Subject: Intermittent Postgres disconnects via maglev.proxy.rlwy.net:47180 (ECONNRESET → connect timeouts) from external client

Service: Railway Managed Postgres
Host/Port:maglev.proxy.rlwy.net:47180
DB name:railway
Region: Us California
Client app: n8n hosted on Render ($7 Starter web service)
Client driver/runtime: Node 18, pg 8.12.0 via pg-pool 3.6.2

Summary

We see stable DB operation for ~10–20 minutes, then a burst of:

1. Connection terminated unexpectedly

2. read ECONNRESET

3. repeated timeout exceeded when trying to connect (1–10+ minutes)

During those windows, new connections fail and existing ones drop. After we manually restart the DB, it works again until the next burst. Load is very light.

Client connection settings

• SSL enabled (rejectUnauthorized=false)

• Pool size: 3

• Connect timeout: 70,000 ms

• Idle connection timeout: 70,000 ms

• App ping/keepalive: lightweight query every 5s

• Workload: small n8n metadata/executions reads; no long-running queries

Timeline (UTC) from client logs (2025-09-09)

16:07:35 OK (periodic query)
16:08:35 OK
16:09:19 ERROR Connection terminated unexpectedly
16:09:27 ERROR read ECONNRESET (twice)
16:10:29 ERROR timeout exceeded when trying to connect
16:10:35 ERROR timeout exceeded when trying to connect
16:11:31/12:33/13:35/14:35/15:35 Repeated connect timeouts
16:14:48 ERROR Failed to hard-delete executions (root cause: timeout exceeded when trying to connect)

Railway Postgres logs after our manual restart (same day)

16:25:39 starting PostgreSQL 17.6
16:25:40 database system was interrupted; last known up at 16:18:35
16:25:40 database system was not properly shut down; automatic recovery in progress
16:25:40 redo starts at 0/32F5850
16:25:40 invalid record length at 0/32F5990: expected at least 24, got 0
16:25:40 redo done; checkpoint end-of-recovery
16:25:40 database system is ready to accept connections

Notably, client-side errors begin at 16:09:19, while the DB reports last known up at 16:18:35 and an unclean shutdown at restart time. That suggests proxy/LB issues or backend host/container problems that preceded the final DB interruption.

What we tried

• Tuned pool size/timeouts and added 5s app-level pings.

• Restarted DB; problem recurs later.

• Traffic is minimal; other app calls (HTTP, Google Sheets OAuth refresh) succeed—only DB path fails.

Hypotheses

• Maglev proxy/LB instability in our region causing RSTs, then “black-holed” connects.

• DB container/host events (OOM/host drain/migration) leading to unclean shutdown.

• Aggressive TCP/NAT idle timeouts on the proxy path despite app pings.

• Less likely: per-client limits at very low connection counts.

Requests to Railway

1. Check proxy/LB logs for maglev.proxy.rlwy.net:47180 between 16:08–16:16 UTC and surrounding minutes for RSTs, health flaps, backend detachments for our DB.

2. Check DB container/host events around 16:09–16:19 UTC (OOM kills, node drain, maintenance).

3. Confirm effective TCP/idle timeouts and recommended Postgres/pg client keepalive settings (tcp_keepalives_*), or other best-practice values for external clients.

4. Advise on any known incidents in this region during that window.

5. Is there a way to obtain a direct endpoint (bypassing shared proxy) or recommended PgBouncer approach for cross-provider clients?

Context

We may migrate this DB to Render Postgres (private network) to avoid cross-provider hops, but we’d like root cause on the Railway side to decide future usage and to help you triage a potential platform issue.

Happy to provide full client/server logs or run a short reproduction window if needed.