User Connection Routing

A client connection lives on the Gateway node that accepted it, while post-commit delivery can run on a different Channel Leader node. User routing locates a UID's concrete connection owner through bounded, expiring memory state and prevents old Leaders, processes, or Sessions from receiving new traffic.

Presence Authority does not hold concrete TCP Sessions or write online routes into Slot Raft. It is a high-frequency memory view rebuilt from Gateway activity and client reconnects.

Activation registers a local pending Session before registering its route with the current UID authority. Registration failure rolls back local state. Device conflicts require acknowledged old-owner actions before the pending route can commit. The final MarkActive rechecks that the Session still exists, closing the race where authority accepted a route after the local connection disappeared.

Every registration, touch, lookup, and unregister carries an exact target:

(HashSlot, SlotID, LeaderNodeID, LeaderTerm, ConfigEpoch)

The routing revision accompanies the target as freshness evidence. When a node gains authority for a physical hash slot, it installs a fresh identity. A changed Leader term or configuration epoch clears old active and pending state. A route-revision-only update under the same Raft identity can retain the directory while advancing its observed revision.

Losing authority removes all memory state for that hash slot. Old calls receive not-leader or stale-target instead of serving expired routes.

An online endpoint carries at least:

• UID, DeviceID, Device Flag, and Device Level;
• owner NodeID and process BootID;
• SessionID and listener;
• monotonic OwnerSeq, connection time, and last activity.

BootID separates processes that reuse a node ID, SessionID separates connections, and OwnerSeq prevents delayed register or touch from overwriting a newer unregister tombstone. UID plus node ID alone could deliver into a restarted or replaced connection.

Each client Ping marks dirty activity in the owner-local registry instead of producing an authority RPC. An app worker drains a bounded set, groups by current authority target, and sends TouchRoutes batches. Failed routes are requeued only when the same owner route remains current.

Authority uses activity-second buckets and an expiry index, scanning only due buckets instead of every online user. TTL expiry removes an active route but preserves explicit unregister tombstones and monotonic owner sequence.

Post-commit delivery finds recipients as follows:

1. Channel append hashes UIDs in one subscriber page across the 256 physical hash slots.
2. One route snapshot aligns UIDs with complete authority targets and groups by target.
3. EndpointsByTargets validates each group fence; one stale group does not erase successful siblings.
4. Returned routes coalesce by owner node into bounded local or remote push batches.
5. A remote owner-push RPC transports routes and frame only; the target owner revalidates UID, Boot ID, Session ID, and owner sequence.
6. The concrete Session writes RECV through an entry-independent handle and binds owner-local ACK state when required.

Only the sender's exact SEND Session is suppressed; its other devices can still receive synchronization. Duplicate recipient rows can preserve duplicate-write semantics, while endpoint lookup stays within a bounded plan.

A new connection with the same UID and Device Flag can conflict with an old route. A Master level can replace same-kind device connections, while a Slave level generally replaces only the same Device ID. Authority retains the old active route and creates a pending candidate until owner action succeeds, avoiding two advertised conflicting routes before the old connection is actually closed.

Presence is in memory, so failover has a route-rebuild window. Callers must tolerate temporary unknown or offline state and recover through client reconnect plus durable message sync instead of turning online presence into a high-frequency global Raft log.