Generally, the training of a model is a time-consuming phase, therefore it is advisable to carry it out long before the execute phase.
It is advisable to write the output on a type different from Reporting Layer as the data processing is provided with a deletion option for the previous execution that deletes the entire output. If the execute data processing is run by several users (e.g. each user for their own cost centre) the last execution will prevail. Whereas, if you write on the Reporting Layer, thanks to the functionalities provided by DTP and Data Source Mapping, you can select a specific part of the prediction.
It is advisable not to export those Analytical Workspaces that use more than 5 models.