This is one of those things you simply don't expect to test when you're benchmarking performance. I've been trying to cut away chunks of 20-50ms on a side project I've been working on hosted on a .io domain, and I'm seriously considering switching because of this. Great article for sure. I also wonder at what point the trade off is between a vanity domain vs a more performant domain. For a CDN this makes perfect sense, but would it be better to launch with a less-notable url vs a more performant one?
I'd say the inverse is more likely. If you're going to fire a single request to a domain only you are using and you're running a full local resolver, it may make a difference.
For a public CDN: your browser already has the file cached. If it doesn't then it has the domain cached. If it doesn't then the dhcp-provided resolver has it. If it doesn't, then at least it already has the TLD nameserver available immediately, and the TLD can serve that response from very hot cache. It's CDNs job to make sure this happens.
With a vanity domain, you can fully control the TTL values. All my sites use a vanity domain because it doesn't tie me to a particular CDN, and they have 86400 TTL.
When you have a vanity domain like cdn.example.com, the recursive resolver already knows the nameservers for example.com, so this actually reduces the additional DNS lookups.
On the first request. All subsequent requests will be cached.
Oh, and if the requestor has a large DNS cache upstream, it's already done.
Oh, and if the browser used a pre-fetch, that's already done.
Oh, and if you have already invested your branding effort across 100 sites, maybe you don't want to re-do all that?
Oh, and if you need to cut 50ms from your first time page load, have you considered dropping all the trackers and analysis JS loads? Can you deliver your first page without any JS at all? Can you do it without a database lookup?
Those are all things you should do before killing anyone.
If you haven't done already, take a look at don't delivery improvements. Serve don't files locally, subset them, and use variable fonts and woff2. I'm a micro performance enthusiast myself and it's my #1 optimization with the biggest gain.
I'd like to point out that these are empty platitudes. Imagine going to your boss to justify not moving tlds and saying, "well sdan on hackernews said it was fine!"
