Support having plugins as dependencies in shareable config #3458

Open
sindresorhus opened this Issue Aug 20, 2015 · 95 comments

Projects

None yet
@sindresorhus
Contributor

My shareable config uses rules from an external plugin and I would like to make it a dependency so the user doesn't have to manually install the plugin manually. I couldn't find any docs on this, but it doesn't seem to work, so I'll assume it's not currently supported.

module.js:338
    throw err;
          ^
Error: Cannot find module 'eslint-plugin-no-use-extend-native'
    at Function.Module._resolveFilename (module.js:336:15)
    at Function.Module._load (module.js:278:25)
    at Module.require (module.js:365:17)
    at require (module.js:384:17)
    at /usr/local/lib/node_modules/eslint/lib/cli-engine.js:106:26
    at Array.forEach (native)
    at loadPlugins (/usr/local/lib/node_modules/eslint/lib/cli-engine.js:97:21)
    at processText (/usr/local/lib/node_modules/eslint/lib/cli-engine.js:182:5)
    at processFile (/usr/local/lib/node_modules/eslint/lib/cli-engine.js:224:12)
    at /usr/local/lib/node_modules/eslint/lib/cli-engine.js:391:26

I assume it's because you only try to load the plugin when the config is finished merging.

Other shareable configs that depend on a plugin instructs the users to manually install the plugin too and they have it in peerDependencies. I find this sub-optimal though and I don't want the users to have to care what plugins my config uses internally.

The whole point of shareable configs is to minimize boilerplate and overhead, so this would be a welcome improvement.

Want to back this issue? Post a bounty on it! We accept bounties via Bountysource.

@eslintbot

Thanks for the issue! We get a lot of issues, so this message is automatically posted to each one to help you check that you've included all of the information we need to help you.

Reporting a bug? Please be sure to include:

  1. The version of ESLint you are using (run eslint -v)
  2. The source code that caused the problem
  3. The configuration you're using (for the rule or your entire config file)
  4. The actual ESLint output complete with line numbers

Requesting a new rule? Please be sure to include:

  1. The use case for the rule - what is it trying to prevent or flag?
  2. Whether the rule is trying to prevent an error or is purely stylistic
  3. Why you believe this rule is generic enough to be included

Requesting a feature? Please be sure to include:

  1. The problem you want to solve (don't mention the solution)
  2. Your take on the correct solution to problem

Including this information in your issue helps us to triage it and get you a response as quickly as possible.

Thanks!

@eslintbot eslintbot added the triage label Aug 20, 2015
@lo1tuma
Member
lo1tuma commented Aug 20, 2015

We already discussed this in #2518 with the conclusion that a
peerDependency is the way to go.
On Aug 20, 2015 9:09 AM, "ESLint Bot" notifications@github.com wrote:

Thanks for the issue! We get a lot of issues, so this message is
automatically posted to each one to help you check that you've included all
of the information we need to help you.

Reporting a bug? Please be sure to include:

  1. The version of ESLint you are using (run eslint -v)
  2. The source code that caused the problem
  3. The configuration you're using (for the rule or your entire config
    file)
  4. The actual ESLint output complete with line numbers

Requesting a new rule? Please be sure to include:

  1. The use case for the rule - what is it trying to prevent or flag?
  2. Whether the rule is trying to prevent an error or is purely
    stylistic
  3. Why you believe this rule is generic enough to be included

Requesting a feature? Please be sure to include:

  1. The problem you want to solve (don't mention the solution)
  2. Your take on the correct solution to problem

Including this information in your issue helps us to triage it and get you
a response as quickly as possible.

Thanks!


Reply to this email directly or view it on GitHub
#3458 (comment).

@sindresorhus
Contributor

Ugh, that's such a leaky abstraction. I guess I won't use plugins then...

The user shouldn't have to care what plugins I use for the rules. This is like requiring to manual install of Lodash when you want ESLint. A shareable config is a node module and should act like it.

@lo1tuma
Member
lo1tuma commented Aug 20, 2015

A shareable config is a node module and should act like it.

We use require to load shareable configs or plugins, what would make it act more like a node module than that?

This issue should be also solved when using npm version 3 which installs all subdependencies in the top-level node_modules folder.

@sindresorhus
Contributor

We use require to load shareable configs or plugins, what would make it act more like a node module than that?

Let the shareable config provide the plugin as an object:

module.exports = {
    plugins: [
        require('eslint-plugin-no-use-extend-native')
    ],
    env: {
        node: true
    },
    rules: {
        'comma-dangle': [2, 'never'],
        'no-cond-assign': 2
};

This issue should be also solved when using npm version 3 which installs all sub-dependencies in the top-level node_modules folder.

That's an implementation detail and not always guaranteed. Nobody should ever depend on that. npm@3 promises flatter dependency tree, not flat. If there are conflicts, there will be nesting.

@feross
Contributor
feross commented Aug 20, 2015

The option to require the plugin in the config itself would allow users to use my shareable config without needing to manually install two other plugins.

I like this proposal.

@lo1tuma
Member
lo1tuma commented Aug 20, 2015

@sindresorhus good point about npm 3, let's forget about this.

I kind of like your proposal, but it has a few problems:

  1. Eslint needs to know the name of the plugin. This could be solved easily by providing an object with the name e.g plugins: [ { 'eslint-plugin-foo' : require('eslint-plugin-foo')} ]
  2. Eslint caches plugins once they are loaded. This could be a problem if you use different shareable configs in different .eslintrc files where the shareable configs require the same plugin, but in a different version. Possible solution would be to avoid caching in this case.
@BYK
Member
BYK commented Aug 20, 2015

Possible solution would be to avoid caching in this case.

Or we can prefix plugins provided by shareable configs with the name of the config?

@lo1tuma
Member
lo1tuma commented Aug 20, 2015

@BYK how would you reference the rules then? configname/pluginname/rulename? But I guess we would have the same problem if we avoid caching. We can't determine to which version of the plugin the rule belongs to.

That said, I think we should first decide if we want this feature in ESLint.

@BYK
Member
BYK commented Aug 20, 2015

@BYK how would you reference the rules then? configname/pluginname/rulename?

Yep.

That said, I think we should first decide if we want this feature in ESLint.

Agreed. Might be worth piggy backing on npm 3 instead of introducing this complexity.

@nzakas
Member
nzakas commented Aug 20, 2015

A few things:

  1. npm 3 doesn't solve this problem, the relationship between a config and a plugin remains a peer relationship. The fact that npm 3 flattens dependencies doesn't fundamentally change that relationship, is just an implementation quirk that allows dependencies to be treated as peers in certain situations. That's not a solution, is a gamble.
  2. Configs should not contain executable code, that's very far outside of the responsibilities of configs.
  3. To keep this in context: we are talking about a one-time setup cost rather than ongoing pain.

While I can understand the desire to have one install that works, I don't see a path towards that without introducing a new type of shareable thing that could encapsulate this functionality.

@sindresorhus
Contributor

Then maybe introduce a universal sharing thing that can contain multiple plugins/configs/whatever. It could even in the future allow extending ESLint in some ways, with hooks, but I don't want to start that discussion. Just showing the possibilities with something like this.

JSCS supports it like this: https://github.com/wealthfront/javascript/blob/f1f976e9c75a8d141ec77a5493d9d965d951d4a6/jscs/index.js

I just want the user to be able to npm install one module and have the needed config and plugins without having to care about how anything works internally. That's the beauty of normal npm packages.

@IanVS
Member
IanVS commented Aug 21, 2015

I agree that the current method becomes unwieldy when you begin sharing configs which use other shared configs and/or plugins. For example, the installation instructions for my own personal config (which extends from Standard) is:

npm install --save-dev eslint-plugin-standard eslint-config-standard eslint-config-ianvs 

It would be much nicer UX to only need:

npm install --save-dev eslint-config-ianvs 

That said, I have no idea how that could be accomplished, and in the end it's a pain I can live with until/unless a better solution is found.

@nzakas
Member
nzakas commented Aug 24, 2015

We could extend plugins to allow the inclusion of configs, as plugins were always intended to be a dumping ground of stuff. Thoughts:

  1. How do we specify them in extends? eslint-plugin-foo.configs.whatever? Something else?
  2. We could, in theory, just expose the file system so you could do eslint.plugin-foo/configs/whatever.
  3. This is a bit lousy because we lose the nice eslint-config-* convention for configurations, so it ends up blurring what is a configuration and what is not.
  4. What if a config wants to depend on a plugin that it doesn't own? What does that look like?
  5. What is the expected behavior when a plugin is install directly and the same plugin is installed indirectly via another plugin?
@cjihrig cjihrig referenced this issue in hapijs/lab Aug 24, 2015
Merged

Use a shareable config for ESLint #429

@feross
Contributor
feross commented Aug 30, 2015

@nzakas

Configs should not contain executable code, that's very far outside of the responsibilities of configs.

Could you elaborate on this? It seems like this is a philosophical rather than practical objection. From a user's perspective, an eslint config is just an npm package that they need to install and extend in their .eslintrc. They don't care if there's executable code in there or not. Why complicate things for users?

@sindresorhus sindresorhus referenced this issue in sindresorhus/eslint-config-xo Aug 31, 2015
Closed

Using babel-eslint as parser for esnext #3

@nzakas
Member
nzakas commented Aug 31, 2015

@feross Allowing executable objects arbitrarily in configs would complicate things for users. What I'm saying is let's not complicate it by ensuring that configs remain static regardless of their form.

@joakimbeng
Contributor

Let the shareable config provide the plugin as an object

👍 would love to have this functionality!

We use require to load shareable configs or plugins, what would make it act more like a node module than that?

The problem is that the plugin name is not left as is, but instead parsed and prepended with eslint-plugin-. If ESLint didn't do this one could have solved the problem by adding plugins by their full paths, e.g. plugins: [path.join(__dirname, 'node_modules', 'eslint-plugin-babel')], not fancy but it would probably work.

@nzakas
Member
nzakas commented Sep 1, 2015

We don't have a good answer for this now. We'll revisit once we've finished some 2.0.0 tasks.

@ilyavolodin
Member

Related to #3659

@davidmason

It seems as though this is the case for configs as well, unless I am mistaken. For configs at least, is it possible to change how extends are loaded so that nested extends are processed in the module context that they come from?

This could at least solve the issue for configs, which do not have the issue of executable code.

e.g.

  • say I am making a shareable config for my team's projects, eslint-config-myteam
  • I want to base myteam config on another shareable config eslint-config-goodstyles with a few modifications.
  • in one of my team's projects ("myproject"), I npm install eslint eslint-config-myteam and create .eslintrc that extends: myteam
// eslint-config-myteam/index.js
module.exports = {
  extends: 'goodstyles'
}
# myproject/.eslintrc
extends: myteam

So when eslint is processing myproject/.eslintrc and finds extends: myteam it will locate node_modules/eslint-config-myteam.

At the moment I think it blindly reads that in, then fails when it hits the nested extends: goodstyles because that is not available at the top level. Could it instead keep track of which module it found the myteam config in, and if it finds an extends in there, search in that module for the config it extends. There are a few options for how to search:

  1. look only in the specific module that the extending config is from
  2. look first in the specific module that the extending config is from, then look at the higher level if it is not found there
  3. look first in the module where eslint was run, then in the specific module if the config is missing form there

The question is whether people should be able to override configs by name (on purpose or otherwise) in their extending config. Overriding configs by accident would be possible with 3, so I would rule that out. 1 would not allow peer-dependency configs, so I think 2 is the best option - if someone wants to make other configs peer dependencies they can just not include them in their package.json, but there is the option to include them and make life easier for consumers of their shared config.

@nzakas
Member
nzakas commented Oct 20, 2015

This issue isn't really about extending configs, it's about bundling configs with plugins.

@IanVS
Member
IanVS commented Oct 20, 2015

Or is it about bundling plugins with configs? From the first post in this issue by @sindresorhus:

My shareable config uses rules from an external plugin and I would like to make it a dependency so the user doesn't have to manually install the plugin manually.

It is currently painful to extend a config which uses rules from a plugin. Or am I missing a larger picture?

@nzakas
Member
nzakas commented Oct 20, 2015

Semantics. To me "configs with plugins" means "config packages contain dependencies on plugin packages".

@davidmason

@nzakas should I copy the "configs with configs" thing to a separate issue?

@nzakas
Member
nzakas commented Oct 21, 2015

@davidmason no need. That's a level of complexity that I don't want to build. If you want to extend an existing config in your own, just require it, make whatever modifications you want, then export it. It will work.

@davidmason

@nzakas makes sense, I'll do that. I can check the eslint code to make sure I merge my changes in the same way so there are no surprises.

@phated phated referenced this issue in gulpjs/eslint-config-gulp Oct 23, 2015
Closed

Extend node-style-guide base #6

@btmills btmills added a commit to btmills/eslint-config-btmills that referenced this issue Nov 16, 2015
@btmills btmills Move eslint-plugin-react to peerDependencies
It's optional unless you're using React, but npm@3 won't complain.

See eslint/eslint#3458
4e75783
@jkrems jkrems referenced this issue in airbnb/javascript Nov 16, 2015
Merged

Use require.resolve to allow nested extend #582

@michaelcontento

Any news on this topic? I do understand that, from some point of view, peerDependencies might be the "theoretical right" spot to declare it.

But if we take a step back and think about the user, a more flexible / babeljs@6-like solution would be wunderful! Custom bundles of eslint rules/plugins are currently cumbersome to both

  • install it in the first place (npm install -D eslint-config-acme-corp eslint-config-dep-a eslint-config-dep-b eslint-plugin-dep-c ... .. uff)
  • keeping track of updates
    • automated checks (like http://greenkeeper.io) will notify me about updates available for dependencies I (the regular user) don't care about
    • what if eslint-config-acme-corp drops one dependency? or replaces it with a whole new one? it becomes a instant pain for all users of the bundle

The issues above are the things I, as the creator of the bundle, like to avoid for all my users. And this is currently, AFAIK, not possible. So from a user point of view a good solution to create, share and update "eslint presets" would totally be a benefit. Or to put it differently: Those are the core requirements of a good "eslint presets" feature, they are currently missing and so is the whole feature.

A personal short side-note: The require.resolve-hack used by groupon and proposed to airbnb is nothing eslint should encourage you to! Also it only works for extends and parser - plugins does not allow this "trick".

@nzakas
Member
nzakas commented Dec 2, 2015

@michaelcontento no updates. This is quite a bit more difficult than Babel plugins because you can directly reference plugins in config files. If a shareable config depends on a plugin foo, and a user has manually installed that plugin as well, then when the end user references foo, what is the expected behavior? Which instance of the plugin should it refer to?

We will take another look after 2.0.0, by this is the main problem that needs resolving.

@michaelcontento

Thank you for the honest response!

Is there a target release date for 2.0.0?

@nzakas
Member
nzakas commented Dec 2, 2015

Targeting January.

@mysticatea
Member

FYI: Shareable configs can include some plugins in those dependencies field.

For example, my config is depending on eslint-plugin-mysticatea, eslint-plugin-node, and eslint-plugin-react. But user (me!) does not need to install the plugins manually (e.g. package.json).
Because npm does flatten dependencies, so user's project can require() the plugins.

For npm 2.x, we can use peerDependencies.

@nzakas
Member
nzakas commented Dec 4, 2015

@mysticatea that only works when the user hasnt also installed the plugins (maybe a different version). The risk is still there if you ever do that in the future.

Also keep in mind that npm 3 only flattens dependencies when there's no conflict amongst dependency versions for all dependents. If there is a conflict, the npm 2 approach is used.

@lukekarrys lukekarrys referenced this issue in andyet/eslint-config-andyet Dec 5, 2015
Open

Plugin dependencies #2

@boneskull

npm 3 doesn't solve this problem, the relationship between a config and a plugin remains a peer relationship. The fact that npm 3 flattens dependencies doesn't fundamentally change that relationship, is just an implementation quirk that allows dependencies to be treated as peers in certain situations. That's not a solution, is a gamble.

👍 to this. ask anyone on npm team; if you need a package, you need to specify it as a dependency, flattened node_modules or no, or you're just asking for trouble

@boneskull

if anyone wants to mess around with this, here's a quick-and-dirty monkeypatch to avoid having to npm install n packages to get one config and the plugins it uses.

@izaakschroeder izaakschroeder added a commit to izaakschroeder/eslint that referenced this issue Dec 18, 2015
@izaakschroeder izaakschroeder New: Enhanced local plugin resolution.
When writing presets sometimes it is desirable to include plugins as dependencies, since the preset may be explicitly for that dependency and that means one less thing users need to include and it offers an easy way to pin versions of a plugin with your preset. Unfortunately, the default `eslint` behavior is incapable of loading these local-style plugins for a couple of reasons:

 * `require` only resolves things relative to local eslint installation,
 * absolute paths are not supported,
 * developers don't want any code in `preset` definition files.

So with all those caveats in mind, the behavior has been altered to simply start the preset resolution process at the _caller_'s module (i.e. the preset) instead of the _callee_'s module (i.e. `eslint`).

This fixes eslint#3458.
11dc566
@izaakschroeder izaakschroeder added a commit to izaakschroeder/eslint that referenced this issue Dec 18, 2015
@izaakschroeder izaakschroeder New: Enhanced local plugin resolution. (fixes #3458)
When writing presets sometimes it is desirable to include plugins as dependencies, since the preset may be explicitly for that dependency and that means one less thing users need to include and it offers an easy way to pin versions of a plugin with your preset. Unfortunately, the default `eslint` behavior is incapable of loading these local-style plugins for a couple of reasons:

 * `require` only resolves things relative to local eslint installation,
 * absolute paths are not supported,
 * developers don't want any code in `preset` definition files.

So with all those caveats in mind, the behavior has been altered to simply start the preset resolution process at the _caller_'s module (i.e. the preset) instead of the _callee_'s module (i.e. `eslint`).
9832120
@10xjs 10xjs added a commit to metalabdesign/eslint-config-metalab that referenced this issue Jan 16, 2016
@10xjs 10xjs Include plugins in `peerDependencies` for `npm^2`
Plugins as dependencies in configs are not supported by `npm^2`. By adding the plugins to `peerDependencies`, `npm^2` will install the plugins in the root `node_modules` where `eslint` looks.

See: eslint/eslint#3458
609ec85
@10xjs 10xjs referenced this issue in metalabdesign/eslint-config-metalab Jan 16, 2016
Merged

Move plugins to `peerDependencies`. #22

@10xjs 10xjs added a commit to metalabdesign/eslint-config-metalab that referenced this issue Jan 16, 2016
@10xjs 10xjs Move plugins to `peerDependencies`
Plugins as `dependencies` in configs are not supported by `eslint`.

See: eslint/eslint#3458
c73c2d0
@10xjs 10xjs added a commit to metalabdesign/eslint-config-metalab that referenced this issue Jan 16, 2016
@10xjs 10xjs Move plugins to `peerDependencies`
Plugins as `dependencies` in configs are not supported by `eslint`.

See: eslint/eslint#3458
2110e80
@callumlocke

Sorry to ping everyone, but has there been any progress on this?

I want to make a shareable config that extends xo and then adds a few custom in-house rules. Is the situation still that this is not possible?

@IanVS
Member
IanVS commented Feb 23, 2016

You can, but you need to treat xo as a peerDependency. Take a look at semistandard for an example of another config doing something similar.

@nzakas
Member
nzakas commented Feb 25, 2016

@callumlocke our current priority is making autofix better. When there are updates for this issue, we will comment.

@ta2edchimp ta2edchimp added a commit to ta2edchimp/eslint-config-ta2edchimp that referenced this issue Apr 18, 2016
@ta2edchimp ta2edchimp fix: Make `eslint-config-idiomatic` a peer dep.
This is due to the underlying architecture, how ESLint will resolve the extending config.
See eslint/eslint#3458
fdf4ed2
@ta2edchimp ta2edchimp referenced this issue in ta2edchimp/eslint-config-ta2edchimp Apr 18, 2016
Merged

fix: Make `eslint-config-idiomatic` a peer dep. #4

@ta2edchimp ta2edchimp added a commit to ta2edchimp/eslint-config-ta2edchimp that referenced this issue Apr 18, 2016
@ta2edchimp ta2edchimp fix: Make `eslint-config-idiomatic` a peer dep. (#4)
This is due to the underlying architecture, how ESLint will resolve the extending config.
See eslint/eslint#3458
659a455
@hulbert hulbert referenced this issue in TakeScoop/eslint-config-scoop Apr 19, 2016
Closed

eslint-plugin-require-path-exists should maybe be peer dependency? #1

@johanbrook johanbrook added a commit to johanbrook/meteor-publication-collector that referenced this issue May 23, 2016
@johanbrook johanbrook Declare babel-eslint as dependency for linting
Because of eslint/eslint#3458.
8719bf5
@vjpr vjpr added a commit to live-js/eslint-config-live that referenced this issue Jun 20, 2016
@vjpr vjpr remove all airbnb stuff because of eslint/eslint#3458 9b9369d
@niksy niksy added a commit to niksy/eslint-config-niksy that referenced this issue Jun 23, 2016
@niksy niksy Check for "dependencies" and "peerDepenendencies"
Fix for following issues:

- eslint/eslint#3458
- eslint/eslint#2518
b84cebb
@markelog
Member

Since this was possible in JSCS and it is needed for good amount of plugin writers (including me) i think we can treat it as priority.

If no one minds i'd like to talk about this on the next TSC meeting

@markelog markelog referenced this issue in gruntjs/grunt Jul 19, 2016
Merged

Switch to ESLint from JSHint+JSCS #1534

@markelog markelog added a commit to markelog/eslint-config-grunt that referenced this issue Jul 19, 2016
@markelog markelog Treat pabigot plugin as peer dep
Because of the eslint/eslint#3458
685eb15
@markelog markelog added the tsc agenda label Jul 19, 2016
@nzakas nzakas removed the tsc agenda label Jul 21, 2016
@nzakas
Member
nzakas commented Jul 21, 2016

Per TSC meeting (21-Jul-16), it would be good to get an overview of the JSCS plugin use case plus a description of how JSCS currently solves this problem. Then, we need a proposal for how to overcome the problems discussed earlier in this thread (primarily, what happens when two different versions of a plugin are required by different configs used in the same project?).

@markelog
Member
markelog commented Jul 21, 2016 edited

Oh, there is bunch of issues then that, will try to document a proposal and tackle them when I finish with couple other issues.

Will put a JSCS label here, not sure if it needed or not though, so feel free to remove it if you feel it is not appropriate here

@markelog markelog added this to the JSCS Compatibility milestone Jul 21, 2016
@gaearon
gaearon commented Jul 26, 2016 edited

Just to add another use case example for this. We recently released Create React App—a tool that creates React apps with no build configuration. Internally we use ESLint but we are intentionally keeping it non-configurable. (There is an escape hatch for people who want configuration: they can “eject” and then all config files get copied into their projects.)

This project also follows the “single dev dependency” approach because many people (especially those getting started) are tired of installing dependencies. I know it sounds funny but it’s true. That create-react-app gained 4k stars in 4 days speaks to the desire for zero configuration and having few dependencies. Of course this doesn’t work for everyone. But our initial users seem fairly happy with this setup.

The problem for us is that while command line linting works fine (since we’re able to specify a custom hidden config), editor integrations are not as smooth.

Having to do this in the generated project is already a tiny bit frustrating:

{
  "name": "my-app",
  "version": "0.0.1",
  "private": true,
  "devDependencies": {
    "react-scripts": "0.1.0"
  },
  "dependencies": {
    "react": "^15.2.1",
    "react-dom": "^15.2.1"
  },
  "scripts": {
    "start": "react-scripts start",
    "build": "react-scripts build",
    "eject": "react-scripts eject"
  },
  "eslintConfig": {
    "extends": "./node_modules/react-scripts/config/eslint.js" // meh!
  }
}

The only reason eslintConfig is needed here is because editors don’t “know” about our custom config, and we have to make it discoverable.

But of course even that doesn’t solve the problem completely because of this issue. If the user is on npm 2, they’re out of luck completely because plugins are not getting discovered. Even if the user is on npm 3, it working feels like a fragile coincidence.

I think the bigger theme here is that as a community, we solved many problems of customizing configuration (and ESLint has been terrific at it), but maybe it’s time for the pendulum to swing back, and for us to start solving the problem of sane zeroconf defaults via opinionated packages.

@gaearon gaearon referenced this issue in facebookincubator/create-react-app Jul 26, 2016
Closed

Proposal: 'configurator' field in package.json #215

@gaearon
gaearon commented Jul 26, 2016

On a related note, I posted a proposal for a configurator field in package.json. It is intended for zero-conf tools like standard and create-react-app, and I think a convention like that could solve this issue. Please let me know what you think: facebookincubator/create-react-app#215

@gaearon gaearon referenced this issue in facebookincubator/create-react-app Jul 26, 2016
Closed

resolve paths in eslint config #173

@markelog markelog self-assigned this Jul 26, 2016
@stefanpenner
stefanpenner commented Aug 15, 2016 edited

cross-posted as per: #6865 (comment)

eslintrc at odds with npm link (and other node resolution friendly usages)

Some examples: (this may address)

Given 2 modules:

  • Module B: the extraction of generic build tool code (the invoker of eslint)
  • Module A: the application which which configures eslint + depends on the generic build tool code module

Working scenario:

  1. A depends on B
  2. B is a descendent of A on the file system
  3. when B "resolves" A's configuration it will walk up the file system and eventually find it
~/projects/A/
~/projects/A/node_modules/B

Failing scenario 1:

  1. A depends on B
  2. B is in an ancestor directories node_module of A on the file system
  3. when B "resolves" A's configuration it will walk up the file system and not find A, rather it may find something else
~/projects/A/
~/projects/node_modules/B

Failing scenario 2 (npm link):

  1. A depends on B
  2. B npm link'd to A on the file system
  3. when B "resolves" A's configuration it will walk up the file system and not find A, rather it may find something else
~/projects/A/
~/projects/node_modules/B -> ~/src/B

In node references to other code, and modules is always related to the including modules position on disk. Not relative to the invoker, or cwd. But fs/system calls are relative to cwd().

// foo/bar/baz.js
require('some-module'); // is a resolve relative to the current file
@Cellule Cellule added a commit to Cellule/eslint-myrules that referenced this issue Aug 21, 2016
@Cellule Cellule Workaround eslint/eslint#3458
Added a monkeypatch to force eslint to resolve my plugins in the scope of this config.
Only accept patches of plugin and eslint as I want to manually inspect each minor/major versions' changelog
f08bce0
@Cellule Cellule added a commit to Cellule/eslint-myrules that referenced this issue Aug 21, 2016
@Cellule Cellule Workaround eslint/eslint#3458
Added a monkeypatch to force eslint to resolve my plugins in the scope of this config.
Only accept patches of plugin and eslint as I want to manually inspect each minor/major versions' changelog
83b3ac2
@Cellule
Contributor
Cellule commented Aug 22, 2016

For those interested, I found a workaround to have your plugins as dependencies in your shareable config.
See Cellule/eslint-myrules@83b3ac2

Essentially this monkey patches eslint plugin resolution by resolving the plugin yourself therefore using your module's resolve paths.
This works with npm link as well.
It heavily depends on the current eslint file structure (works at least between versions 2.0.0 and 3.3.0) thus can be broken at any time.

Until this is officially supported I hope this might help someone else who like me wanted to alleviate the burden of settings up and maintaining eslint plugin versions.

@pirelenito pirelenito referenced this issue in saguijs/sagui Aug 22, 2016
Open

Using airbnb style guide #191

@jdalton
jdalton commented Aug 24, 2016 edited

Just pinging in that I too was bit by this.

@iamstarkov

I think community needs shareable configs to behave as usual npm packages—encapsulate deps with itself

@nzakas nzakas removed this from the JSCS Compatibility milestone Aug 30, 2016
@nzakas
Member
nzakas commented Aug 30, 2016

Thanks everyone for the feedback. We're aware of everyone's desire for this feature and plan on addressing it once we've finished up JSCS compatibility work. As such, I'm going to lock this issue because we understand the request and the use cases, and at this point we're just getting a bunch of "me too" comments.

@nzakas nzakas locked and limited conversation to collaborators Aug 30, 2016
@nzakas
Member
nzakas commented Oct 6, 2016

Summary

I know this is a popular issue, so I'm going to try to summarize where we are and why this is hard. The team is aware that this is a popular request, but unfortunately, popularity does not make the work any simpler or create the extra time needed to investigate how to solve this issue.

History

ESLint supported plugins first, and sometime later, shareable configs were created. At that time, we encouraged people to use peer dependencies through npm in order to define the relationships between shareable configs and plugins. This worked well because npm would install peer dependencies for you and it solved the deduping problem if, for example, a shareable config relied on a different version of a peer dependency than the one you already had installed.

Unfortunately, npm decided they would stop installing peer dependencies by default, instead relying on developers to manually install those dependencies. That was really the beginning of the request for shareable configs to be able to include plugins because developers didn't want to tell users of their shareable configs to manually install a bunch of other stuff. That's fair, but unfortunately, the relationship between shareable configs and plugins was designed, from the start, to be one of peer dependencies. With the npm changes, that relationship was no longer optimal.

Differences with npm Packages

There are a lot of comments that say something along the lines of, "well, I think shareable configs should just act like any other npm package in the way it defines its dependencies." Yes, it would be great if that were possible, however, shareable configs don't work the same way that other npm modules do (that's a feature, not a bug).

One of our early assumptions was that there could only be one instance of a plugin loaded per run. This is important because of the way we refer to plugin assets like rules. For instance, suppose you had eslint-config-myplugin, your config looks like this:

plugins:
  - myplugin
rules:
  myplugin/rule: "error"

First, we load the plugin based on myplugin in the plugins array, and then we reference a rule within the plugin as myplugin/rule. In order to use that convention, we must know with 100% certainty what myplugin refers to, and it must refer to the same thing across all configs in the project in order to ensure each config may override settings that other configs set. That means each config cannot search up the directory structure for plugins like npm package lookup does. We keep an internal cache of the plugins and use that for lookup.

The Hard Case

All that said, the reason this issue hasn't moved forward is because of one very serious edge case: what if two different versions of the same plugin are required by two different configs in a project? For instance, suppose:

  • You have eslint-config-foo that depends on eslint-plugin-foo@2.0.0 because it wants "foo" rules.
  • The user has a dependency on eslint-plugin-foo@1.2.0 directly.

In this case, we have several problems:

  1. How do you determine which eslint-plugin-foo version to load, given that the config file will just say foo?
  2. If we choose to load, say, the highest-up package by default (eslint-plugin-foo@1.2.0) and eslint-config-foo references a rule that doesn't exist until version 2.0.0, ESLint will throw an error because the rule is missing.
  3. If we instead choose to load the package closest to the config that first specified it, we have the same problem in the other direction: a rule the user's config references might have been removed or the options format might have changed, causing ESLint to error.
  4. If we choose to load both packages, then we'll risk naming collisions for rules and environments.

At this point, you'll likely arrive at a number of different solutions:

Throw an error if the same package is attempted to be loaded twiceT

This is deceptively difficult to do. First, we don't have a good way of knowing that we're attempting to load a different version of the plugin rather than just the same one we saw before. Second, it might be beyond the end user's capabilities to fix this problem because they may be using several shareable configs, each of which depends on a different version of the plugin. Penalizing the end user for including configs with contradictory plugin package versions seems like the wrong choice.

Figure out a way to load two different versions of the same package

First, that would mean figuring out a way to load two different versions of the same package, which Node.js isn't designed to do by default, so we'd have to go crawling around the filesystem to do this. Crawling around the filesystem is made much more difficult due to npm node_modules flattening, so we couldn't really know the correct places to look, which would mean a lot of searching around for packages. We'd also need to come up with some way to ensure the end users knows which version of plugin they are configuring at any given time, which likely means we'd need a different naming scheme for configuring plugins that were included via shareable config vs. installed by the user. That's also difficult because shareable configs are treated like any other config inside of ESLint, so that would also be a lot of work.

Create a new shareable thing that works the way we want it to

Of course, when what you created isn't working the way you wanted, it's logical to stop and consider if you want to create a new type of thing that can work the way you want. The big problem here is that we have a big ecosystem of already-existing shareable configs and plugins, and would we really want to impose the burden of creating something new to fulfill that need?

Where We Are

I hope that helps explain why this issue hasn't moved much. Despite the desire of many people to figure out how to make this work, at the end of the day, there are enough complexities that it will take a significant amount of time from someone (on the team or off) to really dig in and try to solve this. As the ESLint team is a group of unpaid volunteers, no one has had enough time or interest to dig into this problem.

Next Steps

If you are interested in pitching in to solve this problem, we are open to hearing design proposals. The only constraint is that any design must address the hard problem of having two different parts of the project include different versions of the same plugin. All of the proposals to this point, both in this thread and through various pull requests, have not addressed it and so cannot be considered.

I'm unlocking this thread to allow for designs to be posted but if I see the thread fill up with +1s and other useless comments, I'll lock it again.

What we need is someone or a group of someones who are willing to dig and do the research and prototyping necessary to solve this problem. If that's you, please let us know.

@nzakas nzakas unlocked this conversation Oct 6, 2016
@feross
Contributor
feross commented Oct 7, 2016 edited

@nzakas

Figure out a way to load two different versions of the same package

First, that would mean figuring out a way to load two different versions of the same package, which Node.js isn't designed to do by default, so we'd have to go crawling around the filesystem to do this.

This issue could potentially be addressed by allowing configs to include code, like this:

module.exports = {
  plugins: {
    foo: require('eslint-plugin-foo')
  },
  rules: {
    'foo/rule': 'error'
  }
}

This way, each config can use the exact plugin version that they list in their package.json dependencies. No need to crawl around the filesystem.

We'd also need to come up with some way to ensure the end users knows which version of plugin they are configuring at any given time.

We could solve this by saying: you can only configure plugin rules if you require that plugin in your config. That way it's clear what version of the plugin you're configuring.

If you re-use a plugin rule that was already defined in an earlier config, then the later rule takes precedence (like how it works today). Note: Both the rule's setting and the rule code from the later config/plugin will take precedence.

Optional convenience feature: We could consider relaxing the rule about depending on the plugin, if the config is just disabling rules from that plugin.

One downside of this approach is that it only works with JS configs, not JSON or YAML.

@gaearon gaearon referenced this issue in facebookincubator/create-react-app Oct 7, 2016
Open

Work around ESLint plugin discovery issue #866

@stoikerty
stoikerty commented Oct 7, 2016 edited

@feross Essentially you're specifying a path, right? I don't see why that would have to be restricted to javascript. Babel does something similar where you can specify presets & plugin paths.

// how babel reads it
{
  "presets": [
    "/my_project/node_modules/babel-preset-es2015/lib/index.js",
    "/my_project/node_modules/babel-preset-react/lib/index.js"
  ],
  "plugins": [
    "/my_project/node_modules/babel-plugin-transform-flow-strip-types/lib/index.js"
  ]
}

// can get compiled into a babel-readable config from any other language, example js:
  presets: [
    require.resolve('babel-preset-es2015'),
    require.resolve('babel-preset-react')
  ],
  plugins: [
    require.resolve('babel-plugin-transform-flow-strip-types')
  ]
@nzakas
Member
nzakas commented Oct 7, 2016

@feross unfortunately, I don't think we can accept a solution that only works in a JS file with passing around objects. There are plenty of folks who don't use JS-based configurations, and we need to be able to support them, as well. We'd like to keep configs as data-only.

@stoikerty that's an interesting idea - we could do an intermediate step that allows you to pass in package by filepath. We'd have to do something to allow you to name that package (since we wouldn't have the package name. Something like what @feross mentioned:

plugins:
  foo: "/some/location/to/index.js"
rules:
  foo/rule: "error"

There is still a concern in that case that someone might include eslint-config-foo in a different config, and that would end up colliding with the foo defined here. Maybe we'd need some special naming convention to differentiate between foo meaning eslint-config-foo vs. foo meaning a custom name chosen for a filepath.

@gaearon
gaearon commented Oct 7, 2016

There is still a concern in that case that someone might include eslint-config-foo in a different config

This might be ugly but could we consider an absolute filepath a valid plugin “identifier”?

const foo = require.resolve('eslint-plugin-foo');

module.exports = {
  plugins: [foo],
  rules: {
    [`${foo}/rule`]: 'error'
  }
}

We would not support relative filepaths which excludes any non-JS configs from supporting this feature, but presumably it's fine to make a subset of configs (JS-based ones) more powerful?

@stoikerty
stoikerty commented Oct 10, 2016 edited

There is still a concern in that case that someone might include eslint-config-foo in a different config, and that would end up colliding with the foo defined here.

When eslint is parsing a config-file, is it able to distinguish between configs and their associated plugins?

Mapping it out, if you have the following:
⦿ pluginConfigA ➤ pluginConfigB ➤ userConfig ◉

where:

  • pluginConfigA uses foo@0.2.0
  • pluginConfigB uses foo@0.3.0 and includes pluginConfigA
  • userConfig uses foo@0.2.4 and includes pluginConfigB

What is the most reasonable way to handle this scenario? Is it possible for each config to reference and use its own foo?

@ilyavolodin
Member

One other requirement that I wanted to mention is that the new way should be backwards compatible. We can't just break hundreds of plugins that are already published.

@einarlove einarlove referenced this issue in unfold/eslint-config Oct 14, 2016
Open

eslint-preset #9

@not-an-aardvark
Member

I think the main issue here is that the current processing of config files doesn't allow contextual identifiers. In other words, eslint-plugin-import has to mean the exact same thing regardless of the config file that it appears in. This is very limiting, because if we want sharable configs to be usable independently from each other, they have to be able to have their own dependencies, and potentially their own versions of eslint-plugin-import.

To fix this, rules and plugins referenced in a config file should be resolved relative to the config file itself. In other words, config files will become more like modules rather than using a global plugin namespace.

Proposal:

  • When a config file references a plugin, the plugin is resolved with require() relative to that config file.
  • A config file can access another config file's references by using a / as a delimiter. This allows a user to override settings specified in a particular config file.

For example, here's eslint-config-foo, which depends on eslint-plugin-bar@1.0.0. eslint-config-foo can specify eslint-plugin-bar as a dependency.

// ./node_modules/eslint-config-foo/index.js

module.exports = {
  plugins: ['bar'], // resolved as if require() was used from this file, yielding eslint-plugin-bar@1.0.0
  rules: {
    'bar/no-horses': 'error',

    // As before, eslint-plugin-foo can have local references to rules as well.
    'tball': './lib/rules/sports-jokes/tball.js'
  }
};

And here's eslint-plugin-bar@1.0.0:

// ./node_modules/eslint-config-foo/node_modules/eslint-plugin-bar/index.js

module.exports = {
  rules: {
    'no-horses': require('./lib/rules/bar-jokes/no-horses.js');
  }
};

Meanwhile, the user would like to use eslint-config-foo, but they would also like to use eslint-plugin-bar@2.0.0 separately. They can do this by installing eslint-config-foo and eslint-plugin-bar as dependencies (or devDependencies), and arranging their config file like this:

// ./.eslintrc.js

module.exports = {
  // `foo` is resolved by using require('eslint-config-foo') from this file
  extends: ['foo'],

  // `bar` is resolved by using require('eslint-plugin-bar') from this file.
  // This will yield eslint-plugin-bar@2.0.0, since the user has that installed as a dependency.
  plugins: ['bar'],

  rules: {
    // Again, references to `bar` are resolved by using `require` from this file.
    // So this will turn on the `tender` rule from eslint-plugin-bar@2.0.0:
    'bar/tender': 'error',

    // As before, `foo` is resolved by using require('eslint-config-foo') from this file.
    // This yields eslint-config-foo, which allows the user to configure rules from that config.
    'foo/tball': 'off',

    // Finally, a slash character can be used to chain references.
    // In this case, `foo` is resolved by using require('eslint-config-foo') from this file.
    // Then `foo/bar` is resolved by using require('eslint-plugin-bar') from eslint-config-foo.
    // This allows the user to toggle the rules from eslint-plugin-bar@1.0.0 that are used by foo:
    'foo/bar/no-horses': 'warn'
  }
};
@BYK
Member
BYK commented Oct 15, 2016

@feross unfortunately, I don't think we can accept a solution that only works in a JS file with passing around objects. There are plenty of folks who don't use JS-based configurations, and we need to be able to support them, as well. We'd like to keep configs as data-only.

Can we offer this as a JS-only feature to at least get started? I understand the concern around restricting configs to only data but if we had the actual JS and module system, it may open other doors. Existing JSON and YAML configurations are already compatible with this, just more limited and in the future if we decide to drop support for these, it should be trivial to write a tool (or may be even include it in eslint) that converts these to JS modules.

@not-an-aardvark
Member

@BYK I think it's worth figuring out a good solution first, rather than implementing if halfway and figuring out the rest later. If we implement it halfway, it will make it much harder to change the system later.

@gaearon
gaearon commented Oct 15, 2016

@not-an-aardvark

It is not clear to me from the team responses so far whether making JS configs more powerful is considered implementing it “halfway”, and why.

This is a hard problem if you treat configs purely as data, precisely because data presumes constant plugin identities whereas the issue is about potentially conflicting plugins.

Other tools, such as Babel, have already successfully solved the exact same problem by making presets code rather than data. I don’t quite understand why ESLint can’t follow the same path, even while allowing backward compatibility with “less powerful” formats. I also don’t quite see why having multiple formats to specify configs is useful, but this is probably due to my ignorance about the topic.

@not-an-aardvark
Member

@gaearon My objection was mostly in response to offering as a JS-only feature "to at least get started". If we determine that JS-only is the solution we want to use, then that's fine, but I don't think we should rush to a temporary solution with plans to change it later, because that will cause backwards-compatibility issues.

I think the solution in #3458 (comment) would be better anyway -- it would allow locally-resolved dependencies without needing to explicitly call require.

@abhisekp abhisekp referenced this issue in abhisekp/cyclic-next Oct 16, 2016
Open

revert(travis): Restore node v0.12 and v4 support #4

@nzakas
Member
nzakas commented Oct 17, 2016

Can we offer this as a JS-only feature to at least get started?

I thought I'd addressed this pretty clearly earlier in the thread, but I'm going to say it once again: no. Continuing to discuss strategies that only work in JS configs is not helping us solve the problem. I'm not going to go into a long explanation of why we have different config file formats, but the fact is that we do and we have a large ecosystem using all different types of formats. Our promise has always been that the formats are interchangeable, and I don't see that changing anytime soon. So let's focus on the problem at hand.

@not-an-aardvark That's an interesting approach, and I really appreciate you taking the time to explain it. Some points that popped into my head while reading it:

  1. It's unclear to me if this approach is breaking. Right now, if someone is using a shareable config that relies on a plugin, they are installing both the plugin and config, presumably at the same level. However, the shareable config has plugins: ['bar'] and the user config does not even though it's configuring rules like bar/rule. I think your proposal would break those configs (Airbnb is the most popular shareable config, and they are setup like this).
  2. There's a level of increasing complexity here that I'm concerned about: what if a shareable config is dependent on another shareable config that, in turn, is dependent on a plugin? How does that affect how you configure rules? (eslint-config-a depends on eslint-config-b which depends on eslint-plugin-p, do rules now become a/b/p/rule?)
  3. Keep in mind that plugins can also export configs that can be inherited from (extended). That means you can't always assume that one member of a slash-delimited identifier is a config vs. a plugin.

I think my overall feedback is that there may be a breakdown with the naming convention once we get into more involved relationships between shareable configs and plugins. Do you have thoughts around that?

@not-an-aardvark
Member

It's unclear to me if this approach is breaking. Right now, if someone is using a shareable config that relies on a plugin, they are installing both the plugin and config, presumably at the same level. However, the shareable config has plugins: ['bar'] and the user config does not even though it's configuring rules like bar/rule. I think your proposal would break those configs (Airbnb is the most popular shareable config, and they are setup like this).

There are two issues to consider here:

  • Currently, plugins and configs are installed at the same level. This part would not result in a breaking change, because resolving require('eslint-plugin-bar') from foo would still be able to find bar, due to Node's node_module resolution algorithm.
  • Users currently can reference rules that were enabled by foo by simply referencing bar/rule, rather than foo/bar/rule. This would be a breaking change according to the current proposal. As a workaround for backwards compatibility, maybe we could handle missing references to bar based on whether another config had a valid reference to bar, and use that reference as a fallback.
// ./node_modules/eslint-config-foo/index.js

module.exports = {
  plugins: ['bar'],
  rules: {
    'bar/rule1': 'error'
  }
};
// ./.eslintrc.js

module.exports = {
  extends: ['foo'],
  rules: {
    // eslint-config-foo had a valid reference to `bar`, and no valid reference can be resolved here.
    // As a fallback, this reference will work as an alias for 'foo/bar/rule1', resolving as the version of
    // `bar` that eslint-config-foo used.
    'bar/rule1': 'warn'
  }
};

While this would preserve backwards-compatibility, it isn't an ideal solution, since it could cause confusing behavior if the user installs something else that depends on a different version of eslint-plugin-bar. If we did decide to follow this pattern, we would probably want to deprecate it in favor of explicitly listing foo/bar/rule1 instead. (Alternatively, I'm open to hearing other workarounds for this.)

There's a level of increasing complexity here that I'm concerned about: what if a shareable config is dependent on another shareable config that, in turn, is dependent on a plugin? How does that affect how you configure rules? (eslint-config-a depends on eslint-config-b which depends on eslint-plugin-p, do rules now become a/b/p/rule?)

Yes, this would be resolved as a/b/p/rule. This does lead to increasing complexity, but I'm not sure this is avoidable; if there are there are multiple different versions of eslint-plugin-p scattered throughout node_modules, it is necessary to explicitly specify which version of eslint-plugin-p the user is referring to.

We could allow shorthands for these references if they're unambiguous (see the backwards-compatibility workaround above), which would allow things like p/rule rather than a/b/p/rule. However, I don't think we should encourage this pattern, since it's liable to unexpectedly break if a different version of eslint-plugin-p gets installed by some other dependency.

Keep in mind that plugins can also export configs that can be inherited from (extended). That means you can't always assume that one member of a slash-delimited identifier is a config vs. a plugin.

True, but I don't think this is a problem; we can simply resolve extended config files the same way as everything else. For example: 'airbnb/airbnb-base/import/no-unresolved': 'off'

I think my overall feedback is that there may be a breakdown with the naming convention once we get into more involved relationships between shareable configs and plugins. Do you have thoughts around that?

I would say that the naming convention works in these relationships, but it doesn't produce an ideal experience for users that want to configure rules from inherited plugins. Admittedly, having to use a path like a/b/p/rule in your config file is annoying.

However, if we want to allow multiple different versions of configs at the same time (which is necessary for a good resolution to this issue), while still allowing users to individually configure inherited rules, I don't think this annoyance is avoidable in the general case. Allowing multiple versions of a plugin to be run simultaneously will require the user to be more specific about which version of the plugin they want to configure.

@nzakas
Member
nzakas commented Oct 21, 2016

Yes, to clarify, I have no problem with the part of the proposal to load plugins relative to shareable configs. In fact, I've never been opposed to that, my only problem has been with how we expect users to address multiple versions of the same plugin in their configs -- to me, that's the hard part of this problem. (Although loading plugins relative to their configs is a nontrival effort to implement, it's pretty straight forward.)

I would say that the naming convention works in these relationships, but it doesn't produce an ideal experience for users that want to configure rules from inherited plugins. Admittedly, having to use a path like a/b/p/rule in your config file is annoying.

Yeah, I think this is a legitimately horrible experience. The problems (off the top of my head):

  1. A lot of cognitive overhead for end users.
  2. It brings the dependency chain of a shareable config into the end-user config. Exposing dependency information like that invites future compatibility problems if a dependency changes or is renamed.
  3. Still doesn't address the fact that currently, rule namespaces are only plugin names (namespace/rule). We'd need a different syntax for namespaces that are config names to avoid naming collisions.

If we are making the argument that shareable configs should work like any other modules, then I think the only real direction is to force shareable configs to re-export whatever it is that they want to expose to end users. That way, the end user only needs to know about the shareable config without any of its dependencies (which is ultimately what we want). So maybe a shareable config has something like:

exportedPlugins:
   foo: react       # Load eslint-plugin-react into identifier foo
   bar: import    # Load eslint-plugin-import into identifier bar

Then in end-user config, you'd do:

extends: myconfig
rules:
   $myconfig/foo/rule1: "error"
   $myconfig/bar/ruleA: "warn"

Where the $ indicates that this is coming from the config rather than a plugin. (Could be any special character, doesn't have to be $.)

The problem with this approach is that it only really works one level down. If a shareable config A depends on another shareable config B, I'm not sure how you'd say to promote the rules from B so they can be accessed from A.

As I've said a few times, the devil is in the details on this effort. :)

@ljharb
Contributor
ljharb commented Oct 22, 2016

Perhaps you could also do:

exportedPlugins:
   foo: react       # Load eslint-config-airbnb into identifier foo
   bar: import    # Load eslint-plugin-import into identifier bar
   baz: import/dep # Load eslint-plugin-import's exported plugin "dep" into identifier baz

In other words, it'd be one level down, but at each level, it could be explicitly re-passed?

@not-an-aardvark
Member

If we are making the argument that shareable configs should work like any other modules, then I think the only real direction is to force shareable configs to re-export whatever it is that they want to expose to end users

I'm not sure I understand this point. I thought a design goal here was that users should always be able to disable a given rule in their config. If we're okay with a rule being impossible to disable from a config since it's not explicitly exported by a plugin, then this changes some of the strategies that we might use to address the problem.

@caesarsol

Importing slightly different rules from different versions of a single plugin seems theoretically logic but practically confusing, IMHO.

Wouldn't it be more straightforward to just fail-fast in a similar scenario, or maybe to make an arbitrary choice (plugin-wide or rule-wide) in case of version conflict?

@ilyavolodin
Member

After reading this thread, I think I actually agree with @caesarsol It seems like we are trying to preemptively fix an issue that isn't a real problem (at least, that's not something we've seen so far). While it's entirely possible to have very complex configuration with multiple configs inherited and deep nesting, it seems like it would be very rare, it also seems that the person that might go through the trouble of setting something like that up, knows what they are doing and should be able to fix it himself. Maybe we don't need to solve the issue of multiple versions of the same plugin?

@ljharb
Contributor
ljharb commented Oct 22, 2016

fwiw, Airbnb's use case for eslint-config-airbnb will be satisfied as long as:

  • when there's no dep conflict (ie, multiple versions of the same plugin), users can install a single package with no peer deps
  • when there is a conflict, it either a) works properly or b) fails immediately and loudly on the first run of eslint
@vvo vvo added a commit to algolia/eslint-config-algolia that referenced this issue Oct 24, 2016
@vvo vvo docs(deps): force installation of eslint*
Since eslint/eslint#3458 is not yet solved.
b18724e
@nzakas
Member
nzakas commented Oct 30, 2016

@ilyavolodin I disagree. How many times have we seen edge cases pop up from multiple people over a period of time that eventually forces us to make a change? I'm pretty exhausted by this discussion, so if we're going to solve this problem, then I want to solve it in such a way that we don't need to think about it ever again. :)

@not-an-aardvark

I'm not sure I understand this point. I thought a design goal here was that users should always be able to disable a given rule in their config.

They'd still be able to disable the rule, but the reference would be through the config rather than directly to the plugin.

@nzakas
Member
nzakas commented Oct 30, 2016

Okay, so I was lying awake in bed last night when I realized something: if we slightly change this discussion from "how can a shareable config bundle plugins as a dependency" to "how can a config bundle plugins as a dependency," then the problem is actually already solved because plugins can contain configs and can also specify other plugins as dependencies.

Example

Suppose Airbnb has a config they want people to use and that config depends on eslint-plugin-react. If the config is exposed in a plugin instead of a shareable config, you can make a plugin that looks like this:

// eslint-plugin-airbnb
const reactPlugin = require("eslint-plugin-react");

// export the es6 config
module.exports.configs = {
    es6: require("./configs/es6.json")
};

// export the rules you want to expose
module.exports.rules = reactPlugin.rules;

The eslint-plugin-airbnb package can depend directly on eslint-plugin-react, so the dependencies are installed automatically with npm i eslint-plugin-airbnb. Then, in your local project config, you can do this:

plugins:
  - airbnb
extends:
  - plugin:airbnb/es6
rules:
  airbnb/jsx-uses-vars: "error"

In this case, the project config is extending the config from eslint-plugin-airbnb and the plugin is re-exporting the jsx-uses-vars rule from eslint-plugin-react, so you have to address it using airbnb/ instead of react/. IMHO, this is kind of nice because it makes the plugin dependency completely opaque. However, if you wanted a bit more transparency, you could change the plugin to prepend react/ to the beginning of each rule:

// eslint-plugin-airbnb
const reactPlugin = require("eslint-plugin-react");

// export the es6 config
module.exports.configs = {
    es6: require("./configs/es6.json")
};

// export the rules you want to expose
module.exports.rules = {};
Object.keys(reactPlugin.rules).forEach(ruleName => {
    module.exports.rules["react/" + ruleName] = reactPlugin.rules[ruleName];
});

Then you could reference the rules as airbnb/react/jsx-uses-vars.

Evaluation

It seems like this approach solves all of the hard problems we've been trying to address with a decent amount of flexibility (in that plugin authors can decide which rules to export and how they should be named):

  1. Allows sharing of configs that depend on other packages without a separate install step.
  2. Allows multiple versions of the same plugin to be included by different plugins/configs without clashing.
  3. Gives plugin authors the ability to control how rules they depend on are represented in their plugins, so they can choose the best approach.
  4. Whether or not a config relies on a plugin is mostly opaque as far as the user is concerned.

Next Steps

As I've said before, shareable configs were really not designed to be used in the way we are trying to get them to be used in this thread, and I think the reason we keep banging into walls is because we're trying to force shareable configs into a paradigm they were never designed for.

Plugins, on the other hand, are made to be more composeable and looking back, it actually looked like we wanted to encourage people to use plugin configs instead of shareable configs when we first implemented it.

It seems like the best next step is for someone to actually try this out and give feedback on the experience. Any volunteers?

@fson
fson commented Oct 30, 2016

@nzakas This sounds like a good solution and I'd be happy to give it a try in Create React App.

If it works, this would solve some persistent problems we've had:

  • Create React App only adds a single build dependency (react-scrips) to the project. But currently using editor integrations for ESLint requires our users to also install all the plugin dependencies we use either globally or in their project.
  • We also expose our ESLint config as a separate eslint-config-react-app package. Users of this shareable config need to install the correct versions of its peer dependencies and we must keep these versions up-to-date in our installation instructions.

I can try this out and give feedback on how it goes.

@ljharb
Contributor
ljharb commented Oct 30, 2016

That's a pretty huge breaking change for all the current users of eslint-config-airbnb - anything that actually solves this problem for me would need to work with users leaving "extends": "airbnb" as-is.

@TheSavior
Contributor

I agree with @ljharb here. Since eslint is very particular about the naming of packages and how plugins / configs are used, it's a pretty big problem that we have to migrate all of our users to a new package and deprecate the old one to solve this.

@fson
fson commented Oct 30, 2016 edited

Would it make sense to allow plugins to define a "default configuration" that would be used when the user has only a name of a plugin in extends (the way shareable configs can be used at the moment)? It seems that this way "extends": "airbnb" could continue to work with a new eslint-plugin-airbnb that would specify a default configuration.

@ljharb
Contributor
ljharb commented Oct 30, 2016

@fson i like that idea but there already exist conflicting eslint-config-foo and "eslint-plugin-foo with a default config" packages all over the place.

@fson fson referenced this issue in facebookincubator/create-react-app Oct 31, 2016
Closed

Move ESLint configuration to an ESLint plugin #993

@ilyavolodin
Member

@ljharb I think it's still worth trying out first. If it proves to be working well with plugins, we could look into ways of either having ESLint handle configs as plugins, or allow ESLint to extend configs and plugins through the same mechanism. So far this is the first suggestion that we've had that seems like it might work, so I wouldn't want to dismiss it without trying it out first.

@ljharb
Contributor
ljharb commented Oct 31, 2016

@ilyavolodin there's certainly nothing wrong with improving plugins. I just want to make it clear that if it requires people to change their eslintrc or not be able to depend on the same eslint-config-airbnb, then it won't actually resolve the issue.

@nzakas
Member
nzakas commented Nov 1, 2016

@fson we actually used to define a default plugin config, but we decided it was too big of a surprise to users to have a plugin automatically enforce a config just because you included it. I don't think we want to go back there.

@ljharb sometimes improvements require breaking changes, unfortunately. It seems like in this case, the change from "airbnb" to "plugin:airbnb/foo" is easily solved with simple find-and-replace, as would be the change from "react/foo" to "airbnb/react/foo" or something else. So you could probably write a small utility to help people convert their existing configs if it proves to be a major pain point.

We can also help the cause by formally deprecating the eslint-config-* format and encouraging people to use plugins from now on (which it seems was my original thinking way back when).

@markelog markelog removed their assignment Nov 1, 2016
@ljharb
Contributor
ljharb commented Nov 1, 2016

@nzakas If you accept that there's no way to make configs solve this problem, then of course it's required - but I don't think that's been proven in any way. If eslint wants to formally deprecate configs, then obviously my hand is forced (altho that's much better than informally deprecating them). (side note: i hope "extends": "plugin:airbnb" would be possible, so people wouldn't have to couple to a plugin's sub-config name)

I'm sure someone could write a utility to convert eslint-config-x into eslint-plugin-y, especially if a plugin can specify a default config (that required explicit extending ofc, not that was applied automatically), but that's a pretty big burden to impose on users.

@g-k g-k referenced this issue in pdehaan/eslint-config-fxa Nov 14, 2016
Closed

Add security rules from mozfreddyb/eslint-config-scanjs #8

@nzakas
Member
nzakas commented Nov 14, 2016

@ljharb I understand your point of view and concerns. I do believe it's been proven we can't do this in configs, and I don't think it's worth the time to continue down that path.

It would be really helpful if you (or anyone on this thread) could try to create a plugin as I described and see if it works for your use case. What we need now is feedback on what I described so we can move forward or not.

@iamstarkov

@ljharb

I'm sure someone could write a utility to convert eslint-config-x into eslint-plugin-y, especially if a plugin can specify a default config (that required explicit extending ofc, not that was applied automatically), but that's a pretty big burden to impose on users.

prob its easy to write codemod for this

@bwrrp bwrrp referenced this issue in FontoXML/eslint-config-fontoxml Nov 29, 2016
Closed

eslint-plugin-react can not be found #1

@not-an-aardvark
Member

I'm in agreement with a few other people on this thread that this would be a very unsatisfying solution.

we actually used to define a default plugin config, but we decided it was too big of a surprise to users to have a plugin automatically enforce a config just because you included it. I don't think we want to go back there.

I agree that it would be confusing for plugins to automatically enforce a config. But that's because enforcing a config by default is precisely the behavior that shareable configs are designed for. If we deprecate shareable configs and tell people to use plugins instead, then we need a good replacement for that behavior.


It feels like we're getting too caught up in the edge cases here, and it's causing us to throw out the idea of usable shareable configs entirely, even when it would make things easier in the vast majority of cases. The whole point of a shareable config is to be able to publish a style guide to npm, and allow users to enforce it without having to worry about configuration details. One of the design goals of ESLint is that users should be able to create custom rules so that core doesn't need to support everyone's requests. But as it currently stands, a shareable config publisher can only use ESLint's core rules, otherwise they incur an installation burden on all their users. For example, I can't use a rule from eslint-plugin-node in my shareable config without forcing users to install eslint-plugin-node manually. The fact that a particular rule comes from eslint-plugin-node, rather than core, is an implementation detail that that shareable configs are supposed to avoid.

This is enough of a problem that people use packages like standard, which is just a wrapper around ESLint that also has a preset config. Standard is popular because it makes it easy to just follow a styleguide, without worrying about configuration. If a beginner wants to download airbnb's style guide and enforce it on their code, they shouldn't have to worry about the distinction between plugins, shareable configs, and plugin-specific configs. They should be able to run npm install eslint-config-airbnb and put one extends: line, and have everything just work.

While I agree that we need to account for edge cases such as multiple versions of the same plugin, our solution shouldn't complicate enforcing a styleguide in general just because a shareable config decides to use a rule from a plugin instead of a core rule. Even implementing an inelegant solution for these edge cases would be better than forcing users to worry about the internals of a shareable config whenever it includes a plugin. While my proposal here has a lot of flaws (some of which might be fixable), I still think it would be a distinct improvement over the current behavior.

@bwrrp bwrrp added a commit to FontoXML/eslint-config-fontoxml that referenced this issue Dec 1, 2016
@bwrrp bwrrp Use eslint-plugin-react as peerDependency.
Requiring plugins as direct dependencies of shareable config is not currently supported in eslint (see eslint/eslint#3458). This plugin needs to be installed manually for now.

Fixes #1
daf157a
@not-an-aardvark
Member
not-an-aardvark commented Dec 13, 2016 edited

Here's an updated proposal.

tl;dr:

  • Configs can have plugins as dependencies.
  • If a reference to a plugin rule is unambiguous, it just works.
  • If a reference to a plugin rule is ambiguous, ESLint throws an error and tells the user to be more specific.
  • References to plugins are resolved such that the end user always has the ability to resolve an ambiguity.

Detailed version

  • A reference to a plugin takes the form plugin/rule, and can optionally be preceded by a chain of dependent configs separated by slashes. For example, if the current config depends on foo, and foo depends on bar, and bar depends on plugin, the rule can be referenced with foo/bar/plugin/rule. (syntax up for bikeshedding)
  • A reference to a plugin is invalid if it has a config chain that doesn't actually map to a dependent config. For example, if the config bar does not depend on any config nonexistent, the reference foo/bar/nonexistent/plugin/rule is invalid. ESLint should throw an error if it encounters an invalid reference.
  • In the reference foo/bar/plugin/rule, it is not required that bar explicitly depend on plugin.
    • For example, if bar depends on baz, and baz depends on plugin, then foo/bar/plugin/rule is considered equivalent to foo/bar/baz/plugin/rule. In this case, foo/bar/plugin/rule is called an abbreviated reference, which expands to the full reference foo/bar/baz/plugin/rule.
    • However, if bar also depends on qux, and qux depends on a different version of plugin, then the reference foo/bar/plugin/rule is ambiguous. (It could refer to either foo/bar/baz/plugin/rule or foo/bar/qux/plugin/rule.) If ESLint encounters an ambiguous reference (defined as an abbreviated reference that could expand to more than one possible full reference), it should throw an error. It might also output a list of possible full references, so that the user can easily clarify which reference they intended to refer to.
  • Rules are configured by using valid, unambiguous references.

Example

Suppose my project depends on eslint-config-airbnb, which uses rules from eslint-plugin-import. I could configure these rules the same way that I currently do:

rules:
  import/no-unresolved: warn
extends:
  - airbnb

Now I decide to also use eslint-config-node, which depends on a different version of eslint-plugin-import. So I install it and add it to my config file:

rules:
  import/no-unresolved: warn
extends:
  - airbnb
  - node

At this point, ESLint will throw an error, because I'm configuring the import/no-unresolved rule but it's unclear which version of eslint-plugin-import I'm referring to. To fix this, I can use a more specific reference:

rules:
  airbnb/import/no-unresolved: warn
extends:
  - airbnb
  - node

Note that I only need to change my own config file, and I don't need to change anything in eslint-config-airbnb. The reference to import in eslint-config-airbnb is still unambiguous, because eslint-plugin-node is not in eslint-plugin-import's dependency chain.

Disadvantages

One disadvantage to this proposal is that if it's implemented precisely as stated above, adding a plugin to a shareable config would be a breaking change, because it could cause a reference to become ambiguous. For example, suppose my config looks like this:

rules:
  import/no-unresolved: warn
extends:
  - airbnb
  - google

Since eslint-config-airbnb depends on eslint-plugin-import, but eslint-config-google does not, this reference is unambiguous, so the config is valid. However, if eslint-config-google adds eslint-plugin-import internally in a minor release, my config will break because the reference import/no-unresolved will become ambiguous.

I don't see this as a major concern, because if eslint-config-google starts depending on eslint-plugin-import, it's likely that the config turns on new rules from eslint-plugin-import, which would be a breaking change anyway.

@ljharb
Contributor
ljharb commented Dec 13, 2016

That proposal sounds awesome and seems like it would solve my issue completely - as an added bonus, most people would just be able to trim down their package.json and wouldn't have to edit a single line of config!

@DamienCassou DamienCassou added a commit to foretagsplatsen/klassified that referenced this issue Dec 13, 2016
@DamienCassou DamienCassou Update compatibility of node.js versions
* Remove old node.js versions as their versions of npm are not
  compatible with eslint-config-ftgp
  eslint/eslint#3458 (comment)

* Add node.js 7
d546f91
@DamienCassou DamienCassou added a commit to foretagsplatsen/klassified that referenced this issue Dec 13, 2016
@DamienCassou DamienCassou Update compatibility of node.js versions
* Remove old node.js versions as their versions of npm are not
  compatible with eslint-config-ftgp
  eslint/eslint#3458 (comment)

* Add node.js 7
c492021
@DamienCassou DamienCassou added a commit to foretagsplatsen/klassified that referenced this issue Dec 14, 2016
@DamienCassou DamienCassou Update compatibility of node.js versions
* Remove old node.js versions as their versions of npm are not
  compatible with eslint-config-ftgp
  eslint/eslint#3458 (comment)

* Add node.js 7
f6e4f7d
@DamienCassou DamienCassou added a commit to foretagsplatsen/widgetjs that referenced this issue Dec 14, 2016
@DamienCassou DamienCassou Update travis configuration file
* Remove old node.js version as their versions of npm are not
  compatible with eslint-config-ftgp
  eslint/eslint#3458 (comment)

* Add node.js versions 6 and 7

* Use gulp
ab960ed
@DamienCassou DamienCassou added a commit to foretagsplatsen/widgetjs that referenced this issue Dec 14, 2016
@DamienCassou DamienCassou Update travis configuration file
* Remove old node.js version as their versions of npm are not
  compatible with eslint-config-ftgp
  eslint/eslint#3458 (comment)

* Add node.js versions 6 and 7

* Use gulp
4b81c7d
@jasonkuhrt jasonkuhrt referenced this issue in SSENSE/eslint-config Dec 21, 2016
Merged

feat(elon-883): create custom eslint config #1

4 of 6 tasks complete
@ilyavolodin
Member

@not-an-aardvark Your proposal address collisions with multiple configs/plugins of different version. It does not address the changes needed to allow configs to have hard dependency on plugins, however. How would you handle (with the same code) scenarios where plugin exposes shareable config and it's own rules and is extended by another shareable config? How would plugin with shareable config refer to it's own rules anyways?

@not-an-aardvark
Member
not-an-aardvark commented Dec 22, 2016 edited

@ilyavolodin Thanks for the feedback on the proposal. However, I don't understand some of the issues you raised.

Your proposal address collisions with multiple configs/plugins of different version. It does not address the changes needed to allow configs to have hard dependency on plugins, however.

I'm not sure I understand the distinction you're making. As far as I can tell, the proposal would allow configs to have a hard dependency on plugins (that was the main design goal). Could you clarify why you don't think it would work?

How would you handle (with the same code) scenarios where plugin exposes shareable config and it's own rules and is extended by another shareable config? How would plugin with shareable config refer to it's own rules anyways?

This would work the same way as it currently does. The plugin's shareable config, and also a config that directly extends the plugin's shareable config, would both refer to the plugin's rules as plugin-name/rule-name.

@mysticatea
Member

Let me introduce an example of shareable configs which are using plugins.
Somebody don't need to install the plugins to use the config.

This config depends on 3 plugins; eslint-plugin-eslint-comments, eslint-plugin-mysticatea, eslint-plugin-node.
However, we can install the config with npm install --save-dev eslint eslint-config-mysticatea then we can use it. Users don't need to install the plugins manually.

  • npm@2 installs the plugins as peerDependencies.
  • npm@>=3 installs the plugins as dependencies and flattens those.

In both cases, it would notify INVALID_PEER errors if it happened version conflicts of the plugins.
So version conflicts of the plugins are notified as installation errors, we can act to resolve it.

I have been using this mechanism while 1.5 years and I have never encountered any problems.
So honestly I wonder if this complex feature requires truly.

@ljharb
Contributor
ljharb commented Dec 22, 2016 edited

@mysticatea relying on the automatic installation of peer deps, while also relying on npm 3+'s flattening, is certainly a very clever trick, but in no way do I think that's a proper solution to this problem. Additionally, because of the way your peer deps are defined, if I already have an existing top-level dep on eslint-plugin-node at, say, 2.x or 4.x - it will satisfy the peerDep requirement, but not satisfy the dependencies requirement - which will cause duplicates to be installed in my tree in npm 2, and the conflict will fail to hoist in npm 3 - which could either break, or silently do the wrong thing. It's fortunate that you've had no problems but this is a very brittle solution - one which is only lightly lessened if your peer+normal deps are pinned.

@nzakas
Member
nzakas commented Dec 22, 2016

If a reference to a plugin rule is ambiguous, ESLint throws an error and tells the user to be more specific.
References to plugins are resolved such that the end user always has the ability to resolve an ambiguity.

I know this is a long thread, but I pointed out earlier that placing this burden on the end user is unfair and not something I support. It shouldn't be up to end users to understand a magical dependency path to get a rule to work -- that burden should be on the shareable config developer alone. Shareable configs should be black boxes to their consumers, anything more is going far beyond what is either expected or useful.

I'm still waiting to hear prototype feedback about using plugins instead. That seems to solve all of the problems on this thread without any changes. I'd like to ask again for some volunteer(s) to try out that approach and post feedback here.

I don't see any reason to keep coming up with other proposals until someone tries out the plugin approach to either prove or disprove it's viability as a solution.

@ljharb
Contributor
ljharb commented Dec 22, 2016

@nzakas i'm still unclear on how i can try out that approach without changes in eslint, since plugins that my plugin depends on still currently need to be top-level deps.

@not-an-aardvark
Member

@nzakas Thanks for the feedback.

It's unclear to me how shareable configs can be considered "black boxes". Even with their current behavior, the user has to know a lot about a config's internal dependencies and structure. For example, the user has to configure a rule based on the name of the plugin that it comes from (e.g. import/no-unresolved). Also, with the current behavior, the user has to manually install all of a shareable config's dependencies (although I realize that this is due to changes in npm and is not by design). If configs were really black boxes, the userule wouldn't have to worry about details like this. I think making the user consider a config's dependencies only in exceptional cases (i.e. when two versions are installed) is much better than making the user consider the dependencies every time they install a shareable config. As I mentioned above, it seems right now we're optimizing for very uncommon edge cases while making the most common use-case significantly less convenient.

@ilyavolodin
Member

@nzakas The problem I have with your approach, is it sort of promotes marketing somebody else's work as your own. While in your example you export rules out as react/ruleName, not everyone might do that since it's not enforced in any way, in that case, to the end user it will look like rules that are configured by current plugin are also written by the same person. This makes for ambiguous ownership as well as might make it harder to report bugs with the rule in appropriate repository.

@Turbo87 Turbo87 referenced this issue in emberjs/ember.js Jan 23, 2017
Merged

Use eslint plugin for custom eslint rules. #14866

@nason nason added a commit to wework/eslint-config-wework that referenced this issue Feb 1, 2017
@nason nason Update eslint deps
List them in dependencies so consumers dont need to install them explicitly.
See eslint/eslint#3458 (comment)
b5e00df
@maxnordlund
Contributor

This bite today, and after reading through this thread I think it would be best if eslint itself shipped any sort of upgrade tool, like babel-doctor/brew doctor, that could handle and changes in syntax. E.g. extends: "airbnb" -> extends: "plugin:airbnb", and since this would be a one-time thing I do think it's OK to force upon end users.

One disadvantage of this is how to know what the equivalent plugin would, and for that the config needs to somehow specify that. That could probably go into the package.json, or exported as a special property.

But what to do with legacy configs? First off, the same doctor command could warn and suggest/fix sharable configs, but unless you go with some sort of rules like @not-an-aardvark suggests. Alternatively, or perhaps complementary, is to analyze the peer dependencies for the end user and sync it with their packages.json. This would allow any config install to be reduce to npm install --save eslint-config-foo; eslint sync/doctor/whatever, or even eslint install --config foo.

Not sure if I like the latter though, but such an approach would "solve" the issue today from the eyes of an end user, and allow smooth upgrading when a good solution exists. (Stripping out the peer dependencies and update the user config).

I'm sure there's many things I missed, but the important point is to ship an official doctor/upgrade command when the time comes.

@mmazzarolo

Shameless plug.
If you need an example I used the hints in this thread to create an eslint-plugin with bundled dependencies: eslint-plugin-react-app.

@ljharb
Contributor
ljharb commented Feb 20, 2017

@mmazzarolo What happens if there's two conflicting copies of one of your plugin dependencies in the tree?

@mmazzarolo
mmazzarolo commented Feb 20, 2017 edited

@ljharb I'll be honest: I don't know, I haven't played with them yet.
I have been using this lib for a while on multiple projects without any issue, but I'm almost sure that my strategy would not suit at all your use case.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment