I've never quite understood the logic behind #3. Why wouldn't I just scale my monolith up and down instead? I could see some instances where being able to do it independently would be useful (e.g. when the feature has unique resource requirements), but I would guess that, for us, 9 out of 10 times, it has been fine to just scale the monolith as a whole up and down as necessary.
Cost is the easy answer - as with any optimization, I guess.
It can also be really annoying sometimes to scale a monolith - a big one can have a lot of different parts like multiple threadpools/ database pools. One component hits a threshold, you add more instances, now your other component is overprovisioned, maybe even failing due to too many DB connections, or you hit weird thread contention issues.
I've had some pain in the past scaling monoliths in similar situations.
