Khronos: A Real-Time Indexing Framework for Time Series Databases on Large-Scale Performance Monitoring Systems

Time series databases play a critical role in large-scale performance monitoring systems. Metrics are required to be observable immediately after being generated to support real-time analysis. However, the commonly used Log-Structured Merge-Tree structure suffers from periodically visible delay spikes when a new segment is created due to the instantaneous index construction pressure.

In this paper, we present Khronos, an asynchronous indexing framework tailored for high-cardinal monitoring data, aiming at reducing the visible delay. Firstly, we analyze the temporal locality nature of time series and propose a complementary index construction algorithm by only indexing series not reported before to relieve indexing workload. Secondly, we design index structures based on Minimum Excluded value function to effectively reuse indexes of previous segments. Thirdly, we take advantage of the non-repetitive feature of complementary indexes and further develop an intermediate query results reusing approach for deduplicating index traversal among segments. Moreover, we propose an index dependency management strategy that cuts off the previous reusing dependency before persistence to avoid extended dependency overhead.

Experimental results show that our framework significantly reduces the visible delay from minutes to milliseconds. Khronos outperforms the state-of-the-art databases InfluxDB and TimeScaleDB with at least 4 times higher write throughput, hundreds of times lower visible delay, and 6 times lower query latency. Khronos has been deployed in production since 2020 and has become the largest performance monitoring database in Alibaba.