Skip to main content

How it works

While the match runs​

The host builds a snapshot of everything marked for migration every Snapshot Interval (3 s by default) and sends it to every client, compressed and in chunks over reliable RPCs. Clients that already have the previous snapshot only get what changed (delta snapshots).

Each snapshot also contains an ordered successor list: the players who may take over, best first (lowest ping by default).

Every host announces itself through the adapter while it hosts. On Steam that is lobby member data.

When the host is lost​

  1. Electing. Clients lose the connection. If the adapter says the host is still in the session, they first try to rejoin it. Otherwise every client walks the same successor list and picks the first player who is still there.
  2. BecomingHost. The successor opens the same map as a listen server and restores the snapshot. Actors are created first and loaded second, so actors that reference each other work in any order.
  3. WaitingForHost / Reconnecting. Everyone else loads the Client Wait Map, waits for the new host's announcement and connects. Connections are retried Max Reconnect Attempts times. If the successor never shows up, they move on to the next candidate.
  4. GracePeriod. The new host waits up to Reconnect Grace Period for the previous players. Each returning player gets their PlayerState and pawn back, matched by unique net id. The game is paused meanwhile (optional).
  5. Running. The match goes on. Players who did not make it can still come back later.
Running -> Electing -> BecomingHost (successor) -> GracePeriod -> Running
                    \-> WaitingForHost (others) -> Reconnecting -> Running
Running -> RejoiningHost (host still in session) -> Running   | or -> Electing if the host leaves

Each migration increases the epoch, so chained migrations (the new host leaves too) work the same way.

Graceful leave​

A host that leaves on purpose with Leave Session (or Begin Graceful Migration) sends a final snapshot and tells everyone to migrate right away. Nothing is lost and nobody waits for a timeout.

What does not need changing​

No custom GameEngine, GameMode, GameInstance or PlayerController class. The subsystem is a GameInstance subsystem and adds a small component to each remote PlayerController for the snapshot RPCs.