[](https://travis-ci.com/dhoulb/multi-semantic-release) [




Proof of concept that wraps semantic-release to work with monorepos.
This package should work well, but may not be fundamentally stable enough for important production use as it's pretty dependent on how semantic-release works (so it may break or get out-of-date in future versions of semantic-release).
One of the best things about semantic-release is forgetting about version numbers. In a monorepo though there's still a lot of version number management required for local deps (packages in the same monorepo referenced in dependencies or devDependencies or peerDependencies). However in multi-semantic-release the version numbers of local deps are written into package.json at release time. This means there's no need to hard-code versions any more (we recommend just using * asterisk instead in your repo code).
``sh`
yarn add multi-semantic-release --dev
`sh`
multi-semantic-release
Configuration for releases is the same as semantic-release configuration, i.e. using a release key under package.json or in .releaserc file of any type e.g. .yaml, .json.
_But_ in multi-semantic-release this configuration can be done globally (in your top-level dir), or per-package (in that individual package's dir). If you set both then per-package settings will override global settings.
multi-semantic-release does not support any command line arguments (this wasn't possible without duplicating files from semantic-release, which I've tried to avoid).
multi-semantic-release default exports a multirelease() method which takes the following arguments:
- packages An array containing string paths to package.json filesoptions
- An object containing default semantic-release configuration options
multirelease() returns an array of objects describing the result of the multirelease (corresponding to the packages array that is passed in).
`js
const multirelease = require("multi-semantic-release");
multirelease([
${__dirname}/packages/my-pkg-1/package.json,${__dirname}/packages/my-pkg-2/package.json
,`
]);
Automatically finds packages as long as workspaces are configured as-per Yarn workspaces. _You don't need to use Yarn but the way they define monorepos seems intuitive, and is likely what NPM will copy when they add this functionality (as rumoured)._
I'm aware Lerna is the best-known tool right now, but in future it seems clear it will be replaced by functionality in Yarn and NPM directly. If you use Yarn workspaces today (January 2019), then publishing is the only remaining feature Lerna is _really_ required for (though it'd be lovely if Yarn added parallel script execution). Thus using multi-semantic-release means you can probably remove Lerna entirely from your project.
Other packages that enable semantic-release for monorepos work by iterating into each package and running the semantic-release command. This is conceptually simple but unfortunately not viable because:
- If a package is published that depends on minor changes that have been made in a sibling package it could cause extremely subtle errors (the worst kind!) — if the project follows semver religiously this should never happen, but it's better to eliminate the _potential_ for errors
- Dependency version numbers need to reflect the _next_ release at time of publishing, so a package needs to know the state of _all other packages_ before it can publish correctly — this central state needs to be coordinated by something
A key requirement is handling local dep version numbers elegantly. multi-semantic-release does the following:
- The next version number of all packages is established first
- If a release has not changed but has local deps that _have_ changed... do a patch bump on that package toopackage.json
- Before packages are released (in semantic-release's prepare step), the correct current/next version number of _all_ local dependencies is written into the file (overwriting any existing value)
- This ensures the package at the time of publishing will be atomically correct with all other packages in the monorepo.
The above means that, possibly, if someone upgrades dependencies and pulls down a package from NPM _during the multirelease_ (before all its deps have been published at their next versions), then their npm install will fail (it will work if they try again in a few minutes). On balance I thought it was more important to be atomically correct (this situation should be fairly rare assuming projects commit their lockfiles).
This is the jankiest part of multi-semantic-release and most likely part to break relies. I expect this to cause maintenance issues down the line. In an ideal world semantic-release will bake-in support for monorepos (making this package unnecessary).
The way I ended up integrating is to create a custom "inline plugin" for semantic-release, and passing that in to semanticRelease() as the only plugin. This then calls any other configured plugins to retrieve and potentially modify the response.
The plugin starts all release at once, then pauses them (using Promises) at various points to allow other packages in the multirelease to catch up. This is mainly needed so the version number of all packages can be established _before_ any package is released. This allows us to do a patch bump on releases whose local deps have bumped, and to accurately write in the version of local deps in each package.json
The inline plugin does the following:
- verifyConditions: _not used_
- analyzeCommits:
- Replaces context.commits with a list of commits filtered to the folder onlyplugins.analyzeCommits()
- Calls to get the next release type (e.g. from @semantic-release/commit-analyzer) patch
- Waits for _all_ packages to catch up to this point.
- For packages that haven't bumped, checks if it has local deps (or deps of deps) that have bumped and returns if that's trueplugins.generateNotes()
- verifyRelease: _not used_
- generateNotes:
- Calls to get the notes (e.g. from @semantic-release/release-notes-generator) dependencies
- Appends a section listing any local deps bumps (e.g. "my-pkg-2: upgraded to 1.2.1")
- prepare:
- Writes in the correct version for local deps in , devDependencies, peerDependencies in package.jsongit push
- Serialize the releases so they happen one-at-a-time (because semantic-release calls asyncronously, multiple releases at once fail because Git refs aren't locked — semantic-release should use execa.sync() so Git operations are atomic)
- publish: _not used_
- success: _not used_
- fail: _not used_
The integration with semantic release is pretty janky — this is a quick summary of the reasons this package will be hard to maintain:
1. Had to filter context.commits object before it was used by @semantic-release/commit-analyzer (so it only lists commits for the corresponding directory).context.commits
- The actual Git filtering is easy peasy: see getCommitsFiltered.js
- But overriding was very difficult! I did it eventually creating an _inline plugin_ and passing it into semanticRelease() via options.pluginsplugins.analyzeCommits()
- The inline plugin proxies between semantic release and other configured plugins. It does what it needs to then calls e.g. with an overridden context.commits — see createInlinePluginCreator.js dependencies
- I think this is messy — inline plugins aren't even documented :(
2. Need to run the analyze commit step on all plugins before any proceed to the publish step
- The inline plugin returns a Promise for every package then waits for all packages to analyze their commits before resolving them one at a time
- If packages have local deps (e.g. in package.json points to an internal package) this step also does a patch bump if any of them did a bump..releaserc
- This has to work recursively! See hasChangedDeep.js
3. The configuration can be layered (i.e. global and then per-directory overrides for individual packages). git tag
- Had to duplicate the internal cosmiconfig setup from semantic release to get this working :(
4. I found Git getting itself into weird states because e.g. is done asynchronouslyexeca()
- To get around this I had to stagger package publishing so they were done one at a time (which slows things down)
- I think calls to in semantic release should be replaced with execa.sync() to ensure Git's internal state is atomic. For an experiment, you may add --execasync CLI flag that makes all calls synchronous through ritm-hook.
Releases always use a tagFormat of my-pkg-1@1.0.1 for Git tags, and always overrides any gitTag set in semantic-release configuration.
I can personally see the potential for this option in coordinating a semantic-release (e.g. so two packages with the same tag always bump and release simultaneously). Unfortunately with the points of integration available in semantic-release, it was effectively impossible when releasing to stop a second package creating a duplicate tag (causing an error).
To make the tagFormat option work as intended the following would need to happen:
- semantic-release needs to check if a given tag already exists at a given commit, and not create it / push it if that's true
- Release notes for multiple package releases need to be merged BUT the Github release only done once (by having the notes merged at the semantic-release level but only published once, or having the Github plugin merge them)
- Make it clear in documentation that the default tag v1.0.0` will have the same effect as Lerna's fixed mode (all changed monorepo packages released at same time)