test=# ALTER TABLE test1 DROP CONSTRAINT test1_t1_key CASCADE; NOTICE: drop cascades to constraint test2_t1_fkey on table test2 ALTER TABLE
It is the same end result as the first two steps of what you are doing below, just a different direction.
No special reason at all: I began with CASCADE, and as things went wrong, I tried to split the process to better figure out the problem
Is there a solution to" alter" the "test2_t1_fkey" constraint so that it uses the "primary key constraint", then to remove the unnecessary unique constraint on table test1
The following solution works but causes me deadlocks problems with BDR:
Is the below wrapped in a transaction?
Yes. The goal is to wrap this upgrade process inside a transaction to be able to abort it in case something was wrong.
Problem is that some tables may be accessed during the upgrade process. May be a solution is to avoid it by only allowing the upgrade backend and bdr to access the tables, but I do not like the idea to make the database readonly (UPDATE pg_database SET datallowconn = false WHERE pid != upgrade_and_bdr ... ):
So is the above a BDR specific enhancement to pg_database or is pid != upgrade_and_bdr just a placeholder for something else?
Sorry, forget all about BDR. In fact, I need to arrange the tables not to be accessed during the upgrade phase, else this leads to deadlocks, and there is no possible magic to avoid it as I was initially dreaming.
In other words, to solve my problem, I think I have 2 solutions :
1) do the necessary job so that only the upgrade process access the tables during constraints changes; other processes will be stopped during the upgrade
2) or in the upgrade process, terminate all processes except the one that does the upgrade, and the bdr workers.
(the "upgrade_and_bdr" pseudo code was not clear, sorry for this)
in case the upgrade process fails, this would requiere require a manual intervention to solve it (upgrade is called if needed by the application).
If I am following correctly then the changes to the tables are being done on a as needed basis based on some external criteria.
In any case for each table it should be a one time operation, right? Also from a practical stand point the FK between test2 and test1 is not actually changing. So why not just change them ahead of time in a process you can monitor directly?
