But it does not fix the issue of e.g. having used a quadratic algorithm which is breaking your server because in production a user is shoving an order of magnitude more data than you expected (or tested for): you used that algorithm in-request because it was fast. Now it's not fast anymore and your production is getting killed. That's all there is to it, until you fix your code the application is not degraded, it's DOS'd every time that user does something.

And if you do consider workers (because you don't have a better algorithm), what happens for the cheap version? Do you offload it to a worker as well, potentially incurring a spin-off cost greater than the cost of the computation itself, or do you end up with two different codepaths (one sync and one async, just to ensure you're getting as complex as you can) depending on the computation's expected duration?