1. Use basic copypasta from the redis implementation
2. Placeholder in place of final key manipulation logic
3. Wrestle with all the dependency issues

1. There are a number of things that I learned from maven so I went back
   and reexamined my assumptions. I also rebased and then forwardported
   some of the logic on a lateest master
2. BigtableIO seems similar to BigQueryIO in that it's not really
   exposed in a way to publish failed elements

1. Errors out on applying feature set with an inactiveRPC error, trying
   on master now

1. Validate that the bigquery pipeline still works for publishing
   features
2. Removed slf4j logger warnings by excluding the other versions

1. I don't know how to use a gax response observer correctly so I
   replaced readrowsasync with readrows, I don't think that it changes
   how the query is executed

1. Move the logic that was in the observer into the decoder
2. Make the structure conform to similar redis connector
3. Delete observer file
4. Latencies at 1k req/min are < 10ms

The feature reference list gets passed on decoder initialization in
order to filter values before they are serialized if they are not needed
in the request. Small latency improvment

maybe a better regex/set of regexes would be better