Compress Time Series with Smaller Error Tolerances

Despite the wealth of time series data collected across various fields, its sheer volume presents significant challenges for efficient usage and management in database systems. By compressing time series data within a predefined error tolerance for each individual data point, it can greatly boost data storage capacity, accelerate data transmission rates, and, most importantly, enable high-quality analytical tasks. In this paper, we introduce novel algorithms for compressing floating-point time series data, ensuring that each decoded data point stays within a predefined error bound. Our proposed methods are based on XOR encoding, a widely adopted approach used by leading lossless floating-point compression techniques, and are further optimized by selecting either the “shortest” XORed expressions or the shortest bits on “mantissa+flags” the permissible tolerance range. Extensive experiments on widely used datasets demonstrate that our algorithms significantly outperform many state-of-the-art max-error bound compression algorithms, in both compression ratio and execution time. Specifically, for smaller tolerable errors, our algorithms achieve compressed data sizes that are approximately 45% and 65% of those produced by leading lossless and lossy floating-point compression algorithms respectively, while incurring only 75% and 30% of the execution costs in compression and decompression when compared to those competitors.