In addition to making it easier to generate random data as described in
ARROW-2329, I think it would be useful to reduce some of the boilerplate associated with writing down explicit test cases. The benefits of this will be especially pronounced when writing nested arrays.
Example code that could be improved this way:
Rather than having a ton of hand-written assertions, we could compare with the expected true dataset. Of course, this itself has to be tested endogenously, but I think we can write enough tests for the JSON parser bit to be able to have confidence in tests that are written with it