Redesigning Redux
medium.com
medium.com
Once you commit to a config based approach, you're basically closing the door to a future you can't imagine yet. I'm so tired of this short-sighted obsession with configuration some developers seem to have.
I think the author has a myopic vision for where Redux should go.
I think it's the other way around: The greatest weakness of Redux is that (unnecessarily) message passing instead of function calls are used.
Why do a `store.dispatch({type: 'INCREMENT', value: 10})` when it could be a `store.count.increment(10)`? In basically all cases every action is handled by exactly one reducer so just let the reducer be an object with methods instead of a big switch function operating on string constants.
It lets static typing do a lot more work for you when using Typescript/Flow. Less to remember and less to break.
Message passing is a really useful architectural choice because it lets you do powerful global operations. It's a lot crummier on the development ergonomics front.
Because this is how the rest of your programming works, you call functions. In Node you don't have `fs.processMessage('readFile', {path, next})` because `fs.readFile(path, next)` is simpler. Message passing is nice if you pass process boundaries or can't know about the function definition but in Redux you pull in the string constants and everything anyway. Message passing also lets you hook several handlers (reducers) to the same message but I have yet to see it significantly used, this is a niche case.
IMHO you want the tools you use (i.e. redux) to be flexible and powerful, and write helpers, etc, to increase your ease-of-use. Otherwise, as the OP said, you're unnecessarily restricting yourself and creating a more difficult future.
Thanks for your input! Do you mind giving 2-3 examples of good uses of multiple reducers for the same action?
Another example is url updates, I use state-driven-routing and when a url update happens many reducers need to respond to get the pieces of state they are concerned about.
Another example is a login status change, many reducers may need to update themselves if a user logs in our logs out.
It's the difference between having tons of highly specific 1:1 actions getting called from the action creator with their relationship to the root action obfuscated vs. having 1 action that many respond to with their relationship to that action clearly denoted.
To take it a bit further, it reminds me of refactoring vs adding another "if" statement - the former requires comprehensive understanding of the holistic purpose of the system with the dependent relationships properly defined (anti-fragile), whereas the latter is a temporary bandaid that can be implemented with less thought (and the sideeffects that come with it).
Wouldn't the components handle this change and dispatch any actions needed?
> a login status change, many reducers may need to update themselves if a user logs in our logs out.
Again, couldn't this be handled at the component level? It seems risky to have multiple sources of truth for data.
There wouldn't be more than a single source of a given truth, the multiple responders pull different truths relevant to them, the combination of which making the complete app state.
The fact that Redux forces you to use events to modify the store is a useless complication which has spawned reams and reams of crazy libraries like redux-thunk, redux-saga, standard action formats, etc. The craziest thing yet is Ducks: https://medium.freecodecamp.org/scaling-your-redux-app-with-.... This is a pattern which tries as hard as possible to make using events not feel like using events, by tying each event intimately to its listener.
This library is a step in the right direction, but wrapping Redux seems unnecessary. Here's a similar approach in 100 LOC: https://github.com/didierfranc/react-stateful
Here's yet another similar approach, also in 100 LOC https://qubitproducts.github.io/tiny-atom/. Been using it for a year in a few projects.
Is there a point to spreading this piece of sage FUD?
It should not be surprising that pure functions + immutable state results in fewer errors.
The "pattern" Redux asks you to use prevents you from shooting your leg off. If you've written a sizeable application that uses global, mutable state the benefits of Redux shouldn't be surprising.
My only gripe with this eco-system is that it constantly comes up with new names and concepts for things that already exist in the FP world. I agree that the math jargon is off-putting but so is having a dozen words for the same concept. I think the latter is worse.
The library described in the article feels less like an abstraction and more like a shuffling around of syntax. Useful but misleading with a title like, Redesigning Redux.
Update: s/grip/gripe
I agree, but I think it's easy to see how it happens. It's an impediment to adoption of the project if people don't understand what you're talking about, so simplistic shorthand is used initially to get people up to speed, and then you find that your temporary names have become engraved in stone for the vast majority of your users. People being how they are, you're just as likely to get them to change editors as change terminology at that point.
Not exactly a blogpost but a nice reading, explains a lot of the functional programming terms and their mathematical analogs.
edit: s/slang/terms
Developing your own solution is one thing, and developing your own solution that's as battle tested as a popular library is another.
Baking your own solution helps you understand the core problem that a library you could have used tried to solve. But generally speaking developers that tend to do that have difficulty learning top to bottom and favor bottom up approach which leads them to developing things from scratch. That's just a personal anecdote though.
There are cases were a developer has been using a library for a while and is experienced enough to come up with better solution and cleaner APIs but I'd argue that's not the common case.
Understanding other people's code/libraries is a skill worth mastering.
https://github.com/reactjs/redux/tree/master/src
The largest and most important part is CreateStore, but at 250 lines long (mainly comments), it's shockingly simple when you take a look at it.
https://hackernoon.com/transmission-tx-a-flux-alternative-fe...
I'm convinced it's better and simpler, but then again, I'm arguably biased.
Damn, that was a hard pill for me to swallow just now.
1. The constant var `const ACTION_NAME = ...`
2. The constant's value `... = Symbol('ACTION_NAME')`
3. The "action creator" function `const doAction = arg => ...` (extra creativity to add a verb here?)
4. The action creator's return struct: `... => ({ type: ACTION_NAME, payload: arg })`
5. The reducer switch clause `case ACTION NAME:`
6. The container action to callback mapping `onClickThing: dispatch(doAction(42))`
7: The jsx/component calling the callback `onClick={onClickThing}`
1. The tag on the Msg type: `type Msg = ActionName arg | ...`
2. The `case` pattern matcher in the update function
3. The view sending said Msg/actionIt's kinda funny how redux succeeded with such heavy concepts, which is normally not the case. This is still a mistery to me.
Oh and don't get me started on HOC. Want a tooltip over your component? Sure, wrap it in tooltip HOC! Sooo much nicer than directives!
Compare this to Elm and the Elm architecture, and it seems worlds apart. Interoperability between libraries that all works in the same straightforward way, checked by the compiler.
It feels like the JS community made a lot of good decisions for the right reasons, but still ended up in a place that isn't great.
Asynchronous actions feel, to me, that they belong at the component layer where niceties such as spinners are being rendered, and then the backing store is updated with the results of the triggered action.
What drives people to put all of that into their store?
Even Dan Abramov did it in his recent Redux presentation.
It's important that we understand why "best practices" exist but to never see them as commandments. Do what makes sense for your case, but it's your own funeral if you're deviating from best practices without having fully thought it through first.
Functional-core, imperative-shell design suggests to then use this functional core within a shell of code that handles IO, handles failure cases--effectively, anything that can fail belongs here. To me, that's the React component; I don't view React as "functional", as it's got multiple types of state even separate from Redux (and this is not a demerit, IMO). The onClick handler of a React button is no different from the command handler for a CLI application. The code in that handler makes requests of its API client (which I typically inject through context and my own HOC--like I said, React's statefulness doesn't bug me at all), it performs business logic functions on the data yielded from its API client, and it stores the result (typically with a onComplete method or the like, so my redux-connected component can dispatch an action).
I'm amenable to other ways to do that data routing (because that's really all any of this is), I just don't see why the state store is where this routing should be done and see a lot of reasons around testing and failure-rescue and general code cleanliness why I wouldn't. If you were to make HTTP requests in your SQL datastore I'd be remarkably concerned for your sanity; I don't think it makes any more sense to do so in a Redux store.
That makes me super nervous.
The benefit of redux-thunk (and there are obviously other alternatives that provide the same value) is that it lets you write plain functions that have nothing to do with the store, and nothing to do with components. If you put all your business logic into these plain functions, then they're really easy to test, reason about, move, refactor, etc.
> What drives people to put all of that into their store?
It's worth pointing out that redux-thunk doesn't move async code into your store. The store (i.e. reducers) is still completely synchronous.
In many React applications, components are separated into independent "controller"-layers and "view"-layers (containers and components are nomenclature I see occasionally, but the important takeaway is that both inherit from the React component class). Business logic has no place in your view components, of course, but is quite appropriate in your controller components. I don't think component implies view layer here. Even redux actions, when used, are appropriately called from the controller layer and not the view layer. Redux doesn't change anything about logical separation of concerns.
When the Flux pattern was originally introduced, it was driven home that it should not be used for everything. redux-thunk seems to encourage that you use it for everything. What do you use when global state is not necessary (or wanted)?
Sure. Where you actually put stuff depends entirely on the complexity of what it is that you're building. If you have an async function that you need to use in several places then it makes sense to keep it separate. If you have several of these then maintaining a separate container component for each view component that needs access to these quickly becomes unmaintainable.
On the other hand, keeping them as separate functions and adding them in via redux `connect` syntax and `mapDispatchToProps` is simple and makes it clear what's going on. In this scenario, the `connect` higher-order-component acts like the container class, but it built up from separate parts depending on what you need.
> redux-thunk seems to encourage that you use it for everything
I don't think this is true. redux-thunk is a piece of middleware and that's pretty much it. It should be used however is most appropriate for your application and your particular use case. Keeping state inside components isn't always appropriate, but isn't always a bad thing either. If one is growing too complicated, then perhaps the other is more appropriate.
Perhaps I'm dense, but you can serialize the current running state of a Promise in redux-thunk?
(Edit) FYI we had actions to set isWaiting, etc.
If an idea of Redux is that serializable state is a good thing, then redux-thunk seems inimical to that.
I wouldn't. That's why I wouldn't put anything related to it in my Redux state!
But, like, I'm willing to be convinced. This is just where I'm at right now. Maybe your method would make more sense than redux-thunk does to me; have you written anything in more detail about this? (I'm not scared of middlewares, I've used Redux and middlewares even in non-web contexts, so intuitively this at least sounds more promising.)
I've created a gist[1] with some snippets from a current project; it's still alpha, but you can get a general feeling for how the data flow works. I would clean it up a bit if I were officially releasing it to the public, but I'm definitely interested in feedback if anyone thinks I'm doing it ALL WRONG. :)
[1]: https://gist.github.com/parkerault/9dc7e825cc9a62b5efa8a4c1c...
Edit: regarding serializing state; I don't know how others handle it but I subscribe a localStorage writer to the store that only gets called when the store is changed, not when an action dispatches, so if an action creator dispatches a function to redux-thunk, the action doesn't get serialized, just the resulting store after the reducers do their thing. Do you have any examples of actions themselves getting serialized? I believe when Abramov talks about keeping the state serializable in the redux docs he's referring to the store itself; as far as I know anything goes in the actions themselves.
When I refer to serializing an action, I mean that, by necessity, the current state of any in-flight asynchronous actions must be recorded and saved--or else I don't think serialization works, its promises fail and you end up with a state whose completeness you can't guarantee. That opinion may change when I look at this tomorrow, of course--you very well may be cleverer than I am!
I use redux-storage for what you describe for persistence, btw. I like it a lot, I can easily select what I'm going to persist, and I don't have to maintain it. I dynamically pick a backend: IndexedDB where supported, LocalStorage where not.
I think you will find of interest how we've used Immutable.js Maps for reducer state and Immutable.js Records as an easy way to create and pass around (guaranteed immutable) action types.
We were using Immutable.js and functional-style JS a lot in general, so it was a good fit.
https://gist.github.com/adamcee/3191762f2af43ec62a4b335b6695...
In backend, you might spin up a thread to do expensive operation like API access, and that thread has it's own database connection, to read and persist data.
With thunk, you dispatch a thunk-ed action to do expensive operation like fetching data from backend, and that action has it's own access to redux in form of `dispatch` / `getState` arguments.
In the end, my component calls `fetchData` function upon mounting. And the logic dealing with making multiple requests due to paging, or re-trying on errors then lives its own life inside a thunk action, independent from the component that spawned it.
This is exactly how thunks work, you just fire them from your component as a Redux action. I use thunks primarily for when an action needs to get data from an API and insert it into the store (e.g. on login).
Inverting the question, why would I want my presentational components to do a bunch of extra work beyond presenting? Calling an API is not presentational.
This point is not revisited. What is the performance of Rematch in comparison to Redux? I mean I guess it's more or less the same, since it's essentially Redux with less boilerplate, but it would be nice to have an explicit answer to this.
I feel like implements an action for each modification of state (or group of modifications) requires a lot of boilerplate code. In MobX, you treat your data tree as a separate entity, and simply modify it. The UI is automatically updated whenever it needs to.
Mobx isn't bad performance-wise, I think, since it only re-renders the components that are listening to the modified variables when a modification occurs.
The only advantage I see is being able to rewind the state store, which can be quite cool when debugging.
I have done some medium-sized application (only games), and I don't think I've run into any performance or readability problem with MobX. I've never done web apps with a team, though, but I don't think Redux would specially make it easier to work with other programmers.
Another big one is being able to restore the state. In Redux, it's just one giant object that can be easily serialized and use in the future. With Mobx, I'm guessing you have to write deserialize logic for your various classes.
One thing I’ve never quite understood about Redux is why we are dispatching pure data instead of a function of state. This would remove the need for reducers, since your action _is_ your reducer. Something like this...
https://gist.github.com/anonymous/a5d741c5dec81be61e0aa70820...
Would this come with any drawbacks? I could see a large application having performance issues because of a large number of function allocations, but for small applications I don’t see a downside, just a simpler API.
I have another problem with this library, which is relying on singleton pattern for communication. This way, you cannot have more than one rematch instance in one application, which can be limiting. It also makes reasoning more difficult.
[1] https://github.com/rematch/rematch/blob/master/plugins/subsc...
It's easier to get started, but you'll likely have really nasty problems later (testing, server side rendering, etc).
> In this case, subscriptions avoid the need of coupling the auth model to the profile model. Profile simply listens to an action.
I guess it’s no different to using “vanilla” redux actions but because actions contain the reducer namespace then it’s inherently coupled to another model isn’t it?
They lose me with their idea of "models" where I have a global dispatch that can be extended, like dispatch.modelName.actionName. That's just old school object-oriented thinking layered on top of redux's functional programming ideals.