TODO.txt

* How do we deal with more complex features?  For example, what about a deeply hierarchical
  grouping structure.  A user has a feature which is the list of groups he is in.  Each group has
  a list of parent groups.  Each group and each user has a list of permissions and the union
  is all of the permissions for the user.  The code for this is not so bad in prod, but what about
  for data analysis?  
  * Idea 1: Cache subjects in RAM, load up every update for each feature (within a time range?),
    and then organize updates in a b-tree for fast lookup of updates for a particular time.  With
    very large and numerous features, you could run out of RAM, but it seems unlikely for most
    use cases.


* Consider not using times.  See http://api.mongodb.org/perl/current/MongoDB/DataTypes.html
Warning: creating DateTime objects is extremely slow. Consider saving dates as numbers and converting the numbers to DateTimes when needed. A single DateTime field can make deserialization up to 10 times slower.