Microsoft SQL Server 2008 Internals: Transactions and Concurrency
- By Kalen Delaney, Conor Cunningham, Adam Machanic, Ben Nevarez, Paul S. Randal, Kimberly L. Tripp
Concurrency can be defined as the ability of multiple processes to access or change shared data at the same time. The greater the number of concurrent user processes that can be active without interfering with each other, the greater the concurrency of the database system.
Concurrency is reduced when a process that is changing data prevents other processes from reading that data or when a process that is reading data prevents other processes from changing that data. I use the terms reading or accessing to describe the impact of using the SELECT statement on your data. Concurrency is also affected when multiple processes attempt to change the same data simultaneously and they cannot all succeed without sacrificing data consistency. I use the terms modifying, changing, or writing to describe the impact of using the INSERT, UPDATE, MERGE, or DELETE statements on your data. (Note that MERGE is a new data modification statement in SQL Server 2008, and you can think of it as a combination of INSERT, UPDATE, and DELETE.)
In general, database systems can take two approaches to managing concurrent data access: optimistic or pessimistic. Microsoft SQL Server 2008 supports both approaches. Pessimistic concurrency was the only concurrency model available before SQL Server 2005. As of SQL Server 2005, you specify which model to use by using two database options and a SET option called TRANSACTION ISOLATION LEVEL.
After I describe the basic differences between the two models, we look at the five possible isolation levels in SQL Server 2008, as well as the internals of how SQL Server controls concurrent access using each model. We look at how to control the isolation level, and we look at the metadata that shows you what SQL Server is doing.
In either concurrency model, a conflict can occur if two processes try to modify the same data at the same time. The difference between the two models lies in whether conflicts can be avoided before they occur or can be dealt with in some manner after they occur.
With pessimistic concurrency, the default behavior is for SQL Server to acquire locks to block access to data that another process is using. Pessimistic concurrency assumes that enough data modification operations are in the system that any given read operation is likely affected by a data modification made by another user. In other words, the system behaves pessimistically and assumes that a conflict will occur. Pessimistic concurrency avoids conflicts by acquiring locks on data that is being read, so no other processes can modify that data. It also acquires locks on data being modified, so no other processes can access that data for either reading or modifying. In other words, readers block writers and writers block readers in a pessimistic concurrency environment.
Optimistic concurrency assumes that there are sufficiently few conflicting data modification operations in the system that any single transaction is unlikely to modify data that another transaction is modifying. The default behavior of optimistic concurrency is to use row versioning to allow data readers to see the state of the data before the modification occurs. Older versions of data rows are saved, so a process reading data can see the data as it was when the process started reading and not be affected by any changes being made to that data. A process that modifies the data is unaffected by processes reading the data because the reader is accessing a saved version of the data rows. In other words, readers do not block writers and writers do not block readers. Writers can and will block writers, however, and this is what causes conflicts. SQL Server generates an error message when a conflict occurs, but it is up to the application to respond to that error.