Re: [HACKERS] pg_upgrade failed with error - ERROR: column "a" inchild table must be marked NOT NULL

It is not the first time that this topic shows up. See for example
this thread from myself's version of last year:
https://www.postgresql.org/message-id/CAB7nPqTPXgX9HiyhhtAgpW7jbA1iskMCSoqXPEEB_KYXYy1E1Q@mail.gmail.com
Or this one:
http://www.postgresql.org/message-id/21633.1448383428@sss.pgh.pa.us

And the conclusion is that we don't want to do what you are doing
here, and that it would be better to store NOT NULL constraints in a
way similar to CHECK constraints.

> How about: ".. if some parent has the same"
>
> +               heap_close(parent, AccessShareLock);
>
> Maybe, we shouldn't be dropping the lock so soon.

Yes, such things usually need to be kept until the end of the
transaction, and usually you need to be careful about potential lock
upgrades that could cause deadlocks. This patch looks wrong for both
of those things.