Error outputs can obviously be used to improve reliability, but they also have an important part to play in terms of scalability as well
The capability to recover rows is perhaps the most useful course of action.
The less work you perform, the greater the performance and therefore scalability.
In summary, the more expensive the verification, the more bad rows you require to justify adding the validation and fix to the main flow. For fewer bad rows, or a more expensive validation procedure, you have increased justification for keeping the main flow simple and for using the error flow to perform the corrective work. The overall number of rows should also influence your design, because any advantages or disadvantages are amplified with a greater number of rows, regardless of the ratio. For a smaller number of rows, the fixed costs may outweigh the implied benefits, as any component has a cost to manage at runtime, so a more complicated Data Flow may not be worthwhile with fewer overall rows.