MSG 666 when running an insert query on 80M-row indexed table

Strangely, my stored procedure started to receive Msg 666 for some input data.

Stored procedure fails on the last step when it tries to insert a row into a table with the following structure:

Columns:
A_Id: PK, int
B_Id: PK, FK, int
C_Id: PK, FK, int
D_Id: PK, smallint

This is essentially a table that connects all referenced entities together.

Indexes:
IX_TableName_D_id - Clustered index on D_id column
PK_TableName - Unique non-clustered index on all columns (A_Id, B_Id, C_Id, D_Id)

Fragmentation for both indexes is low (<25%). However PK_TableName fragmentation quickly grows, since the amount of operation on the table is quite intense.

Table size:

Row count: ~80,000,000 rows

So, when I try to run a veeery simple query, for some of D_Id's I get the following message:


Msg 666. The maximum system-generated unique value for a duplicate group was
exceeded for index with partition ID 422223771074560. Dropping and
re-creating the index may resolve this; otherwise, use another
clustering key.

Query example:

INSERT INTO TableName
(A_Id,B_Id,C_Id,D_id)
VALUES (1,1,1,14)

For example, when I set D_Id to some values - it fails, '14' for example. If I set D_ID to other values (1,2,3,...13, 15,16,...), the query runs fine.

I suspect there's something really bad going on with indexes... But I cannot get to the bottom of this... :( Why it fails?

The low selectivity issue mentioned by Remus is not sufficient on its own to cause the problem on that size table.

The uniqueifier starts at 1 and can go up to 2,147,483,646 before actually overflowing the range.

It also requires the right pattern of repeated deletes and inserts to see the issue.

CREATE TABLE T
(
X SMALLINT,
Y INT IDENTITY PRIMARY KEY NONCLUSTERED
)

CREATE CLUSTERED INDEX IX ON T(X)

INSERT INTO T VALUES (1),(1),(1),(2),(2)

Gives

+---+---+-------------+
| X | Y | Uniqueifier |
+---+---+-------------+
| 1 | 1 |             |
| 1 | 2 |           1 |
| 1 | 3 |           2 |
| 2 | 4 |             |
| 2 | 5 |           1 |
+---+---+-------------+

Then running

DELETE FROM T 
WHERE Y IN (2,3)

INSERT INTO T VALUES (1),(1)

Gives

+---+---+-------------+
| X | Y | Uniqueifier |
+---+---+-------------+
| 1 | 1 |             |
| 1 | 6 |           3 |
| 1 | 7 |           4 |
| 2 | 4 |             |
| 2 | 5 |           1 |
+---+---+-------------+

Showing in that case the uniqueifier did not reuse the values from the deleted rows.

However then running

DELETE FROM T 
WHERE Y IN (6,7)
WAITFOR DELAY '00:00:10'
INSERT INTO T VALUES (1),(1)

Gives

+---+---+-------------+
| X | Y | Uniqueifier |
+---+---+-------------+
| 1 | 1 |             |
| 1 | 8 |           1 |
| 1 | 9 |           2 |
| 2 | 4 |             |
| 2 | 5 |           1 |
+---+---+-------------+

Showing that the high water mark can be reset after deleting the duplicate with the highest uniqueifier value. The delay was to allow the ghost record cleanup process to run.

Because life is too short to insert 2 billion duplicates I then used DBCC WRITEPAGE to manually adjust the highest uniqueifier to 2,147,483,644

I then ran

INSERT INTO T VALUES (1)

multiple times. It succeeded twice and failed on the third attempt with error 666.

This was actually one lower than I would have assumed. Meaning that the highest uniqueifier inserted was 2,147,483,646 rather than the maximum int size of 2,147,483,647

StackExchange Database Administrators Q#49430, answer score: 18