At Box, we’re modernizing our front-end architecture with frameworks such as react and redux, as well as webpack for the build process. We’ve tackled many performance issues with the new tools and libraries adopted along the way, and would like to share how we were recently able to improve our webapp build time drastically.
TL;DR: We experimented with multiple configurations to accelerate the webpack build performance for a webapp that supports 25 locales, we found that with
- a single CPU
- enabling caching on the babel-loader improves performance by 26%
- enabling caching and parallel processing with uglifyjs-webpack-plugin v1 improves performance by 45%
- NOT exporting an array of configurations from webpack.config.js improves performance by 28%
- 10 CPUs
- leveraging node.js Child Process API improves performance by 67%
Result: build performance improved by 95% (from 3 hours 21 minutes to 9 minutes)
The Box webapp supports 25 locales (and growing), to provide international users with the best experience. However, this brings challenges to our build performance, since each locale needs to have its own bundle. At the worst time, the entire build would take as long as 3 hours and 21 minutes. Knowing that poor build performance hurts the release flexibility, and can be a huge distraction to the release engineer, we started the investigation on our webpack build. We learned some lessons that would apply to most webpack projects that support multiple locales or build multiple variations of configurations.
If your build machines have limited memory, don’t export an array of configurations in your webpack.config.js file
Webpack allows you to export multiple configurations as an array from the webpack.config.js file. Previously we were taking advantage of this feature to export an array of 25 configurations, one for each locale, so when we run webpack all 25 locales’ bundles are built at the same time. This could be a useful feature if you only have a small set of configurations, however, in our case, it made the build process very memory intensive and unoptimized. Initially, we set the --max_old_space_size (memory limit) of the node process running webpack to 4GB, by doubling the limit to 8GB, we improved the performance by 50% (build time down from 3.35 hours to 1.7 hours).
However, with the same memory limit, when switched to running webpack on each configuration sequentially, we improved the performance by another 28% compared to processing multiple configurations at the same time.
Enable caching on the babel-loader
Babel-loader caching can be easily enabled with the cacheDirectory option. With
babel-loader caches its results in node_modules/.cache/babel-loader by default. After enabling caching with babel-loader we improved the build performance by another 26%. Although configurations vary among locales, most of the modules are the same. By enabling caching, expensive babel recompilations are avoided for subsequent locales being processed.
The new uglifyjs-webpack-plugin v1 makes a huge difference
You’re probably already using the UglifyJS plugin that ships with webpack 3, webpack.optimize.UglifyJsPlugin, either by adding it explicitly into the configuration or running webpack with -p flag. It is identical to uglifyjs-webpack-plugin v0.4.6, which has UglifyJS v2 at its core.
The new uglifyjs-webpack-plugin v1 uses UglifyJS v3 under the hood and is scheduled for webpack 4. Its new features such as multi-process parallel running support and caching capability improved the build performance significantly by 45%.
In order to use uglifyjs-webpack-plugin v1 today with webpack 3, you need to
- add "uglifyjs-webpack-plugin": "1.0.1" to devDependencies
- remove -p flag passed to webpack, to avoid uglifyjs-webpack-plugin v0.4.6 from running unnecessarily
Then add the plugin with cache and parallel options enabled and you’re good to go!
Take advantage of your multi-core build machine
In case your build machine has multiple CPUs, you can take advantage of those for your webpack build by using node.js Cluster and Child Process API, or worker-farm. By enabling 9 webpack configurations being processed in parallel, we managed to shorten the build process for processing all 25 configurations by 67%.
Reliability is important for asset building, we want to make sure that in case where build for a single locale fails, the entire build should fail loudly. Also, to prevent potential race conditions, we should disable babel-loader and uglifyjs-webpack-plugin cache options since multiple processes running in parallel are all trying to read/write to the file system simultaneously.
With worker-farm, the script for building all locales looks like the following
Before we went with our own multi-core solution we also tried parallel-webpack which promises to run webpack configurations in parallel, however it didn’t make much improvement to our build performance. Having to work with another interface that’s slightly different from webpack is also undesirable.