Scaling TypeScript compilation

October 28, 2020

I've now spent a few more years on our project to bring TypeScript to Google that I wrote about before. As I wrote there, TypeScript itself already works, and our challenges have been primarily around adapting TypeScript to the specifics of our environment. One way in which our apps are different from typical users of TypeScript is their size: how do you make TypeScript scale to projects with millions of lines of code?

In this post I will explain a bit about some approaches we've taken, and in particular about the project I've been working on for the last few months. I write about it here in public in part because the relevant code is open source (with important caveats, described later). Also, these approaches are generally not specific to TypeScript, so I hope you might find them useful when considering the design of other languages.

I have one additional caveat. One of the hardest constraints we engineers have on design is migration. If we were designing from scratch we might make different design tradeoffs, but in practice massive codebases don't spring into existence to fit some new design, and the evolutionary process often ends up tying our hands. Often the answer to "why not X instead of Y" is that Y was something we had on hand, and the relative benefit of X doesn't overcome the relative cost of implementing it over Y.

Parallelizing with Bazel

TypeScript's typical compilation model is more or less a program at a time: you give the compiler all of your source, and it spits out a pass/fail for typechecking and JavaScript output. This is the obvious model for a language when projects are small, but it means the work of the compiler is proportional to the size of the project and slows as the project grows.

If there's a repeated theme in this post, it's that to make compilation scale you must find ways to not do work proportional to the size of the whole program. On its face, that is a little silly — obviously a compiler must visit the whole program. But what you can do is break that work up into smaller pieces that can be run in parallel or cached.

We do this using the Bazel build system. Bazel's lineage is building C++/Java code, and in the C++ world this is already more or less solved: you "just" invoke the compiler separately on each source file. (It's worth noting that even in C++ you eventually must link your .o files, and interesting optimizations happen between those .o files. The state of the art is still experimenting with how to parallelize that work. Thankfully, TypeScript from a C++ perspective is more or less only a compiler and not a linker, so the separable C++ model mostly works.)

The Bazel website better says what it does, but for the purposes of this post I'll review a few high level concepts. Bazel splits code into a notion of "targets", each which may "depend" on other targets, and you provide Bazel that graph as an explicit input, external to the source code. (Bazel targets are typically sets of source files, but for the purposes of this post you could imagine that targets are 1:1 with source files and I think it'll mostly work out the same.)

For a running example, you might tell Bazel:

Here are some benefits Bazel brings:

  1. Parallelization of builds. We can type check all of the widgets (which don't depend on each other) in parallel, either using multiple CPU cores locally or in cloud build farms. If someone else at the company already built the same code you built, we can reuse their build results without doing the work again.

  2. Accurate tracking of dependencies. Suppose the above app has separate tests for each of its screens, and the carousel widget is only used by the home page screen. If you make a change to the carousel widget, Bazel computes that you only need to rebuild and run the test for the home page, because the other screens are unaffected. One nice way to frame this is that the build/test work is proportional to the size of the change made, rather than the size of the entire app.

  3. Cross-language dependencies. The hypothetical app I described above is maybe all TypeScript, but in Bazel we also express that these build steps depend on the TypeScript compiler (itself a target), which depends on nodejs (another target), which depends on the C++ compiler (target). Or we might have a portion of the app that depends on some RPC code generator written in Java (target), which itself has more Java code targets beneath it. A change to some library within the Java code generator still triggers rebuilding exactly the affected downstream code.

I mention the above not to sell Bazel to you — as the author of a build system myself, I am a little mixed on Bazel, though my objections have more to do with the implementation than the conceptual model — but rather to describe our environment. Our work with TypeScript has been fitting it into the Bazel model, which we do in the following way.

Bazel and TypeScript

To build a Bazel target, we feed its source to the TypeScript compiler in the --declarations mode, which causes it to type check the code, generate JavaScript, and also generate a .d.ts file that summarizes the exported API of that target. This .d.ts summary is then used as an input to any dependent compilation.

If you're not familiar with TypeScript's d.ts. generation, you can see it on the TypeScript playground by clicking on the ".D.TS" tab on the right. In this example on the Playground, you can see the "internal" function stays hidden, and also the inferred return type becomes explicit. (A common analogy to use here is this is like a C++ header file, though it is automatically extracted from the source.)

To make this more concrete, suppose target A.ts depends on target B.ts depends on C.ts. (In Bazel, targets are not source files, but it makes the exposition here simpler to just talk about source files.) To build under Bazel, we first build C.ts, producing some C.d.ts that describes its exported API. Then we build B.ts, giving it C.d.ts as an input. Finally, we build A.ts, giving it B.d.ts and C.d.ts as inputs.

Using these d.ts summaries has a few benefits. For one, they're typically smaller than the original source code, because they don't contain implementation, which means they're less work for the compiler to load. They also include the results of type inference (again see the above Playground example), so we can have the compiler infer the types found in C once, and reuse that result in A and B. Another more subtle benefit is reducing unneeded work: a change to C.ts that doesn't affect its exported API will leave C.d.ts unchanged, which means we don't need to recheck A or B. (I previously wrote about a similar trick using C++ shared objects to avoid relinking).

Importantly in our builds we tweak the TS compiler to not type check the d.ts of dependencies, but instead only to check the direct sources of the current target. Doing this is not easily available via the TypeScript command-line, but was done by writing a different frontend that uses the TS compiler as a library, which we needed to do anyway for other Bazel reasons.

I think this pattern of separating interface definitions from source is pretty common across programming languages. The main design tradeoff here is whether such definitions are written by hand (like C++ headers or say .mli files in OCaml) vs extracting them automatically from source (as we've done here, and as found in other languages).

In all, this mechanism is similar to project references in TypeScript, though I have little familiarity with them — they were added to the language after we built the above. One contrast is that our splits are more fine-grained; in a big project we might have hundreds of thousands of these build steps. There is a tradeoff here, where the more fine-grained the splits get, the more fine-grained your rebuilds can get in terms of not rebuilding unaffected code, but also the more build steps and inevitable redundant work (like booting the compiler at all and parsing inputs) those build steps will do.

Much like how TypeScript is building this feature, I find many programming languages slowly slide their way towards becoming build systems — witness also whether it's possible to use Rust without Cargo — and though it feels inevitable it is also disappointing, because a language-specific build system rarely captures things like cross-language dependencies, testing, and so on. (See for example how "go build" captures much of what you want to build with Go, but also that many Go projects end up with a Makefile around it anyway.) But the question of which layer best captures build problems is always a difficult tradeoff.

Serialization of build steps

Going back to the example of an app made of widgets, which build steps can run in parallel? Clearly at least the widgets, which don't depend on each other. But when target A depends on target B, it seems they must necessarily serialize: we need to wait to finish extracting the d.ts from B before starting to build A. Another way of saying this is that while it appears siblings in the DAG can build in parallel, parents must wait for children.

However, there's a trick: if extracting the API shape is cheap, we can do that up front pass (potentially as build steps that themselves can run in parallel across the whole app), and then we have all the data we need to type check all the targets in parallel, regardless of their interdependencies. As an existence proof, this pattern already works in C++, where the API shape is available before compilation even starts because humans wrote the header files. (I describe this more in the "as a warm-up" bit of this post.)

I believe Bazel uses such a technique for making Java builds more parallel, with a tool called "turbine" that extracts the API shape of Java classes up front. However, though I see turbine code in the Bazel Git repo I'm unclear how/whether it's actually used.

In any case, for TypeScript, there is unfortunately a language feature that prevents doing this: inference of return types. In TypeScript it's common to write code like

export function f(x: boolean) {
  return x ? 3 : g();

and TypeScript will infer the return type of f for you (see d.ts output on the Playground). In the limit you need the whole power of the compiler to compute these return types; for example, consider when g() is a function found in another module, and so on.

In other languages, where the compiler does additional work like optimization during codegen, it might still be worth splitting out API extraction from the compile step, but in TypeScript's case I suspect it's about as much work to have it compute these return types as it is to do basically the whole process of compilation, so it appears to me not worth it to split things to make this more parallel.

One idea we've toyed with is mandating that all our code has explicit return types, because it could unlock a lot of parallelism in the build. I don't understand much about the Flow type checker for JavaScript, but I noticed they have recently been doing work around enforcing more type annotations at module boundaries for performance reasons as well. (I am particularly impressed by how they built language tools to auto-upgrade their users.) However, requiring return types is a pretty big divergence from how users expect the language to work, so for us it's probably just an idea to keep in the back pocket rather than a definite plan.

Transitive dependencies

In the above, I mentioned that building A.ts requires both B.d.ts and C.d.ts as inputs. Using B is obvious — A imports it — but you might wonder why C is involved, and the answer is subtle and deep.

To start with, suppose B.ts contains code like this:

import {Version} from './C';
export function getVersion(): Version { ... }

To type check a user of B, that user must also see the definition of Version, so it must see the definition of C. And this reasoning works recursively; the topmost target in the tree, the one that imports all the rest, potentially may depend on types from all transitively depended-upon .d.ts files to build.

Further, these dependencies may not even follow 'import' statements, due to globals! Suppose C contains code like:

declare const iAmAGlobal: number;

Per the rules of TypeScript, even if A doesn't import C, and even if B doesn't import C, it's still the case that A needs to see this definition to type check.

In practice this means that, until recently, we always needed to feed the whole transitive dependency tree of dependencies into the compiler for any compile. This means that the higher up in the dependency tree a target gets — the more targets it transitively depends upon — the more source code the compiler must process as input. In a bad target we would see a few kb of source and megabytes of transitively loaded d.ts. Again, to make compilation scale, you need to do something about this, and this is the project I've worked on for the last few months.

One broad approach to solve this would be to "roll up" transitive imports. When generating the API shape of B, you could imagine you can inline the parts of C that it B uses directly into its output. This would means when building any target, you would only need its directly imported d.ts as input, as those would have inlined their transitive dependencies already. (I believe a technique like this is used by Go.)

In our experience this approach has ended poorly, because these rolled-up intermediates themselves accumulate in size as you work down the DAG, especially in the presence of diamond dependencies. For example, think back to my initial description of an app comprised of dependencies on multiple screens. In a world where you rolled up dependencies, each of those screens could include in their extracted API their own redundant copy of the (shared) web framework. We had a system in Google that worked like this with JavaScript that we are in the process of moving away from due to scaling problems.

Another approach you can imagine taking is some mechanism for "partial" definitions. For example, in C++ you can forward-declare the Version type, which means users can use the B API and treat Version as opaque. Then if someone wants to actually use Version they must explicitly import C themselves. Alas, there currently isn't a good mechanism for this in TypeScript, and — in a change since I last did a lot of C++ at Google — Google's C++ wizards also now recommend against doing this in C++, for interesting reasons.

Transitive dependencies and transitive imports

A final approach you can take is to observe that, for some long chain of dependencies — A deps on B deps on … Z — it's likely that at some intermediate target no part of Z was actually exposed to users. Which is to say, to type check A you probably need B, maybe parts of C, and so on, but likely not every last transitive dependency Z, which might just be used as an internal implementation detail of one of these libraries and not something that is exposed to A.

To make this more concrete, imagine the code snippet from the previous section instead looked like this:

import {getVersion} from './C';
export function getVersionString() {
  return 'v' + getVersion().asString();

Here, users of this module do not interact with any type from C. In fact, the generated d.ts we use in compilation will reflect that users of this library don't need C to type check; it will look this this, without any import of C:

declare export function getVersionString(): string;

So if you "just" let the compiler follow import statements, which in TypeScript is controlled by the --noResolve flag, it can itself lazily pull in exactly just the modules needed for type checking. And that solves the bulk of the problem — except for globals.

One way I have come to see globals in terms of type checking is that they are like side effects. The type checker follows import statements to gather definitions of exported symbols, but any place that declares a global has a type system side effect of registering additional symbols. Another way to see this is that if you have an import statement that imports some symbol, then it's usually safe to remove that import statement if you no longer use the symbol, but that property doesn't hold if the imported module declares a global.

From that framing, TypeScript has multiple mechanisms for type-level side effects. One is declaration merging. This is particularly common in the lit-html / Polymer ecosystem, where code often registers itself to overload the return type of document.createElement():

declare global {
  interface HTMLElementTagNameMap {
    'my-new-tag': MyNewTag;

Another global-like feature is the declare module 'someString' syntax, which we rely on heavily for our Closure JS interop.

My solution for these has been a mixture of approaches that probably aren't too interesting for this blog post. Some approaches include:

For our JS interop with declare module, I did some pretty gnarly work to reconstruct the implied imports that I won't go into more here because nearly nobody cares about it except Google.

Open source

You might note that that last link is to code on GitHub. Historically we've open-sourced most of the relevant code, in part because we publish other open source projects, such as Angular, that use this whole stack of software. However, as I wrote in my post about TypeScript at Google, the world we live in is unfortunately very far away from the non-Google world, and in practice a lot of this code ends up not being useful for others.

And further, for better or for worse, publishing code on GitHub carries the user expectation that you will address their bug reports or help them, and realistically we just do not have time to do that; we barely have the time to support our own internal users. So I regret to say that we (barely) ended up not publishing the code of this last work I've done, not out of any particular desire to keep it secret but rather just because it's just too much work. It is my hope that this blog post, describing the concepts, ends up more useful than the code would have been anyway.


Though some of the above work is my own invention, a lot of the TypeScript machinery was built before I got involved with it — primarily by Martin Probst, Alex Eagle, and Rado Kirov.