A review is given of existing efforts and future challenges in large-scale ocean modeling on unstructured meshes. Because of large integration time the large-scale ocean circulation models require more attention with respect to conservation and accuracy than their coastal counterparts, and deal with different dynamics. Numerous discretizations of finite-element and finite-volume type have been proposed and explored, but only simplest approaches (like P1-P1 of FESOM or cell-vertex of FVCOM) are realized as working tools. While the results of first applications performed in global context are very encouraging and show the feasibility of local refinement by a factor of 30-50 in a global setup, they also indicate that further efforts are needed if one aims at simulating the ocean in eddying regimes. These efforts must concentrate on numerical efficiency of unstructured-mesh codes and ensure much higher accuracy of advection schemes than currently available on linear elements. Since the selection of discretization introduces obvious limitations on unstructured meshes, one has to re-consider it from the perspective that takes into account the already available lessons, and not only the representation of linear wave dynamics in the shallow-water context. In particular, too large velocity spaces of many mixed discretizations as a rule lead to difficulties in the performance of momentum advection in eddying regimes, and continuous representation of scalar quantities creates many inconveniences in hydrostatic codes. While the need for improved advection schemes tells on its own in favor of either discontinuous Galerkin methods or high-order reconstructions with finite volumes, the key question is how to implement them without significant loss of efficiency. The area of large-scale ocean simulations is almost fully dominated by structured-mesh models which are currently much more accurate and efficient per degree of freedom than models on unstructured meshes. The acceptance of unstructured-mesh technology on this `large-scale background' depends on our progress with this question as well as ability to show on practical examples that refinement is more practical and consistent than nesting.