@osd/optimizer is a package for building OpenSearch Dashboards platform UI plugins (and hopefully more soon).
OpenSearch Dashboards Platform plugins with "ui": true in their opensearch_dashboards.json file will have their public/index.ts file (and all of its dependencies) bundled into the target/public directory of the plugin. The build output does not need to be updated when other plugins are updated and is included in the distributable without requiring that we ship @osd/optimizer 🎉.
The Webpack config is designed to provide the majority of what was available in the legacy optimizer and is the same for all plugins to promote consistency and keep things sane for the operations team. It has support for JS/TS built with babel, url imports of image and font files, and support for importing scss and css files. SCSS is pre-processed by postcss, built for both light and dark mode and injected automatically into the page when the parent module is loaded (page reloads are still required for switching between light/dark mode). CSS is injected into the DOM as it is written on disk when the parent module is loaded (no postcss support).
Source maps are enabled except when building the distributable. They show the code actually being executed by the browser to strike a balance between debuggability and performance. They are not configurable at this time but will be configurable once we have a developer configuration solution that doesn't rely on the server (see #615).
To make front-end code easier to debug the optimizer uses the BROWSERSLIST_ENV=dev environment variable (by default) to build JS and CSS that is compatible with modern browsers. In order to support all browsers that we support with the distributable you will need to specify the BROWSERSLIST_ENV=production environment variable or build a distributable for testing.
The @osd/optimizer is automatically executed from the dev cli, the OpenSearch Dashboards build scripts, and in CI. If you're running OpenSearch Dashboards locally in some other way you might need to build the plugins manually, which you can do by running node scripts/build_opensearch_dashboards_platform_plugins (pass --help for options).
You can limit the number of workers the optimizer uses by setting the OSD_OPTIMIZER_MAX_WORKERS environment variable. You might want to do this if your system struggles to keep up while the optimizer is getting started and building all plugins as fast as possible. Setting OSD_OPTIMIZER_MAX_WORKERS=1 will cause the optimizer to take the longest amount of time but will have the smallest impact on other components of your system.
We only limit the number of workers we will start at any given time. If we start more workers later we will limit the number of workers we start at that time by the maximum, but we don't take into account the number of workers already started because it is assumed that those workers are doing very little work. This greatly simplifies the logic as we don't ever have to reallocate workers and provides the best performance in most cases.
Bundles built by the optimizer include a cache file which describes the information needed to determine if the bundle needs to be rebuilt when the optimizer is restarted. Caching is enabled by default and is very aggressive about invalidating the cache output, but if you need to disable caching you can pass --no-cache to node scripts/build_opensearch_dashboards_platform_plugins, or set the OSD_OPTIMIZER_NO_CACHE environment variable to anything (env overrides everything).
When a bundle is determined to be up-to-date a worker is not started for the bundle. If running the optimizer with the --dev/--watch flag, then all the files referenced by cached bundles are watched for changes. Once a change is detected in any of the files referenced by the built bundle a worker is started. If a file is changed that is referenced by several bundles then workers will be started for each bundle, combining workers together to respect the worker limit.
In order to dramatically reduce the size of our bundles, and the time it takes to build them, bundles will "ref" other bundles being built at the same time. When the optimizer starts it creates a list of "refs" that could be had from the list of bundles being built. Each worker uses that list to determine which import statements in a bundle should be replaced with a runtime reference to the output of another bundle.
At runtime the bundles share a set of entry points via the __osdBundles__ global. By default a plugin shares public so that other code can use relative imports to access that directory. To expose additional directories they must be listed in the plugin's opensearch_dashboards.json "extraPublicDirs" field. The directories listed there will also be exported from the plugins bundle so that any other plugin can import that directory. "common" is commonly in the list of "extraPublicDirs".
NOTE: We plan to replace the
extraPublicDirsfunctionality soon with better mechanisms for statically sharing code between bundles.
When a directory is listed in the "extraPublicDirs" it will always be included in the bundle so that other plugins have access to it. The worker building the bundle has no way of knowing whether another plugin is using the directory, so be careful of adding test code or unnecessary directories to that list.
Any import in a bundle which resolves into another bundles "context" directory, ie src/plugins/*, must map explicitly to a "public dir" exported by that plugin. If the resolved import is not in the list of public dirs an error will be thrown and the optimizer will fail to build that bundle until the error is fixed.
SASS imports in bundles are automatically converted to CSS for one or more themes. In development we build the v7light and v7dark themes by default to improve build performance. When producing distributable bundles the default shifts to * so that the distributable bundles will include all themes, preventing the bundles from needing to be rebuilt when users change the active theme in OpenSearch Dashboards's advanced settings.
To customize the themes that are built for development you can specify the OSD_OPTIMIZER_THEMES environment variable to one or more theme tags, or use * to build styles for all themes. Unfortunately building more than one theme significantly impacts build performance, so try to be strategic about which themes you build.
Currently supported theme tags: v7light, v7dark, v8light, v8dark
Examples:
# start OpenSearch Dashboards with only a single theme
OSD_OPTIMIZER_THEMES=v7light yarn start
# start OpenSearch Dashboards with dark themes for version 7 and 8
OSD_OPTIMIZER_THEMES=v7dark,v8dark yarn start
# start OpenSearch Dashboards with all the themes
OSD_OPTIMIZER_THEMES=* yarn startTo run the optimizer from code, you can import the OptimizerConfig class and runOptimizer function. Create an OptimizerConfig instance by calling it's static create() method with some options, then pass it to the runOptimizer function. runOptimizer() returns an observable of update objects, which are summaries of the optimizer state plus an optional event property which describes the internal events occuring and may be of use. You can use the logOptimizerState() helper to write the relevant bits of state to a tooling log or checkout it's implementation to see how the internal events like WorkerStdio and WorkerStarted are used.
Example:
import { runOptimizer, OptimizerConfig, logOptimizerState } from '@osd/optimizer';
import { REPO_ROOT } from '@osd/utils';
import { ToolingLog } from '@osd/dev-utils';
const log = new ToolingLog({
level: 'verbose',
writeTo: process.stdout,
})
const config = OptimizerConfig.create({
repoRoot: Path.resolve(__dirname, '../../..'),
watch: false,
dist: true
});
await lastValueFrom(
runOptimizer(config).pipe(logOptimizerState(log, config))
);This is essentially what we're doing in script/build_opensearch_dashboards_platform_plugins and the new build system task.
The optimizer runs webpack instances in worker processes. Each worker is configured via a WorkerConfig object and an array of Bundle objects which are JSON serialized and passed to the worker as it's arguments.
Plugins/bundles are assigned to workers based on the number of modules historically seen in each bundle in an effort to evenly distribute the load across the worker pool (see assignBundlesToWorkers).
The number of workers that will be started at any time is automatically chosen by dividing the number of cores available by 3 (minimum of 2).
The WorkerConfig includes the location of the repo (it might be one of many builds, or the main repo), wether we are running in watch mode, wether we are building a distributable, and other global config items.
The Bundle objects which include the details necessary to create a webpack config for a specific plugin's bundle (created using webpack.config.ts).
Each worker communicates state back to the main process by sending WorkerMsg and CompilerMsg objects using IPC.
The Optimizer captures all of these messages and produces a stream of update objects.
Optimizer phases:
'initializing'- Initial phase, during this state the optimizer is validating caches and determining which builds should be built initially.
'initialized'- Emitted by the optimizer once it's don't initializing its internal state and determined which bundles are going to be built initially.
'running'- Emitted when any worker is in a running state. To determine which compilers are running, look for
BundleStateobjects with type'running'. 'issue'- Emitted when all workers are done running and any compiler completed with a
'compiler issue'status. Compiler issues include things like "unable to resolve module" or syntax errors in the source modules and can be fixed by users when running in watch mode. 'success'- Emitted when all workers are done running and all compilers completed with
'compiler success'. 'reallocating'- Emitted when the files referenced by a cached bundle have changed, before the worker has been started up to update that bundle.
Workers have several error message they may emit which indicate unrecoverable errors. When any of those messages are received the stream will error and the workers will be torn down.
For an example of how to handle these states checkout the logOptimizerState() helper.