Consider a simple banking application where two users have access to the funds in a particular account. Bob reads the account and finds there is 1000 dollars in it, so he withdraws 500. Jane reads the same account before Bob has changed it, sees 1000 dollars, and withdraws 800. The account should be 300 dollars overdrawn, however, depending on which transaction gets processed first it will contain either 500 or 200 dollars. This poses a serious problem and needless to say, any database system with multi-user access needs some sort of system to deal with these scenarios.
The techniques used to solve this and other related problems are known in the database industry as concurrency control.
Traditional products used locks which stated that a particular transaction was going to modify a record. Once the lock was placed, no one else could read or modify the data until the lock was released. The lock may block changes to a single record, a page (a group of records stored together on disk) of records, or every record examined by a particular transaction, depending on the lock resolution. Lock resolution is a tradeoff between performance and accuracy—by blocking updates at the page level, for example, some updates will be blocked which do not in fact conflict with updates made by other transactions, but performance will be improved in comparison with record level locks.
Locking becomes an even bigger problem when combined with another feature common to all such systems, isolation. This is because transactions typically involve both a read and a write—in this example, to read the value of the account and then change it. In order to show an isolated view of the data the entire transaction, including records read but never written to, must be locked in many database servers.
In InterBase, readers do not block writers. Instead, each record in the database can exist in more than one version. For instance, when Bob and Jane read the accounts they would both get "version 1", reading 1000 dollars. When Bob then changes the account to make his withdrawal the data is not overwritten, but instead a new "version 2" will be created with 500 dollars. Jane's attempt to make her 800 dollar withdrawal will notice that there is a new version 2, and her attempt to make a withdrawal will fail.
This approach to concurrency control is called multiversion concurrency control. InterBase's implementation of multiversion concurrency control is commonly called its multi-generational architecture. InterBase was the second commercial database to use this technique; the first was DEC's Rdb/ELN.
Multiversion concurrency control also makes true snapshot transaction isolation relatively simple to implement. A transaction with snapshot isolation in InterBase shows the state of the database precisely as it was at the instant the transaction began. This is very useful for backups of an active database, long-running batch processes, and the like.