The term ACID conveys the role transactions play in mission-critical applications. Coined by transaction processing pioneers, ACID stands for atomicity, consistency, isolation, and durability.
These properties ensure predictable behavior, reinforcing the role of transactions as all-or-none propositions designed to reduce the management load when there are many variables.
A transaction is a unit of work in which a series of operations occur between the BEGIN TRANSACTION and END TRANSACTION statements of an application. A transaction executes exactly once and is atomic — all the work is done or none of it is.
Operations associated with a transaction usually share a common intent and are interdependent. By performing only a subset of these operations, the system could compromise the overall intent of the transaction. Atomicity eliminates the chance of processing a subset of operations.
A transaction is a unit of integrity because it preserves the consistency of data, transforming one consistent state of data into another consistent state of data.
Consistency requires that data bound by a transaction be semantically preserved. Some of the responsibility for maintaining consistency falls to the application developer who must make sure that all known integrity constraints are enforced by the application. For example, in developing an application that transfers money, you should avoid arbitrarily moving decimal points during the transfer.
A transaction is a unit of isolation — allowing concurrent transactions to behave as though each were the only transaction running in the system.
Isolation requires that each transaction appear to be the only transaction manipulating the data store, even though other transactions may be running at the same time. A transaction should never see the intermediate stages of another transaction.
Transactions attain the highest level of isolation when they are serializable. At this level, the results obtained from a set of concurrent transactions are identical to the results obtained by running each transaction serially. Because a high degree of isolation can limit the number of concurrent transactions, some applications reduce the isolation level in exchange for better throughput.
A transaction is also a unit of recovery. If a transaction succeeds, the system guarantees that its updates will persist, even if the computer crashes immediately after the commit. Specialized logging allows the system's restart procedure to complete unfinished operations, making the transaction durable.