Sample operator

[okennedy]

This branch adds a new Operator called DrawSample, which makes sampling a first-class primitive (as a step towards addressing #362). Currently support exists for two forms of sampling:

• Uniform: Draw a flat % of the input data.
• Stratified: Indicate a column and a value -> % map. Pass through the specified % of records, depending on the value in the indicated column.

Both of these sampling are probabilistic (you get approximately the specified %, rather than exactly that amount).

The choice to make DrawSample a distinct operator (rather than folding it into, say Select) was motivated by the following:

• Mimir and Spark handle functions differently. Both % based sampling strategies rely on seeded calls to rand() for consistency... at present, this has the potential to create inconsistent sample sets depending on where rand() is executed (see #361)
• Implementing stratified sampling in Mimir is tricky, since the sampler needs to track the value -> % map. This is implemented in Spark by inlining the map into a closure through spark's udf method. It would be possible in Mimir if we had support for dictionary types... but we don't. Thus, the only way to implement this (without resorting to repeated string parsing in a tight loop) is directly in RAToSpark... which means we need something on the order of an Operator to pull it off.

I would like to revisit this decision, ideally once we add a more robust typesystem into Mimir (e.g., #367).

A few more notes:

• The sampling operators are currently not exposed to the outside world. Initially, I think the right way to do this would be through the Vizier API. I want to get the Mimir model rewrite in place before exposing it through the Mimir SQL API. @mrb24 : I need some help figuring out how to muck with that code, since it looks generated and I don't know where the source lives.
• Adding deterministic sampling (i.e., give me EXACTLY # rows) is going to be tricky... the easiest way would be some sort of shuffle + limit...
• The current stratified sampling operator is precise, but a little user-hostile. An ideal version of the operator would first do a group-by count to find the # in each group, and then select sampling rates to ensure equal representation from each group. This would be best implemented using the new model framework.

[mrb24]

@okennedy Create view statements were not parsing after merging this pull. I made a change to MimirSQL.scala to fix it. Was there a reason for having what I removed in this commit? commit: 6e91f3bdb36f50567b71be815e58533556ebdc1e