MV-IDX: Multi-Version Index in Action

Multi-Versioning DBMS (MV-DBMS) represent a very good match to the properties of Flash storage and the combination of both offers conceptual advantages. Yet, the specifics of indexing in MV-DBMS on Flash have been widely neglected. Although an index in a MV-DBMS references multiple versions of a data item, it is only allowed to return a single (at most one) version of that data item ”visible” to the current index operation. Logically separating version visibility checks from the index structure and operations, as in the traditional version-oblivious index, leads to version management overhead: to determine the appropriate version of a data item, the MV-DBMS first fetches all versions that match the search criteria and subsequently discards invisible versions according to the visibility criteria. This involves unnecessary I/Os to fetch tuple versions that do not need to be checked. We propose the idea that version-aware indexing has additional responsibility to recognize different tuple versions of a single data item and to filter invisible tuple versions in order to avoid unnecessary I/Os. In this work we demonstrate an approach called Multi-Version Index (MV-IDX) that allows index-only visibility checks which significantly reduce the amount of I/O as well as the index maintenance overhead. MV-IDX is implemented in the PostgreSQL open source MV-DBMS. We demonstrate that the MV-IDX achieves significantly lower response times and higher transactional throughput on OLTP workloads than the version-oblivious approach. We showcase latency and throughput improvements by utilizing the DBT2 TPC-C benchmarking tool and report saved I/Os. We also showcase how the proposed approach performs better on SSDs.