The problem is devs who don't understand what they're doing and just think that they can use floats in every situation and it'll just work out fine. This is not helped by many popular scripting languages who just default to floats when a result doesn't fit an integer (something more reasonable languages like Common Lisp don't do for instance).
For instance, to speak of videogames again, very tight precision isn't usually an issue but loss of granularity when numbers get very big can cause problems, especially if you have very large levels. That being said rationals wouldn't really help you here, you'd have the same problem except now you have to keep two numbers within bound instead of one. Imagine having a very small offset in a complex operation and ending up with a number like 100000000000000000000000000/100000000000000000000000001 !