This paper studies the interaction between coordination and social learning in a dynamic regime change game. Social learning provides public information to which players overreact due to the coordination motive. So coordination affects the aggregation of private signals through players' optimal choices. Such endogenous provision of public information results in inefficient herds with positive probability, even though private signals have an unbounded likelihood ratio property. Therefore, social learning is a source of coordination failure. An extension shows that if players could individually learn, inefficient herding disappears, and thus coordination is successful almost surely. This paper also demonstrates that along the same history, the belief convergence differs in different equilibria. Finally, social learning can lead to higher social welfare when the fundamentals are bad.