Distributed databases

A hotel chain has hotels all over the world. One central database means a receptionist in Tokyo waits on a server in London, and one broken link takes everyone down. A distributed database splits the data across several sites on a network, yet still works as one database to the people and programs that use it.

• Fragmentation: the database is cut into pieces and each piece is stored at one site (for example each hotel's own bookings). Together the pieces make the whole.
• Replication: copies of the data are kept at more than one site (or the whole database at every site). Data stays available if a site fails, and reads are fast and local, but every change has to reach every copy.
• Partitioning is how the pieces are chosen. Range partitioning uses ranges of a key (bookings by date). Hash partitioning applies a hash function to the key for an even, unpredictable spread. List partitioning names which values go where (country to site). Round-robin deals rows out in turn when there is no natural key.

Replication can be arranged in different ways: a master database takes all the changes and passes them on to the others, or several masters all accept changes and share them, or every site is a peer of the others.

• Concurrency control. Many users at many sites change data at once, so the system must keep to the ACID properties (A3.3.6). Pessimistic control locks data before use, assuming clashes will happen. Optimistic control does not lock but checks for a clash at the moment of saving, assuming they are rare.
• Data consistency. All copies must agree on the latest values. Strong consistency makes every update visible everywhere at once; weaker models allow copies to be briefly out of step in return for speed.
• Transparency. Users should not need to know where data is (distribution transparency), that copies exist (replication transparency) or that a site failed (failure transparency): another takes over unnoticed.
• Fault tolerance. Replication means the system carries on when parts fail.
• Security. More sites and more network traffic mean more to protect: encrypt stored and moving data, use roles and permissions, and use strong authentication such as multi-factor.

Distributed databases can be homogeneous (every site uses the same database system and schema, which is easier to manage) or heterogeneous (different systems or schemas, needing translation between them).

Advantages: faster processing, because the work and data are spread and kept near the users; easier expansion, by adding sites; greater reliability, through replication; and local control of local data.

Disadvantages: complexity, particularly consistency and transactions that cross sites; dependence on a fast, stable network; extra storage and update effort for replication; and a larger attack surface.