Private network
The servers of a cluster share a private network, so every app and database reaches the others by name on any server, as if they all ran on one.
The servers of a cluster share a private network. Deplo sets it up the moment a server joins, and you configure nothing:
- Every app reaches its environment's apps and databases by name, on whichever server they run.
The same names and ports as on one server:
db-shop,deplo-web, or a Compose service's own name likeapiorredis. - It stays up while Deplo or the cluster's leader is down. The servers talk to each other directly; only adding or removing a server waits for Deplo.
- It is encrypted end to end and only the cluster's own servers are on it.
What reaches what
| From an app on server-a | To, in the same environment, on server-b | Reachable |
|---|---|---|
| Any app | A database, as db-shop | Yes |
| Any app | A single-image app, as deplo-web | Yes |
| Any app | A service of a Compose app, as api (or its hostname) | Yes |
| Any app | Any of those, on any TCP port it listens on | Yes |
| Any app | Anything of another environment, team or cluster | No |
Nothing changes in your code after a move or a failover: the name follows the app to its new server.
TCP only
Names carry TCP connections, which is what HTTP, databases, Redis and queues use. UDP between two servers is not carried.
Two Compose apps of one environment that both have a service called redis share that name, as
they would on one server: the deploy warns about it. Across servers, the other server reaches
one of them. Give the services different names to be sure which one answers.
Direct or relayed
Each pair of servers connects directly when it can: on the same office network, on public addresses, and through most home routers. When two servers cannot reach each other at all, their traffic goes through Deplo's relay instead. It still works, with a little more delay.
The server's card in the cluster says when that happens:
| On the card | What it means |
|---|---|
| Nothing | It reaches every other server of the cluster directly |
| Relayed | "Reaches eu-2 through Deplo's relay, with more delay" |
| Can't reach a server | "Cannot reach eu-2 now": both are online, yet nothing gets through |
A pair switches between direct and relayed on its own, without dropping open connections. A relayed pair needs Deplo's machine to be up; a direct one does not.
Making a relayed pair direct
Servers connect directly over UDP port 51820. When a firewall between two servers blocks it, let UDP 51820 through between them and the card clears within a few minutes. Nothing else needs opening, and the port answers no one who is not a server of the cluster.
The database page
A database whose apps run on another server says so only when that pair is relayed: "web reaches this database through Deplo's relay. It is faster on the same server."
Older servers
A server whose agent is too old for the private network stays off it, and its traffic goes through Deplo's relay as before. Deplo updates every agent on its own. Until then, a Compose app's services on that server are reachable by name from its own server only.
Over the API
ClusterMember.network { relayed unreachable } lists the servers a member reaches only through the
relay, or not at all. Over MCP, get_cluster returns the same field for each server.
Did this page help you?