Skip to content

Latest commit

 

History

History
226 lines (170 loc) · 8.46 KB

on-stable-dom-handlers.md

File metadata and controls

226 lines (170 loc) · 8.46 KB

This page describes an advanced technique to avoid Reagent's "anonymous callback" inefficiency.

This technique is only useful when you are writing performance-critical components. Indeed, you may never need it, but understanding it will deepen your knowledge about Reagent.

A performance-critical component would need to satisfy all of the following:

  1. be repeatedly used on a page, perhaps within the rows of a list, or the cells of a grid, or the branches of a tree
  2. and then be re-rendered a lot
  3. and involve creating anonymous callback functions in re-renders

Equality Tests

Equality tests are at the heart of the issue, so let's start by considering this one:

(= [1 :2 "3"]
   [1 :2 "3"])

Question: will it evaluate to true or false?
Answer: true. Because the elements #!clj 1, #!clj :2 and #!clj "3" are identities, so different instances test equal, and because vector equality is determined by element equality.

Implications For Reagent

Reagent uses equality checks on props to avoid doing unnecessary work.

Consider this view:

(defn parent
  [x]
  [child 1 :2 "3"])

Imagine that parent has already rendered once. And then something causes it to re-render (perhaps its parent re-renders and supplies a different x).

When parent re-renders, it will re-render child and supply three props: #!clj 1, #!clj :2 and #!clj "3". But these props will be the same as the ones supplied with the "last render" and, as a result, Reagent calculates that it doesn't need to perform the re-render of child.

Reagent assumes that child is rendered by a pure function and it deduces that giving a pure function the same props/args as last time, results in the same result (the same hiccup). And so Reagent avoids doing this unnecessary work. It is an efficiency thing.

Question: But how does Reagent check for this "sameness" between the props supplied "this render" vs the "last render"?
Answer: by testing the props using #!clj =

So:

  • #!clj 1 (last time) is #!clj = to #!clj 1 (this time)
  • #!clj :2 is #!clj = to #!clj :2
  • and #!clj "3" too is equal to last time

All three props are = to last time, which allows Reagent to skip the re-render of child.

But Wait, Functions

Next step: we introduce anonymous callback functions.

Consider this test:

(= (fn [n] (+ n 1))
   (fn [n] (+ n 1)))

Question: will it evaluate to true or false?
Answer: false. You might be able to see that the two anonymous functions are equal, but = says "no". Anonymous functions are not like #!clj 1 or #!clj :2

Armed with this knowledge, what happens if we add a new prop for child and make it an anonymous callback function? Like this:

(defn parent
  [x]
  [child 1 :2 "3"  (fn [] (dispatch [:something]))])

So, that's now four props given to child. The first three of them are as before. But the new last one is an anonymous function.

What a difference that new one makes!! Now, every time parent is re-rendered, child will also be re-rendered. Every time. Reagent's efficiency checks which try to avoid unnecessary work are broken when one prop is an anonymous function because the value "this time" does not test = to last time, even though we can see they are the same function.

Fixing It

Surely the way to fix this is simple: don't use an anonymous function.

First we create:

(defn callback
  []
  (dispatch [:something]))

Then we perform a test:

(= callback callback)

Question: what will it evaluate to? true or false?
Answer: true. callback is a symbol which is bound to a function. So we are comparing the value bound to callback (a function), and that will test = because it is the same function on both sides of =.

Let's use that;

(defn parent
  [x]
  [child 1 :2 "3"  callback])   ;; <--- used here

Now all four props for child will test = to last time. And child won't be re-rendered.

In fact, we can make this simpler by using a Form-2 function:

(defn parent
  [x]
  (let [callback   #(dispatch [:something])]   ;; <-- create here
    (fn [x]
      [child 1 :2 "3"  callback])))

The outside function creates the anonymous function, which means it doesn't get recreated and change on each render, and so callback will test = to "last time" in all renders.

Problem fixed? Yes, but unfortunately, only for the simple case. Which really means "no, not fixed enough".

Where It Breaks Down

Say, parent has an argument id. And, say, we need to use that in the callback.

(defn parent
  [id]
  (let [callback  (fn [] (dispatch [:something id]))]
    (fn [id]
      [child 1 :2 "3" callback])))

Does this work? Sadly, no. Can you spot the bug?

callback has closed over the original value for id - the one given to the outer function. callback will never dispatch the "latest" value of id provided to the re-render function (notice that id is an argument to both the outer and inner functions in the Form-2 Component).

The Technique

So, enough preamble and explanation, here's the real solution - and it has three parts:

(defn parent
  [id]
  (let [callback          (fn [the-id] (dispatch [:something the-id]))   ;; <-- Note 1
        callback-factory  (callback-factory-factory callback)]           ;; <-- Note 2
    (fn [id]
      [child 1 :2 "3"  (callback-factory id)])    ;; <-- Note 3

Notes:

  1. callback no longer tries to "close over' id. Instead, it takes id as an argument.
  2. callback-factory-factory has a long name, and it performs the trick. It returns a function which, when called, will always return the same function, which wraps callback. (code supplied below)
  3. Each time we call callback-factory, it will return the same function. But it does it in a way which allows for id to vary on each render.

Which only leaves me to show you the hero in our story, which I offer
without further explanation for you to read and understand:

(defn callback-factory-factory
  "returns a function which will always return the `same-callback` every time 
   it is called. 
   `same-callback` is what actually calls your `callback` and, when it does,
   it supplies any necessary args, including those supplied at wrapper creation
   time and any supplied by the browser (a DOM event object?) at call time." 
  [the-real-callback]
  (let [*args1        (atom nil)
        same-callback (fn [& args2]
                        (apply the-real-callback (concat @*args1 args2)))]
    (fn callback-factory
      [& args1]                   
      (reset! *args1 args1)
      same-callback)))               ;; <-- always returns the same function

More Advanced Again

Sometimes the callback will need to accept a DOM event argument. In the following code, notice on-change:

(defn some-input-view 
  [_]
  (let [on-change         (fn [id dom-event]    ;; <-- Note 1
                            (dispatch [:changed id (-> dom-event .-target .-value)]))
        on-change-factory (callback-factory-factory on-change)]

    (fn [id text]
       [:input {:type      "text" 
                :value     text
                :on-change (on-change-factory id)}] ])))   ;; <-- Note 2

Note:

  1. this time the callback, on-change, takes two arguments. The id and the DOM event dom-event
  2. But, when we call the factory we only supply one of these two arguments, almost like this is a "partial". The browser will be supplying dom-event later when it calls the callback.

Summary

This page introduces an advanced technique.

Reagent's method for avoiding unnecessary re-renders is subverted by "anonymous functions" and, specifically, callbacks which are generated by successive renders. These functions do not test equal to each other, even though we programmers can see they are the same each time.

Most of the time, we ignore this issue. We happily accept the small inefficiency this potentially causes because we get to write clear, terse code, and we value that highly.

But, in some rare situations, this small inefficiency can accumulate into a problem when there are large numbers of the same components simultaneously on a page, and when they are unnecessarily redrawn a lot.

When it does become a problem, this page showed you a technique for fixing the issue, but it comes at the cost of you having to write more code, and introduce an abstraction based on callback-factory-factory.