Cloudflare is one control plane, but it's 1) anycast 2) over 100 locations 3) multiple servers per site. The SPOF with Cloudflare is either your account getting hacked, or (much less likely) a CF control plane/configuration outage.

Someone could probably build a better DNS provider than Cloudflare, at least for their specific site, but it would be difficult. I don't know of any other DNS provider I'd consider more robust, particularly against high load, DoS, etc.

(Disclaimer: worked for Cloudflare 2014-2016)

This is rather weird claim. Getting your account hacked is the least of your worries, when every single thing is maintained by a single organization, a single team, on a single autonomous system, with lots of reliance on single hardware and software components. This is basically three nines level stuff, with hours of downtime per year on average. It's relatively easy to do better than that with do-it-yourself CDN using DNS-routed multiple independent hosting providers for edge nodes, getting to four, five nines level. (and DNS separately is not worth talking about, because it has built in redundancy, anything will always be more reliable than a centralized DNS provider)

> with lots of reliance on single hardware and software components

Are your saying you have inside knowledge and that's the architecture? (I'm guessing no)

I've never seen any company use completely different software or hardware for each redundant instance of a service, so it's a pretty good bet CloudFlare isn't doing this either.

That would imply a SPOF. One component of any kind that can take down the whole thing. The only such “component” I can think of is a human, not a technical one.

I meant single kind of component, not one thing handling everything.

If you use juniper routers, for example, a juniper specific software, hardware problem or a mistaken assumption about juniper specifics could take down all routers and an entire network.

Without saying too much, cf has a fairly diverse network in terms of hardware. Partly because the sprawl makes it hard to keep every single colocation running the same hardware, when a single colo can consist of hundreds of independent nodes (metals).

Usually any kind of such issue wouldn’t take out every component of a type unless user error is involved. Like deploying the wrong policy to all devices. Which again comes back to humans being the weakest link. I can’t really imagine a particular hardware problem that would kill all components at once. This being said I probably didn’t experience all possible problems so far :).

Sure, humans are the weakest link, who cares? Regardless of the cause of the failure, catastrophic cascades are more likely if the systems and operations involved are homogeneous. That's why evolution prefers diversity.

Natural evolution. Because it follows vastly different rules. I still fail to see any example given for how one “cascade failure” takes down all the routers in an infrastructure like Cloudflare’s. I can think of a dozen ways a non-homogeneous system can fail (like incompatible updates and config, confusions between the types of components, etc.). In the end you usually end up with massive expenditure that introduces just as many problems as it solves.

You may be onto something that nobody else spotted but so far the arguments are too generic (“failure”) or incompatible with the premise (rules of natural evolution applied to very artificial systems).

Don’t get me wrong, I’m happy to learn something new and understand a different point of view but so far the arguments don’t help me.

I actually used that example because they had this exact problem in the past, here's the link: https://blog.cloudflare.com/todays-outage-post-mortem-82515/

And Google had a similar issue recently where they didn't test their firmware upgrade auto rollback, so a boned firmware upgrade bricked a region until they could manually recover it. If they had rolled out to all regions at once, and all the devices were the same, it would have been chaos.

Often the only thing preventing mass upgrades from becoming mass outages is some part of the network just happens to work differently.

Well... sure. Just as long as you keep in mind that a "not really tested [something]" will very easily get replicated on multiple vendor devices. It's not like they did it because they hate Vendor1 but will make sure to test all the corner cases with Vendor2. Some scenarios will by design ensure you can't have all of them fail because they are not susceptible to the same failure modes. But to compensate you introduce additional failure modes (especially human error) by using a heterogeneous environment. Not to mention the expense.

I was actually hoping for an interesting human independent failure. A cosmic radiation caused bit flip that would occur on a particular device. :) Human error can overcome any setup.

> bigger redundancy in NS servers (4 instead of 2 at Cloudflare)

I don't think this is really all that important, as typically all these servers are running anycast. If you really wanted redundancy, you would have NS records from at least two different providers.

That's a good point. A backup DNS on another service is on my bucket list once sites get a little bigger and more relevant (right now it's not worth it).

You're paying for a service that's less robust because they have more IP addresses serving their database?

For the time beeing I'm not sure which is more or less robust because I have not seen any DNS outages in both.