Bundle cjs and es formats

[TrySound]

Sorry, I have confused branches. I think this will be a breaking change and should be merge to next branch.

[gaearon]

Can you show build output comparison?

[TrySound]

https://gist.github.com/TrySound/987213f06dd9b29ca2cecbdea55d8047

[timdorr]

I'm not sure what this gets us though. We only need to compile out ES.next stuff for these builds, as they're only used in the context of a module-aware system (node, webpack, rollup, etc). We don't need to bundle for these systems.

[TrySound]

It's about optimization for bundlers and node.
/cc @Rich-Harris

[timdorr]

Sure, but that will show up in stack traces as just redux.js, which doesn't map back to the source code in the repo, which would make it harder to trace errors and develop fixes.

Also, if the bundle is bad in some way, we wouldn't know, as the tests run against the code as-is (transpiled through babel), not in its bundled form.

Optimization should come at the end of the bundle process, not at the library level. Our early optimization might conflict with a later optimization or prevent one from occurring. We should just ship code and let the user bundle and optimize however they want.

I get where you're coming from, but unfortunately, this is going to cause more problems than it might solve.

[gaearon]

I dunno, isn't this exactly what we're doing for React?

[TrySound]

That's why I opened this PR. I think debug is not so significant here. Redux is small library. Function names are more than descriptive. Producing smaller packages, smaller in count of files, we are forcing the most optimal versions of libraries independent of build tools. Better UX
You care about bundle quality? I guess uglify is the most blackboxed item of stack today.
That's my opinion.

[gaearon]

I agree with @TrySound and I would like this to be reconsidered. There is very little surface area in Redux, and IMO the argument around tests just means we need better integration tests for result bundles. Personally I’m coming to the conclusion that the end user can’t be trusted to optimize efficiently, and also doesn’t have enough leverage (e.g. can’t easily remove junk code at the module boundaries if they use Webpack).

[gaearon]

@TrySound Can you suggest a testing approach that would work well? IMO it would be nice to have a way to run all existing tests on result bundles in addition to the source. They should be testing the public API anyway.

[timdorr]

I dunno, isn't this exactly what we're doing for React?

No?

I'm still a pretty firm 👎 on this change. I don't know if any other library bundling for anything other than UMD.

[TrySound]

@timdorr It's. Try [email protected]

[gaearon]

No?

Sorry, I meant React 16. I thought we were pretty vocal about moving to Rollup and flat bundles :P

[Rich-Harris]

Forgive me for weighing in — have been passively aware of this conversation since I was @-ed yesterday.

My view is that libraries should always bundle their own source code unless they have a compelling reason not to — @gaearon rightly points out that generating the best possible output (fewest bytes, quickest startup) requires a certain level of expertise, and for the app developer to be using particular tools. It's quite easy to end up with various unnecessary overhead such as unnecessarily transpiled import declarations, or the cruft that most bundlers introduce at module boundaries.

By pointing main and module at a bundle you're guaranteeing that Redux users get the best possible outcome, regardless of their level of expertise.

@timdorr you raise two excellent points:

• Debugging — for Redux developers, this can be solved by generating a sourcemap and installing source-map-support before tests run. For Redux consumers diagnosing errors in their app, it's a little harder because they would have to deal with sourcemaps themselves (and you would have to include them in the repo). It's probably not a huge deal in Redux's case since it's only a few hundred LOC anyway
• Testing — Personally I've come round to the view that tests should cover the distributable code rather than individual modules anyway, because that excludes the possibility of bugs in your build process causing issues that show up in people's apps and not in your test suite (I've been there...). It does mean you're limited to testing the public API rather than unit tests of internal modules, but then again the API is the only thing that really matters!

Just my 2c — I don't have a dog in this fight but thought I'd try and articulate why I think this is a good direction for the ecosystem to go in.

[timdorr]

Sorry, I meant React 16

Ah, OK. I haven't yet tried out the alphas myself. But I would point out React has different design and performance goals than Redux. To be honest, the "well, it's what React does" argument doesn't ever sway me. React can afford to jump through certain hoops because the benefits are often multiplied as they make their way downstream (both in code and the ecosystem surrounding it).

And yes, Redux is orders of magnitude smaller and simpler, so there is less potential "damage" from things like this. However, I don't see the benefit in making the build, testing, and debugging processes more complex for no practical benefit to any stakeholder. So, since this is central to my concerns, I'll ask it more directly: what specific and practical gains does library bundling get us?

Anecdotally, I would never apply this to a library like React Router, as one of the bundle-slimming techniques we use is encouraging direct imports (import Router from 'react-router/Router') as sort of a poor man's tree shaking. Bundling everything together, including unused components, would have a negative effect on bundle sizes in that case.

[gaearon]

since this is central to my concerns, I'll ask it more directly: what specific and practical gains does library bundling get us?

Clean, readable code without junk and indirections.

Output like this is suboptimal:

if (!(0, _isPlainObject2['default'])(action)) {
  throw new Error('Actions must be plain objects. ' + 'Use custom middleware for async actions.');
}

Especially in hot paths (we have to dereference ['default'] every time).

This also doesn't look great:

	exports.__esModule = true;
	exports.compose = exports.applyMiddleware = exports.bindActionCreators = exports.combineReducers = exports.createStore = undefined;

	var _createStore = __webpack_require__(2);

	var _createStore2 = _interopRequireDefault(_createStore);

	var _combineReducers = __webpack_require__(7);

	var _combineReducers2 = _interopRequireDefault(_combineReducers);

	var _bindActionCreators = __webpack_require__(6);

	var _bindActionCreators2 = _interopRequireDefault(_bindActionCreators);

	var _applyMiddleware = __webpack_require__(5);

	var _applyMiddleware2 = _interopRequireDefault(_applyMiddleware);

	var _compose = __webpack_require__(1);

	var _compose2 = _interopRequireDefault(_compose);

	var _warning = __webpack_require__(3);

	var _warning2 = _interopRequireDefault(_warning);

	function _interopRequireDefault(obj) { return obj && obj.__esModule ? obj : { 'default': obj }; }

when it all could’ve been top-level functions referencing each other.

I’m not sure I understand the concerns about testing and debugging. How often do we release Redux updates? How fast is it changing? I can understand the concern if there is a lot of churn, but I don’t quite see why this is an issue here. Testing the output bundles will only make the process more solid (and help avoid release bugs).

[TrySound]

I guess internal and interface testing are both possible with babel. We can import interface part from entry point and internal stuff from source files. That's what I did here.

[TrySound]

One other point for bundling is babel. It adds helpers in every file. Rollup adds 'em only once. So for end user is not possible to make optimal build other then babel source files of the library. And this is awful UX. Because of this I reduced component based library recompose in two times. It's not small redux.

[timdorr]

Clean, readable code without junk and indirections.

Ok, now I'm starting to get it 😄

I’m not sure I understand the concerns about testing and debugging.

I'm only concerned with stack traces pointing back to code that's different than what they see in the repo. Namely, if they see createStore.js:123 vs redux.js:456, they might have a hard time learning what's going on. Again, it's less of a concern for Redux where the codebase is so small.

I guess a lot of my concerns stem from couching this discussion on code running in the browser. But I'm running of steam wanting to talk through transpiled code compatibility issues...

Anyways, I guess this looks good to me. Even though we don't guarantee it, this will break anyone using direct imports. Would we want to break this behavior on a minor or major? If major, then can I rebase this PR against next. @TrySound? I don't think there are too many potential conflicts, so it should be save to do so.

[TrySound]

Yeah, I think it should be released as major.

[gaearon]

Even though we don't guarantee it, this will break anyone using direct imports

That’s the only part that’s concerning me. I’d like to have a good way forward there. What should libraries depending on specifically createStore use? Should we tell them “use Webpack 2” and make sure tree shaking works (in ES build)? Also how do we make an ES build?

[timdorr]

OK, I'm going to edit this to be against next. There will likely be conflicts because I haven't yet rebased against the latest master changes (particularly, rollup for UMD). I'll try to get to that shortly, so don't worry about resolving those if they come up.

[timdorr]

Oh yeah, ignore all that interim junk 😄

[timdorr]

OK, there we go. Merge conflicts resolved.

What should libraries depending on specifically createStore use?

We could keep transpiling individual files into lib as an interim solution. The bundled versions would be preferred, based on package.json. But those wanting direct imports could still do as before. And when they can upgrade to a tree-shaking bundler, they can do away with direct imports then.

Also how do we make an ES build?

I'm confused, isn't that already in this? It's just bundled together, but the import/exports remain untouched. The reason for the ES builds is to support tree-shaking bundlers, so they wouldn't be doing direct imports and would be using whatever the module entry point specifies.

[timdorr]

Fixed up some merge conflicts and got this using the latest rollup config syntax.

For sizing comparison on the current next branch:

 ⇶ wc lib/* lib/utils/*
     63     308    2323 lib/applyMiddleware.js
     50     286    2038 lib/bindActionCreators.js
    144     840    6317 lib/combineReducers.js
     35     134     927 lib/compose.js
    266    1373    9873 lib/createStore.js
     45     200    1944 lib/index.js
      0       0       0 lib/utils
     13      61     381 lib/utils/actionTypes.js
     24     103     694 lib/utils/warning.js
    640    3305   24497 total

And this PR:

 ⇶ wc lib/*
  566  3100 22032 lib/redux.js

So, a nice size reduction, and therefore a bit of complexity reduction too.

Here's the latest output: https://gist.github.com/timdorr/f6d76d41508bf5a352846ba685322a6c

[timdorr]

OK, I'm feeling good on this. It's well-tested and has a sensible reasoning for the change. We'll be sure to communicate that change to users when it lands in 4.0. Thanks, @TrySound!