Open Collections

Abstract

To meet the demand for increasingly accurate sensor monitoring and forecasting, time series datasets have grown to take larger, more detailed samples with more frequent sampling rates. As a result, the size of time series datasets has grown larger and now require more efficient indexing methods to manage properly. Time series databases have unique traits that allow for efficient indexing structures to be used. The timestamps are always increasing in an append-only fashion, a characteristic which can be exploited to create more efficient indexing structures. In this research, two different indexing algorithms for time series databases are evaluated. The spline index model uses existing points of the time series data to form a series of linear approximations. The other index model examined is the piece-wise geometric model (PGM), which forms fully independent lines that approximate the underlying time series data. Experimental results show both the Spline and PGM learned indexes outperform conventional indexes for time series data. Performance metrics for binary search and simpler single line approximations are also included for comparison.