I’ve noticed that the reactDOM.renderToString()
method starts to slow down significantly when rendering a large component tree on the server.
Background
A bit of background. The system is a fully isomorphic stack. The highest level App
component renders templates, pages, dom elements, and more components. Looking in the react code, I found it renders ~1500 components (this is inclusive of any simple dom tag that gets treated as a simple component, <p>this is a react component</p>
.
In development, rendering ~1500 components takes ~200-300ms. By removing some components I was able to get ~1200 components to render in ~175-225ms.
In production, renderToString on ~1500 components takes around ~50-200ms.
The time does appear to be linear. No one component is slow, rather it is the sum of many.
Problem
This creates some problems on the server. The lengthy method results in long server response times. The TTFB is a lot higher than it should be. With api calls and business logic the response should be 250ms, but with a 250ms renderToString it is doubled! Bad for SEO and users. Also, being a synchronous method, renderToString()
can block the node server and backup subsequent requests (this could be solved by using 2 separate node servers: 1 as a web server, and 1 as a service to solely render react).
Attempts
Ideally, it would take 5-50ms to renderToString in production. I’ve been working on some ideas, but I’m not exactly sure what the best approach would be.
Idea 1: Caching components
Any component that is marked as ‘static’ could be cached. By keeping a cache with the rendered markup, the renderToString()
could check the cache before rendering. If it finds a component, it automatically grabs the string. Doing this at a high level component would save all the nested children component’s mounting. You would have to replace the cached component markup’s react rootID with the current rootID.
Idea 2: Marking components as simple/dumb
By defining a component as ‘simple’, react should be able to skip all the lifecycle methods when rendering. React already does this for the core react dom components (<p/>
, <h1/>
, etc). Would be nice to extend custom components to use the same optimization.
Idea 3: Skip components on server-side render
Components that do not need to be returned by the server (no SEO value) could simply be skipped on the server. Once the client loads, set a clientLoaded
flag to true
and pass it down to enforce a re-render.
Closing and other attempts
The only solution I’ve implemented thus far is to reduce the number of components that are rendered on the server.
Some projects we’re looking at include:
- React-dom-stream (still working on implementing this for a test)
- Babel inline elements (seems like this is along the lines of Idea 2)
Has anybody faced similar issues? What have you been able to do?
Thanks.
4
Answers
It’s not a complete solution
I had the same issue, with my react isomorphic app, and I used a couple of things.
Use Nginx in front of your nodejs server, and cache the rendered response for a short time.
In Case of showing a list of items, I use only a subset of list. For example, I will render only X items to fill up the viewport, and load the rest of the list in the client side using Websocket or XHR.
Some of my components are empty in serverside rendering and will only load from client side code (
componentDidMount
).These components are usually graphs or profile related components. Those components usually don’t have any benefit from SEO point of view
About SEO, from my experience 6 Month with an isomorphic app. Google Bot can read Client side React Web page easily, so I’m not sure why we bother with the server side rendering.
Keep the
<Head>
and<Footer>
as static string or use template engine (Reactjs-handlebars), and render only the content of the page, (it should save a few rendered components). In case of a single page app, you can update the title description in each navigation insideRouter.Run
.Using react-router1.0 and react0.14, we were mistakenly serializing our flux object multiple times.
RoutingContext
will callcreateElement
for every template in your react-router routes. This allows you to inject whatever props you want. We also use flux. We send down a serialized version of a large object. In our case, we were doingflux.serialize()
within createElement. The serialization method could take ~20ms. With 4 templates, that would be an extra 80ms to yourrenderToString()
method!Old code:
Easily optimized to this:
In my case this helped reduce the
renderToString()
time from ~120ms to ~30ms. (You still need to add the 1xserialize()
‘s ~20ms to the total, which happens before therenderToString()
) It was a nice quick improvement. — It’s important to remember to always do things correctly, even if you don’t know the immediate impact!Idea 1: Caching components
Update 1: I’ve added a complete working example at the bottom. It caches components in memory and updates
data-reactid
.This can actually be done easily. You should monkey-patch
ReactCompositeComponent
and check for a cached version:You should do this before you
require('react')
anywhere in your app.Webpack note: If you use something like
new webpack.ProvidePlugin({'React': 'react'})
you should change it tonew webpack.ProvidePlugin({'React': 'react-override'})
where you do your modifications inreact-override.js
and exportreact
(i.e.module.exports = require('react')
)A complete example that caches in memory and updates
reactid
attribute could be this:I think fast-react-render can help you. It increases the performance of your server rendering three times.
For try it, you only need to install package and replace ReactDOM.renderToString to FastReactRender.elementToString:
Also you can use fast-react-server, in that case render will be 14 times as fast as traditional react rendering. But for that each component, which you want to render, must be declared with it (see an example in fast-react-seed, how you can do it for webpack).