Time Series Management Systems: A Survey

Analysis

Why This Paper Matters

As the Internet of Things and Cyber-Physical Systems proliferate, the volume and velocity of time series data have grown dramatically. General-purpose Database Management Systems often struggle with the specific requirements of time series data, such as high write throughput, efficient time-based queries, and specialized compression. This survey is significant because it systematically organizes the fragmented landscape of Time Series Management Systems (TSMSs), providing a clear taxonomy based on architecture. For AI practitioners, understanding these systems is crucial because time series data is a primary input for many machine learning models, from predictive maintenance to anomaly detection. The paper's classification helps practitioners select the right TSMS for their workload and informs the design of data pipelines that feed AI models.

Technical Contributions

The paper's main technical contribution is its classification of TSMSs into architectural categories, which include:

• Native TSMSs: Systems built from the ground up for time series (e.g., InfluxDB, TimescaleDB).
• Extended DBMSs: General-purpose databases with time series extensions (e.g., PostgreSQL with TimescaleDB).
• Stream Processing Engines: Systems focused on real-time processing (e.g., Apache Flink, Spark Streaming).
• Hybrid Systems: Combining storage and stream processing capabilities.

The survey also provides a detailed comparison of each system's storage model (row vs. columnar), indexing techniques (e.g., B-trees, LSM-trees), query language support (SQL vs. NoSQL), and support for Approximate Query Processing (AQP) and stream processing. This structured comparison is a valuable resource for system designers.

Results

The survey does not present new experimental results but synthesizes findings from existing literature. It notes that native TSMSs typically achieve higher write throughput and better compression than extended DBMSs, often by an order of magnitude. For example, systems like InfluxDB can handle millions of data points per second on commodity hardware. The paper also highlights that AQP capabilities are still nascent in most TSMSs, with only a few systems offering sampling or sketching techniques. The summary of research directions points to the need for better integration of machine learning for query optimization and adaptive storage.

Significance

This survey has had a lasting impact on the field by providing a common vocabulary and framework for discussing TSMSs. It has been cited over 200 times, indicating its role as a standard reference. For AI practitioners, the paper's vision for next-generation TSMS—incorporating machine learning for self-tuning, predictive caching, and automated data lifecycle management—directly aligns with the trend toward AI-driven data infrastructure. The survey also underscores the importance of stream processing for real-time AI applications, where low-latency ingestion and querying are critical. By clarifying the trade-offs between different system designs, this work helps practitioners make informed decisions when building time series pipelines for AI workloads.