Engine-native failover — active calls re-home to a surviving node, not to voicemail.
Every node owns one fixed RTP port per core; calls demux by 4-tuple. A call’s media home is just node : port — unlimited calls, and re-homing needs no per-call port allocation.
Heartbeats and a leader lease ride a dedicated Redpanda control plane — never the media path. Call ownership registers on established, clears on hangup.
On the answering 200 OK, everything needed to reconstruct the dialog — Call-ID, tags, route set, CSeq, remote media, codec — journals to Redis, off-node.
The leader detects the dead node, rendezvous-hashes each owned call to a survivor, and the takeover node rebuilds the dialog from the journal, opens a fresh media leg and recovers the call with an in-dialog re-INVITE.
Failover knowledge flows over Redpanda while media keeps streaming. The design target is a sub-2-second blip — the caller hears a hiccup, not a dial tone.
Telequick runs every layer of voice — you bring the models.