Skip to content
Rui Yang edited this page Jun 27, 2015 · 41 revisions

The following assumes familiarity with core.async.

While the beginner and intermediate tutorials are good for getting a feel for the Om primitives they are not representative guides on how to structure an Om application for scale.

Cursors are a powerful abstraction but often not the tool to reach for when structuring larger applications. In the worst case cursors are a quick trip back to non-modular programs where every component has full knowledge and direct dependence on the global state of the application.

In order to recover modularity we must reach for other tools. We'll see how core.async channels and reference cursors allow us to regain modularity in complex applications.

The following is a loose set of patterns, real applications may need more or less than what is described. Still the following is a good starting point for doing your own explorations around organizing larger Om programs.

  • The Request Channel
  • The Publish & Notification Channels
  • Reference Cursors

The Request Channel

The Problem

In many applications you have some component several levels deep in the render tree. The question is, how do we get data to this component without having to pass data to everyone in between the root of the application and where the component actually lives?

Solution

The Request Channel pattern allows us to avoid passing data through components uninterested in the data. The strategy is a familiar one - treat a component as if it were a client browser. Instead of receiving its data directly it will request it when it mounts. In order for this to work we need to establish a global service channel that can process requests from different parts of your application.

Let us assume we'll have a table view somewhere whose contents will be requested via an XHR request. In the following we'll just mock out the remote data for simplicity.

The global serve will be a multimethod. By using multimethods we can extensibly serve different "client" components in our application. Again this is a familiar pattern - a routing table.

Note that serve takes a map that represents the request analogous to how you might process form parameters or a JSON API request server side. In order to respond to the client one of the arguments is a response channel. We simply write out the data the client needs onto that channel:

(def remote-data (vec (map #(str "Item " %) (range 100))))

(defmulti serve :op)

(defmethod serve :data
  [{:keys [start per-page res]}]
  (put! res (subvec remote-data start per-page)))

When we establish our application root we'll do something like the following. You can ignore notif-chan and pub-chan for now, they form the other pattern we'll talk about momentarily:

(defn main []
  (let [req-chan     (chan)
        pub-chan     (chan)
        notif-chan   (pub pub-chan :topic)]

    ;; server loop
    (go
      (while true
        (serve (<! req-chan))))

    (om/root
      (fn [app-cursor owner]
        (reify
          om/IRender
          (render [_]
            ...)))
      app-state
      {:shared {:req-chan   req-chan
                :notif-chan notif-chan
                :pub-chan   pub-chan}
       :target (. js/document (getElementById "app"))})))

Your table view component might look something like the following:

(defn table-view [_ owner]
  (reify
    om/IInitState
    (init-state [_]
      {:start 0 :per-page 10 :data nil})
    om/IDidMount
    (did-mount [_]
      (let [res (chan)]
        (go
          (>! (:req-chan (om/get-shared owner))
              {:op :data :start start
               :per-page per-page
               :res res})
          (om/set-state! owner :data (<! res)))))
    om/IRenderState
    (render-state [_ {:keys [data]}]
      (if data
        (apply dom/table nil (map table-row data))
        (dom/div "Loading ...")))))

Now any component that needs data can get at it. No need to share global state all over your application.

The Publish & Notification Channels

The Problem

Component B and Component C do not share a common parent Component A yet they need to communicate to each other. How do we synchronize state without again passing the global app state around so that B & C can mutate some value they synchronize on? Component B needs to tell Component C something important without having a direct reference to Component C.

The Publish Channel

Component B should publish a topic onto pub-chan.

(defn widget-b [data owner]
  (reify
    om/IRender
    (render [_]
      (dom/button
        #js {:onClick
             #(put! (:pub-chan (om/get-shared owner))
                    {:topic :hello :data "Hi there!"})}
        "Click me!"))))

But how will Component C get this message?

The Notification Channel

When Component C mounts it subscribes to topics it cares about. This looks something like the following:

(defn widget-c [data owner]
  (reify
    om/IInitState
    (init-state [_]
      {:message nil})
    om/IDidMount
    (did-mount [_]
      (let [events (sub (:notif-chan (om/get-shared owner)) :hello (chan))]
        (go
          (loop [e (<! events)]
            (om/set-state! owner :message (:data e))
            (recur (<! events))))))
    om/IRenderState
    (render-state [_ {:keys [message]}]
      (if message
        (dom/p nil message)
        (dom/p nil "Waiting ... waiting ... waiting ...")))))

Reference Cursors

Building truly modular and reusable components is a time consuming task. In most cases this level of abstraction is outside the time constraints of the problem at hand. Introducing channels to do something as simple as modifying a collection introduces manual resource management (of the go loops) and a considerable amount of ceremony.

From version 0.8.0, Om offers a new tool for this - reference cursors. Reference cursors are like cursors but they have the novel property of representing an identity in the application state that you can observe. Observation is similar to React event handlers - there is no need to worry about observing multiple times nor is there any concern about needing to unobserve. Reference cursors fully support the Om state management model - time travel properties are preserved.

Using reference cursors is simple and natural and allows components to be organized around a shared API they can just call out to.

For example suppose we want a logical collection from a vector in our app state. We can now write an API for it using reference cursors:

(def app-state
  (atom {:items [{:text "cat"} {:text "dog"} {:text "bird"}]}))

(defn items []
  (om/ref-cursor (:items (om/root-cursor app-state))))

In a subview we can now simply write the following.

(defn sub-view [{:keys [title]} owner]
  (reify
    om/IRender
    (render [_]
      (let [xs (om/observe owner (items))]
        (dom/div nil
          (dom/h2 nil title)
          (apply dom/ul nil
            (map #(dom/li nil (:text %)) xs)))))))

This is very natural and straightforward, the component did not need to take its data from a parent nor did it need to receive it via a channel.

Now imagine we have a parent view that looks like the following:

(defn main-view [_ owner]
  (reify
    om/IRender
    (render [_]
      (let [xs (items)]
        (dom/div nil
          (om/build sub-view {:title "View A"})
          (om/build sub-view {:title "View B"})
          (dom/button
            #js {:onClick
                 (fn [e] (om/transact! xs #(assoc % 1 {:text "zebra"})))}
            "Switch To Zebra!"))))))

This parent is blissfully unaware that two children components also rely on the same logical collection. Notice that it can transact! on the reference cursor the same as any other cursor.

Many existing convoluted patterns around the manipulation of application data can be circumvented simply by using reference cursors and designing a simple API that any component can call into.

Further Investigation Needed

The three patterns above eliminate many cases where we may have relied on passing around too much of the application state for coordination. Nothing in the tutorial is conclusive, rather treat it as small set of guidelines to build up more expressive functionality suitable for the large application you are building.