Persistent Data Structures for Incremental Join Indices

Join indices are used in relational databases to make join operations faster. Join indices essentially materialise the results of join operations and so accrue maintenance cost, which makes them more suitable for use cases where modifications are rare and joins are performed frequently. To make the maintenance cost lower incrementally updating existing indices is to be preferred. The usage of persistent data structures for the join indices were explored. Motivation for this research was the ability of persistent data structures to construct multiple partially different versions of the same data structure memory efficiently. This is useful, because there can exist different versions of join indices simultaneously due to usage of multi-version concurrency control (MVCC) in a database. The techniques used in Relaxed Radix Balanced Trees (RRB-Trees) persistent data structure were found promising, but none of the popular implementations were found directly suitable for the use case. This exploration was done from the context of a particular proprietary embedded in-memory columnar multidimensional database called FastormDB developed by RELEX Solutions. This focused the research into Java Virtual Machine (JVM) based data structures as the implementation of FastormDB is in Java. Multiple persistent data-structures made for the thesis and ones from Scala, Clojure and Paguro were evaluated with Java Microbenchmark Harness (JMH) and Java Object Layout (JOL) based benchmarks and their results analysed via visualisations