I use DO's load balancers in a couple of projects, and they don't list Cloudflare as an upstream dependency anywhere that I've seen. It's so frustrating to think you're clear of a service then find out that you're actually in their blast radius too through no fault of your own.
I find stuff like this all the time, railway.com recently launched an object storage service, but it's simply a wrapper for wasabi buckets under the hood, and they don't mention this anywhere... not even the subprocessors page https://railway.com/legal/subprocessors - customers have no idea they are using wasabi storage buckets unless they dig around the dns records. so i have to do all this research to find upstream dependencies and go subscribe to status.wasabi.com alerts etc.
Hey, I'm the person that was responsible for adding object storage to Railway. It was my onboarding project, basically a project I was able to choose myself and implemented in 3 weeks in my 3rd month after joining Railway.
Object Storage is currently in Priority Boarding, our beta program. We can and will definitely do better, document it and add it to the subprocessor list. I'm really sorry about the current lack of it. There was another important project that I had to do between the beta release of buckets and now. I'm oncall this week, but will continue to bring Buckets to GA next week.
So, just to give this context. There's no intentional malevolence or shadiness going on, it's simply because there's 1 engineer (me) working on it, and there's a lot of stuff to prioritize and do.
It's also super important to get user feedback as early as possible. That's why it's a beta release right now, and the beta release is a bit "rushed". The earlier I can get user feedback, the better the GA version will be.
On the "simply a wrapper for wasabi buckets" - yes, we're currently using wasabi under the hood. I can't add physical Object Storage within 3 weeks to all our server locations :D But that's something we'll work towards. I wouldn't say it's "simply" a wrapper, because we're adding substantial value when you use Buckets on Railway: automatic bucket creation for new environments, variable references, credentials as automatic variables, included in your usage limits and alerts, and so on.
slight off topic: I used DO LBs for a little while but found myself moving away from that toward a small droplet with haproxy or nginx setup. Worked much better for me personally!
The point of an LB for these projects is to get away from a single point of failure, and I find configuring HA and setting up the networking and everything to be a pain point.
These are all low-traffic projects so it's more cost effective to just throw on the smallest LB than spend the time setting it up myself.
I use the LB's for high availability rather than needing load balancing. The LB + 2 web back-ends + Managed DB means a project is resilient to a single server failing, for relatively low devops effort and around $75/mo.
I have a couple of instances of this same pattern for various things that have been running for 5+ years, none of them have suffered downtime caused by the infrastructure. I use ansible scripts for the web servers, and the DO API or dashboard to provision the Load Balancer and Database. You can get it all hooked up in a half hour, and it really doesn't take any maintenance other than setting up good practices for rotating the web servers out for updates.
They wouldn't survive DO losing a DC, they're not so mission critical that it's worth the extra complexity to do that, and I don't recall DO losing a DC in the past 10 years or so.
They did stay up during this outage, which was apparently mostly concentrated on a different product called the 'global load balancer', which ironically is exactly the extra complexity I mentioned to in theory survive a DC outage.
Keep in mind these are "important" in the sense that they justify $100/mo on infra and monitoring, but not "life critical" in that an outage is gonna kill somebody or cost millions of bucks an hour. Once your traffic gets past a certain threshold, DO's costs don't scale that well and you're better off on a large distributed self-managed setup on Hetzner or buying into a stack like AWS.
To me their LB and DB products hit a real sweet spot -- better reliability than one box, and meaningfully less work than setting up a cluster with floating IP and heartbeats and all that for a very minimal price difference.
You can't just "turn off CDN" on the modern internet. You'd instantly DDOS your customers' origins. They're not provisioned to handle it, and even if they were the size of the pipe going to them isn't. The modern internet is built around the expectation that everything is distributed via CDN. Some more "traditional" websites would probably be fine.
If you are DO you could, you just decided not to bother. They control the origins it's spaces (s3), so they could absolutely spin up further gateways or a cache layer and then turn the CDN off.
You outsource it because clouflare have more locations than you so offer lower latency and can offer it at a cost that's cheaper or the same price as doing it yourself.
To the contrary, CDN pricing will usually beat cloud provider egress fees.
Common example: you can absolutely serve static content from an S3 bucket worldwide without using a CDN. It will usually scale OK under load. However, you're going to pay more for egress and give your customers a worse experience. Capacity isn't the problem you're engineering around.
For a site serving content at scale, a CDN is purpose-built to get content around the world efficiently. This is usually cheaper and faster than trying to do it yourself.
They will be doing a mix of peering both across free PNIs and very low cost IXP ports, with the reminder going down transit like Colt or cogent. Probably average cost of the order of about $1 per 20TB of egress in Europe and NA markets.
The thing is with edge capacity is that you massively overbuild on the basis that;
It's generally a long ISH lead time to scale capacity (days not minutes or hours).
Transit contracts are usually 12-60 months
Traffic is variable, not all ports cover all destinations
Redundancy
So if you are doing say 100Gbps 95%ile out of say London then you will probably have at least 6+ 100Gb ports, so you do have quite a bit of latent capacity if you just need it for a few hours.
Yes all sites showing the CloudFlare error due to the massive outage. Seems their outages are getting more frequent and taking down the internet in new ways each time.
Are you using their "reserved IPs"? I was thinking of starting to use them, but now I wonder if it is part of their load balancing stack under the hood.
