Date: Fri, 27 Jan 2006 21:24:49 -0800 (PST)
From: Stephan Szabo <sszabo@megazone.bigpanda.com>
To: "Kashmira Patel (kupatel)" <kupatel@cisco.com>
Cc: Michael Fuhr <mike@fuhr.org>, pgsql-sql@postgresql.org
Subject: Re: Question about check constraints
In-Reply-To: 
 <18AE59788A3FC640A367E5652E664D800173133E@xmb-sjc-237.amer.cisco.com>
Message-ID: <20060127210947.K43706@megazone.bigpanda.com>
References: 
 <18AE59788A3FC640A367E5652E664D800173133E@xmb-sjc-237.amer.cisco.com>
MIME-Version: 1.0
Content-Type: TEXT/PLAIN; charset=US-ASCII


On Fri, 27 Jan 2006, Kashmira Patel (kupatel) wrote:

> Both concerns.
> 1) There are actually more than two columns with such checks, and each
> one calls a few functions which execute some more queries. So I would
> like to invoke these checks only when necessary.
> 2) The bigger concern is the side effect: Here's my schema:
>
> CREATE TABLE vm_device
> (
>     device_id      INTEGER      UNIQUE NOT NULL
>                                 REFERENCES device_table(device_id)
>                                 ON UPDATE CASCADE
>                                 ON DELETE CASCADE,
>     preference     VARCHAR(1)   NOT NULL DEFAULT 'U'
>                                 CHECK (CASE WHEN preference = 'U'
>                                             THEN true
>                                             ELSE validate_preference()
>                                        END),
>     enabled        BOOLEAN      NOT NULL DEFAULT false
>                                 CHECK (CASE WHEN enabled = false
>                                             THEN true
>                                             ELSE
> validate_system_enabled() AND
>
> validate_enabled(device_id)
>                                        END),
>     attach_vm      BOOLEAN      NOT NULL DEFAULT false
>                                 CHECK (CASE WHEN attach_vm = false
>                                             THEN true
>                                             ELSE validate_attach_vm()
>                                        END),
>
>     PRIMARY KEY (device_id)
> ) WITHOUT OIDS;
>
> This table contains some information about a device in my system.  The
> issue is with the enabled column. It basically enables/disables the
> device.  The device can be enabled only when the two check conditions
> pass. But once it is enabled, the conditions of the system might change
> such that if executed again, these conditions might not pass. We want to
> allow such situations. The problem arises when we want to change the
> value of some other column, say attach_vm. Although the check
> constraints for the attach_vm column pass, those for enabled column
> fail, and I cannot complete my updates.

In that case check constraints are probably not the correct tool.  IIRC
the theoretical model from the spec implies that all check constraints in
the system are supposed to be satisfied at their check time (statement end
in most cases) and that it doesn't matter what the change was, and as
such, the conditions of the system shouldn't have been allowed to change
such that a row with enabled=true existed when its constraint would be
violated at this moment. I don't think that's practically reasonable to
enforce in general, but we do the best we can which is fail the later
update.

> Any suggestions on the best way to overcome this?

I'd second Michael's suggestion of a trigger.


> -----Original Message-----
> From: Michael Fuhr [mailto:mike@fuhr.org]
> Sent: Friday, January 27, 2006 4:40 PM
> To: Kashmira Patel (kupatel)
> Cc: pgsql-sql@postgresql.org
> Subject: Re: [SQL] Question about check constraints
>
> On Fri, Jan 27, 2006 at 03:06:26PM -0800, Kashmira Patel (kupatel)
> wrote:
> >   I have a table where two columns have two different check
> > constraints associated with them. When I update one column, the check
> > constraint on the other column is also executed. Is there a way to
> > avoid this? I want to check only for the condition defined for the
> column being updated.
>
> I don't think you can change this behavior: each CHECK constraint is
> evaluated for the new row regardless of whether a particular column
> changed or not.  However, you could enforce the constraints with a
> trigger and skip checks where NEW.column is the same as OLD.column.
>
> Why the concern?  Are the checks expensive?  Do they have side effects?
> What do they do?