A few machines, one set of services.
Swarm is built into every Docker Engine. You no longer say "start this container there", but "I want three of this service", and the swarm decides where they run and starts them again when a node falls away. The app shows what is in a swarm, runs it, and helps you start one on your own servers — with a question first wherever one click reaches more than one machine.
New to swarms?Read the comic: a restaurant, a head chef, and a cook who goes home at nine →A swarm that won’t come together?A join that times out, WSL off the network, half the requests lost: what solved each →What you can do
In a swarm, and as what
Every machine card says whether it is a manager, a worker or neither, and machines in the same swarm share a mark. A locked swarm shows up in the status bar.
Services per stack
How many run of how many should, where each task runs and why one stopped. The logs of all tasks of a service together, in one view.
Every task at once
The Tasks tab lists every task of every service: its node, its state, since when, and why it stopped or never started. A task waiting for a node says per node why none fits. Filter by node or service, and switch on the tasks the swarm let go.
Scale, roll out, roll back
Scale a service, roll out a new image with the rolling update from its spec, roll it back or out again. Adjust environment, ports, limits, restart policy and placement, or create a small one — for mysql or postgres the usual variables are a click away, and a password stays masked or comes from a secret of the swarm. The digest of the image is pinned, as docker service create does, and a service without one says so.
Resilience of a stack
One replica of a service without a volume, and an update that does not roll back when it fails, each with a link to scale or roll out.
Stacks from their compose file
Deploy a stack from its file or from a template in the catalog, see where it differs from that file, deploy it again with what changes shown first, or remove it. Volumes stay. Before a deploy the app checks the names against this machine: a volume of a Compose project with the same name holds the button until you confirm. With the option to remove what is no longer in the file, the services that would go are named first. A folder on C: works on a manager in WSL too, and a host folder on a swarm with more nodes gets a warning first.
Nodes
Role, state, leader, address and labels, and for a node that is not ready the reason: no heartbeat, a new leader, or looking for a manager. Drain one after seeing which tasks move and what has nowhere to go; pause, activate, promote or demote with a warning about the quorum — and against promoting a node that went down twice this week. A WSL node that is still active is pointed out: traffic between machines does not reach it, so it belongs on drain.
Secrets and configs
Create one from text or a file, see which services use it, and replace it in one step that switches every service over. What a secret holds goes straight to the swarm and is never kept or shown. A password in the form for a service can come from one, as a file.
Start one, or join one
Start a swarm on the machine you are on, or join one through a manager that is also in the app — it fetches the token itself. Pasting a token works too. A join that hangs says through which address it tries, and Stop and join again lets go of it and starts over.
The overview first
On a manager the page opens on the Overview: the leader, the nodes, and the join tokens, each with a menu to show, copy, copy the join command or renew. Autolock is a switch that asks first; its key is shown once. How often nodes report to the leader is a choice from 5 to 60 seconds, and how many old tasks the swarm keeps per replica one from 1 to 10. The last manager only leaves after you type the machine’s name.
What the swarm uses
CPU, memory and network of all swarm tasks together, next to the Overview, with per node its tasks and what they use. Docker has no such number for a swarm: the app asks every node that is also a machine in the app for the containers of its tasks, and says for any other node why it is not measured.
A task is not a container
On Containers, a container that belongs to a service stands with its stack, points to its service, and asks before you stop what the swarm would only replace. Renaming one is off: the swarm finds a task by its name. The tasks the swarm stopped or replaced stay folded under their stack, and the app can remove their containers after an hour.
Events and notices
Services, nodes, secrets and configs on the Events page. A notice when a node goes down, the leader changes, an update is rolled back or stalls, or a service stays below its count, with the reason per node when there is nowhere it can run — and one per service for tasks that keep failing.
A swarm chapter in the checkup
One manager or an even number, a WSL node that is not drained, ports open to every node through ingress, services without a restart policy or health check, nodes drained for days, secrets nobody uses.
"scale web to 3"
Ctrl+K knows services, stacks and nodes. Type "scale web to 3" or "drain node2", in any of the five languages.
Your own servers
The address is the hard part, so the app says which one.
Under WSL the address of the distro cannot be reached from other machines without mirrored networking, and under Docker Desktop neither. The app looks at where the engine runs, suggests the address others can reach — Wi-Fi and Ethernet before a VPN — and lists the ports a firewall has to let through: 2377, 7946 and 4789. Under WSL with mirrored networking it reads the Hyper-V firewall and gives the rules to paste, for those ports and for the ports your services publish, and says when the laptop itself only reaches such a port through localhost. Joining through a manager in Docker Desktop it refuses up front.
One click, more than one machine
The question says what moves before anything does.
Draining a node moves everything on it; the last manager leaving takes the whole swarm with it. Those buttons first say what will happen, with names and numbers, and the most dangerous ones ask you to type the name of the machine.
Nothing new to install
Swarm mode is part of the engine you already run.
No agent and no extra daemon. The app talks to a manager the way the docker CLI does, and stacks go through docker stack deploy on the machine you are on, whatever the CLI itself points at.
Only the swarm
The same swarm section, as an app of its own.
Docker Swarm MX is this swarm section on its own, for Windows, for whoever runs a swarm from day to day and needs nothing else of Docker Client MX. On top of what is here, it backs up the swarm and its volumes on a schedule, sends a message when something breaks, opens a root shell on any node, shows the images and the disk of every node, keeps an audit log of every change, and protects production per context.
All features