We have observed that some queries that operate on relatively small data volumes do not need to be executed in parallel to still run quickly. Queries processing even a few thousand rows can run well on just one or two cores, and actually run slower when given more cores, due to the inter-process communication overheads incurred.

So if your queries process only a small number of rows, consider making the max_parallelism_level settings very low from the outset if you are targeting large numbers of concurrent users.