TS3D: A Temporal Multimodal Dataset for Distributed Database System Analysis

Distributed databases have become foundational infrastructure across a wide range of industries. Unfortunately, the strict data privacy and security regulations often prevent practitioners from accessing real operational data, limiting the reliability of their analysis. While a few public log datasets for distributed databases are available, relying solely on them for downstream tasks still introduces performance bottlenecks. A major limitation is that they typically provide either textual logs or temporal values, but rarely both, making it difficult to fully understand system behavior. Being able to leverage the multimodal information inherent in distributed databases can drastically improve the accuracy of downstream analysis. In this paper, we propose TS3D, a temporal multimodal dataset derived from a distributed database. The dataset is the largest to date, with 300 million numerical data points and 90 million contextual data records. We showcase its utility through two representative downstream tasks. First, we propose a multimodal anomaly detection framework that fuses logs and numerical time series data, thereby enhancing both existing log-based models and time series-based models, and we design a hybrid evaluation metric capable of jointly assessing both point-level and segment-level anomalies. Second, we present a SQL-level analysis framework that constructs temporal-dependency and table-dependency graphs to trace anomaly root causes and performs pod-level SQL optimization across a database cluster. We further show how to formally validate these auxiliary methods derived from the dataset. Our empirical results show that, using our multimodal dataset, the efficacy of distributed anomaly detection improves by up to 26% while distributed root cause localization improves by up to 35%.