Long-Term Time series forecasting is an essential task that uses historical data to forecast future trends and period patterns, such as in traffic planning and weather forecasting. Recently, there has been a boom in patch-based forecasting models which divide the input sequence into subseries-level patches to better extract temporal patterns. However, real-world time series data usually present unique compositions and complex multi-scale temporal patterns. Thus, setting a fixed patch length may not adequately capture their periodic characteristics. In light of this finding, we propose ScaleMixer, a completely MLP-based design, which combines the adaptive selection of patch length and the MLP-Mixer structure along with a downsampling process. By conducting comprehensive experiments on five time series datasets from real world, ScaleMixer effectively achieves performance that compares favorably with that of state-of-the-art models.