Collect will store the entire output in a List in memory. This solution is acceptable for "Little Data" problems although if the entire problem fits in the memory of a single machine there is less motivation to use Spark.
Most problems which benefit from Spark are large enough that even the data assigned to a single partition will not fit into memory. In my special case the output now is in the 0.5 - 4 GB range but in the future might get to 4 times that size - something a single machine could write but not hold at one time. I find that for most problems a file like Part-0001 is not what the next step wants to use - the minute a step is required to further process that file - even move and rename - there is little reason not to let the spark code write what is wanted in the first place. I like the solution of using toLocalIterator and writing my own file
