Comment by blandflakes

I have a similar experience, though it's not necessarily related to shortcomings in the STM implementation (which I always felt was quite competent in the space). To expand, I think a few things are true:

1. I've come to feel that actually managing concurrency is a "below the fold" competency for most engineers. I'm not saying it should be, and people need to understand what a data race is. 2. As you pointed out, most state in services that I work on is not in need of coordinated updates, so atoms are where you want to be, or single-writer queues. 100% agreed 3. Structured concurrency approaches, fibers, and other abstracted runtimes seem to be the best way to get concurrent and parallel work, given points 1 and 2 about how coordinated state really isn't that common for most work on my team

A lot of times when we're updating state we're more having to think about it in the database, employing either a last-write-wins strategy or when coordination is in need, some form of pessimistic or optimistic locking.