Users modifying data can affect other users who are reading or modifying the same data at the same time. These users are said to be accessing the data concurrently. If a data storage system has no concurrency control, users could see the following side effects:
Uncommitted dependency (dirty read)
Inconsistent analysis (nonrepeatable read)
Missing and double reads caused by row updates
Lost updates occur when two or more transactions select the same row and then update the row based on the value originally selected. Each transaction is unaware of the other transactions. The last update overwrites updates made by the other transactions, which results in lost data.
For example, two editors make an electronic copy of the same document. Each editor changes the copy independently and then saves the changed copy thereby overwriting the original document. The editor who saves the changed copy last overwrites the changes made by the other editor. This problem could be avoided if one editor could not access the file until the other editor had finished and committed the transaction.
Uncommitted dependency occurs when a second transaction selects a row that is being updated by another transaction. The second transaction is reading data that has not been committed yet and may be changed by the transaction updating the row.
For example, an editor is making changes to an electronic document. During the changes, a second editor takes a copy of the document that includes all the changes made so far, and distributes the document to the intended audience. The first editor then decides the changes made so far are wrong and removes the edits and saves the document. The distributed document contains edits that no longer exist and should be treated as if they never existed. This problem could be avoided if no one could read the changed document until the first editor does the final save of modifications and commits the transaction.
Inconsistent analysis occurs when a second transaction accesses the same row several times and reads different data each time. Inconsistent analysis is similar to uncommitted dependency in that another transaction is changing the data that a second transaction is reading. However, in inconsistent analysis, the data read by the second transaction was committed by the transaction that made the change. Also, inconsistent analysis involves multiple reads (two or more) of the same row, and each time the information is changed by another transaction; thus, the term nonrepeatable read.
For example, an editor reads the same document twice, but between each reading the writer rewrites the document. When the editor reads the document for the second time, it has changed. The original read was not repeatable. This problem could be avoided if the writer could not change the document until the editor has finished reading it for the last time.
Phantom reads occur when an insert or delete action is performed against a row that belongs to a range of rows being read by a transaction. The transaction's first read of the range of rows shows a row that no longer exists in the second or succeeding read as a result of a deletion by a different transaction. Similarly, the transaction's second or succeeding read shows a row that did not exist in the original read as the result of an insertion by a different transaction.
For example, an editor makes changes to a document submitted by a writer, but when the changes are incorporated into the master copy of the document by the production department, they find that new unedited material has been added to the document by the author. Similar to the nonrepeatable read situation, this problem could be avoided if no one could add new material to the document until the editor and production department finish working with the original document.
Missing a updated row or seeing an updated row multiple times
Transactions that are running at the READ UNCOMMITTED level do not issue shared locks to prevent other transactions from modifying data read by the current transaction. Transactions that are running at the READ COMMITTED level do issue shared locks, but the row or page locks are released after the row is read. In either case, when you are scanning an index, if another user changes the index key column of the row during your read, the row might appear again if the key change moved the row to a position ahead of your scan. Similarly, the row might not appear if the key change moved the row to a position in the index that you had already read. To avoid this, use the SERIALIZABLE or HOLDLOCK hint, or row versioning. For more information, see Table Hints (Transact-SQL) and Row Versioning-based Isolation Levels in the Database Engine.
Missing one or more rows that were not the target of update
When you are using READ UNCOMMITTED, if your query reads rows using an allocation order scan (using IAM pages), you might miss rows if another transaction is causing a page split. This cannot occur when you are using read committed because a table lock is held during a page split and does not happen if the table does not have a clustered index, because updates do not cause page splits.