Psycopg cuts away part of IPv6 address

[stinovlas]

Problem

MWE:

from psycopg import Connection
Connection.connect("postgresql://user:password@[2001:1488:fffe:20::229]:5432/dbname")

Output:

Traceback (most recent call last):
  File "/home/jmusilek/git/others/auctions/psycopg_mwe.py", line 3, in <module>
    Connection.connect("postgresql://user:password@[2001:1488:fffe:20::229]:5432/dbname")
  File "/home/jmusilek/.virtualenvs/auctions/lib/python3.12/site-packages/psycopg/connection.py", line 748, in connect
    raise last_ex.with_traceback(None)
psycopg.OperationalError: connection failed: 1488:fffe:20::229", port 5432 failed: server closed the connection unexpectedly
        This probably means the server terminated abnormally
        before or while processing the request.

As you can see from the traceback, psycopg tries to connect to 1488:fffe:20::229, not 2001:1488:fffe:20::229. The first part is cut off. This happens with all IPv6 addresses and also when I specify a hostname in conninfo, which is then resolved to an IPv6 address. This does not happen with IPv4 addresses.

Unfortunatelly, I wasn't able to track down where this happens.

[dlax]

It seems to me that it's the error message that is wrong (includes truncated address), because of

psycopg/psycopg/psycopg/pq/misc.py

Line 105 in 558cbe7

bmsg = bmsg.split(b":", 1)[-1].strip()

If you put a print(f"{obj.error_message=}") above this line, you should see that the full address is actually used to connect. E.g.:

>>> import psycopg
>>> psycopg.connect(host="::1")
obj.error_message=b'connection to server at "::1", port 5432 failed: fe_sendauth: no password supplied\n'
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/home/denis/src/psycopg3/psycopg/build/__editable__.psycopg-3.2.0.dev1-py3-none-any/psycopg/connection.py", line 118, in connect
    raise last_ex.with_traceback(None)
psycopg.OperationalError: connection failed: :1", port 5432 failed: fe_sendauth: no password supplied

[dvarrazzo]

Uh, right. Maybe we should have the stripping stricter. There are a few words that are used, such as ERROR, WARNING... but they also undergo l10n. Maybe there is some pattern we can use though (such as stripping a single uppercase word).