Three SSE-routes (solo, backlog, notifications) each create a long-
running pg.Client that LISTENs on scrum4me_changes. On abrupt close
(Fast Refresh, browser refresh, Vercel function recycle) the
pgClient.end()-await sometimes hangs silently, leaving the underlying
socket connected to Postgres. The connection stays in 'idle' on Neon's
side and after ~10-20 reconnects the connection-pool fills up — new
SSE connects fail with ERR_INCOMPLETE_CHUNKED_ENCODING in the browser.
Fix: shared `closePgClientSafely` helper that races client.end()
against a 2 s timeout; on timeout it force-destroys the underlying
socket so the OS releases the FD and Postgres notices the disconnect.
Validated by direct DB inspection: 18 stale 'idle LISTEN'-connections
were piled up before the fix; after manual pg_terminate_backend cleanup
the SSE-stream stabilised. This change makes the pile-up impossible
going forward.
- new lib/realtime/pg-client-cleanup.ts
- 3 routes use the helper instead of bare `await pgClient.end()`
- 3 unit tests for the helper (timely-end, hang-falls-back-to-destroy,
end-rejection-is-swallowed)
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Janpeter Visser
6c6c8b96b7