How to optimally serve and load JavaScript files?

Question

I'm hoping someone with more experience with global-scale web applications could clarify some questions, assumptions and possible misunderstandings I have.

Let's take a hypothetical site (heavy amount of client-side / dynamic components) which has hundreds of thousands of users globally and the sources are being served from one location (let's say central Europe).

1. If the application depends on popular JavaScript libraries, would it be better to take it from the Google CDN and compile it into one single minified JS file (along with all application-specific JavaScript) or load it separately from the Google CDN?
2. Assetic VS headjs: Does it make more sense to load one single JS file or load all the scripts in parallel (executing in order of dependencies)?

My assumptions (please correct me):

Compiling all application-specific/local JS code into one file, using CDNs like Google's for popular libraries, etc. but loading all of these via headjs in parallel seems optimal, but I'm not sure. Server-side compiling of third party JS and application-specific JS into one file seems to almost defeat the purpose of using the CDN since the library is probably cached somewhere along the line for the user anyway.

Besides caching, it's probably faster to download a third party library from Google's CDN than the central server hosting the application anyway.

If a new version of a popular JS library is released with a big performance boost, is tested with the application and then implemented:

• If all JS is compiled into one file then every user will have to re-download this file even though the application code hasn't changed.
• If third party scripts are loaded from CDNs then the user only has download the new version from the CDN (or from cache somewhere).

Are any of the following legitimate worries in a situation like the one described?

• Some users (or browsers) can only have a certain number of connections to one hostname at once so retrieving some scripts from a third party CDN would be result in overall faster loading times.
• Some users may be using the application in a restricted environment, therefore the domain of the application may be white-listed but not the CDNs's domains. (If it's possible this is realistic concern, is it at all possible to try to load from the CDN and load from the central server on failure?)

Answer 1

Compiling all application-specific/local JS code into one file

Since some of our key goals are to reduce the number of HTTP requests and minimize request overhead, this is a very widely adopted best practice.

The main case where we might consider not doing this is in situations where there is a high chance of frequent cache invalidation, i.e. when we make changes to our code. There will always be tradeoffs here: serving a single file is very likely to increase the rate of cache invalidation, while serving many separate files will probably cause a slower start for users with an empty cache.

For this reason, inlining the occasional bit of page-specific JavaScript isn't as evil as some say. In general though, concatenating and minifying your JS into one file is a great first step.

using CDNs like Google's for popular libraries, etc.

If we're talking about libraries where the code we're using is fairly immutable, i.e. unlikely to be subject to cache invalidation, I might be slightly more in favour of saving HTTP requests by wrapping them into your monolithic local JS file. This would be particularly true for a large code base heavily based on, for example, a particular jQuery version. In cases like this bumping the library version is almost certain to involve significant changes to your client app code too, negating the advantage of keeping them separate.

Still, mixing request domains is an important win, since we don't want to be throttled excessively by the maximum connections per domain cap. Of course, a subdomain can serve just as well for this, but Google's domain has the advantage of being cookieless, and is probably already in the client's DNS cache.

but loading all of these via headjs in parallel seems optimal

While there are advantages to the emerging host of JavaScript "loaders", we should keep in mind that using them does negatively impact page start, since the browser needs to go and fetch our loader before the loader can request the rest of our assets. Put another way, for a user with an empty cache a full round-trip to the server is required before any real loading can begin. Again, a "compile" step can come to the rescue - see require.js for a great hybrid implementation.

The best way of ensuring that your scripts do not block UI painting remains to place them at the end of your HTML. If you'd rather place them elsewhere, the async or defer attributes now offer you that flexibility. All modern browsers request assets in parallel, so unless you need to support particular flavours of legacy client this shouldn't be a major consideration. The Browserscope network table is a great reference for this kind of thing. IE8 is predictably the main offender, still blocking image and iFrame requests until scripts are loaded. Even back at 3.6 Firefox was fully parallelising everything but iFrames.

Some users may be using the application in a restricted environment, therefore the domain of the application may be white-listed but not the CDNs's domains. (If it's possible this is realistic concern, is it at all possible to try to load from the CDN and load from the central server on failure?)

Working out if the client machine can access a remote host is always going to incur serious performance penalties, since we have to wait for it to fail to connect before we can load our reserve copy. I would be much more inclined to host these assets locally.

Answer 2

1. Many small js files is better than few large ones for many reasons including changes/dependencies/requirements.
2. JavaScript/css/html and any other static content is handled very efficiently by any of the current web servers (Apache/IIS and many others), most of the time one web server is more than capable of serving 100s and 1000s requests/second and in any case this static content is likely to be cached somewhere between the client and your server(s).
3. Using any external (not controlled by you) repositories for the code that you want to use in production environment is a NO-NO (for me and many others), you don't want a sudden, catastrophic and irrecoverable failure of your whole site JavaScript functionality just because somebody somewhere pressed commit without thinking or checking.