Two server instances tail the same PostgreSQL, so the WAL LSN means the same thing on both. When instance A drains for a deploy, it sends a GoAway; the client re-dials the load balancer, lands on B, and resumes from its last LSN — a delta, not a full snapshot — then closes A cleanly. Hit drain A.
Resume is at-least-once and idempotent (apply by primary key); the new instance replays from the client's LSN, so the handoff is gapless — see guides/fan-out.