java - Is MERGE an atomic statement in SQL2008?

Question

I am using a MERGE statement as an UPSERT to either add a new record or update the current one. I have multiple threads driving the database through multiple connections and multiple statements (one connection and statement per thread). I am batching the statements 50 at a time.

I was very surprised to get a duplicate key violation during my tests. I expected that to be impossible because the MERGE will be performed as a single transaction, or is it?

My Java code looks like:

private void addBatch(Columns columns) throws SQLException {
  try {
    // Set parameters.
    for (int i = 0; i < columns.size(); i++) {
      Column c = columns.get(i);
      // Column type is an `enum` with a `set` method appropriate to its type, e.g. setLong, setString etc.
      c.getColumnType().set(statement, i + 1, c.getValue());
    }
    // Add the insert as a batch.
    statement.addBatch();
    // Ready to execute?
    if (++batched >= MaxBatched) {
      statement.executeBatch();
      batched = 0;
    }
  } catch (SQLException e) {
    log.warning("addBatch failed " + sql + " thread " + Thread.currentThread().getName(), e);
    throw e;
  }
}

The query looks like this:

MERGE INTO CustomerSpend AS T 
USING ( SELECT ? AS ID, ? AS NetValue, ? AS VoidValue ) AS V 
ON T.ID = V.ID 
WHEN MATCHED THEN 
    UPDATE SET T.ID = V.ID, T.NetValue = T.NetValue + V.NetValue, T.VoidValue = T.VoidValue + V.VoidValue 
WHEN NOT MATCHED THEN 
    INSERT ( ID,NetValue,VoidValue ) VALUES ( V.ID, V.NetValue, V.VoidValue );

The error reads:

java.sql.BatchUpdateException: Violation of PRIMARY KEY constraint 'PK_CustomerSpend'. Cannot insert duplicate key in object 'dbo.CustomerSpend'. The duplicate key value is (498288              ).
at net.sourceforge.jtds.jdbc.JtdsStatement.executeBatch(JtdsStatement.java:944)
at x.db.Db$BatchedStatement.addBatch(Db.java:299)
...

The key on the table is a PRIMARY key on the ID field.

Answer 1

MERGE is atomic meaning that either all changes are committed or all changes are rolled back.

It does not prevent duplicate keys in case of high concurrency. Adding holdlock hint will take care of that.

MERGE INTO CustomerSpend WITH (HOLDLOCK) AS T 
USING ( SELECT ? AS ID, ? AS NetValue, ? AS VoidValue ) AS V 
ON T.ID = V.ID 
WHEN MATCHED THEN 
    UPDATE SET T.ID = V.ID, T.NetValue = T.NetValue + V.NetValue, T.VoidValue = T.VoidValue + V.VoidValue 
WHEN NOT MATCHED THEN 
    INSERT ( ID,NetValue,VoidValue ) VALUES ( V.ID, V.NetValue, V.VoidValue );