2010/8/11 Marko Tiikkaja <marko.tiikkaja@cs.helsinki.fi>:
> On 8/11/10 8:31 AM +0300, Peter Eisentraut wrote:
>
>> Thinking about SQL assertions (check constraints that are independent of
>> one particular table), do you think it would be reasonable to implement
>> those on top of constraint triggers? On creation you'd hook up a
>> trigger to each of the affected tables. And the trigger function runs
>> the respective check expression. Conceptually, this doesn't seem to be
>> very far away from foreign key constraints after all.
>
> I thought the point of ASSERTIONs was that you could write a thing such as:
>
> CREATE ASSERTION foo CHECK ((SELECT count(*) FROM tbl) = 4);
>
> Enforcing that kind of constraints without true serializability seems
> impractical.
Exactly what I thought when I read this. Without true serializability,
the view of the database at any moment during a transaction doesn't
have to be the same as the view that a newly started transaction gets.
Therefore, checking that the assertion holds after changing something
doesn't necessarily guarantee that it will hold for any other
transactions.
To elaborate on a variant of Marko's example, where the "=" is
replaced with "<=". Assume "non-true SERIALIZABLE" transactions:
* The table has 3 rows.
* T1 inserts a row, and concurrently, T2 also inserts a row; after
each statement, the assertion is not violated for the corresponding
transaction's snapshot.
* The assertion is now violated for a subsequent transaction T3
(because it sees 5 rows).
Nicolas