allow you to check your query syntax against an actual database to ensure that your syntax is correct (and so they don't have to implement a compliant parser for every dialect/subdialect of SQL). You could also envision many different worlds where people would do behavior that sounds suspect (such as downloading API schema from remote servers), but it provides such a huge developer experience advantage that the tradeoff would be worth the downsides. Everything's a tradeoff.

A victim of success

Procedural macros are not free. They take nonzero amounts of time to run because they are effectively extending the compiler with arbitrary extra behavior at runtime. This gives you a lot of power to do things like what serde does, but as more of the ecosystem uses it more and more, it starts taking nontrivial amounts of time for the macros to run. This causes more and more of your build time being spent waiting around for a proc macro to finish crunching things, and if the proc macro isn't written cleverly enough it will potentially waste time doing the same behavior over and over again.

This can slow down build times, which make people investigate the problem and (rightly) blame serde for making their builds slow. Amusingly enough, serde is used by the Rust compiler rustc and package manager cargo. This means that the extra time compiling proc macros are biting literally everyone, including the Rust team.

Mara is hmm
<Mara> Keep in mind though that the Rust compiler is already very damn fast. One of the standard benchmarks we use across hardware is the "how fast do you compile xesite" test. Xesite is a fairly complicated Rust program that uses a bunch of crates and weird language features like the procedural macro maud to generate HTML. If you want to run the benchmark for yourself, install hyperfine and run the following command:
hyperfine --prepare "cargo clean" "cargo
build --release"
Here's the results on our new MacBook Pro M2 Max:
$ hyperfine --prepare "cargo clean" "cargo build --release"
Benchmark 1: cargo build --release
Time (mean ± σ):     41.872 s ±  0.295 s    [User: 352.774 s, System: 22.339 s]
Range (min … max):   41.389 s … 42.169 s    10 runs
In comparison, the homelab shellbox machine that production builds are made on scores this much:
hyperfine --prepare "cargo clean" "cargo build --release"
Benchmark 1: cargo build --release
Time (mean ± σ):     103.852 s ±  0.654 s    [User: 1058.321 s, System: 42.296 s]
Range (min … max):   102.272 s … 104.843 s    10
Procedural macros are plenty fast, it's always a tradeoff because they always could be faster. For additional timing information about xesite builds, look at the timing report.

The change

In essence, the change makes serde's derive macro use a precompiled binary instead of compiling a new procedural macro binary every time you build the serde_derive dependency. This removes the need for that macro to be compiled from source, which can speed up build times across the entire ecosystem in a few cases.

Cadey is coffee
<Cadey> To be fair, this precompiled binary fiasco only affects x86_84/amd64 Linux hosts. The majority of CI runs on the planet use x86_64 Linux hosts. Given how much of a meme "Rust has slow compile times" has become over the last decade, it makes sense that something had to give. It would be nice if this affected more than cold CI runs (IE: ones without a pre-populated build cache), but I guess this is the best they can do given the constraints of the compiler as it exists today.

However, this means that the most commonly used crate is shipping an arbitrary binary for production builds without any way to opt-out. This could allow a sufficiently determined attacker to use the serde_derive library as a way to get code execution on every CI instance where Rust is used at the same time.

Aoi is wut
<Aoi> Can't you do this anyways with a proc macro given that it's a dynamic library in the compiler?
Cadey is coffee
<Cadey> Well, yeah, sure. The main difficulty is that when you're doing it in a proc macro you have to have the code in a human-readable format somewhere along the line. This would allow users to discover that the version of the code distributed with the crate differs from the version inside source control fairly trivially. Compare this to what you'd have to do in order to determine if a binary is compiled from different source code. That requires a completely different set of skills than comparing source code.

Combine that with the fact that the Rust ecosystem doesn't currently have a solid story around cryptographic signatures for crates and you get a pretty terrible situation all around.

hero image blog/2023/serde/gpg-ux

But this does speed things up for the cost of using serde as a weapon to force ecosystem change.

In my testing the binary they ship is a statically linked Linux binary:

$ file ./serde_derive-x86_64-unknown-linux-gnu
./serde_derive-x86_64-unknown-linux-gnu: ELF 64-bit LSB pie executable, x86-64, version 1 (SYSV), static-pie linked, BuildID[sha1]=b8794565e3bf04d9d58ee87843e47b039595c1ff, stripped

$ ldd ./serde_derive-x86_64-unknown-linux-gnu
        statically linked

Mara is hacker
<Mara> Note: you should never run ldd on untrusted executables. ldd works by setting the environment variable LD_TRACE_LOADED_OBJECTS=1 and then executing the command. This causes your system's C dynamic linker/loader to print all of the dependencies, however malicious applications can and will still execute their malicious code even when that environment variable is set. I've seen evidence of applications exhibiting different malicious behavior when that variable is set. Stay safe and use virtual machines when dealing with unknown code.
Numa is delet
<Numa> To misquote James Mickens, the best way to find out what something does is by executing it to discover more clues.

Out of date "file not found" error with a friend using cargo2nix

Cadey is coffee
<Cadey> This is out of date. The friend of mine in question has since rebooted their system and cannot reproduce this problem. We assume rac's machine got bitflipped or something.

Frustratingly, a friend of mine that uses cargo2nix is reporting getting a "file not found" error when trying to build programs depending on serde. This is esepecially confusing given that the binary is a statically linked binary, but I guess we'll figure out what's going on in the future.

Aoi is wut
<Aoi> Wait, but if the proc macro binary exists how could the file not be found?
Mara is hacker
<Mara> That's the fun part. That error message doesn't just show up when you ask the computer to run a binary that doesn't exist. It also shows up when the binary is loading and the kernel is loading the dynamically linked dependencies. So the program binary can exist but if a dynamic dependecy doesn't, it'll bail and fail like that.
Cadey is coffee
<Cadey> Yeeep, this is one of the worst errors in the Linux ecosystem. Don't feel bad about it being confusing, this bites everyone eventually. The first time I encountered it, I spent more time than I'm comfortable admitting figuring it out. I had to resort to using strace. I felt like a massive idiot when I figured it out.

There's also additional concerns around the binary in question not being totally reproducible, which is slightly concerning from a security standpoint. If we're going to be trusting some random guy's binaries, I think we are in the right to demand that it is byte-for-byte reproducible on commodity hardware without having to reverse-engineer the build process and figure out which nightly version of the compiler is being used to compile this binary blob that will be run everywhere.

I also can't imagine that distribution maintainers are happy with this now that Rust is basically required to be in distribution package managers. It's unfortunate to see turn from a source code package manager to a binary package manager like this.

Numa is delet
<Numa> Nah, trust me bro. It's totes a legit binary, don't think about it so much and just run this arbitrary code on your system. What could go wrong?
Aoi is coffee
<Aoi> Uhhhh, a lot??? Especially if this becomes a common practice that is validated by the biggest project using it. This feels like it could have a massive chilling effect across the entire ecosystem where this behavior becomes more normalized and expected. I don't know if I'd want to see that become a reality.

This doesn't even make build times faster

The most frustrating part about this whole affair is that while I was writing the majority of this article, I assumed that it actually sped up compliation. Guess what: it only speeds up compilation when you are doing a brand new build without an existing build cache. In many cases this means that you only gain the increased build speed in very limited cases: when you are doing a brand new clean build or when you update serde_derive.

Aoi is wut
<Aoi> I guess these are some semi-common usecases where this would be useful, but I don't think this is worth the extra threat vector.

This would be much more worth the tradeoff if it actually gave a significant compile speed tradeoff, but in order for this to make sense you'd need to be building many copies of serde_derive in your CI builds constantly. Or you'd need to have every procedural macro in the ecosystem also follow this approach. Even then, you'd probably only save about 20-30 seconds in cold builds on extreme cases. I really don't think it's worth it.

The middle path

Everything sucks here. This is a Kobayashi Maru situation. In order to really obviate the need for these precompiled binary blobs being used to sidestep compile time you'd need a complete redesign of the procedural macro system.

Cadey is angy
<Cadey> Or, you'd need the proper compile-time reflection support that ThePHD was going to work on until the whole RustConf debacle happened. This would entirely obviate the need for the derive macro serde uses in its current form. We could have had nice things.

One of the huge advantages of the proc macro system as it currently exists is that you can easily use any Rust library you want at compile time. This makes doing things like generating C library bindings on the fly using bindgen trivial.

Aoi is wut
<Aoi> How does that work though? It can't do something awful like parsing the C/C++ headers manually, can it?
Numa is happy
<Numa> That's the neat part, it actually does do that by using clang's C/C++ parser!
Aoi is coffee
<Aoi> That's horrifying.
Mara is hacker
<Mara> It is yeah, but this is what you have to do in the real world to get things working. It's worth noting that you don't have to always do this at compile time. You can commit the intermediate code to your git repo or write your bindings by hand, but I think it's better to take the build speed loss and have things get generated for you so you can't forget to do it.

Maybe there could be a lot of speed to be gained with aggressive caching of derived compiler code. I think that could solve a lot of the issues at the cost of extra disk space being used. Disk space is plenty cheap though, definitely cheaper than developer time. The really cool advantage of making it at the derive macro level is that it would also apply for traits like Debug and Clone that are commonly derived anyways.

I have no idea what the complexities and caveates of doing this would be, but it could also be interesting to have the crate publishing step do aggressive borrow checking logic for every supported platform but then disable the borrow checker on crates downloaded from The borrow checker contributes a lot of time to the compilation process, and if you gate acceptance to on the borrow checker passing then you can get away without needing to run the extra borrow checker logic when compiling dependencies.

Aoi is wut
<Aoi> Yeah but when the borrow checker changes behavior slightly within the same Rust edition, what happens? What if there is a bug that allows something to pass muster in one version of rustc that shouldn't be allowed, making the code in fundamentally wrong?
Cadey is coffee
<Cadey> I claimed ignorance of the problems for a reason! I realize that this would nearly impossible in practice, but I feel like this could be more of a viable option than telling people it's okay to put binaries in the mostly source-code based package store that is

Tangent about using WebAssembly

WASM for procedural macros?

Aoi is wut
<Aoi> Wait, how is this relevant here? This seems like a nonsequitor, doing proc macro compliation/running with WebAssembly would undoubtedly be slower, right? If only going by the rule that a layer of abstraction is by definition more overhead than not having it?
Cadey is coffee
<Cadey> The maintainer of serde is also the creator of watt, a runtime for executing precompiled procedural macros with WebAssembly. Adopting a solution like this would vastly improve the security, isolation, and reproducibility of procedural macros. I really wish this was more widespread. With optimizations such as adopting wasmtime for executing these proc macros, it could be made a lot faster on standard development/production environments while also not leaving people on obscure targets like rv64-gc in the dust.

I'm also pretty sure that there is an easier argument to be made for shipping easily replicatable WASM blobs like Zig does instead of shipping around machine code like serde does.

One of the core issues with procedural macros is that they run unsandboxed machine code. Sandboxing programs is basically impossible to do cross-platform without a bunch of ugly hacks at every level.

I guess you'd need to totally rewrite the proc macro system to use WebAssembly instead of native machine code. Doing this with WebAssembly would let the Rust compiler control the runtime environment that applications would run under. This would let packages do things like:

  • Declare what permissions it needs and have permissions changes on updates to the macros cause users to have to confirm them
  • Declare "cache storage" so that things like derive macro implementations could avoid needing to recompute code that has already passed muster.
  • Let people ship precompiled binaries without having to worry as much about supporting every platform under the sun. The same binary would run perfectly on every platform.
  • More easily prove reproducibility of the proc macro binaries, especially if the binaries were built on the registry server somehow.
  • Individually allow/deny execution of commands so that common behaviors like bindgen, pkg-config, and compiling embedded C source code continue working.

This would require a lot of work and would probably break a lot of existing proc macro behavior unless care was taken to make things as compatible. One of the main pain points would be dealing with C dependencies as it is nearly impossible* to deterministically prove where the dependencies in question are located without running a bunch of shell script and C code.

Cadey is coffee
<Cadey> *If you are using Nix, this is trivial, but sadly we aren't at a place where Nix is used by everyone yet.

One of the biggest headaches would be making a WebAssembly JIT/VM that would work well enough across platforms that the security benefits would make up for the slight loss in execution speed. This is annoyingly hard to sell given that the current state of the world is suffering from long compilation times. It also doesn't help that WebAssembly is still very relatively new so there's not yet the level of maturity needed to make things stable. There is a POSIX-like layer for WebAssembly programs called WASI that does bridge a lot of the gap, but it misses a lot of other things that would be needed for full compatibility including network socket and subprocess execution support.

Mara is happy
<Mara> There is an extension to WASI called WASIX that does solve nearly all of the compatibility problems, but WASIX isn't standard yet and my runtime of choice wazero doesn't have out-of-the box support for it yet. Hopefully it will be supported soon! I just wish it wasn't associated to the wasmer mark of Cain.

This entire situation sucks. I really wish things were better. Hopefully the fixes in serde-rs/serde#2580 will be adopted and make this entire thing a non-issue. I understand why the serde team is making the decisions they are, but I just keep thinking that this isn't the way to speed up Rust compile times. There has to be other options.

I don't know why they made serde a malware vector by adding this unconditional precompiled binary in a patch release in exchange for making cold builds in CI barely faster.

The biggest fear I have is that this practice becomes widespread across the Rust ecosystem. I really hate that the Rust ecosystem seems to have so much drama. It's scaring people away from using the tool to build scalable and stable systems.

Cadey is percussive-maintenance
<Cadey> I mean at some level, to be in a community is to eventually cause conflict. I'm not tired of the conflicts existing, I'm tired of the conflicts being poorly handled and spilling out into GitHub hellthreads that leave everyone unhappy. Let's hope this event doesn't spill out into even more intelligent and highly capable people burning out and leaving.

This article was posted on M08 26 2023. Facts and circumstances may have changed since publication. Please contact me before jumping to conclusions if something seems wrong or unclear.

Tags: rust serde plt

This post was not WebMentioned yet. You could be the first!

The art for Mara was drawn by Selicre.

The art for Cadey was drawn by ArtZora Studios.

Some of the art for Aoi was drawn by @Sandra_Thomas01.