Postgresql – pg_upgrade fails with an unspecified error

You'll need to re-initdb the new database with LC_CTYPE set to de_DE (i.e. use a non-Unicode encoding); check the old DB's encoding and use the same one when you initdb the new database.

Alternately, use pg_dump and pg_restore to do a dump and reload of the old DB's contents into the new DB. This requires that you have the old DB running.

I hacked the problem by running (as root):

mv /usr/bin/pg_ctl{,-orig}
echo '#!/bin/bash' > /usr/bin/pg_ctl
echo '"$0"-orig "${@/unix_socket_directory/unix_socket_directories}"' >> \
     /usr/bin/pg_ctl
chmod +x /usr/bin/pg_ctl

Run pg_upgrade as intended, then undo the hack:

    mv -f /usr/bin/pg_ctl{-orig,}

The problem is that pg_upgrade executes the program pg_ctrl with arguments that specify files in the old "unix_socket_directory" rather than the new "unix_socket_directories" (note the second is plural). This hack renames the original /usr/bin/pg_ctl to /usr/bin/pg_ctl-orig, and then creates a shell script in its place that simply calls the original pg_ctl program, passing all arguments with any strings "unix_socket_directory" changed to "unix_socket_directories".

In bash, one can change a portion of a string, say from bar to baz in a variable $foo, by using ${foo/bar/baz} (note this does not change the variable, but rather returns the variable's modified contents). Arrays can also be used with ${x/y/z} to retrieve an array with all of its contents replaced, all at once. The variable $@ is an array that contains all arguments passed to the program/script/function, so the new pg_ctl script executes the old one with all arguments changed from the old directory name to the new one.