The centralized coded caching scheme is a technique proposed by Maddah-Ali and Niesen as a solution to reduce the network burden in peak times in a wireless system. Later Yan et al. reformulated the problem as designing a corresponding placement delivery array, and proposed two new schemes from this perspective. These schemes above significantly reduce the transmission rate $R$, compared with the uncoded caching scheme. However, to implement the new schemes, each file should be cut into $F$ pieces, where $F$ increases exponentially with the number of users $K$. Such constraint is obviously infeasible in the practical setting, especially when $K$ is large. Thus it is desirable to design caching schemes with constant rate $R$ (independent of $K$) as well as small $F$. In this paper we view the centralized coded caching problem in a hypergraph perspective and show that designing a feasible placement delivery array is equivalent to constructing a linear and (6, 3)-free 3-uniform 3-partite hypergraph. Several new results and constructions arise from our novel point of view. First, by using the famous (6, 3)-theorem in extremal combinatorics, we show that constant rate caching schemes with $F$ growing linearly with $K$ do not exist. Second, we present two infinite classes of centralized coded caching schemes, which include the schemes of Ali-Niesen and Yan et al. as special cases, respectively. Moreover, our constructions show that constant rate caching schemes with $F$ growing sub-exponentially with $K$ do exist.