IPv6 Feature Flag causes ENETUNREACH when connecting to hosts with AAAA records

Service: Python FastAPI (Railway Preview environment)

Feature flag: IPv6 (enabled via Settings → Feature Flags)

---

What happened

I enabled the IPv6 feature flag and simultaneously switched my Supabase DATABASE_URL from the connection pooler to the

direct DB host. The service immediately started failing on every database call with:

OSError: [Errno 101] Network is unreachable

The failure occurs at the TCP socket layer (SQLAlchemy → asyncpg → socket connect), before any SSL or auth negotiation.

The service crash-looped across multiple container restarts — the error was persistent, not transient.

Troubleshooting

I resolved the two hostnames from my local machine:

dig AAAA db.xhszgkrasrskvlfhuavt.supabase.co

→ AAAA 2600:1f16:1cd0:3320:348a:c3ba:f8a2:59c4 (real IPv6 address)

dig AAAA aws-1-us-east-2.pooler.supabase.com

→ CNAME to an AWS ELB with no AAAA record (IPv4 only)

Root cause

With IPv6 enabled, the Railway container gains an IPv6 interface. Python's getaddrinfo() returns the AAAA record for the

direct Supabase DB host, asyncpg attempts IPv6 first, and gets ENETUNREACH — Railway has no route to that AWS IPv6 address

(2600:1f16::.../us-east-1).

The pooler URL has no AAAA record, so Python falls back to IPv4 and connects successfully.

Confirmation

Reverting DATABASE_URL to the pooler (IPv4-only) URL fixed the issue without disabling the IPv6 feature flag. IPv6 is

still enabled and the service is healthy. This confirms the issue is specifically Railway's IPv6 routing being unable to

reach AWS IPv6 addresses — not a misconfiguration on my end.

Expected behavior

Either:

- IPv6 routing should have a working path to AWS IPv6 addresses, or

- asyncpg/Python should fall back to IPv4 when IPv6 yields ENETUNREACH (happy eyeballs), or

- the docs should note that IPv6 routing is incomplete and hosts with AAAA records may be unreachable

Workaround

Use a DATABASE_URL that resolves to IPv4 only (e.g. Supabase's connection pooler instead of the direct DB host).