tmp-postgres currently creates temporary PostgreSQL instances using (long) absolute paths. This can hit Unix domain socket path length limits (often 108 characters, 104 for macOS), especially in environments like Nix builds where paths are deeply nested.
This causes failures when trying to start or connect to the database. For example, socket paths like:
/nix/var/nix/builds/nix-build-........./tmp-postgres-xxxxxxxxxxxxxx/.s.PGSQL.5432
can exceed the limit and create failure during a test harness execution.
Suggested workaround
Change the working directory (chdir) to the temp data directory before starting Postgres and using socket paths. This way, relative paths (like ./socket) stay short and avoid the issue.
See how Lix did it here: https://git.lix.systems/lix-project/lix/src/commit/801567adf0ccf41114e24a6b4f182c0da1ccf503/lix/libutil/unix-domain-socket.cc#L40-L101.
This would make tmp-postgres more reliable in build-heavy environments without filesystem namespacing abilities (i.e. this doesn't happen that much on Linux because you can hide the long prefix with kernel namespaces, which macOS cannot).
Original report came from https://git.lix.systems/lix-project/lix/issues/913#issuecomment-13322.
FWIW, Lix will workaround this problem in general, but fixing it here also makes sense.
tmp-postgrescurrently creates temporary PostgreSQL instances using (long) absolute paths. This can hit Unix domain socket path length limits (often 108 characters, 104 for macOS), especially in environments like Nix builds where paths are deeply nested.This causes failures when trying to start or connect to the database. For example, socket paths like:
can exceed the limit and create failure during a test harness execution.
Suggested workaround
Change the working directory (
chdir) to the temp data directory before starting Postgres and using socket paths. This way, relative paths (like./socket) stay short and avoid the issue.See how Lix did it here: https://git.lix.systems/lix-project/lix/src/commit/801567adf0ccf41114e24a6b4f182c0da1ccf503/lix/libutil/unix-domain-socket.cc#L40-L101.
This would make
tmp-postgresmore reliable in build-heavy environments without filesystem namespacing abilities (i.e. this doesn't happen that much on Linux because you can hide the long prefix with kernel namespaces, which macOS cannot).Original report came from https://git.lix.systems/lix-project/lix/issues/913#issuecomment-13322.
FWIW, Lix will workaround this problem in general, but fixing it here also makes sense.