Question

I can’t find anything discussing the case where there are multiple possible indexes that could be used to backup a FK constraint.

It seems from the test below that at FK creation time the FK gets bound to a specific index and this will always be used to validate the FK constraint irrespective of whether a new better index gets added later.

Can any one point to any resources confirming or denying this?

CREATE TABLE T1(
    T1_Id INT IDENTITY(1,1) PRIMARY KEY CLUSTERED  NOT NULL,
    Filler CHAR(4000) NULL,
) 

INSERT INTO T1 VALUES ('');


CREATE TABLE T2(
    T2_Id INT IDENTITY(1,1) PRIMARY KEY NOT NULL,
    T1_Id INT NOT NULL CONSTRAINT FK REFERENCES T1 (T1_Id), 
    Filler CHAR(4000) NULL,
)

/*Execution Plan uses clustered index - There is no NCI*/ 
INSERT INTO T2 VALUES (1,1) 

ALTER TABLE T1 ADD CONSTRAINT
    UQ_T1 UNIQUE NONCLUSTERED(T1_Id) 

/*Execution Plan still use clustered index even after NCI created*/    
INSERT INTO T2 VALUES (1,1) 

SELECT fk.name,
       ix.name,
       ix.type_desc
FROM sys.foreign_keys fk
JOIN sys.indexes ix ON ix.object_id = fk.referenced_object_id
AND ix.index_id = fk.key_index_id
WHERE fk.name = 'FK'


ALTER TABLE T2 DROP CONSTRAINT FK
ALTER TABLE T2  WITH CHECK ADD  CONSTRAINT FK FOREIGN KEY(T1_Id)
REFERENCES T1 (T1_Id)    

/*Now Execution Plan now uses non clustered index*/    
INSERT INTO T2 VALUES (1,1)    

SELECT fk.name,
       ix.name,
       ix.type_desc
FROM sys.foreign_keys fk
JOIN sys.indexes ix ON ix.object_id = fk.referenced_object_id
AND ix.index_id = fk.key_index_id
WHERE fk.name = 'FK'

DROP TABLE T2
DROP TABLE T1

Answer 1

Martin, I apologize that this isn’t much of an answer, but since you made me curious as well, I did some digging. The information that I can share is:

It is unlikely in current versions, including Denali, that an alternative would ever be considered in this situation.