In response to the query throughput bottlenecks and extreme query latency issues faced by current leading algorithms such as FreshDiskANN in query-update mixed load scenarios, we draw inspiration from the successful application of log-structured merge (LSM) trees in secondary indexes. This work introduces an LSM-based update-friendly disk vector index framework.

Building on the architecture of FreshDiskANN, we designed and implemented a three-layer architecture that includes a disk intermediate layer. We also introduced a dynamic mechanism for determining disk component search parameters, as well as a re-layout algorithm for the merge operation deletion phase. These innovations significantly reduce query latency and I/O overhead during the deletion phase.

The primary modifications in this work are located under the following directories:

• include/gp
• include/src
• src/gp
• src/lsm
• tests/lsm

In addition, some minor modifications were made to the original FreshDiskANN code base to accommodate the experimental setup.

To reproduce the experiments, please follow the environment setup instructions from the FreshDiskANN repository:
FreshDiskANN Repository

You can find the corresponding shell scripts for reproduction under the tests/lsm directory.