As the others have already noted, there's no specific recommended best practice percentages because what's normal and what's too much varies per subnet.
As I noted (years ago) in the one reference that @balaji.bandi provides, when such bandwidth limiters are exceeded, "good" traffic is usually impacted too. So, to your question to what level can switches function without issue, is an "it depends". Once they trigger, legitimate traffic may be impacted but there may have been no need to trigger. Conversely, if not triggered when actually needed, you have issues too. It's offen difficult to find the "just right" values; which, again, may be different on each subnet.
Keep in mind, unicast or multicast flooding often isn't as impactful as broadcast flooding. The latter is so impactful because the host needs to accept and analyze broadcast frames/packets; they cannot be (address) filtered by hardware or as easily filtered by software.
Switch interfaces, too, need analyze broadcasts, but something more common than possibly years ago is CoPP to rate limit traffic to the switch's CPU.
