Re: Review: Non-inheritable check constraints

On Mon, Dec 19, 2011 at 12:07 PM, Nikhil Sontakke <nikkhils@gmail.com> wrote:
>> It seems hard to believe that ATExecDropConstraint() doesn't need any
>> adjustment.
>
> Yeah, I think we could return early on for "only" type of constraints.

It's not just that.  Suppose that C inherits from B which inherits
from A.  We add an "only" constraint to B and a non-"only" constraint
to "A".   Now, what happens in each of the following scenarios?

1. We drop the constraint from "B" without specifying ONLY.
2. We drop the constraint from "B" *with* ONLY.
3. We drop the constraint from "A" without specifying ONLY.
4. We drop the constraint from "A" *with* ONLY.

Off the top of my head, I suspect that #1 should be an error; #2
should succeed, leaving only the inherited version of the constraint
on B; #3 should remove the constraint from A and leave it on B but I'm
not sure what should happen to C, and I have no clear vision of what
#4 should do.

As a followup question, if we do #2 followed by #4, or #4 followed by
#2, do we end up with the same final state in both cases?

Here's another scenario.  B inherits from A.  We a constraint to A
using ONLY, and then drop it without ONLY.  Does that work or fail?
Also, what happens we add matching constraints to B and A, in each
case using ONLY, and then remove the constraint from A without using
ONLY?  Does anything happen to B's constraint?  Why or why not?

Just to be clear, I like the feature.  But I've done some work on this
code before, and it is amazingly easy for to screw it up and end up
with bugs... so I think lots of careful thought is in order.

-- 
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company