Bridging the gap between algorithmic and learned index structures

Index structures such as B-trees and bloom filters are the well-established petrol engines of database systems. However, these structures do not fully exploit patterns in data distribution. To address this, researchers have suggested using machine learning models as electric engines that can entirely replace index structures. Such a paradigm shift in data system design, however, opens many unsolved design challenges. More research is needed to understand the theoretical guarantees and design efficient support for insertion and deletion. In this thesis, we adopt a different position: index algorithms are good enough, and instead of going back to the drawing board to fit data systems with learned models, we should develop lightweight hybrid engines that build on the benefits of both algorithmic and learned index structures. The indexes that we suggest provide the theoretical performance guarantees and updatability of algorithmic indexes while using position prediction models to leverage the data distributions and thereby improve the performance of the index structure. We investigate the potential for minimal modifications to algorithmic indexes such that they can leverage data distribution similar to how learned indexes work. In this regard, we propose and explore the use of helping models that boost classical index performance using techniques from machine learning. Our suggested approach inherits performance guarantees from its algorithmic baseline index, but at the same time it considers the data distribution to improve performance considerably. We study single-dimensional range indexes, spatial indexes, and stream indexing, and show that the suggested approach results in range indexes that outperform the algorithmic indexes and have comparable performance to the read-only, fully learned indexes and hence can be reliably used as a default index structure in a database engine. Besides, we consider the updatability of the indexes and suggest solutions for updating the index, notably when the data distribution drastically changes over time (e.g., for indexing data streams). In particular, we propose a specific learning-augmented index for indexing a sliding window with timestamps in a data stream. Additionally, we highlight the limitations of learned indexes for low-latency lookup on real- world data distributions. To tackle this issue, we suggest adding an algorithmic enhancement layer to a learned model to correct the prediction error with a small memory latency. This approach enables efficient modelling of the data distribution and resolves the local biases of a learned model at the cost of roughly one memory lookup.Open Acces