The primary guideline for selecting a primary index column set is to achieve an even distribution of rows across the AMPs.
The more singular the values of a column chosen as the primary index, the more even the distribution of table rows across the AMPs of a system. A unique index is ideal for ensuring optimal distribution. When distribution is optimized, so is parallel processing.
While it is true that the ideal primary index for a table both optimizes retrieval and distribution, the reality is that you are often faced with trading one off against the other. It is not exceedingly rare for a particular primary index to provide maximal access, but poor distribution (or vice versa).