Transactions are critical to maintaining data integrity. They are used to group operations into units of work that act in an all-or-nothing fashion. Transactions have the following qualities:
Atomicity. Atomicity refers to the all-or-nothing property of transactions. Either every data update in the transaction completes successfully, or they all fail, leaving the data store in its original state. A transaction cannot be only partially successful.
Consistency. Each transaction takes the data store from one consistent state to another consistent state.
Isolation. Transactions are isolated from each other. When you are reading persistent data in one transaction, you cannot "see" the changes that are being made to that data in other uncompleted transactions. Similarly, the updates you make in one transaction cannot conflict with updates made in other concurrent transactions. The form of conflict resolution employed depends on whether you are using pessimistic or optimistic transactions. Both types are described later in this chapter.
Durability. The effects of successful transactions are durable; the updates made to persistent data last for the lifetime of the data store.
Together, these qualities are called the ACID properties of transactions. To understand why these properties are so important to maintaining data integrity, consider the following example:
Suppose you create an application to manage bank accounts. The application includes a method to transfer funds from one user to another, and it looks something like this:
public void transferFunds (User from, User to, double amnt)
{
from.decrementAccount (amnt);
to.incrementAccount (amnt);
}
Now suppose that user Alice wants to transfer 100 dollars to user Bob. No problem; you simply invoke your transferFunds method, supplying Alice in the from parameter, Bob in the to parameter, and 100.00 as the amnt. The first line of the method is executed, and 100 dollars is subtracted from Alice's account. But then, something goes wrong. An unexpected exception occurs, or the hardware fails, and your method never completes.
You are left with a situation in which the 100 dollars has simply disappeared. Thanks to the first line of your method, it is no longer in Alice's account, and yet it was never transferred to Bob's account either. The data store is in an inconsistent state.
The importance of transactions should now be clear. If the two lines of the transferFunds method had been placed together in a transaction, it would be impossible for only the first line to succeed -- either the funds would be transferred properly or they would not be transferred at all and an exception would be thrown. Money could never vanish into thin air; the data store could never get into an inconsistent state.
There are two major types of transactions: data store, or pessimistic, transactions and optimistic transactions. Each type has both advantages and disadvantages.
Pessimistic transactions generally lock the data store records they act on, preventing other concurrent transactions from using the same data. This avoids conflicts between transactions, but consumes a lot of database resources. Additionally, locking records can result in deadlock, a situation in which two transactions are both waiting for the other to release its locks before completing. The results of a deadlock are data store-dependent; usually one transaction is forcefully rolled back after some specified time out interval, and an exception is thrown.
Optimistic transactions consume less resources than pessimistic transactions, but only at the expense of reliability. Because optimistic transactions do not lock data store records, two transactions might change the same persistent information at the same time, and the conflict will not be detected until the second transaction attempts to commit. At this time, the second transaction will realize that another transaction has concurrently modified the same records (usually through a timestamp or versioning system), and will throw an appropriate exception. Note that optimistic transactions still maintain data integrity; they are simply more likely to fail in heavily concurrent situations.