Matthew Perry

@groutr we could also try to make this `np.unique` approach optional so that users could opt-in to this behavior if they knew their datasets would benefit from it.

@groutr Would you like to split the percentile computation into a separate PR? That seems like a clear win to me and I don't want to hold that part up...

@j08lue we don't have any plans to add features in the short term. But we would gladly accept PRs and work with you to get them merged!

This is one scene that fails on it: [`LC80250322016238LGN00`](http://landsat-pds.s3.amazonaws.com/L8/025/032/LC80250322016238LGN00/index.html). Note that it's all clouds and water, no land. This will be common enough that it's an issue. 

Strange, I'm not able to reproduce this after trying this and a few other scenes. Gonna need to dig deeper to find a failing test case.

Units on the TOA products would be "Reflectance" but only if we use raw units or provide a scaling factor to covert back.

Note, let's also set the scaling factor in the metadata tags

Looking at this more closely, `riomucho` which has control of the destination dataset object so that will require some changes to that lib. [Ticket here](https://github.com/mapbox/rio-mucho/issues/14) meanwhile bumping this from the...

Just a +1, it would greatly help implement the workflow I'm trying to achieve: 1. Edit `schemas/app.sql` 2. Run something _like_ `pg-schema-diff plan ...` but instead of writing to stdout,...