The nodelay option for limit_req can delay excessive requests but it’s not desired in some situations. I just found an intuitive way to show the difference while tweaking the GitHub Avatar proxy.

Every grid is an image loaded from GitHub avatar.

With nodelay:

Without nodelay:

So in my use case with nodelay option can make clients feel faster loading. But may hit request limit more easily. Without nodelay keep excessive requests in the burst bucket and load them in sequence. But clients may feel slow.