December 2020 Summaries

A major cable company initially built a virtual customer support agent on Amazon Aurora, but faced issues due to its single-master architecture and vulnerability to failures. The company sought a new database with always-on availability, strong consistency, and low-latency reads. They switched to CockroachDB for its resiliency, consistency, and low-latency reads, resulting in a multi-region deployment that can survive datacenter failures.

CockroachDB has introduced support for spatial data storage and processing, compatible with PostGIS. The new feature enables the use of spatial indexing in a horizontally scalable, dynamically sharded database. This is achieved by using a divide the space approach, which divides the space being indexed into buckets of various sizes. CockroachDB's implementation uses the S2 library from Google for cell covering and ID assignment. The new feature also includes two new distributed query processors, inverted filterer and inverted joiner, that apply to spatial SELECT and JOIN queries. These processors evaluate general set expressions derived from the expressions being evaluated and can be distributed for scalable evaluation.

In response to the high demand for spatial data and indexing capabilities, CockroachDB introduced spatial data storage in its 20.2 release, aligning its features with the popular PostGIS extension while maintaining CockroachDB's scalability and resilience. The development involved choosing a "divide the space" approach using Google's S2 library to handle spatial indexing, as opposed to PostGIS's "divide the objects" method, which was incompatible with CockroachDB's dynamic horizontal scaling. This approach leverages a quad-tree structure for indexing and utilizes distributed query processors to efficiently handle spatial queries by reducing false positives through precise evaluations. The blog post details the mechanics of spatial indexing, including the benefits of the chosen method, such as simplified bulk ingestion and memory-efficient compactions, and outlines future enhancements like geo-partitioned indexes and query optimizations for other data types.