If you normalize your data to have the same magnitude, the two are indistinguishable. Sometimes it is desirable to ignore the magnitude, hence cosine similarity is nice, but if magnitude plays a role, dot product would be better as a similarity measure. Note that neither of them is a “distance metric