Eleven years of running a free DNS service

This is part eight, and the last part, of the Bancuh DNS story. Part one starts with a Raspberry Pi in 2015.

This week I brought every Bancuh DNS server up to date, one at a time.

Some of them needed it badly. The two servers in Singapore were still on Ubuntu 16.04, and had been running since November 2017. They’re the same two servers I brought back after the user in Vietnam wouldn’t let me turn the service off. The servers in Tokyo and Dallas were on a version of Alpine Linux that had been out of support for almost a year. And on one of the Tokyo servers, I found that Docker had never been set to start on boot. I’d started it by hand when I built the server, about a thousand days earlier, and it had simply never been restarted since. If that server had ever rebooted, it would have come back without its DNS.

Two days and a lot of care later, all seven are on current systems. And the thing that struck me most wasn’t what was broken. It was how little had gone wrong while I wasn’t looking. Despite everything in the last seven posts, Bancuh DNS mostly just runs.

It’s been eleven years since the Raspberry Pi. This is a look back at the whole thing.

Where it is now

Bancuh DNS today is seven small servers in four places: two in Singapore, two in Tokyo, two in Paris and one in Dallas. Each one blocks about seven million domains, answers plain DNS, DNS-over-TLS and DNS-over-HTTPS, and resolves everything itself. Most of them have 1 GB of memory, and use less than half of it.

It costs me about USD 56 a month: USD 13 for Singapore, USD 16 for Tokyo and Dallas, about USD 26 for Paris, and the domain. That’s a bit over USD 650 a year. A third of the Paris bill is just the two IPv4 addresses.

I don’t know how many people use it. There are no accounts, and I’ve never tried to count. My guess is not many. But some do, and a few of them have been here for years.

Most months, I don’t touch it. The servers run well on their own, with very little attention. The work comes in bursts, when I change something properly: a rewrite, a new front end, or a week like this one.

What it gave me

What do I get out of it? In money or recognition, nothing. I keep it running because people depend on it.

That’s not the whole truth, though. Bancuh DNS is where I learned Rust, which became the language I use for almost everything I build. It taught me how DNS actually works, from recursion to certificates. It taught me most of what I know about running something for other people. And it’s on my own phone: I use its DNS-over-TLS through Android’s Private DNS setting every day.

What the eleven years taught me

Looking back over the series, a few things come up again and again.

Users kept it alive, and shaped it. The service exists because a stranger tracked me down on Facebook, and most of what it does started as someone asking for it.

Most of my outages were my own updates. The code that changes a service is riskier than the code that runs it.

Privacy isn’t something you get once. I had it, lost it in a rewrite, and needed users with a leak test to make me earn it back.

Some things can’t be taken back. Once people have typed a server’s name into their phones, that name has to keep working.

And small servers shaped the engineering. Many of the big changes, from capping BIND’s cache and adding swap to replacing BIND altogether, came from trying to fit the service into less memory and less money.

The promise I didn’t keep

In 2015, when I first offered the service to strangers, I wrote:

I won’t block download sites, torrents, or do any kind of censorship. For the moment, I am sticking to adservers only.

Bancuh DNS today is a strict family filter. It blocks ads and trackers, but also adult sites, gambling, VPNs, other public DNS services, and since December 2024, torrent sites and trackers. I promised in the first post to explain how that happened.

The honest answer is trust. I was focused on the engineering, and I had neither the time nor the knowledge to build a good blocklist. Tomatoide did. Tomatoide started by opening issues and pull requests to tune the lists, and at some point I simply gave them push access. Today, Tomatoide has made more changes to Bancuh’s blocklist than I have. The list drifted from “ads only” towards a family filter one reasonable request at a time, and I let it, because I trusted the person doing it.

It did mean breaking a promise I’d made in public, even if I’d only made it “for the moment”. But trusting Tomatoide, and the many others who opened issues to report a broken site or suggest a list, made Bancuh far better than anything I’d have built on my own. And the filtering page says plainly what it blocks, so nobody has to take my old promise on trust.

Saying no

Over the years I’ve said no to some things users clearly wanted. A lighter server that only blocked ads (#211). Unblocking Google Ads and Analytics because they broke some sites (#214).

The reasons were rarely about the idea itself. Some features aren’t worth the cost when there’s a good workaround: if Bancuh is too strict for one device, you can point that device somewhere else. Some just can’t be engineered well with what I have. And some I couldn’t maintain. A second, lighter list would need someone to curate it the way Tomatoide curates this one, and I don’t know how to do that. In #211 I said I’d revisit it if I ever became significantly richer. That hasn’t happened yet.

For #214, the reason was simpler: people chose Bancuh expecting it to block ads and analytics, and I didn’t want to change that under them.

The last year

Most of the recent posts end with me working on something with Claude Code: the leak test bug, the review of the front end, and this week’s updates across all seven servers. That’s not a coincidence.

For most of its life, Bancuh DNS was bigger than I could properly look after in the time I had. I kept it running, but I rarely looked underneath. Working with Claude Code changed that. I can make bigger changes, more safely and more smoothly, keep everything up to date, and actually fix the things I used to work around. It’s what makes the service feel sustainable to me now. I love what it lets me do.

What happens next

I’ll keep it running for as long as I can. I can afford it now. If money ever gets tight, I might have to ask for help. Two people have already suggested a “buy me a coffee” link for simplesolat, my prayer times app, and the same idea would apply here. Asking for money for a free service still feels wrong to me, but maybe they have a point. I haven’t decided.

There’s also a known problem left to fix: after every restart, a server answers for about a minute without its blocklist while it compiles the first one (#220). There always seems to be one more.

A note to 2015

If I could send one message back to the version of me with a Raspberry Pi and a borrowed ad list, it would be short.

Plan the architecture before strangers depend on it. Think about how much memory and CPU each piece will need when the list is a hundred times bigger, because it will be. BIND was never meant to hold millions of blocked domains, and I paid for that for years in bigger servers and swap. Maybe something lighter, like dnsmasq, would have been cheaper. I’ll never know.

And good luck. You’ll need it.