High Availability

DVTDS is designed for uninterrupted operation in 24/7 mode. This is achieved by a number of essential features:

• The whole administration can be done in online mode. There is no downtime for OAM tasks like schema changes, performance tuning, backup and restore or storage management.
• The server maintains a number of copies. It replicates all data changes automatically to other locations. This is based on configuration and dynamic tuning.
• Replication is configurable per category of data: What to copy. Where to copy. At which point in time.
• Synchronous and asynchronous replication are both supported.
• Multi master paradigm with configurable number of copies.
• Automatic resolution of distributed conflicts in real time.

DVTDS nodes can be configured to execute incoming update requests locally and replicate them to neighboring nodes. The neighborhood configuration can be symmetrically, this means: Any node replicates any update request to any other node in the compound thus setting up a fully meshed high availability network. With the exception of the target node IP address a replication compound appears from the client point of view as a single database. The client always sees the same content regardless which node it connects to. Any particular object has the same layout and content as the corresponding object in the other nodes of the mirrored system. From a logical point of view it is "the same". From the physical point of view it has of course a different memory address. Transaction semantics are fully supported and the ACID paradigm is ensured across all replicas.

When client requests are sent in a bundle the WAN latency shows up only once for the bundle instead of affecting every single request. This also holds true for transactions, making them almost as fast as normal requests. See our intercontinental replication benchmark for details about this technique. There we reached more than one million updates per second fully synchronized between sites located on three different continents.

In the event that clients apply different updates to the same object on different nodes we have a distributed conflict. DVTDS resolves such conflicts in real time by applying precedence rules based on the request type, the user that issued the requests and the server number of the affected nodes. Conflict resolving is internally supported by a combination of two phase commit and two phase locking.