In-depth articles on technology shaping what comes next.

DNS Blocking Orders Are Forcing Resolvers Out of Countries

Quad9 pulled its DNS resolver out of France after a court ordered blocking. Here are the mistakes that make DNS infrastructure fragile — and the fixes.

A still canary in a cage hanging in a dark coal mine tunnel, lit by a distant lamp.
Quad9's exit from France is the canary: the air in DNS-land is getting toxic.

A hospital in Lyon doesn't think about its DNS resolver until the day name lookups just stop working. That day is coming for a lot of French organizations: Quad9, the Swiss non-profit behind 9.9.9.9, is pulling its free public resolver out of France after a court imposed daily financial penalties for refusing to block domains tied to alleged copyright infringement. Government offices, schools, utilities — all of them pointed at an IP address that will soon answer with silence, and Quad9 has no way to warn them.

This isn't a story about piracy. It's a story about what happens when courts treat the Internet's address book as a content enforcement tool — and about how fragile our DNS architecture turns out to be when someone leans on it. I've spent years running production DNS, and the failure modes here are embarrassingly familiar. Let's walk through the mistakes, because most teams are making at least two of them right now.

Mistake 1: Treating Legal Jurisdiction as Someone Else's Problem

When engineers design infrastructure, we obsess over redundancy, latency, and failure domains. Almost nobody puts "legal failure domain" on the whiteboard. That's a mistake, and it's the one that just bit Quad9.

Quad9 is headquartered in Switzerland specifically because Swiss law offers strong due-process protections. That was deliberate jurisdiction selection — the legal equivalent of picking a region with good peering. But it wasn't enough. French rightsholders sued in French courts over service delivered to French users, won, and the court attached daily penalties to non-compliance. Quad9 faced a choice: implement blocking that would compromise the integrity of its resolver, or leave. It left.

And it's not alone. Cisco previously withdrew OpenDNS from France, Portugal, and (temporarily) Belgium under similar pressure. A Belgian court has since ordered Cloudflare, Google, and Cisco to start blocking through their public resolvers by early January 2027 or face substantial penalties. The pattern is clear: rightsholders have discovered that public resolvers are a soft chokepoint, and they're working through jurisdictions one at a time.

The fix: put jurisdiction on your architecture diagram. If you operate a resolver, your legal team's threat model matters as much as your SRE team's. If you consume a resolver, ask the uncomfortable question in your vendor review: what's your supplier's exposure to blocking orders in the countries where you have users? "We'd rather exit than filter" is a legitimate answer, but you need to know it's coming.

Mistake 2: Assuming Anycast Makes You Jurisdiction-Proof

There's a lazy assumption floating around that anycast — announcing the same IP from dozens or hundreds of points of presence worldwide — makes a service geographically untouchable. The routing table doesn't care about borders, right?

Wrong. Anycast doesn't reduce your legal surface area; it multiplies it. Every PoP is a presence, and every presence is an argument that you "operate in" that country. When a French court wants leverage over Quad9, the fact that French users get answered by infrastructure reachable in France is the whole case. The technical property we built for performance — users automatically hit the nearest node — is precisely what makes each jurisdiction's demands locally enforceable.

This inverts the usual SRE instinct. We think of more PoPs as more resilience. Under legal attack, more PoPs means more courts with a plausible claim on you. The withdrawal mechanics are grimly interesting, too: to leave France, an anycast operator has to stop answering French queries — which it can't do surgically by user, only by routing and topology. The crude version is withdrawing announcements or null-routing from French networks, and the blast radius of that is exactly the hospitals-and-utilities problem above.

A stylized network map with one glowing European node entangled in dark chains.
Every point of presence is also a point of legal exposure.

The fix: if you run global anycast services, model legal withdrawal as a routing event. Know, per country, exactly what "stop serving here" looks like in your BGP policy, how long it takes to propagate, and what collateral damage it causes. Test it the way you'd test a region failover — because functionally, that's what it is.

Mistake 3: Hard-Coding One Resolver and Calling It Redundant

Here's the part that should embarrass us as an industry. Quad9's French users include hospitals and utilities, and when the service goes dark, their Internet will "fail with no explanation." Not because the Internet is down. Because someone typed 9.9.9.9 into a DHCP scope or a router config three years ago and nobody has thought about it since.

A single hard-coded resolver is a single point of failure with extra steps. Worse, the failure mode is treacherous: resolution doesn't error cleanly, it hangs and retries. Applications stall with timeouts that look like everything except DNS. If you've ever chased a "the API is slow" alert that turned out to be five-second resolution delays on every outbound call, you know this pain. We've written before about how DNS configuration keeps breaking production; this is that same fragility, now triggered by a court order instead of a bad config push.

The failure modes to design against, in order of how often I see them in the wild:

  1. One resolver, no fallback. Resolver dies, everything hangs. This is the French hospital scenario.
  2. Two resolvers from the same operator. 9.9.9.9 and 149.112.112.112 feel redundant. They're one organization, one legal entity, one court order.
  3. Fallback configured but never tested. systemd-resolved's FallbackDNS only kicks in under specific conditions. Have you ever actually pulled the primary and watched what happens?
  4. Silent semantic changes. Your resolver starts returning NXDOMAIN or a block page for domains you legitimately need. Nothing is 'down,' but your integration with a vendor whose domain got swept into a blocking order is.

That last one deserves emphasis. When the threat is censorship rather than outage, monitoring for availability isn't enough — you need to monitor for correctness. A resolver that answers every query but lies about some domains passes every ping check you have.

Mistake 4: Believing DNS Blocking Actually Blocks Anything

Now the policy-level mistake, which engineers keep failing to explain to lawmakers: DNS blocking doesn't work, and the collateral damage is enormous.

A resolver-level block is defeated by changing one setting. Switch to a different resolver, use DNS-over-HTTPS to a provider outside the jurisdiction, run a VPN, or — the nuclear option that's trivially easy — look up the IP once and hard-code it. Anyone motivated enough to pirate content will route around a DNS block in under a minute. Meanwhile, the blocking infrastructure catches bystanders: domains get added to orders with sloppy scoping, shared hosting means one domain's block can poison thousands of sites, and there is essentially no mechanism for a wrongfully blocked party to get fast relief.

There's a darker architectural consequence, too. When trustworthy neutral resolvers get squeezed out of a country, users don't stop resolving — they migrate to whatever's left. That means sketchy free resolvers with unknown logging, unknown malware filtering, and unknown data practices. The policy intended to make the Internet marginally safer for rightsholders makes it materially less safe for everyone else. We've seen this movie before with safety laws that become censorship tools: the filtering layer built for one purpose always gets repurposed.

A DNS block is a locked screen door: it stops nobody who wants in, but it tells everyone the neighborhood is getting worse.

The fix: if you're in a position to influence policy — and if you run infrastructure, you are, whether you like it or not — make the technical case early and in writing. Enforcement at the hosting and registrar layer targets the content. Enforcement at the resolver layer targets the phone book, and torching the phone book doesn't close the speakeasy.

Mistake 5: No Exit Plan for Your DNS Dependencies

You have a runbook for your cloud region going down. Do you have one for your resolver going dark — or going compliant-with-a-court-order? Almost nobody does. Here's the minimum I now consider table stakes.

First, inventory where resolver addresses actually live in your estate. It's worse than you think: DHCP scopes, router configs, Kubernetes dnsConfig blocks, container base images, IoT firmware, that one printer. Second, deploy diversity that means something — at least two resolvers from different organizations, incorporated in different jurisdictions. Third, verify resolution content, not just resolution success. A small cron-driven check catches both outage and censorship:

#!/usr/bin/env bash
# Canary-check your resolvers for both availability AND honesty.
# Alert if a resolver dies OR starts lying about known-good domains.
RESOLVERS=("9.9.9.9" "1.1.1.1" "8.8.8.8")
CANARIES=("example.com" "cloudflare.com" "github.com")
for r in "${RESOLVERS[@]}"; do
for d in "${CANARIES[@]}"; do
answer=$(dig +short +time=2 +tries=1 "@$r" "$d" A 2>/dev/null)
if [[ -z "$answer" ]]; then
echo "ALERT: resolver $r returned nothing for $d"
# page someone — this is either an outage or a block
fi
done
done

Fourth — and this is the one that actually survives a splinternet — give yourself the option of not depending on anyone's public resolver at all. A local recursive resolver like Unbound talks directly to the authoritative nameservers. There's no middleman to sue:

server:
verbosity: 1
do-tcp: yes
do-udp: yes
# Resolve from the root down; no upstream forwarder to be court-ordered
root-hints: "/etc/unbound/root.hints"
qname-minimisation: yes
harden-dnssec-stripped: yes
cache-min-ttl: 60
# Serve the local network
interface: 10.0.0.2
access-control: 10.0.0.0/8 allow

Self-hosting recursion has real costs — you lose Quad9's excellent threat-intel filtering, you own the cache-poisoning and amplification hardening, and cold-cache latency is worse. But for critical infrastructure, the trade is increasingly worth it. The threat model has shifted from "what if the resolver is attacked" to "what if the resolver is legislated."

Many small lighthouses shining independently while one large central lighthouse goes dark.
Local recursive resolvers are lighthouses you own: dimmer, but nobody can switch them off remotely.

The Splinternet Is a Routing Problem Too

Step back from the individual mistakes and the shape of the thing comes into focus. Quad9's CTO described a "general dread" among the DNS community at IETF, and he's right to feel it. Each country that imposes resolver blocking with local penalties creates a fork in what "the Internet" means depending on where you resolve from. France today, Belgium with a deadline in early 2027, and the template now exists for any rightsholder in any court.

One counterargument deserves an honest answer: some engineers argue DNS blocking is good precisely because it's easy to bypass — it steers the casual majority while the motivated minority routes around it, and everyone gets what they want. I don't buy it. That logic builds a two-tier Internet where the regulated layer is a lie and the working layer is unregulated by accident. Worse, it normalizes the enforcement machinery; the orders get broader, the penalties steeper, and one day the "easy bypass" resolvers are all gone from your jurisdiction too. Quad9 leaving France is what that slope looks like from the top.

So treat this as the drill it is. This week: find every place 9.9.9.9, 1.1.1.1, or 8.8.8.8 is hard-coded in your infrastructure. Add a second resolver from a different legal entity. Stand up a canary-domain check that catches silent blocking, not just silence. And put one line in your next architecture review template: what happens to this system when a court, not an outage, takes a dependency away?

The hospitals in France are going to learn the answer the hard way. You don't have to.