Robustness Analysis of Distributed Databases via Fault Injection

As the importance of reliable data storage increases, various techniques are applied to obtain the reliabilities. The usage of the assumed fault tolerant cloud providers increases, even though not a sufficient amount of research is performed on their fault tolerance. On cloud servers, a distributed Database Management System (DBMS) can be used to achieve safe data storage. Various DBMSs are making claims about its fault tolerance, a measure of robustness, they provide. However, the claims are rarely validated. This research targets to validate the fault tolerance of the DBMS. The research scope is limited to test a DBMS on its data integrity, if only original inserted data is returned to the client. The research focusses on verifying the claimed fault tolerance of the Apache Cassandra DBMS, a popular, fault tolerant column store DBMS. A fault injection framework is proposed and implemented to create a suitable and flexible test environment. Fault injection experiments on the DBMS are performed with and without replication. Overall, Cassandra can be considered fault tolerant. When targeting data files only, most faults are detected immediately and no invalid data is returned to the client. Modified system files did not return an immediate result. When modifying the index files, faults were undetected in the short term analyses of the experiments, resulting in missing data both with and without replication. However, replication did reduce the number of faults returned by Cassandra.