Expected results in tests were written some time ago, at that time pandas were using as default different (less precise) representation of float numbers when read with read_csv method. pandas 1.2.0 changed that default, but float_precision='legacy' optional argument allows to use the same precision as in older versions.