Storage size of the index itself, insert/update performance. Also there may be very little gain from adding an index to a column with only few distinct values or a very skewed distribution (partitions may be a better approach in those cases).