A quick look at the source code, seems like you are going for a Numpy+Pandas implementations (evidenced by the DataFrame implementation, which closely follows the way Pandas implements it). Just FYI, Wes McKinney, the original Pandas author and main contributor, has written an [article](https://wesmckinney.com/blog/apache-arrow-pandas-internals/) listing what he considered the biggest problems with Pandas (3 years ago already) -- and I've seen him somewhere else describing how the memory allocation scheme (keeping all ints together, all floats together, etc.) is a really bad decision -- that it is much better in practical dataframe use to manage each column individually (this is only shortly touched in the blog post I linked to).
Also, it's probably a good idea to take a look at [ggplotnim's dataframe](https://github.com/Vindaar/ggplotnim#data-frame) which is (a) already implemented (b) on top of arraymancer and (c) is integrated with, and derived from, an excellent plotting library; so that when you want to do some graphics, you don't have to reimplement plotnine of matplotlib.
