When Are Constraints Checked
Under the model constraints are conceptually checked at all statement boundaries (and only at statement boundaries). By default the same is true of SQL. However, SQL does not support the "multiple assignment" concept, described in the theory book, for database updates. For that reason it has to include an alternative method of addressing the problems that multiple assignment addresses. SQL does so by allowing the checking of specified constraints to be temporarily deferred and reinstated later-but never across a transaction boundary. As a result, it is possible for the database to appear to be inconsistent, but only to the user whose as yet uncommitted transaction has given rise to that state of affairs. As a consequence of deferred constraint checking, SQL code that depends on consistency with declared constraints is obviously exposed to that assumption of consistency being false when the code is executed while checking is deferred.
For example, the table expression SELECT Name FROM IS_CALLED WHERE StudentId = 'S1' might be expected never to result in a table containing more than one row, thanks to the key constraint applying to IS_CALLED; thus it might be used in a scalar subquery. However, if the checking of that key constraint is temporarily deferred and two or more rows with StudentId equal to 'S1' temporarily appear in that table, then the scalar subquery will give rise to a run-time exception. Fortunately, SQL does allow a constraint to be declared as NOT DEFERRABLE, and that is the default option.