Durability is one of the four key ACID attributes required to ensure the accurate and reliable operation of a transactional database. Durability refers to the ability to maintain database consistency and availability in the face of external problems, such as hardware or operating system failure. Durability is provided by four features of VoltDB: snapshots, command logging, K-safety, and disaster recovery through database replication.
Snapshots are a "snapshot" of the data within the database at a given point in time written to disk. You can use these snapshot files to restore the database to a previous, known state after a failure which brings down the database. The snapshots are guaranteed to be transactionally consistent at the point at which the snapshot was taken. Chapter 9, Saving & Restoring a VoltDB Database describes how to create and restore database snapshots.
Command Logging is a feature where, in addition to periodic snapshots, the system keeps a log of every stored procedure (or "command") as it is invoked. If, for any reason, the servers fail, they can "replay" the log on startup to reinstate the database contents completely rather than just to an arbitrary point-in-time. Chapter 10, Command Logging and Recovery describes how to enable, configure, and replay command logs.
K-safety refers to the practice of duplicating database partitions so that the database can withstand the loss of cluster nodes without interrupting the service. For example, a K value of zero means that there is no duplication and losing any servers will result in a loss of data and database operations. If there are two copies of every partition (a K value of one), then the cluster can withstand the loss of at least one node (and possibly more) without any interruption in service.
Database Replication is similar to K-safety, since it involves replicating data. However, rather than creating redundant partitions within a single database, database replication involves creating and maintaining a complete copy of the entire database. Database replication has a number of uses, but specifically in terms of durability, replication lets you maintain two copies of the database in separate geographic locations. In case of catastrophic events, such as fires, earthquakes, or large scale power outages, the replica can be used as a replacement for a disabled cluster.
Previous chapters described snapshots and command logging. The next chapter describes how you can use database replication for disaster recovery. This chapter explains how K-safety works, how to configure your VoltDB database for different values of K, and how to recover in the case of a system failure.
K-safety involves duplicating database partitions so that if a partition is lost (either due to hardware or software problems) the database can continue to function with the remaining duplicates. In the case of VoltDB, the duplicate partitions are fully functioning members of the cluster, including all read and write operations that apply to those partitions. (In other words, the duplicates function as peers rather than in a master-slave relationship.)
It is also important to note that K-safety is different than WAN replication. In replication the entire database cluster is replicated (usually at a remote location to provide for disaster recovery in case the entire cluster or site goes down due to catastrophic failure of some type).
In replication, the replicated cluster operates independently and cannot assist when only part of the active cluster fails. The replicate is intended to take over only when the primary database cluster fails entirely. So, in cases where the database is mission critical, it is not uncommon to use both K-safety and replication to achieve the highest levels of service.
To achieve K=1, it is necessary to duplicate all partitions. (If you don't, failure of a node that contains a non-duplicated partition would cause the database to fail.) Similarly, K=2 requires two duplicates of every partition, and so on.
What happens during normal operations is that any work assigned to a duplicated partition is sent to all copies (as shown in Figure 11.1, “K-Safety in Action”). If a node fails, the database continues to function sending the work to the unaffected copies of the partition.