There is a whole chapter in Designing Data Intensive Applications dedicated specifically to handling atomic transaction concurrency issues while avoiding the performance trade-off you're describing here, you will definitely find it interesting.

https://www.oreilly.com/library/view/designing-data-intensiv...