Restore access to Docker containers over the tailnet after a Tailscale update
The remedy is well sourced. I pulled the cited page (tailscale.com/docs/reference/messages/client/docker-stateful-filtering) and it carries the warning string verbatim and lists tailscale set --stateful-filtering=false, tailscale up --netfilter-mode=off and dockerd --iptables=false as the fixes, so the commands are real and current. The Omarchy framing checks out against the quattro tree and this machine: docker, docker-compose and ufw-docker are all in omarchy-base.packages, install/config/enable-services.sh runs systemctl enable docker.socket, and install/config/firewall.sh does ufw default deny incoming and installs ufw-docker's after.rules block. THREE DEFECTS. (1) THE CAUSE IS STALE. It says "Nothing on your side changed; the filtering default did." Stateful filtering was on by default only in Tailscale 1.66.0 to 1.66.3. The changelog entry for v1.66.4 (2024-05-20) reads "Linux: Stateful filtering is now off by default" precisely because it broke container DNS, and it was never re-enabled: ipn/prefs.go in tailscale main still has NoStatefulFiltering: opt.NewBool(true) in the defaults with the comment "The default is to not apply stateful filtering." So on any client from mid-2024 onward this is not something an upgrade turns on. It is something the operator turned on. (2) The docker.service override drops the --containerd=/run/containerd/containerd.sock argument that Arch's shipped unit passes, so dockerd stops using the system containerd.service it still Wants/Afters and spawns its own. That is a second, avoidable breakage bolted onto an already-advanced step. (3) The verify reads tailscale debug prefs | grep -i statefulfilter "shows it disabled", but the pref is the inverted NoStatefulFiltering, so the correct state prints true. As written a reader is likely to read the output backwards. Also worth flagging for the record's provenance: the third source URL, github.com/basecamp/omarchy/blob/master/install/config/firewall.sh, 404s. master is the Omarchy 3 tree. The file exists only on quattro.
Disabling stateful filtering means the machine will accept unsolicited inbound traffic from other tailnet nodes to whatever it forwards, rather than only replies to connections it initiated. That is fine on a tailnet you control with ACLs in place, and a meaningful loosening on a shared tailnet. Review your ACLs before doing it. Editing docker.service to pass --iptables=false will break container networking and any ufw-docker rules if the rest of the ruleset is not written by hand. Do not do that as a first move.
Symptom
Containers that were reachable from other devices on the tailnet stop answering after a Tailscale upgrade. A self-hosted service on 100.x.y.z:8080 times out from the phone even though the container is up and reachable from the host itself. Containers may also stop resolving DNS. The Tailscale client surfaces the message: Stateful filtering is enabled and Docker was detected; this may prevent Docker containers on this host from resolving DNS and connecting to Tailscale nodes.
Cause
Tailscale's stateful filtering only lets traffic through that is part of a connection it already tracked. Docker's own iptables rules move packets between the bridge network and the host in a way Tailscale does not see as part of an established flow, so return traffic to containers is dropped and container DNS can fail. This is not something a Tailscale upgrade does to you any more: it was on by default only in 1.66.0 through 1.66.3, and v1.66.4 (2024-05-20) turned it back off specifically because it broke containers. The current default in ipn/prefs.go is still NoStatefulFiltering: true, i.e. filtering off. So on a modern client you are seeing this because the node was explicitly brought up with --stateful-filtering=true, or because a pref set during that 2024 window has persisted in the node's state ever since. On Omarchy the warning fires readily because docker, docker-compose and ufw-docker are all in the base package set and docker.socket is enabled at install, so a bridge network is there for Tailscale to detect.
Fix
Check whether the warning is present and what state you are actually in:
tailscale status
sudo tailscale debug prefs | grep -i statefulfilter
docker network ls
Read that pref carefully. It is inverted. The field is NoStatefulFiltering, so "NoStatefulFiltering": true means stateful filtering is **off** (the default), and false means it is **on** and is what you are hitting.
Turn stateful filtering off:
sudo tailscale set --stateful-filtering=false
That is persistent. It does not need re-applying after a reboot or a tailscale down/up cycle. Then confirm from another tailnet device:
# from your phone or another machine
curl -v http://100.x.y.z:8080/
If you would rather Docker not manage netfilter at all (advanced, and you then own every rule), keep the rest of Arch's shipped ExecStart intact and only add the flag, or dockerd will also stop using the system containerd.service and start its own:
sudo systemctl edit docker.service
[Service]
ExecStart=
ExecStart=/usr/bin/dockerd -H fd:// --containerd=/run/containerd/containerd.sock --iptables=false
Or the mirror image: Tailscale stops writing rules and you manage them:
sudo tailscale up --netfilter-mode=off
On Omarchy, remember ufw-docker has already installed its own block in /etc/ufw/after.rules. If you disable Docker's iptables management you break those protections too, so prefer the single --stateful-filtering=false change unless you have a specific reason not to.
Verify: sudo tailscale debug prefs | grep -i statefulfilter prints "NoStatefulFiltering": true, and the client no longer emits the Docker warning. From another tailnet device, curl http://<tailscale-ip>:<port>/ against a containerised service returns a response. Inside a container, getent hosts archlinux.org resolves.
Verify
sudo tailscale debug prefs | grep -i statefulfilter shows it disabled, and the Tailscale client no longer emits the Docker warning. From another tailnet device, curl http://<tailscale-ip>:<port>/ against a containerised service returns a response. Inside a container, getent hosts archlinux.org resolves.