Access row/col data via attributes

[ghost]

• Supercedes ENH: Series item access via attribute #1903 #1904
• Namespaces rows/cols as df.r.<row label>, df.c.<col label>
• get/set, multindex in both axes
• Designates df. (7531038) for future removal.
• No perf hit on attribute lookup, if you don't use it (raised in ENH: Series item access via attribute #1903 #1904).

Open issues:

• .r and .c (which i like), don't generalize to dim>2. Use a numbered scheme instead? (a0, a1,etc')
• no panel support. No reason, I just never use them.
• Setting columns has a caveat, need your numpy-fu help here (see bottom).

Demo:

In [1]: import pandas as pd
   ...: from pandas.util.testing import makeCustomDataframe as mkdf
   ...: 
   ...: pd.options.display.notebook_repr_html=False

In [2]: df=mkdf(4,2,r_idx_nlevels=1)

In [3]: df
Out[3]: 
C0      C_l0_g0 C_l0_g1
R0                     
R_l0_g0    R0C0    R0C1
R_l0_g1    R1C0    R1C1
R_l0_g2    R2C0    R2C1
R_l0_g3    R3C0    R3C1

In [4]: df.r.R_l0_g1 # tab-complete
Out[4]: 
C0
C_l0_g0    R1C0
C_l0_g1    R1C1
Name: R_l0_g1, dtype: object

In [5]: df.r.R_l0_g1 = df.r.R_l0_g0

In [6]: df
Out[6]: 
C0      C_l0_g0 C_l0_g1
R0                     
R_l0_g0    R0C0    R0C1
R_l0_g1    R0C0    R0C1
R_l0_g2    R2C0    R2C1
R_l0_g3    R3C0    R3C1

In [7]: df.c.C_l0_g1 # # tab-complete
Out[7]: 
R0
R_l0_g0    R0C1
R_l0_g1    R0C1
R_l0_g2    R2C1
R_l0_g3    R3C1
Name: C_l0_g1, dtype: object

In [8]: df.c.C_l0_g1 = df.c.C_l0_g0

In [9]: df
Out[9]: 
C0      C_l0_g0 C_l0_g1
R0                     
R_l0_g0    R0C0    R0C0
R_l0_g1    R0C0    R0C0
R_l0_g2    R2C0    R2C0
R_l0_g3    R3C0    R3C0

Multindex example (note recursive syntax):

In [1]: import pandas as pd
   ...: from pandas.util.testing import makeCustomDataframe as mkdf
   ...: 
   ...: pd.options.display.notebook_repr_html=False

In [2]: df=mkdf(4,2,r_idx_nlevels=2,c_idx_nlevels=2)

In [3]: df
Out[3]: 
C0              C_l0_g0 C_l0_g1
C1              C_l1_g0 C_l1_g1
R0      R1                     
R_l0_g0 R_l1_g0    R0C0    R0C1
R_l0_g1 R_l1_g1    R1C0    R1C1
R_l0_g2 R_l1_g2    R2C0    R2C1
R_l0_g3 R_l1_g3    R3C0    R3C1

In [4]: df.r.R_l0_g1.r.R_l1_g1 # tab-complete. twice
Out[4]: 
C0       C1     
C_l0_g0  C_l1_g0    R1C0
C_l0_g1  C_l1_g1    R1C1
Name: R_l1_g1, dtype: object

In [5]: df.c.C_l0_g0.c.C_l1_g0 # tab-complete. twice
Out[5]: 
R0       R1     
R_l0_g0  R_l1_g0    R0C0
R_l0_g1  R_l1_g1    R1C0
R_l0_g2  R_l1_g2    R2C0
R_l0_g3  R_l1_g3    R3C0
Name: C_l1_g0, dtype: object

In [6]: df=mkdf(4,2,r_idx_nlevels=2,c_idx_nlevels=2)

In [7]: df.c.C_l0_g0.c.C_l1_g0=df.c.C_l0_g1.c.C_l1_g1

In [8]: df
Out[8]: 
C0              C_l0_g0 C_l0_g1
C1              C_l1_g0 C_l1_g1
R0      R1                     
R_l0_g0 R_l1_g0    R0C1    R0C1
R_l0_g1 R_l1_g1    R1C1    R1C1
R_l0_g2 R_l1_g2    R2C1    R2C1
R_l0_g3 R_l1_g3    R3C1    R3C1

Suggest a fix for this? I'm sure I'm just lacking in numpy-fu here.

In [7]: df.r.R_l0_g0.r.R_l1_g0 = df.r.R_l0_g1.r.R_l1_g1 # this works
In [8]: df.r.R_l0_g0.r.R_l1_g0 = df.r.R_l0_g1 # this works
In [3]: df.c.C_l0_g0.c.C_l1_g0=df.c.C_l0_g1.c.C_l1_g1 # this works
In [5]: df.c.C_l0_g0.c.C_l1_g0=df.c.C_l0_g1 # this doesn't

[ghost]

designated 0.12, need to update release.txt beforee merging.

[ghost]

Add issue to remove df. in future release if this is merged
and removal has support.

[wesm]

I'm -1 on removing df.foo_col. Mainly because I find it really useful. The perf hit is acceptable

[ghost]

It wasn't the perf hit (is there any? getattr is a fallback) I was thinking of
but neatness.
I also find the mixing of object properties and data accessors in the same ns
messy, and so prefer seperating things out in a way that also treats cols
and rows on an equal footing.

I'd prefer closing this rather then introducing yet another similar but different
way to do the same thing which will coexist forever.

This wasn't meant as just a feature. it's a cleanup.

[wesm]

I'm for adding a general way to do what we're describing. The df. is very convenient and harmless enough to stay imho (plus it's familiar compared with record arrays). I'll take a closer look later

[ghost]

After more thought, df.colX is heavily used in bool indexing and would be bad to give up.
frames with labeled rows that are also identifiers (no spaces, punctuation) occur, but are
not that common, unlike column names.
This doesn't generalize well to other types of index, specificlly DateTimeIndex with Timestamp
labels that that begin with a digit (year).

Withdrawn.