Transactions
Databases approach errors in a completely different way. Instead of making error recovery explicit, databases make error recovery implicit through the use of transactions; see An Introduction to Database Systems by C.J. Date (Addison-Wesley, 1991).
In this Transact-SQL example, the transaction begins with BEGIN TRAN T1 and ends with COMMIT TRAN T1:
BEGIN TRAN T1 UPDATE table1 ... UPDATE table2 ... SELECT * from table1 UPDATE table3 ... COMMIT TRAN T1
Transactions are units of work that either completely succeed, or can be rolled back completely. If an error such as a disk full, a key violation, or user error occurs during the transaction, it simply rolls back (undoes) all of the operations to the beginning of the transaction. This is safe and robust because the data is guaranteed to be left in its initial consistent state. While recovering from exceptions is difficult in C++, it is extremely easy in databases with transactions.
The Atomic Library (calumgrant.net/atomic) I present here implements transactions in C++, based on database transactions. Here is the same example using a transaction:
atomic::vector<std::string> list1, list2;
void f(const std::string &s)
{
atomic::transaction tr;
list1.push_back(s);
list2.push_back(s);
tr.commit();
}
The transaction tr defines the scope of the transaction, namely the entire function f(). list1 and list2 are of type atomic::vector, which is a special container that interacts with atomic::transaction. If f() exits without calling tr.commit(), then all of the changes in f() are undone.
The Atomic C++ Library for transactional programming provides a range of containers that can be rolled back to any point in their past by transactions.
Transactional containers are designed in such a way that deletions can be safely undone without the possibility of failure when the container is rolled back, even if it involves inserting data back into the data structure.
