2.0.1 • Published 5 years ago

@use-cdn/karma v2.0.1

Weekly downloads
-
License
MIT
Repository
github
Last release
5 years ago

Configuration

Create a use-cdn.conf.js file in the cwd of your test runner. (This often is the top directory of your project.) It must export an object with the following options:

  • cdn: string (optional): The CDN to use for those packages that don't have a cdn field. This effectively lets you define a default CDN for the whole configuration.

  • cdns: object (optional): this object allows configuring the various CDNs that you use in your configuration. A key of this object is the name of a CDN and the values are objects that contain configuration settings. See CDN Configuration below.

  • resolvers: object (optional): this object allows configuring the various version resolvers that you use in your configuration. A key of this object is the name of a version resolver and the values are objects that contain configuration settings. See Version Resolver Configuration below.

  • packages: object: an array of package configuration. See Package Configuration below.

CDN Configuration

A CDN configuration may contain the following settings:

  • url: string (optional): you would use this to override the default URL associated with the CDN. This may be useful if you want to use a mirror of the CDN, for instance.

  • resolver: string (optional): The version resolver to use for this CND. If a resolver is not provided, the npm resolver is used.

Version Resolver Configuration

A version resolver configuration may contain the following settings:

  • url: string (optional): you would use this to override the default URL associated with the resolver. This may be useful if you want to use a mirror of the version resolver, for instance.

Package Configuration

Each package configuration is an object with the following fields:

  • package: string: the name of the package. The specific format is dependent on the CDN you use.

  • resolveAs: string (optional): This is the package name that should be used while resolving the package. Unfortunately, it is possible for a package to be published to a CDN under a name and to a resolver under a different name. A good example of this is Bootstrap, which is published to Cdnjs as twitter-bootstrap and to NPM as bootstrap. A configuration that fetches Bootstrap from Cdnjs but uses NPM as the version resolver would have to use package: "twitter-bootstrap", resolveAs: "bootstrap".

  • cdn: string (optional): the CDN to use to fetch the package. Default: unpkg.

  • version: string: the version of the package. The specific format is dependent on the CDN you use.

  • files: a list of files to get from the CDN.

NOTE: use-cdn does not allow fetching the same package from two different CDNs in the same configuration. This is invalid:

{
  packages: [{
    package: "foo",
    cdn: "unpkg",
    [...],
  }, {
    package: "foo",
    cdn: "cdnjs",
    [...],
  }],
}

The name "foo" is duplicated. You also cannot have a resolveAs parameter which has the same value as another resolveAs or package field. This is invalid:

{
  packages: [{
    package: "foo",
    resolveAs: "bar",
    cdn: "unpkg",
    [...],
  }, {
    package: "bar",
    cdn: "cdnjs",
    [...],
  }],
}

The name "bar" is duplicated.

What's a version resolver? Why do I need one?

Suppose I have a package that I advertise as working with "the latest version of jQuery". In my test suite I want to test my code wiht "the latest version of jQuery". The problem is that "the latest version of jQuery" is a moving target. I'd like it if my test suite would just automatically load the latest version rather than require me to manually update the version number loaded in the tests every time a new jQuery version is released. Note that if I don't want to automatically get the latest version but rather to manually update the version myself, I can specify a "jquery" package with a specific version number. But if I do want the latest version, I need a way to tell use-cdn that I want the latest. Or if I want the latest "rc", I'd have update version numbers manually. The same thing happens if I want to setup a test suite that runs with, say, the latest version of jQuery 2. If use-cdn allowed only a literal and complete version number, then I'd have to chase new releases.

A version resolver solves these issues. The npm cli tool is a good example of a tools that provides good version resolution. You can specify a package version with a tag like "latest" or "dev" or "rc". And you can specify a package version like "2", which means "the latest version in the "2.x" series".

The problem for use-cdn is that some CDNs peform version resolution, and some don't. unpkg.com, for instance, performs version resolution in the same way npm does. However, cdnjs.com does not perform version resolution. So for CNDs that don't provide their own resolution method, an external version resolver can be used to provide some useful resolving semantics.

Configuration Examples

Here is an example that uses the unpkg.com CDN (which is the default), but overrides the default URL to use https://mirror/:

const semver = require("semver");

module.exports = {
  cdns: {
    unpkg: {
      url: "https://mirror/",
      resolver: "native",
    },
  },
  resolvers: {
    unpkg: {
      url: "https://mirror/",
    }
  }
  packages: [{
    package: "jquery",
    version: "latest",
    files: [
      "dist/jquery.js",
    ],
  }, {
    package: "bootstrap",
    version: "latest",
    files: [
      version => `dist/js/bootstrap.${semver.intersects(version, ">=4") ?
"bundle." : ""}js`,
      "dist/css/bootstrap.css",
    ],
  }],
};

This configuration fetches the JavaScript for the latest version of jQuery and fetches the JavaScript and CSS for the latest version of Bootstrap. The configuration for Bootstrap shows how a function can be used. Bootstrap 4 changed the name of the JavaScript file that should be loaded so the function tests the version number to determine which file to request.

If you need a configuration that only contains the packages array, you can reduce the configuration exported by your file to just that array, as in the following example:

const semver = require("semver");

module.exports = [{
  package: "jquery",
  version: "latest",
  files: [
    "dist/jquery.js",
  ],
}, {
  package: "bootstrap",
  version: "latest",
  files: [
    version => `dist/js/bootstrap.${semver.intersects(version, ">=4") ?
"bundle." : ""}js`,
    "dist/css/bootstrap.css",
  ],
}];

CDNs Supported

(PRs for adding other CDNs are welcome.)

unpkg

The package available on this CDN are those published to npmjs.com.

This CDN has a native version resolver which resolves version numbers in the same way the npm version resolver does. The version string is any single version number supported by NPM. For instance,

  • 1 refers to the latest version is the 1.x series.

  • latest refers to whatever version the latest dist-tag resolves to.

cdnjs

The package available on this CDN are those specifically published to this CDN.

This CDN does not have a native version resolver.

Version Resolvers Supported

npm

This is a resolver that uses registry.npmjs.org to resolve versions and tags. This is the default resolver for ALL CDNs for which you do not specify another resolver.

null

This is a resolver that just returns the version or tag passed to it. If asked to resolve "1", it will return "1". This resolver may be useful if you want to set specific versions manually. Using the "null" resolver for such cases allows you to skip the network requests that would be made for resolving.

unpkg

This is the resolver you get when specifying the "native" resolver on the "unpkg" CDN. This resolver queries unpkg.com to resolve tags and version numbers. In theory, this resolver performs the same resolution as the npm resolver since unpkg simply exposes the packages available on npm. However, this resolver uses unpkg.com whereas the npm resolver uses registry.npmjs.org.

Test Runners Supported

(PRs for adding other runners are welcome.)

Karma

To use use-cdn with Karma you need to install:

npm install @use-cdn/cli @use-cdn/karma

You need to create a configuration file as described above. In your Karma configuration you need:

  frameworks: ["use-cdn", ...],
  plugins: [
    "karma-*",
    "@use-cdn/karma",
  ],

You must list the plugin in plugins because Karma does not autodetect plugins in NPM scopes. Also, note how the plugin name is @use-cdn/karma but the framework name is use-cdn.

The use-cdn framework unshifts the files it adds to Karma into Karma's files configuration array. So the files it adds will always be in front of those your list in files. The files it adds all have the following flags:

included: true,
served: true,
watched: false,

You must also modify your build procedure so that the use-cdn CLI tool runs before Karma. For instance, your package.json could have a script:

  "test": "use-cdn && karma start --single-run",

Features

Caching

This library caches from run to run the file contents it gets from the CDNs. If a file that is needed is in the cache, then the CDN is not accessed again to fetch this file.

THIS LIBRARY ASSUMES THAT THE CONTENT OF A VERSION IS IMMUTABLE. It does not verify. It simply assumes. If foo@1.2.3 is in the cache and it is requested again, then this library will just get it from the cache. In theory, this can turn out to be an issue of a package is released through a package manager and served through a CDN that allow releasing a version number and then fixing it. Suppose Alice releases version foo@1.2.3. Bob accesses it through this library: it is cached. Then Alice realises a mistake and releases a new, and different foo@1.2.3. If Bob tries to get it through this library again, he'll get the old release, not the new one.

This is a problem unlikely to happen in practice because:

  • Sensible package managers don't allow re-releasing the same version.

  • Even if a package manager and CDN happen to allow re-releasing a version. It is terrible practice and should not be done. Sensible developers don't do it.

This assumption simplifies the caching code. If a version is on disk, it is available, period. We don't need to check with the CDN whether somehow it has changed.

Version integrity

When accessing a CDN naively, it is possible to run into a race condition if a new release happens to be published at the same time the accesses are done. Suppose I need to get foo.js and foo.css from foo@latest. When I grab foo.js when foo@latest is foo@1.2.3, but foo@2.0.0 happens to be published before I grab foo.css. So now I have a foo.js and a foo.css from different versions of the pacakge.

That's a problem that happens only rarely but this library does prevent it, because it is rather easy to prevent. It resolves tags to a specific version number once per run and uses the same version number for the whole run. So in the scenario above, version 1.2.3 would be used. Then on subsequent runs latest would resolve to the new version number.

Note, however, that version resolution is performed anew with each invocation of use-cdn. Say you run use-cdn once and it resolves foo@1 to foo@1.2.3 or foo@latest to foo@1.2.3. If you run use-cdn again, it will resolve the version again. If version 1.3 was released before your second run, then foo@1 and foo@latest will resolve to this new version.

Contributing

PRs are welcome. Make sure that:

  1. You add tests to maintain coverage.

  2. You lint your code.

Implementation Notes

The initial vision for this library was a Karma plugin consisting of a framework and middleware. The middleware would wait until the suite actually makes a request to a file to be served from a CDN to access the CDN and serve the file. Karma does not handle this well. When a file is added to the files configuration of Karma, the file is handled in one of these ways:

  • pattern is an absolute URL, in which case the pattern is taken as a literal URL. This pattern focibly has included: true and all the other options false. This absolute URL completely bypasses Karma at test time. So if the URL is absolute, no middleware can serve it.

  • pattern is not an absolute URL, in which case Karma scans the disk for files matching pattern. If there is no file on disk which matches, then this file has no further effect on the test run: Karma just prints a warning but otherwise ignores it.

Our framework cannot use an absolute URL because such URLs entirely bypass the framework: this does not allow the framework to caching responses from run to run, nor does it ensure version integrity.

If we use a non-absolute URL, the problem is that it has no effect unless there's a file on disk that Karma can use.

In order to implement the initial vision, there would have to be a way to get Karma to generate a script element for a resource found at a non-absolute URL and which does not yet exist on disk when Karma processes the files configuration. This is not yet possible.

Another approach was considered whereby a Karma pluging consisting only of a framework would fetch all the files prior to to Karma processing files. This requires that Karma handle frameworks that operate asynchronously because the network operations for fetching files are asynchronous. Unfortunately, Karma does not handle asynchronous frameworks. (It is possible to launch asynchronous operations in a framework but Karma won't wait for them to be done.)

2.0.1

5 years ago

2.0.0

5 years ago

1.0.0

6 years ago