Silent port conflict goes undetected

Start by reproducing the Docker and Postgres.app setup in the issue, then inspect the source references to SO_REUSEPORT and SO_REUSEADDR and compare the context from #676. Done means a conflicting listener is detected and reported reliably without allowing traffic to reach different clusters unexpectedly.

I ran into a confusing scenario where I think Postgres.app should have reported "Port in use" and failed to start, but instead managed to bind and steal traffic from a process that was already listening on that port.

I have a Docker container running PostGIS and exposing container port 5432 on host port 5432. I was able to open a psql shell on port 5432, but was confused because the database didn't have the right data. Eventually I figured out that Postgres.app was also listening on port 5432:

$ sudo lsof -i :5432
COMMAND     PID     USER   FD   TYPE             DEVICE SIZE/OFF NODE NAME
com.docke  1029 wchargin  128u  IPv6 0x97e222976b6cf6c5      0t0  TCP *:postgresql (LISTEN)
postgres  45709 wchargin    7u  IPv6 0x91b1855dc1c99cac      0t0  TCP localhost:postgresql (LISTEN)
postgres  45709 wchargin    8u  IPv4 0xa012493f18604ea9      0t0  TCP localhost:postgresql (LISTEN)

$ ps -p 45709 -o comm
COMM
/Applications/Postgres.app/Contents/Versions/16/bin/postgres

I searched the source for SO_REUSEPORT / SO_REUSEADDR and found the comment leading to #676. I'm glad that that issue is resolved, but it seems like this issue is an unfortunate consequence of the fix.

I do think that it may be somewhat confusing to use this socket option, since it subverts the application that if an application is listening on a port then it will be able to receive packets sent to that port. Notably, this confusion only occurs because my Docker container is listening on all interfaces. If I expose 127.0.0.1:5432:5432 instead of just 5432:5432, then Postgres.app properly reports "Port in use" or the Docker container fails with "Ports are not available". So it feels a bit clumsy and unpredictable that Postgres.app behaves this way.

Might there be a different way of solving the original issue that doesn't cause this side-effect?

Steps to reproduce

1. Make sure that Postgres.app is stopped.
2. Run docker run --rm -e POSTGRES_PASSWORD=password -p 5432:5432 -it postgres, and leave that running in its own terminal.
3. Run psql -h localhost -p 5432 -U postgres, and execute (say) CREATE TABLE tab(), SELECT count(1) FROM tab.
4. Start Postgres.app. Note that it moves to the "Running" state without any error.
5. Run psql -h localhost -p 5432 -U postgres again. Note that SELECT count(1) FROM tab now fails: no such relation.

If you have the psql shells from (3) and (5) open at the same time, you can note that their \conninfos report identical results:

You are connected to database "postgres" as user "postgres" on host "localhost" (address "::1") at port "5432".

…yet they are actually talking to different clusters!

(Then you can Ctrl-C in the terminal from step 2, which will destroy the container.)