Split out graph Expr code from Dataframe Expr code

[mrocklin]

Closes #142
Supercedes #158

This is the smallest step I could think of to start this change. I may want to play around a little with arrays this week in anticipation of hanging out with @dcherian next week. I figured it would be good to split this off now if possible.

(sorry for the poor timing, I know it's an active week for this repository)

cc @phofl

[mrocklin]

from dask_expr import _core as core, _expr as expr
everything = set(dir(object))
c = set(dir(core.Expr)) - everything
d = set(dir(expr.Expr)) - everything - c

What's in core.Expr

{'__dask_graph__',
 '__dask_keys__',
 '__dict__',
 '__module__',
 '__weakref__',
 '_defaults',
 '_depth',
 '_find_similar_operations',
 '_layer',
 '_lower',
 '_meta',
 '_name',
 '_node_label_args',
 '_parameters',
 '_remove_operations',
 '_required_attribute',
 '_simplify_down',
 '_simplify_up',
 '_task',
 '_to_graphviz',
 '_tree_repr_lines',
 'dependencies',
 'find_operations',
 'lower_completely',
 'lower_once',
 'operand',
 'pprint',
 'rewrite',
 'simplify',
 'substitute',
 'substitute_parameters',
 'tree_repr',
 'visualize',
 'walk'}

What's in expr.Expr

{'__add__',
 '__and__',
 '__bool__',
 '__getattr__',
 '__getitem__',
 '__invert__',
 '__mul__',
 '__neg__',
 '__or__',
 '__pos__',
 '__radd__',
 '__rand__',
 '__rge__',
 '__rgt__',
 '__rle__',
 '__rlt__',
 '__rmul__',
 '__ror__',
 '__rsub__',
 '__rtruediv__',
 '__rxor__',
 '__sub__',
 '__truediv__',
 '__xor__',
 '_combine_similar',
 '_combine_similar_branches',
 '_divisions',
 '_is_length_preserving',
 'abs',
 'align',
 'all',
 'any',
 'apply',
 'astype',
 'clip',
 'columns',
 'combine_first',
 'combine_similar',
 'count',
 'divisions',
 'dtypes',
 'fillna',
 'idxmax',
 'idxmin',
 'index',
 'isna',
 'isnull',
 'known_divisions',
 'mask',
 'max',
 'mean',
 'memory_usage_per_partition',
 'min',
 'mode',
 'nbytes',
 'ndim',
 'npartitions',
 'nunique_approx',
 'optimize',
 'prod',
 'rename_axis',
 'replace',
 'round',
 'size',
 'std',
 'sum',
 'to_timestamp',
 'var',
 'where'}

[mrocklin]

There are other things we could think about doing, like renaming _expr.Expr to Frame or something, or moving stuff within a dataframe directory, but I think that all of that could also come later and doesn't necessarily need to block this first step. My hope is that this first step is pretty innocuous.

[phofl]

Do you have an idea how we can solve the issue that defining things like addition and multiplication on the expression gets more complex if we split this up? This is why we postponed the initial pr IIRC

[mrocklin]

I think that we have array.Add and dataframe.Add. I intentionally didn't move the __add__ and similar methods over. If we ever mix types then we must add a ConvertToArray or ConvertToDataframe operation beforehand.

[phofl]

Beforehand meaning before calling this on the expression level or on the expression level? I think this is better suited on the collection level generally speaking.

[mrocklin]

All I mean is that the answer with Add and Mul and such is not to share, at least short term. Regardless, I think that this change doesn't get into that question. I think that this is a safe change to make. Maybe we should chat live sometime today?