pandas groupby cummax just assigning original values instead of updating the max-so-far

Question

I have this dataframe:

    type      run   corrected_episode   Reward
0   notsweet    0   0                   35.0
1   notsweet    0   100                 20.0
2   notsweet    0   200                 20.0
3   notsweet    0   300                 22.0
4   notsweet    0   400                 20.0

I want to create a new column, best_so_far, that has a monotonically increasing value for the corresponding Reward grouped by type, run, and corrected_episode. Easy enough, right? Except the following happens when I try to use groupby and cummax:

foo['best_so_far'] = foo.groupby(['type','run','corrected_episode']).Reward.cummax() yields:

type            run corrected_episode   Reward  best_so_far
0   notsweet    0   0                   35.0    35.0
1   notsweet    0   100                 20.0    20.0
2   notsweet    0   200                 20.0    20.0
3   notsweet    0   300                 22.0    22.0
4   notsweet    0   400                 20.0    20.0

The "best so far", well, isn't the best. I get the same results if I use foo['best_so_far'] = foo.groupby(['type','run','corrected_episode']).Reward.apply(lambda x: x.cummax())

I know this is possible because I've done this dozens of times with other dataframes, there's just something weird about this one that this simple procedure just doesn't work.

Answer 1

You can try remove corrected_episode

foo['best_so_far'] = foo.groupby(['type','run']).Reward.cummax()

Answer 2

After posting this of course I discovered what happened, but I'm going to share what I did to fix this here because this is the kind of Violation of the Principle of Least Astonishment that pandas is prone to.

The solution was to do this, instead:

foo['best_so_far'] = foo.groupby(['type','run']).Reward.cummax()

That is, I over specified the columns by including corrected_episode that had the unintended effect of just executing cummax() for that one element. However, I had originally included corrected_episode to ensure that the order of the rows was correct -- i.e., the dataframe was actually the result of massaging a lot of data (you are seeing a teeny tiny subset), and the order of the data wasn't necessarily sane for the cummax() to work as I envisioned.