Fairness assumptions are needed to verify liveness properties of concurrent systems. In this paper we explore the so-called unconditional fairness in Markov decision processes (MDPs), which is a prerequisite for quantitative assume- guarantee reasoning. Unconditional fairness refers to execu- tions where all processes are guaranteed to participate. We prove that realisability of unconditional fairness coincides with the absence of partial deadlocks, i.e., end components where a process suffers from starvation. We propose a weak variant of the stubborn set method to reduce MDPs, while preserving the realisability of unconditional fairness and maximal probabili- ties of reaching bottom end components under fair schedulers.