SPST-Index: A Self-Pruning Splay Tree Index for Caching Database Cracking

In database cracking, a database is physically self-organized into cracked partitions with cracker indices boosting the access to these partitions. The AVL Tree is the data structure of choice to implement cracker indices. However, it is particularly cache-inefficient for range queries, because the nodes accessed only for a few times (i.e, “Cold Data”) and the most accessed ones (i.e, “Hot Data”) are spread all over the index. In this paper, we present the Self-Pruning Splay Tree (SPST) data structure to index database cracking and reorganize “Hot Data” and “Cold Data” to boost the access to the cracked partitions. To every range query, the SPST rotates to the root the nodes pointing to the edges and to the middle value of the predicate interval. Eventually, the most accessed tree nodes remain close to the root improving CPU and cache activity. On the other hand, the least accessed tree nodes remain close to the leaves and are pruned to improve updates. Our experimental evaluation shows 37% more Instructions per Cycle and 75.9% less cache misses in L1 for lookup operations in the SPST compared to the AVL tree. Our data structure outperforms the AVL tree for lookups and maintenance costs in three major data access patterns: random, sequential and skewed. The SPST outperforms the AVL in 4% even in the worst case scenario with mixed workloads with lookups and batch updates.