7.3 KiB
Eek! Performance Problems
Table Of Contents
- 1. Is It The
debug
Interceptor? - 2.
=
On Big Structures - 3. Are you Using a React
key
? - 4. Callback Functions
- A Weapon
1. Is It The debug
Interceptor?
This first one is something of a non-problem.
Are you are using the re-frame.core/debug
Interceptor?
You should be, it's useful. But you do need to be aware of its possible performance implications.
debug
reports what's changed after an event handler has run by using
clojure.data/diff
to do deep, CPU intensive diff on app-db
.
That diff could be taking a while, and leading to apparent performance problems.
The good news is this really isn't a production problem. debug
should only be
present in an Interceptor Chain at development time, and it should be removed
from production using this technique.
Also related, anything which writes large data structures, or strings, to the js console, will be slow. So press F12, pull up devtools console, and have a good look at what's happening in there.
2. =
On Big Structures
Reagent uses =
to compare the previous value of a prop with the
new value of that prop, when it determines if a component needs
rerendering. Make sure you have a good understanding of this..
In the worst case, if those props are big data structures which differ only in some
tiny, leaf aspect, then a lot of CPU cycles will be spent doing
the =
comparison only to eventually work out that, indeed, the
answer is false
.
This problem is exacerbated when components return a lot of hiccup, because
lots of hiccup normally means lots of props which, in turn, means lots of =
work to do on each of those props. Any rerender with those characteristics
could end up chewing a lot of CPU cycles.
An Example Of Problem 2
Imagine you were rendering a 19 by 19 "Go" board.
And imagine that you have a high level board renderer component which creates hiccup for the 361 sub components (19 x 19 grid), and that it provides 3 props to each child:
- grid x cord
- grid y coord
- a chunk of data representing the current game state, from which each of the 361 individual grid components is expected to extract the data they need to render their grid position.
This arrangement could be slow.
First, you have a parent component returning hiccup for 361 sub-components
and that's a lot of hiccup!! Sure, it might not be much code - just a couple
of nested for
, but the hiccup data structure built will be substantial.
Second, after the board renderer returns all this hiccup, for every
one of those 361 sub-components, Reagent must then check the 3 props to
see if they are =
to the value last rendered (to determine if they, in turn,
need to be rerendered), and the comparison on the 3rd prop (game state)
might be deep and expensive. Worse, we do the same expensive check 361
times in a row, and every time we get a false
(because games state
is not =
to last time).
Third, because Reagent gets 361 falses
, it will further rerender
all 361 sub-components even though 360 of them produce the same
hiccup as last time - only one position in the gird has changed.
So, when a new stone is placed on the board, and the game state changes, that triggers a large amount of unnecessary calculation, just to figure out that there's only a rendering change at one point in the 19x19 grid.
So, that's how you can get a performance problem: lots of hiccup,
mixed with time consuming =
tests on big props.
Solutions To Problem 2
The solution is to not do the unnecessary work. Duh!
Produce only the hiccup that is needed. Don't unnecessarily pass around big complicated state in props, unless you really need to.
In the Go example described above, for each new stone placed, only one point in the Go board actually needs to be rerendered, and yet our code asked Reagent to chew a lot of CPU to figure that out.
These kinds of tweaks would improve performance:
-
don't give the entire game state to each of the 361 sub components and then ask them to extract what they need. Instead, give each just the state it needs, and nothing more. That will make the
=
process faster. It will also allow for Reagent to figure out that 360 of the sub components have the same props as last time, and don't need rerendering. And, so, only one sub-component will be rerendered when the parent "board level" component rerenders. -
Also, could you render the board row by row? So that less hiccup is produced by any one component? Can those rows
subscribe
to just the data for their row, so they only rerender when the row-data changes; they only generate hiccup when something really has changed?
3. Are you Using a React key
?
Correctly using React keys
can also make a huge difference to performance.
Some resources:
- http://stackoverflow.com/questions/27863039/key-property-inside-component-function
- http://stackoverflow.com/a/37186230/5215391
- https://groups.google.com/d/msg/reagent-project/J1ELaLV20MU/iutebA-JEgAJ
4. Callback Functions
Look at this div
:
[:div {:on-mouse-over (fn [event] ....) } "hello"]
Every time it is rendered, that :on-mouse-over
function will be regenerated,
and it will NOT test =
to the last time it rendered. It will appear to be a new function.
It will appear to React that it has to replace the event handler.
Most of the time, this is not an issue. But if you are generating a LOT of DOM this small inefficiency can add up.
To work around the problem, lift the function out of the render. Use a Form-2 function like this:
(defn my-component
[]
(let [mouse-over-cb (fn [event] ....) ] ;; created once
(fn [] ;; rendered many times
[:div {:on-mouse-over mouse-over-cb}]))
Now, React will see that mouse-over-cb
is the same as last time. It won't think
the event handler has been replaced.
But like I say, don't be too paranoid about this, it is unlikely to be an issue unless you have something like a table with a lot of rows.
A Weapon
Of course, the way to really track down what is going on is to use the OFFICIAL debugging technique. See the four dominoes play out in the console. You may be surprised by what you find is happening.
Be aware that tracing adds its own performance drag - there's the
overhead of all that stuff getting written on the js console.
Especially if the data getting traced is big - for example,
tracing all of app-db
in the console can take a while and force
Chrome devtools to take masses of RAM. So you may want to selectively
add tracing when poking about.