Where Docker gets subnets from
172.17.0.0/16 default bridge network, docker0
172.18.0.0/16 … 172.31.0.0/16 14 more networks
192.168.0.0/20 16th network: 192.168.0.x … 192.168.15.x
192.168.16.0/20 … and onward in chunks of 16 /24 subnets
Every compose stack creates at least one network, <project>_default, and a stack with a database usually
has two. Together with the shared proxy, the fifteenth user-defined network shows up as early as the seventh
or eighth stack with a database, and it gets 192.168.0.0/20 — which covers the office network
192.168.10.0/24. Later chunks reach the router’s guest network 192.168.89.0/24 too
(Guest network on MikroTik). For the server, that subnet is now local: a reply to an address in it goes into the
br-… bridge instead of to the gateway, and the client sees a hung connection.
How to set your own pool
Add the key to /etc/docker/daemon.json — next to the log limit from Docker on Ubuntu, and don’t forget
the comma after the previous key:
"default-address-pools": [
{ "base": "172.20.0.0/14", "size": 24 }
]
172.20.0.0/14 is 172.20.0.0–172.23.255.255: homes and offices use 172.16.0.0/12 less often than
192.168 and 10. size: 24 gives each network a /24, 254 addresses: plenty for a stack, and the pool holds
1024 networks. If a network on your path to the server is in 172.20–172.23, take another piece of
172.16.0.0/12.
sudo dockerd --validate --config-file /etc/docker/daemon.json
# configuration OK
sudo systemctl restart docker
systemctl restart docker restarts every container — do it at a quiet time.
The pool applies only to new networks: existing ones keep their subnet. Recreate a stack’s network with
docker compose down && docker compose up -d in its directory. Docker removes a shared network like proxy
only when no containers are left on it: stop the stacks on it, then
docker network rm proxy && docker network create proxy, and bring the stacks back up.
If it doesn’t work
Another stack came up, and SSH to the server from the office or a neighboring home network hangs
- Cause
- The new Docker network got a subnet containing the client’s address: the server’s reply went into the Docker bridge instead of to the gateway.
- Fix
- Run
ip routeon the server and look for a line like192.168.…/20 dev br-…. Setdefault-address-poolsand recreate that network.
A device on the home network doesn’t open from a container, but opens from the host
- Cause
- The container’s network took the same subnet: to the container that address is a neighbor on the bridge, and the packet never reaches the router.
- Fix
- The same pool in
daemon.json, then recreate the container’s network.
could not find an available, non-overlapping IPv4 address pool among the defaults to assign to the network
- Cause
- The default pools ran out: they hold 31 subnets, and the server already has that many networks.
- Fix
default-address-poolswithsize: 24— the pool then holds over a thousand networks.docker network pruneremoves unused networks, butproxygoes with them if every stack on it is stopped right now.
How to check
The subnets of every network on the server:
docker network ls -q --filter driver=bridge | xargs docker network inspect --format '{{.Name}} {{range .IPAM.Config}}{{.Subnet}}{{end}}'
# bridge 172.17.0.0/16
# proxy 172.20.0.0/24
# umami_internal 172.20.1.0/24
A new network gets a /24 from your pool:
docker network create pooltest
docker network inspect pooltest --format '{{range .IPAM.Config}}{{.Subnet}}{{end}}'
docker network rm pooltest
From the machine that lost the server — ssh web1 again.