I. Introduction
Hello! Ndeewo! Molweni! Салам! Across the world, people are coming online with smartphones, smart watches, and other small, low-resource devices. The technology industry needs an internationalization solution for these environments that scales to dozens of programming languages and thousands of human languages.
Enter ICU4X. As the name suggests, ICU4X is an offshoot of the industry-standard i18n library published by the Unicode Consortium, ICU (International Components for Unicode), which is embedded in every major device and operating system.
This week, after 2½ years of work by Google, Mozilla, Amazon, and community partners, the Unicode Consortium has published ICU4X 1.0, its first stable release. Built from the ground up to be lightweight, portable, and secure, ICU4X learns from decades of experience to bring localized date formatting, number formatting, collation, text segmentation, and more to devices that, until now, did not have a suitable solution.
Lightweight: ICU4X is Unicode's first library to support static data slicing and dynamic data loading. With ICU4X, clients can inspect their compiled code to easily build small, optimized locale data packs and then load those data packs on the fly, enabling applications to scale to more languages than ever before. Even when platform i18n is available, ICU4X is suitable as a polyfill to add additional features or languages. It does this while using very little RAM and CPU, helping extend devices' battery life.
Portable: ICU4X supports multiple programming languages out of the box. ICU4X can be used in the Rust programming language natively, with official wrappers in C++ via the foreign function interface (FFI) and JavaScript via WebAssembly. More programming languages can be added by writing plugins, without needing to touch core i18n logic. ICU4X also allows data files to be updated independently of code, making it easier to roll out Unicode updates.
Secure: Rust's type system and ownership model guarantee memory-safety and thread-safety, preventing large classes of bugs and vulnerabilities.
How does ICU4X achieve these goals, and why did the team choose to write ICU4X over any number of alternatives?
II. Why ICU4X?
You may still be wondering, what led the Unicode Consortium to choose a new Rust-based library as the solution to these problems?
II.A. Why a new library?
The Unicode Consortium also publishes ICU4C and ICU4J, i18n libraries written for C/C++ and Java. Why write a new library from scratch? Wouldn’t that increase the ongoing maintenance burden? Why not focus our efforts on improving ICU4C and/or ICU4J instead?
ICU4X solves a different problem for different types of clients. ICU4X does not seek to replace ICU4C or ICU4J; rather, it seeks to replace the large number of non-Unicode, often-unmaintained, often-incomplete i18n libraries that have been written to bring i18n to new programming languages and resource-constrained environments. ICU4X is a product that has long been missing from Unicode's portfolio.
Early on, the team evaluated whether ICU4X's goals could have been achieved by refactoring ICU4C or ICU4J. We found that:
ICU4C has already gone through a period of optimization for tree shaking and data size. Despite these efforts, we continue to have stakeholders saying that ICU4C is too large for their resource-constrained environment. Getting further improvements in ICU4C would amount to rewrites of much of ICU4C's code base, which would need to be done in a way that preserves backwards compatibility. This would be a large engineering effort with an uncertain final result. Furthermore, writing a new library allows us to additionally optimize for modern UTF-8-native environments.
Except for JavaScript via j2cl, Java is not a suitable source language for portability to low-resource environments like wearables. Further, ICU4J has many interdependent parts that would require a large amount of effort to bring to a state where it could be a viable j2cl source.
Some of our stakeholders (Firefox and Fuchsia) are drawn to Rust's memory safety. Like most complex C++ projects, ICU4C has had its share of CVEs, mostly relating to memory safety. Although C++ diagnostic tools are improving, Rust has very strong guarantees that are impossible in other software stacks.
For all these reasons, we decided that a Rust-based library was the best long-term choice.
II.B. Why use ICU4X when there is i18n in the platform?
Many of the same people who work on ICU4X also work to make i18n available in the platform (browser, mobile OS, etc.) through APIs such as the ECMAScript Intl object, android.icu, and other smartphone native libraries. ICU4X complements the platform-based solutions as the ideal polyfill:
Some platform i18n features take 5 or more years to gain wide enough availability to be used in client-side applications. ICU4X can bridge the gap.
ICU4X can enable clients to add more locales than those available in the platform.
Some clients prefer identical behavior of their app across multiple devices. ICU4X can give them this level of consistency.
Eventually, we hope that ICU4X will back platform implementations in ECMAScript and elsewhere, providing a maximal amount of consistency when ICU4X is also used as a polyfill.
II.C Why pluggable data?
One of the most visible departures that ICU4X makes from ICU4C and ICU4J is an explicit data provider argument on most constructor functions. The ICU4X data provider supports the following use cases:
Data files that are readable by both older and newer versions of the code; for more detail on how this works, see ICU4X Data Versioning Design
Data files that can be swapped in and out at runtime, making it easy to upgrade Unicode, CLDR, or time zone database versions. Swapping in new data can be done at runtime without needing to restart the application or clear internal caches.
Multiple data sources. For example, some data may be baked into the app, some may come from the operating system, and some may come from an HTTP service.
Customizable data caches. We recognize that there is no "one size fits all" approach to caching, so we allow the client to configure their data pipeline with the appropriate type of cache.
Fully configurable data fallbacks and overlays. Individual fields of ICU4X data can be selectively overridden at runtime.
III. How We Made ICU4X Lightweight
There are three factors that combine to make code lightweight: small binary size, low memory usage, and deliberate performance optimizations. For all three, we have metrics that are continuously measured on GitHub Actions continuous integration (CI).
III.A. Small Binary Size
Internationalization involves a large number of components with many interdependencies. To combat this problem, ICU4X optimizes for "tree shaking" (dead code elimination) by:
Minimizing the number of dependencies of each individual component.
Using static types in ways that scope functions to the pieces of data they need.
Splitting functions and classes that pull in more data than they need into multiple, smaller pieces.
Developers can statically link ICU4X and run a tree-shaking tool like LLVM link-time optimization (LTO) to produce a very small amount of compiled code, and then they can run our static analysis tool to build an optimally small data file for it.
In addition to static analysis, ICU4X supports dynamic data loading out of the box. This is the ultimate solution for supporting hundreds of languages, because new locale data can be downloaded on the fly only when they are needed, similar to message bundles for UI strings.
III.B. Low Memory Usage
At its core, internationalization transforms inputs to human-readable outputs, using locale-specific data. ICU4X introduces novel strategies for runtime loading of data involving zero memory allocations:
Supports Postcard-format resource files for dynamically loaded, zero-copy deserialized data across all architectures.
Supports compile-time linking of required data without deserialization overhead via DataBake.
Data schema is designed so that individual components can use the immutable locale data directly with minimal post-processing, greatly reducing the need for internal caches.
Explicit "data provider" argument to each function that requires data, making it very clear when data is required.
ICU4X team member Manish Goregaokar wrote a blog post series detailing how the zero-copy deserialization works under the covers.
III.C. Deliberate Performance Optimizations
Reducing CPU usage improves latency and battery life, important to most clients. ICU4X achieves low CPU usage by:
Writing in Rust, a high-performance language.
Utilizing zero-copy deserialization.
Measuring every change against performance benchmarks.
The ICU4X team uses a benchmark-driven approach to achieve highly competitive performance numbers: newly added components should have benchmarks, and future changes to those components should avoid regressing on those benchmarks.
Although we always seek to improve performance, we do so deliberately. There are often space/time tradeoffs, and the team takes a balanced approach. For example, if improving performance requires increasing or duplicating the data requirements, we tend to favor smaller data, like we've done in the normalizer and collator components. In the segmenter components, we offer two modes: a machine learning LSTM segmenter with lower data size but heavier CPU usage, and a dictionary-based segmenter with larger data size but faster. (There is ongoing work to make the LSTM segmenter require fewer CPU resources.)
IV. How We Made ICU4X Portable
The software ecosystem continually evolves with new programming languages. The "X" in ICU4X is a nod to the second main design goal: portability to many different environments.
ICU4X is Unicode's first internationalization library to have official wrappers in more than one target language. We do this with a tool we designed called Diplomat, which generates idiomatic bindings in many programming languages that encourage i18n best practices. Thanks to Diplomat, these bindings are easy to maintain, and new programming languages can be added without needing i18n expertise.
Under the covers, ICU4X is written in no_std Rust (no system dependencies) wrapped in a stable ABI that Diplomat bindings invoke across foreign function interface (FFI) or WebAssembly (WASM). We have some basic tutorials for using ICU4X from C++ and JavaScript/TypeScript.
V. What’s next?
ICU4X represents an exciting new step in bringing internationalized software to more devices, use cases, and programming languages. A Unicode working group is hard at work on expanding ICU4X’s feature set over time so that it becomes more useful and performant; we are eager to learn about new use cases and have more people contribute to the project.
Have questions? You can contact us on the ICU4X discussion forum!
Want to try it out? See our tutorials, especially our Intro tutorial!
Interested in getting involved? See our Contribution Guide.
Want to stay posted on future ICU4X updates? Sign up for our low-traffic announcements list, icu4x-announce@unicode.org!
Over 144,000 characters are available for adoption
to help the Unicode Consortium’s work on digitally disadvantaged languages
![[badge]](https://www.unicode.org/announcements/ynh-infinity.png)
![[beta image]](https://www.unicode.org/announcements/cldr40-alpha-annc.png)
The Unicode CLDR v42 Beta is now available for specification review and integration testing. The release is planned for Oct 19, but any feedback on the specification needs to be submitted well in advance of that date. The specification is available at Draft LDML Modifications. The biggest change is the new Person Names Formatting section.
![[Nag Mundari image]](https://www.unicode.org/announcements/u15-rel-annc.jpg)
Version 15.0 of the Unicode Standard is now available, including
the core specification, annexes, and data files. This version adds 4,489
characters, bringing the total to 149,186 characters. These additions include
two new scripts, for a total of 161 scripts, along with 20 new emoji characters,
and 4,193 CJK (Chinese, Japanese, and Korean) ideographs.
The new scripts and characters in Version 15.0 add support for
modern language groups including:![[Image credit Noto Emoji]](https://www.unicode.org/announcements/u15-emoji-annc-large.png "[Image credit Noto Emoji: https://github.com/googlefonts/noto-emoji]")
Posts
