Experiments are conducted in an open-channel flow where half of the section is smooth and the other half consists of an array of cubes, which are either submerged or emergent. A shear layer featuring large-scale Kelvin–Helmholtz structures develops between the two subsections. The flows are first analysed in the framework of the double-averaging method (averaging of the flow both in time and space). Double averaging could be performed thanks to an experimental set-up (three-dimensional, two-component telecentric scanning particle image velocimetry) that allows to measure the velocity field in a large volume, including the interstices between the cubes. A momentum balance performed on the smooth subsection indicates that the loss of momentum towards the rough subsection has the same order of magnitude than the momentum loss through bed friction. This lateral momentum flux occurs nearly exclusively through turbulent shear stress, whereas secondary currents plays a minor role and dispersive shear stress is negligible. A pattern recognition technique is then applied to investigate statistically the large-scale Kelvin–Helmholtz structures that develop in the shear layer. The structures appear to be coherent over the water depth and to be strongly inclined in the vertical, the top part being ahead. The educed coherent structure is responsible by itself for the shape of the velocity profile across the shear layer and for a large part of the turbulence (up to 60 % for the turbulent shear stress). Finally, a coupling is identified between the passage of the Kelvin–Helmholtz structures and the instantaneous wake flow around the cubes at the interface.