More Efficient Feature Generation

[bitsofbits]

We should be able to speed up feature generation a lot even without using windowing by:

• Passing in or sniffing an end_date for old data. Choose start_date to be dt(1?, 7?)
  days earlier than the date we care about.

• Compute features only from start_date onward

• Merge the features with old data

• All of this may get simpler by switching to the simpler 5 minute rule that David uses. Just
  pick one point out of every 5 minute interval of the day. (Perhaps use median across values,
  or point most at Median time-wise).

---

It may be better / simpler to switch to sharding by date, then have a utility that converts date shards into mmsi shards (for some date range).