(return to home)

Application architecture and state management with react based applications.

This is a brief summary of the architectural decisions that I have taken when I have started the development of uxbox and with time I found that the result is scaling pretty well for relatively big project, so I decided to write about it.

I will use ClojureScript as base lenguage but the main ideas explained here can be done with plain javascript with minor adaptations.

TL;DR

A quick summary of the design decisions (almost all explained in the article):

  • Unidirectional state transformation flow.
  • Event based system with three well defined event types: pure state transformations, async state transformations and side effectful actions.
  • A client-server architecture on the frontend-only application with strong separation of logic from UI.
  • FRP streams as a glue between UI and the state transformation logic.
  • State modeled for fast lookup (like a database) instread of to be UI driven.

The event system

Let start with the first piece: the event system. As I have said previously, there are three different events:

  • pure state transformations
  • asynchronous actions that can or not be state transformations
  • side effects actions

The concept of an event is modeled using a basic ClojureScript constructions such are records for define new types and protocols for define the different event interfaces. For understand it better, let see some code. This is how a pure state transformation event is defined:

(defprotocol UpdateEvent
  "Defines a simple state transformation.
  It receives a state and should return the
  transformed state."
  (-apply-update [event state]))

And this is how an implementation of that event will look in an hipotetical toy project that manages one unique counter:

(defrecord Increment [n]
  UpdateEvent
  (-apply-update [_ state]
    (update state :counter (fnil #(+ n %) 0))))

(defn increment
  "Create an instance of Increment event."
  ([] (increment 1))
  ([n] (Increment. n)))

If you don't need a named types, you can use reify for create anonymous objects, you are not restricted to use records.

Evidently, for small things like this, it seems like too much boilerplate, but for big applications having well defined events makes things more clear to understand and allows to have the application logic encapsulated in small, testeable events.

I have said that there are three different kind of events, this is how the rest two events can be defined:

(defprotocol WatchEvent
  "Defines a asynchronos state transformation.
  It receives the state and the reference to the current
  stream of events and should return an other stream
  of same kind of events (that can be of `UpdateEvent`,
  `WatchEvent` or `EffectEvent`."
  (-apply-watch [event state stream]))

(defprotocol EffectEvent
  "Defines a side effectfull action.
  It receves the state and the reference to the current
  stream of events. The return value will be ignored."
  (-apply-effect [event state stream]))

Do not worry about them, they will be explained later with better examples.

Stream Loop

In order to start playing with events, there are two additional functions, init and emit!.

(defonce ^:private bus (rx/bus))

(defn emit!
  "Emits an event or a collection of them."
  ([event]
   (rx/push! bus event))
  ([event & events]
   (run! emit! (cons event events))))

(defn init
  "Initializes the stream event loop and
  return a stream with model changes."
  [state]
  ;; [implementation ommited]
  ))

The init funtion just initializes something that I call streamloop, it receives a initial state and returns a stream of state transformations:

(defn get-initial-state
  []
  {:counter 1})

(defonce stream
  (init (get-initial-state)))

It's responsability is pretty simple: process events and emit state snapshots after each transformations. At this moment you should not care how init is implemented, because is important to understand its purpose to be here (the complete implementation of that is referenced on the end of this article).

And later, the emit! variadic function, that just does that you will expect, emit event instances into the stream:

(emit! (increment 1)
       (increment 1))
;; => nil

If we subscribe to the stream returned by init function before emiting any events, we will obtain a stream of state snapshots after each applied transformation:

(require '[beicon.core :as rx])

(rx/subscribe stream (fn [state]
                       (println "state:" state)))
;; [console] state: {:counter 1}
;; [console] state: {:counter 2}

Obviously, every approach has its advantages and tradeoff's. In this case the advantadge and the tradeoff at the same time is that the state lives in the stream, so it only can be modified emiting events.

In order to start using that state in your react based application, you have to materialize the state to something that components can understand.

Let see a little example using rum:

(require '[rum.core :as rum])

(defonce state (rx/to-atom stream))

(rum/defc counter < rum/reactive
  []
  (let [counter (:counter (rum/react state))
        increment #(emit! (increment 1))]
    [:div.label {:on-click increment}
     [:span counter]]))

(rum/mount (counter) js/document.body)

The state is materialized into atom and then used in a example rum component.

Independently of the explained tradeoff, this approach is quite flexible because it does not dictate how you should pass the state to the UI components. In fact this approach can be used without problems with om or reagent based applications.

Async state transformations & side effects

Pure state transformations are nice for show in examples, but real applications are full of side effects and asynchronous stuff.

Before going deeper, let's go to see an example event using WatchEvent protocol. Imagine that you need to load the counter value from some kind of backend or database that has asynchronous interface:

(defrecord CounterLoaded [value]
  UpdateEvent
  (-apply-update [_ state]
    (assoc state :counter value)))

(defrecord LoadCounter [path]
  WatchEvent
  (-apply-watch [_ state stream]
    (->> (http/get path)
         (rx/from-promise)
         (rx/map ->CounterLoaded))))

(defn load-counter []
  (LoadCounter. "/my/counter"))

As you can see, if you want to perform asynchronous stuff, you should implement two events: one for fetch the counter and other for set it in the state.

However, the flexibility of clojurescript allows us implement the UpdateEvent protocol directly to the javascript function, allowing the use of plain functions as pure transformations event instances. So the previously code can be expressed in this manner:

(defrecord LoadCounter [path]
  WatchEvent
  (-apply-watch [_ state stream]
    (->> (http/get path)
         (rx/from-promise)
         (rx/map (fn [value]
                   #(assoc % :counter value))))))

(defn load-counter []
  (LoadCounter. "/my/counter"))

The first most common problem that you will found using this approach is: How I can syncronize two or more asynchronous events?.

A good use case for the demostration purposes is: load some data and then redirect user to other page. Using the previously defined events, let see how it can be approached:

(defrecord GoToCounter []
  WatchEvent
  (-apply-watch [_ state stream]
    (rx/merge
      (rx/of (load-counter))
      (->> stream
           (rx/filter #(instance? CounterLoaded %))
           (rx/map (fn [_] #(assoc % :location :counter)))))))

The third parameter on -apply-watch is a reference to the main stream where emit! publish events. So you can subscribe to changes that will hapen in some future and react in terms of them. As streams are cancellable by default, implement something like autocompleting using this kind of constructions is quite simple.

You should expect that two events passed to emit! function will be processed by the streamloop in order and pure state transformations will be executed in expected order, but the operations of resulting async operations will be executed in completly arbitrary order and the explained approach serves for synchronize two or more asynchronous events.

The state model

An other very important part of the application architecture is how you have plan to model your state and how your components will have access to it. There are a lot of solutions to this and I think no solution is better than another because everything depends on type of the application.

In my project I've taken the approach where the whole state is visible to all components. The coupling with state problem is solved in this way:

  • Small resuable components usually receives everything that they needs by parameters from its parent components.
  • Big and/or not reusable components just access directly to the state or uses a lenses approach just for create a focused vision of the state.

In fact, in uxbox, the direct access to the state is used in very little portion of code and almost always lenses are used instead. Maybe you are now asking yourself: but that is this so called lenses? As a very simplistic description: lenses are the functional approach to focus to some portion of data structure (tree or seequence like) for easy access to it or transform it.

The most interesting and important part is the ability to derive atoms from other atoms:

(require '[lentes.core :as l])

(defonce counter
  (-> (l/key :counter)
      (l/from-atom state)))

That allows create new derived atoms that focuses on exactly state that the component need:

(rum/defc counter < rum/reactive
  []
  (let [counter (rum/react counter)
        increment #(emit! (increment 1))]
    [:div.label {:on-click increment}
     [:span counter]]))

One of the big advantages of using lenses in reactive components is because they does not trigger the component rerendering if other portion of the atom is modified.

Also, the lenses are not limited to tree like structures, in fact you can provide just a plain function that receives the state as first arguent and should return the focused data. In uxbox the state is strictly organized like a database using data normalization and indexing for fast lookup by id. Using lenses allows have different versions of state just adapted for the UI.

An other very interesting property of lenses is that they are composable in the same way as transducers: using plain function composition facilities such as comp:

(defonce sample-state (atom {:foo {:bar 2}}))

(defonce derived-state
  (-> (comp (l/key :foo)
            (l/key :bar))
      (l/focus-atom sample-state)))

@derived-state
;; => 2

(swap! derived-state inc)

@derived-state
;; => 3

Although, the derived atoms are mutable in the same way as normal atoms, the use of the explained approach of state management they works as read-only focused values because the main state atom is just a materialization of the state that lives in the stream.

Conclusion

This approach works pretty well in my project but I can't assure that it will fits for other projects. Everything depends on the application that you need to build.

I think is pretty flexible, and with little modifications surelly it can work well with different that mine (such as using DataScript for manage the state or anything else).

Referecences

Comments