Bundler Design Trade-offs: The Rationale Behind Creating Rspack

Zack Jackson
17 min readAug 30, 2023

--

Before embarking on the development of Rspack, we explored various build tools and frameworks, including extensive use of Webpack, Vite, esbuild, and Rollup in real-world production environments.

To provide some context, our team, known as the Web Infra Team, is responsible for overseeing the company’s suite of front-end build tools and frameworks. Some of these are open-source, while others are proprietary. Our portfolio includes:

  • ModernJS Builder: A build engine for common front-end applications
  • Garfish & Vmok: Universal solutions for micro frontends
  • ModernJS Framework: A progressive React framework
  • PIA: A high-performance web native development framework
  • Module Tools: A scheme for building common libraries
  • Rspress: A documentation solution
  • Lynx Speedy: A build tool for the cross-platform Lynx framework
  • Web Doctor: A diagnostic analysis tool for builds

This background informs our perspective on the design trade-offs involved in bundler technology and why we decided to create Rspack.

The Complexity of Underlying Build Tools: Daily Challenges and Operational Differences

The underlying build tools in all these frameworks and utilities are complex. A significant portion of our daily on-call responsibilities involves addressing user issues related to these build processes.

Key Differences Between Internal Infra Team and Open-Source Community Operations:

  1. Scope of Responsibility: Open-source teams often focus on single-point solutions like Next.js or React-Native. In contrast, our team has a broader mandate. We aim to manage multiple solutions cost-effectively, minimize user switching costs between frameworks and tools, and facilitate the integration of various solutions, such as supporting both SSR and micro frontends.
  2. On-Call Obligations for Business Teams: Unlike just addressing issues, our on-call duty requires rapid business response. Most problems are resolved within 24 hours, and almost all are solved within a week. This dual focus necessitates high-speed iterations for our solutions while also enabling us to address business-side issues in a cost-effective manner.

Originally posed here by @hardfist_1

Challenges with Webpack: Performance and Debugging

Our initial large-scale deployment of build tools heavily relied on Webpack, a trend that continues with our open-source Modern.js project. Webpack’s scalability is its strongest suit, accommodating almost all our build requirements. However, it comes with its own set of challenges.

Debugging Woes

Webpack operates much like a black box, making debugging a cumbersome process. When the business team encounters build-related issues, troubleshooting becomes a significant hurdle. This often requires intervention from the Infra Team, increasing the on-call pressure on them. This was a key motivator behind the development of Web Doctor, aimed at alleviating this on-call stress.

Performance Limitations

Performance has always been a sticking point for Webpack. We’ve experimented with various optimization strategies, including swc-loader, esbuild-loader, thread-loader, cache-loader, MFSU, and Persistent Cache. While these solutions may offer some relief, they fall short in handling large projects efficiently. Moreover, these optimizations often make the build process even more opaque. For example, Persistent Cache relies heavily on well-configured business build dependencies, esbuild-loader lacks support for ES5 downgrades, and cache-loader can lead to outdated products if the cache isn’t cleared properly.

Challenges with Dual-Engine Approach: Webpack and Vite

The dual-engine approach, where the underlying engine can switch between Vite and Webpack, seemed promising initially. It offered a unified configuration and plugin layer, which partially addressed the performance issues we faced with Webpack. However, this approach introduced a new set of complexities.

Plugin Reusability

The plugin mechanisms for Rollup (used by Vite) and Webpack are fundamentally different. While there are attempts to create universal plugins, like unplugin, these solutions are still in their infancy and lack the robustness to support complex frameworks like Modern.js. This leads to a codebase filled with conditional logic to load different plugins based on the configuration, making it less maintainable.

Performance in Large Projects

Vite’s performance in large-scale projects is not up to the mark. The overhead from thousands of network requests during development can lead to significant delays, especially during hot module reloading (HMR). Rollup’s performance also leaves room for improvement, sometimes even falling behind Webpack, which has the advantage of a persistent cache.

Inconsistency Between Development and Production

Many internal teams end up using Vite for development and Webpack for production due to Rollup’s limitations in product optimization. This creates a significant divergence between development and production environments, making it harder to ensure consistent behavior.

In summary, while the dual-engine approach solves some problems, it introduces new challenges that make it less than ideal for our needs. We’re still in search of a more holistic solution that can offer both performance and flexibility without compromising on either.

Rollup: Strengths and Weaknesses

We’ve extensively used Rollup in library build and early Lynx build scenarios, revealing both its strengths and weaknesses.

Strengths:

  1. Clean Product Format: Rollup generates extremely clean code.
  2. TreeShaking Support: The tool is highly conducive to TreeShaking, making it efficient for eliminating dead code.

Weaknesses:

  1. CommonJS Support: Despite the community’s shift towards ESM, our projects still have a significant number of CommonJS dependencies. Rollup’s architecture makes it difficult to achieve full compatibility with CommonJS. For example, non-strict CJS is always converted to strict mode ESM. This results in frequent on-call interventions to address CommonJS-related issues, which is highly inconvenient.
  2. Compile Performance: Rollup’s performance is comparable to Webpack’s, but it lacks features like Persistent Cache, making secondary cold starts slower. Additionally, Rollup doesn’t support Hot Module Replacement (HMR). While HMR isn’t crucial for library builds, Rollup’s performance with the watch feature is mediocre.

Esbuild

Esbuild has effectively addressed two major shortcomings we encountered with Rollup: CommonJS support and performance. Esbuild considers CommonJS as a first-class citizen, providing robust support for it. Additionally, esbuild’s performance is excellent, making it a strong alternative to Rollup, particularly for library builds. This is why tools like tsup, which is built on esbuild, are considered better alternatives to tsdx, which relies on Rollup. It’s also worth noting that our ModernJS Module Tools are currently using esbuild as the underlying layer.

However, our experience with esbuild has also revealed several challenges.

  • First, its API is extremely streamlined, which limits its plugin extension capabilities, especially for application building. For example, the absence of an onTransform hook makes it difficult to chain different transform extensions. This limitation also affects post-processing chunks, as there's no renderChunk hook for tasks like custom minification.
  • Second, esbuild’s performance in client-side scenarios is not without issues. The number and size of chunks are sensitive factors, and esbuild lacks the deep customization abilities that Webpack offers for optimizing chunks. This is particularly relevant given the diverse file-loading scenarios we encounter, including different browsers and cross-platform containers.
  • Third, while esbuild’s cold start performance is excellent, its rebuild performance can be problematic when more JS plugins are used. Unlike Webpack, which only re-triggers the loader for the changing module, esbuild triggers a full rebuild for all onLoad and onResolve hooks. This results in an O(n) complexity problem in large projects.
  • Fourth, esbuild lacks built-in support for Hot Module Replacement (HMR). To implement HMR, you have to inject an HMR runtime during the Load phase, which introduces compile-time overhead.You can see these struggles in remix
  • Fifth, esbuild’s minimalistic design makes it difficult to implement advanced features like Module Federation, which are almost turnkey solutions in Webpack.

Due to these challenges, we’ve decided to focus on developing a Rust-based bundler. The goal is to create a tool that combines the strengths of Webpack, Rollup, and esbuild while addressing their respective weaknesses, providing a more efficient and robust build process.

How We Developed Rspack

Initially, our approach to developing the Rust Bundler was not to emulate Rust Webpack. Instead, we aimed to address some of the existing issues with esbuild. Our starting point was essentially a Rust version of esbuild and Rollup, but with built-in support for HMR, CommonJS, and Bundle Splitting. Given that we had already encapsulated a complete framework based on esbuild and had run it in a production environment for a year, we wanted to maintain compatibility with esbuild’s interface while resolving its limitations in HMR and bundle splitting. The ‘legacy’ branch of Rspack still retains this original design, aimed at being compatible with the esbuild interface.

However, as we delved deeper into the development process, completing the first version of our Rust Bundler based on the Rollup architecture and reaching the project’s Proof of Concept stage, we encountered a series of issues. These challenges forced us to reevaluate our architectural choices, ultimately leading us to adopt an architecture more aligned with Webpack.

This shift was driven by the need for a more robust and flexible solution, capable of addressing the complexities and varied requirements we had identified. The Webpack architecture offered us the extensibility and customization options that were lacking in our initial approach, making it a more suitable foundation for Rspack.

First-Class Citizen Approach in a Language-Agnostic Manner

Rollup’s architecture is designed to support only ESM as a first-class citizen. This means that other module systems like CommonJS, or even non-JS module systems, need to be converted to ESM to be compatible. This conversion process introduces a host of issues. One prevalent problem is the inconsistent resolve logic across different module systems. Other inconsistencies include default values for sideEffects, chunk generation logic, and more.

Because Rollup converts all modules into ESM, it doesn’t adequately distinguish between different types of modules at its core layer. This lack of differentiation forces developers to rely on plugins and employ hacky logic to implement functionalities. For example, each conversion from CommonJS to ESM often requires annotating the original CommonJS code to ensure compatibility and functionality.

This approach is problematic for several reasons. It not only complicates the build process but also makes it difficult to maintain and extend the system. It’s a significant issue when dealing with large, complex codebases that use a mix of different module systems or when trying to integrate third-party libraries that may not be ESM-compatible.

In contrast, our aim with Rspack is to be language-agnostic and to treat all module systems as first-class citizens. This approach allows for greater flexibility and reduces the need for complex conversions or hacky workarounds, making the build process more efficient and robust.

Note
Contrary to the intuition of many people, Webpack is language agnostic like Parcel, while Rollup is only a first-class citizen with
Javascript . This may also be the most overlooked point of Webpack 5 , which supports more first-class citizen modules.

Plugin API Design

From the inception of the Rust Bundler project, one core requirement has remained constant: the need to support plugins written in JavaScript. This requirement stems from our understanding of business needs, particularly the necessity for scalability. A bundler that lacks scalability is difficult to implement effectively in a business environment, making the design of the plugin API a critical issue for us.

The two primary factors we considered in the API design were performance and composability. Upon evaluation, we found that Rollup’s API fell short of meeting these needs. Rollup’s architecture, which is more aligned with treating JavaScript as a first-class citizen, didn’t offer the flexibility and performance we required for a diverse and scalable plugin ecosystem.

Our aim is to develop an API that not only allows for high performance but also supports a composable architecture. This would enable plugins to be easily combined and configured, thereby enhancing the bundler’s flexibility and adaptability to various business requirements. This focus on performance and composability is part of our broader strategy to make Rspack a robust, efficient, and scalable solution.

Note
Simple API is useful for adoption but maybe hard for scaling.

Module Conversion

Module transformation is a core feature that all bundler plugins must address. While all bundlers offer plugins to handle this, the implementations differ. For example, Rollup uses transform, while Webpack uses loader for this purpose.

Upon a comprehensive analysis of module transformation functionality, we identified that this is essentially a three-dimensional requirement:

  1. Filter: Determines which modules undergo transformation. This could be based on file extensions, directory location, or other attributes.
  2. Converter (Transformer): Specifies how the filtered modules are to be converted. This could involve transpiling code, optimizing images, or other transformations.
  3. Module Type Conversion: Sometimes it’s necessary to convert a module from one type to another, such as from svg to jsx or vice versa.

Each of these dimensions has its own set of challenges and requirements, and the plugin architecture needs to be flexible enough to accommodate a wide range of use-cases and performance considerations.

Taking the svgr plugin as an example helps illustrate the complexity involved in module conversion. The primary function of the svgr plugin is to convert an SVG file into a React component. Breaking this down into the three elements we discussed:

  1. Filter: The filter is set to /.svg$/, meaning it will only process files with an .svg extension.
  2. Converter (Transformer) : The plugin uses @svgr/core to transform the SVG content into the corresponding JSX component.
  3. Module Type Conversion: After the SVG has been converted, it’s necessary to treat the resulting content as JSX for further handling.

This example underscores the intricacy of module conversion logic. Each step — filtering, converting, and changing the module type — has its own set of challenges and considerations. The plugin architecture must be robust enough to handle such complexities efficiently.

In Rollup, the three dimensions of module conversion — filtering, transforming, and changing the module type — are bundled into a single transform hook. This design choice leads to several significant issues:

  • High-Frequency Callback Communication: The filter logic in Rollup’s transform hook is executed within the hook itself. This means that to perform filtering, every module incurs the overhead of Rust-to-JS communication. In scenarios involving tens of thousands of modules, especially with Hot Module Replacement (HMR), this overhead becomes substantial. Esbuild addresses this by performing filtering first and only then executing the JS callback. Notably, Esbuild uses Golang's regex for filtering to avoid the Golang-to-JS call overhead, although this can lead to user confusion due to differences between Golang and JS regex.
build.onLoad({ filter: /.txt$/ }, async (args) => {
let text = await fs.promises.readFile(args.path, 'utf8');
return {
contents: JSON.stringify(text.split(/\s+/)),
loader: 'json',
};
});
  • Loss of User Flexibility: With Rollup, the filter logic is hard-coded into the transform hook, making it difficult for users to modify the filter logic externally. For instance, if you later decide to process files with different extensions but the same SVG content, you'd have to modify the Rollup plugin itself. In contrast, with Esbuild or Webpack, you can directly modify the filter logic.
  • Loss of Combinability in Module Conversion Logic: Rollup’s architecture, which is geared towards JavaScript, requires additional steps to convert the JSX file generated by @svgr/core into a JS file. This means that the plugin has to handle the JSX-to-JS conversion logic, even if similar functionality exists in other plugins. For example, vite-plugin-svgr uses Esbuild to convert JSX to JS, duplicating functionality and introducing another dependency.

After extensive research on almost all plugin APIs for module conversion, we found that only Parcel and Webpack’s APIs adequately address these issues. Discussion Link

Parcel

Parcel takes a more modular approach to handle the three dimensions of module conversion:

  • Filter: Parcel uses a pipeline to define the filtering logic. You can specify which transformers to use for different types of files directly in the configuration.
{
"transformers": {
"icons/*.svg": ["@company/parcel-transformer-svg-icons", "..."],
"*.svg": ["@parcel/transformer-svg"]
}
  • Converter (Transformer) : Parcel employs a transform plugin to define how the conversion should take place. The plugin API allows you to retrieve the asset's source code and source map, run it through a compiler, and set the results back on the asset.
import {Transformer} from '@parcel/plugin';
export default new Transformer({
async transform({asset}) {
// Retrieve the asset's source code and source map.
let source = await asset.getCode();
let sourceMap = await asset.getMap();
// Run it through some compiler, and set the results
// on the asset.
let {code, map} = compile(source, sourceMap);
asset.setCode(code);
asset.setMap(map);
// Return the asset
return [asset];
}
});
  • Module Type Conversion: Parcel allows you to change the type of the asset directly within the transformer. This is done using asset.type.
import {Transformer} from '@parcel/plugin';
export default new Transformer({
async transform({asset}) {
let code = await asset.getCode();
let result = compile(code);
asset.type = 'js'; // change asset type
asset.setCode(result);
return [asset];
}
});

This modular approach provides a high degree of flexibility and control, allowing for more complex and varied module conversion scenarios. It also avoids the issues seen in Rollup, such as high-frequency callback communication and loss of user flexibility.

Webpack

Webpack

Webpack’s approach to module conversion is also modular but differs in implementation details:

  • Filter: Webpack uses rule.test to specify which files should be processed by a particular loader.
module: {
rules: {
test: /.svgr/,
use: ['@svgr/webpack']
}
}
  • Converter (Transformer) : Webpack employs loaders for the actual conversion logic, such as @svgr/webpack for SVG to React component conversion.
  • Module Type Conversion: Webpack uses inlineMatchResource for this. It's less intuitive compared to directly modifying asset.type, and although there's a proposal for a more intuitive "virtual resource," it hasn't been implemented yet.

AST Reuse

The reuse of Abstract Syntax Trees (ASTs) between different module transformations is a crucial design aspect for performance. Parsing overhead is often a significant bottleneck, and reusing ASTs can substantially mitigate this issue. Here’s how Esbuild and Rollup approach AST reuse:

Esbuild

Esbuild takes the most straightforward approach: it doesn’t support module conversion operations, so the question of AST reuse is moot. Esbuild’s parse, transform, and minify steps all share the same AST. This is a key reason why Esbuild’s performance is superior to other bundlers. The trade-off is limited scalability due to the lack of support for custom transformations.

Rollup

Rollup allows for the return of ASTs in both its load and transform hooks. The ASTs must conform to the standard ESTree format. If a standard ESTree AST is returned, Rollup can internally reuse the AST, avoiding the need for duplicate parsing. This offers a more flexible approach compared to Esbuild but requires that plugins adhere to the ESTree standard for maximum efficiency.

Note
One advantage of rollup to reuse AST is that rollup only supports JavaScript, which means that only the standard ESTree AST
data structure needs to be considered, but this is not tried for Parcel and Webpack.

Parcel

Parcel’s approach to AST reuse is particularly well-conceived. It provides detailed guidelines on how to achieve AST reuse, a design aspect that early versions of Rspack also borrowed from. Parcel’s design addresses a significant challenge in AST reuse: how to handle scenarios where string transformations and AST transformations intersect.

For general transformers, there are four possible cases:

  1. String -> String: Both the input and output are strings.
  2. String -> AST: The input is a string, and the output is an AST.
  3. AST -> AST: Both the input and output are ASTs.
  4. AST -> String: The input is an AST, and the output is a string.

Managing the interplay between these different types of transformations is a complex task. Parcel handles this complexity well by allowing transformers to specify both their input and output types and managing the conversions internally. This ensures that ASTs are reused whenever possible, optimizing performance without sacrificing flexibility.

Parcel’s approach offers a balanced solution, providing both performance optimization through AST reuse and the flexibility to handle various transformation types. This makes it a robust choice for scenarios requiring both performance and adaptability.

Webpack

Webpack does offer the capability to return an AST in its loaders to facilitate AST reuse. However, this feature has not gained much traction in the community due to several limitations:

  1. ESTree Standard: Webpack requires that the returned AST conform to the ESTree standard. Unfortunately, many community-contributed loaders for JavaScript transformations don’t return a standard ESTree AST. This includes popular loaders like babel-loader and swc-loader, making it challenging to implement AST reuse effectively in Webpack.
  2. Limited AST Types: While Webpack’s parser can technically support multiple types of ASTs (e.g., CSS AST and JavaScript AST), it currently only supports JavaScript ASTs. This limitation has led to the feature being available but not widely adopted in practice.

These constraints have resulted in Webpack’s AST reuse feature being underutilized, despite its potential for performance optimization. It’s a feature that exists but hasn’t been effectively operationalized due to these community and technical barriers.

Beyond module conversion and reusing AST, we also considered various plugin design aspects like reducing the Rust-to-JS communication frequency, avoiding redundant parsing overhead and virtual modules. After a comprehensive evaluation, we found that Webpack’s architecture aligns better with our goals for customization and performance in the development of Rust Bundler. We opted for a Webpack-based approach primarily because it’s a lower-risk option; almost no one on the team has experience with Parcel, making Webpack a safer bet.

Webpack Design Exploration

Despite committing to the Webpack architecture, the implementation journey involved several detours, particularly around the concept of first-class citizen support.

Initial Approach and Challenges

Influenced by Esbuild and facing incomplete loader support, we initially extended first-class citizen support to various JavaScript extensions like jsnext, ts, tsx, jsx. While this expedited business-side implementation, it introduced several issues:

  1. Inaccurate Bundle Splitting: The codegen doesn't occur immediately post-AST conversion, making it impossible to obtain the converted code for bundle splitting analysis. This leads to inaccuracies, especially when transformations like injecting babel and swc runtimes cause significant code changes.
  2. Transformation Problems: Defaulting to swc for ts and js files led to transformation errors in some modules, such as core.js.
  3. Secondary Conversion Issues: When users employ swc-loader for transformations, it results in modules undergoing secondary conversions, causing problems.
  4. Uncontrollable TypeScript Compilation: For instance, decorators in TypeScript may undergo multiple compilations, and the default compilation settings may not align with user requirements. Users might prefer different semantics for decorators in different modules, complicating the situation.

Future Considerations

Given these challenges, Rspack is considering retracting the first-class citizen support for ts, tsx, jsx and other such modules. Instead, it plans to allow users to handle compilation via builtins:sw``c-loader. This approach aims to align with Webpack's handling of first-class citizens while ensuring core layer stability.

Codegen Architecture

Initial Approach: AST-Based Codegen

In the early stages, we adopted an AST-based codegen scheme, contrasting with Webpack's dependency-based string replacement approach. While the AST-based scheme offered better performance by avoiding repeated parsing overhead, it diverged significantly from Webpack's architecture.

Challenges

  1. Dependency on String Replacement: Webpack’s Runtime and treeshaking logic are tightly coupled with the dependency information derived from string replacement. Using an AST-based approach disrupted this harmony.
  2. Persistent Cache Implementation: The AST-based codegen complicated the subsequent implementation of a persistent cache, adding layers of complexity that were hard to manage.

Transition to String Replacement: Version 0.3

Given these challenges, we transitioned from an AST-based codegen architecture to a string replacement-based architecture in version 0.3. This move aimed to align more closely with Webpack's architecture, facilitating better integration with features like Runtime, treeshaking, and persistent caching.

TreeShaking Architecture

Initial Approach: AST-Based Scheme

Initially, our TreeShaking approach was influenced by our AST-based codegen scheme. We borrowed from Esbuild's AST-friendly TreeShaking method. However, this approach later revealed several limitations.

Challenges

Reexport and Multi-Entry Optimization: The AST-based approach struggled with TreeShaking for reexport and multi-entry scenarios. Esbuild's TreeShaking optimization lags behind Webpack's, as highlighted in Esbuild GitHub Issue #2049.

Transition to Webpack’s Implementation

Due to these challenges, we transitioned our TreeShaking implementation to align with Webpack's. This move aimed to leverage Webpack's more advanced and in-depth TreeShaking optimization strategies.

Ongoing Exploration

While we’ve made this transition, several questions and challenges remain, and we continue to explore solutions to further refine our TreeShaking architecture.

Beyond Webpack: The Road Ahead for Rspack

Out-of-Box Solution

Webpack’s most glaring issue is its poor development experience, especially for newcomers. Setting up a project from scratch with Webpack can be daunting, especially when compared to the ease of using Vite. With the decline in community support for create-react-app, there are fewer upper-layer solutions built on Webpack. Recognizing the importance of an out-of-the-box experience, we plan to collaborate with Modern.js to enhance this aspect.

Diagnostics and Debugging

Webpack operates as a black box for most users, lacking essential debugging tools. To address this, we’ll continue to refine Web Doctor to improve the debugging experience for both Rspack and Webpack. Additionally, we’ll integrate more debugging information directly into Rspack to offer an out-of-the-box debugging experience.

Optimization

While Webpack is often considered a leader in product optimization, there’s still room for improvement. Webpack’s bundle splitting, code splitting, and tree shaking are all module-based, limiting the optimization techniques that can be applied. We aim to explore function-level granularity for these optimization methods to further enhance runtime performance.

Portable Cache and Remote Build Cache

Within our organization, we have numerous large-scale applications and Monorepo setups. Webpack currently lacks support for portable caching capabilities, making it challenging to implement distributed build cache sharing. This feature is crucial for large applications and Monorepo setups where build speed is a key factor. We plan to focus on developing this capability in the future.

By addressing these areas, we aim to not only serve as a drop-in replacement for Webpack but also to push the boundaries of what’s possible in the Webpack ecosystem to deliver a superior user experience.

--

--

Zack Jackson

Infra Architect @ ByteDance. Creator of Module Federation. Specializing in Bundler and Javascript Orchestration at scale.