Do nodes keep a running probability on how often they fail, so that the scheduler roughly knows how many times to duplicate a job?

I'm a bit confused as to why duplicating jobs on multiple machines helps with slower nodes in the network. I understand why this could speed up a given job, if one machine simply completes faster, but surely there are trade-offs here. How likely is it that a machine that that much faster than another. Would it not be more clever to only duplicate work once other machines go idle. Do these machines ever actually go idle?

how likely is that a machine goes that* much faster than another. Whoops!