article --summary
The summary
The author describes a credit monitoring company whose core jobs once ran for a day or two, by which point the data was already stale. He lays out how he defined the real workload, including millions of monthly queries and credit data that could not be cached, and why that rules out any option relying on caching.
He then explains how the team trialed several candidate databases and why most failed on fit, while SingleStore worked with their existing MySQL-compatible code and handled both analytical and transactional load. Before committing, they ran simulations with fake data at production scale.
The cutover was done by queuing requests, swapping the database, and replaying the queue without dropping anything. Jobs that once took 24 to 48 hours then finished in about 15 minutes, and throughput rose from 60 to 1,000 requests per minute.
Editorial note: This is a summary of an article first published by guywarner.dev. The full piece is on their site.