Large-scale transaction processing systems must balance the need for minimal response times with maintaining the integrity of the data being processed. In theory, transactions are processed sequentially in the order received. Very large systems exceed the capacity of even a supercomputer; therefore, the data processing must be divided among multiple systems. Consider the classic example: an airline reservation system. Multiple agents globally enter hundreds of thousands (or millions) of transactions daily, linking customers to available seats on specific flights. The traditional way to handle such a huge transaction load takes two forms. First, the total database is divided into "shards," that is, segments that are parts of the whole but can be treated independently. Within each shard there may be replicas, that is, multiple instances of the same data in that shard. The redundancy ensures the integrity of the data. In this complex system there remains the requirement of apparent serial processing, that is, transactions handled in chronological sequence. This implies coordination and synchronization among the replicas in a shard because their data cannot be treated independently. A strict synchronization protocol further implies fairly large overhead as replicas communicate with each other to maintain the integrity of the process.
展开▼
