Hacky semantic-release for monorepos
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enough for important production use as it's pretty dependent on how semantic-release works (so it may break or get out-of-date in future versions of semantic-release).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).
- CLI & JS API
- Automated & configurable cross-pkg version bumping
- Provides alpha & beta-branched release flow
- Supports npm (v7+), yarn, pnpm (with limitations), bolt-based monorepos
- Optional packages ignoring
- Linux/MacOs/Windows support
- multi-semantic-release
yarn add multi-semantic-release --dev
npm i multi-semantic-release -D
- Node.js >= 10
- git-notes enabled
multi-semantic-release [options]
npx multi-semantic-release [options]
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 automatically detects packages within workspaces for the following package-managers:
Make sure to have a workspaces
attribute inside your package.json
project file. In there, you can set a list of packages that you might want to process in the msr process, as well as ignore others. For example, let's say your project has 4 packages (i.e. a, b, c and d) and you want to process only a and d (ignore b and c). You can set the following structure in your package.json
file:
{
"name": "msr-test-yarn",
"author": "Dave Houlbrooke <[email protected]",
"version": "0.0.0-semantically-released",
"private": true,
"license": "0BSD",
"engines": {
"node": ">=8.3"
},
"workspaces": [
"packages/*",
"!packages/b/**",
"!packages/c/**"
],
"release": {
"plugins": [
"@semantic-release/commit-analyzer",
"@semantic-release/release-notes-generator"
],
"noCi": true
}
}
Make sure to have a packages
attribute inside your pnpm-workspace.yaml
in the root of your project.
In there, you can set a list of packages that you might want to process in the msr process, as well as ignore others.
For example, let's say your project has 4 packages (i.e. a, b, c and d) and you want to process only a and d (ignore b and c). You can set the following structure in your pnpm-workspace.yaml
file:
packages:
- 'packages/**'
- '!packages/b/**'
- '!packages/c/**'
Note, workspace:
prefix in pkg versions is not supported yet. issues/85
Make sure to have a bolt.workspaces
attribute inside your package.json
project file.
In there, you can set a list of packages that you might want to process in the msr process, as well as ignore others.
For example, let's say your project has 4 packages (i.e. a, b, c and d) and you want to process only a and d (ignore b and c). You can set the following structure in your package.json
file:
{
"name": "msr-test-bolt",
"author": "Dave Houlbrooke <[email protected]",
"version": "0.0.0-semantically-released",
"private": true,
"license": "0BSD",
"engines": {
"node": ">=8.3"
},
"bolt": {
"workspaces": [
"packages/*",
"!packages/b/**",
"!packages/c/**"
]
},
"release": {
"plugins": [
"@semantic-release/commit-analyzer",
"@semantic-release/release-notes-generator"
],
"noCi": true
}
}
There are several tweaks to adapt msr to some corner cases:
Flag | Type | Description | Default |
---|---|---|---|
--sequential-init |
bool | Avoid hypothetical concurrent initialization collisions | false |
--debug |
bool | Output debugging information | false |
--first-parent |
bool | Apply commit filtering to current branch only | false |
--deps.bump |
string | Define deps version update rule.
|
override |
--deps.release |
string | Define release type for dependent package if any of its deps changes.
|
patch |
--deps.prefix |
string | Optional prefix to be attached to the next version if --deps.bump set to override . Supported values: ^ | ~ | '' (empty string) |
'' (empty string) |
--dry-run |
bool | Dry run mode | false |
--ignore-packages |
string | Packages list to be ignored on bumping process (append to the ones that already exist at package.json workspaces) | null |
--ignore-private-packages |
bool | Private packages will be ignored | false |
Examples:
$ multi-semantic-release --debug
$ multi-semantic-release --deps.bump=satisfy --deps.release=patch
$ multi-semantic-release --ignore-packages=packages/a/**,packages/b/**
You can also combine the CLI --ignore-packages
options with the !
operator at each package inside package.json.workspaces
attribute. Even though you can use the CLI to ignore options, you can't use it to set which packages to be released – i.e. you still need to set the workspaces
attribute inside the package.json
.
allowUnknownFlags
is enabled, so the rest of flags will be passed to internal semrel
call as options
argument for all packages.
multi-semantic-release default exports a multirelease()
method which takes the following arguments:
packages
An array containing string paths topackage.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).
const multirelease = require("multi-semantic-release");
multirelease([
`${__dirname}/packages/my-pkg-1/package.json`,
`${__dirname}/packages/my-pkg-2/package.json`,
]);
Multi-semantic release seems to be compatible with many CI/CD systems. At least we are sure about three of them, here are examples of configurations:
- GitHub Actions → https://github.com/qiwi/semantic-release-toolkit
- Travis CI → https://github.com/qiwi/pijma
- AppVeyor → https://github.com/qiwi/masker
When releasing a monorepo you may get a npm ERR! code ETARGET
error. This is caused by npm version
creating a reify update on packages with future dependency versions MSR has not updated yet.
The simplest work around is to set workspaces-update to false either in your .npmrc or manually by running npm config set workspaces-update false
When releasing a monorepos you may get EINVALIDNPMTOKEN
error. The more packages, the more chance of error, unfortunately.
INVALIDNPMTOKEN Invalid npm token.
The npm token (https://github.com/semantic-release/npm/blob/master/README.md#npm-registry-authentication) configured in the NPM_TOKEN environment variable must be a valid token (https://docs.npmjs.com/getting-started/working_with_tokens) allowing to publish to the registry https://registry.npmjs.org/.
Do not rush to change your token. Perhaps this is related to npm whoami
request throttling on your registry (just a hypothesis: semantic-release/npm#416). At this point you can:
- Rerun your build as many times as necessary. You may get lucky in a new attempt.
- Use semrel-extra/npm plugin for npm publishing (recommended).
This error seems to be related to concurrent git invocations (issues/24). Or maybe not.
Anyway we've added a special --sequental-init
flag to queue up these calls.
Automatically finds packages as long as workspaces are configured as-per the workspace-feature of one of the support package managers.
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 too - Before packages are released (in semantic-release's prepare step), the correct current/next version number of all local dependencies is written into the
package.json
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 only - Calls
plugins.analyzeCommits()
to get the next release type (e.g. from @semantic-release/commit-analyzer) - 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
patch
if that's true
- Replaces
- verifyRelease: not used
- generateNotes:
- Calls
plugins.generateNotes()
to get the notes (e.g. from @semantic-release/release-notes-generator) - Appends a section listing any local deps bumps (e.g. "my-pkg-2: upgraded to 1.2.1")
- Calls
- prepare:
- Writes in the correct version for local deps in
dependencies
,devDependencies
,peerDependencies
inpackage.json
- Serialize the releases so they happen one-at-a-time (because semantic-release calls
git push
asynchronously, multiple releases at once fail because Git refs aren't locked — semantic-release should useexeca.sync()
so Git operations are atomic)
- Writes in the correct version for local deps in
- 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:
- Had to filter
context.commits
object before it was used by@semantic-release/commit-analyzer
(so it only lists commits for the corresponding directory).
- The actual Git filtering is easy peasy: see getCommitsFiltered.js
- But overriding
context.commits
was very difficult! I did it eventually creating an inline plugin and passing it intosemanticRelease()
viaoptions.plugins
- The inline plugin proxies between semantic release and other configured plugins. It does what it needs to then calls e.g.
plugins.analyzeCommits()
with an overriddencontext.commits
— see createInlinePluginCreator.js - I think this is messy — inline plugins aren't even documented :(
- 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.
dependencies
in package.json points to an internal package) this step also does apatch
bump if any of them did a bump. - This has to work recursively! See hasChangedDeep.js
- The configuration can be layered (i.e. global
.releaserc
and then per-directory overrides for individual packages).
- Had to duplicate the internal cosmiconfig setup from semantic release to get this working :(
- I found Git getting itself into weird states because e.g.
git tag
is done asynchronously
- 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
execa()
in semantic release should be replaced withexeca.sync()
to ensure Git's internal state is atomic. - Fortunately, another workaround has been implemented.
Synchronizer
is the neat part. It is critical to make the tag and commit publishing phases strictly sequential. Event emitter allows:- To synchronize release stages for all packages.
- To ensure the completeness of checks and the sufficiency of conditions for a conflict-free process.
Releases always use a tagFormat
of [email protected]
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)
Feel free to open any kind of issues: bugs, feature requests or questions. You're always welcome to suggest a PR. Just fork this repo, write some code, add a pinch of tests and push your changes. Any feedback is appreciated.