using .iloc to assign new value

[ndy-cd]

System information

• Linux Ubuntu 19.10
• Modin 0.7.3
• Python 3.7.7

Problem:

When trying to use .iloc to assign a new value there are no changes taking place.

I already left a comment on the issue #147 but it's already closed and I wasn't sure that you can see it.

Code to reproduce:

import modin.pandas as pd #dask
test = pd.DataFrame({'first': range(10)})
test['first'].iloc[0] = 500
#or test['first'].iloc[[0, 1, 2]] = 500
print(test) # there is no changes

[devin-petersohn]

Thanks @ndy-cd, we prefer the new issue since the old one was resolved. This issue is slightly different because of the assignment.

So what is happening is that the test['first'].iloc assignment is happening inplace in pandas. Modin is doing the assignment, but not propagating the test['first'] back to the original dataframe. It will be straightforward to fix, the two types of indexing case in the same statement (first, and iloc) was not handled.

Assignment with loc/iloc will work if you don't access the column first. A simple temporary workaround for the double indexing is to turn it into a single loc call:

import modin.pandas as pd #dask
test = pd.DataFrame({'first': range(10)})
test.loc[0, 'first'] = 500  # 0 is the Index value of the 0th row
#or test.iloc[[0, 1, 2], 'first'] = 500
print(test)

If you prefer to use iloc you can also do this:

import modin.pandas as pd #dask
test = pd.DataFrame({'first': range(10)})
test.iloc[0, 0] = 500  # 0 is the column and row number
#or test.iloc[[0, 1, 2], 0] = 500
print(test)

We have the machinery in place to support this, but it looks like something got lost with the double indexing. Thanks for the detailed report! We will get this fixed.

[ndy-cd]

Thank you very much for detailed answer!

It seems to be a same problem with using .replace()
test['first'].replace({500: 40}, inplace=True)

[devin-petersohn]

Thanks @ndy-cd, that is expected behavior.

Pandas has many ways to do the same thing as you probably know, we have one internal way to do things so that we can better optimize these fundamental operations.