I agree with most of this article, except that - for me - grouping by table columns is a more intuitive and flexible way to control subset application than 'rank'.

k9 is in development, but blending k7 and k9 we'd get something like:

   t: [[]sensor:  1 2 1 2 1 2
         day:     1 1 2 2 1 1
         reading: 1 2 3 4 5 6]
  
   0 1 1 0 * select by sensor, day from t
  sensor day|             
  ------ ---|----------------
  1      1  |[[]reading: 0 0]
  1      2  |[[]reading: ,3]
  2      1  |[[]reading: 2 6]
  2      2  |[[]reading: ,0]