[QST] How to update Table.column value in libcudf Java API

[hiroki-nishimoto-fixstars]

What is your question?

2 Questions.

1. How can I write bellow cudf(Python) code to libcudf(Java) code?
2. Are there any introduction or documents or example codes for libcudf(Java)?

---

Hi. I'm trying to rewrite cudf(Python) code to libcudf(Java) code.

In my understanding, they are implemented diffrenct concept. So, I do not believe that they are completely compatible, of course.

How can I write a process in libcudf(Java) that corresponds to the following process of updating a Series in a cudf.DataFrame?

libcudf(Java) prohibit to access Table.columns and don't have update column API. ex) setColumn(index, value)

import pandas as pd
df = pd.DataFrame({"a":[1,2,3], "b" : [4,5,6]})
df["a"] = [9,8,7]
print(df)


'''
output
  a b
0 9 4
1 8 5
2 7 6
'''

A simple implementation would yield the following code, which does not seem very smart.

Table df = ... // omission

ColumnVector update_a = ColumnVector.fromInts(9,8,7);
int numCols = df.getNumberOfColumns();
ColumnVector cols = new ColumnVector[numCols];
for( int i = 0 ; i < numCols; i++ ){
    if( i != 0 ){
        cols[i] = df.getColumn[i];
    }
    else {
        cols[i] = update_a;
    }
}
df = new Table(cols);

Also, I know there is libcudf(C++) documentation, but I could not find any Java documentation or example code. Is there somewhere I can find it?

Now I'm trying to implement with reading Test code and API implementation...

[GregoryKimball]

@ttnghia would you please share your thoughts on the best way to update a column using the libcudf Java API?

[ttnghia]

libcudf(Java) prohibit to access Table.columns

No, cudf Java provides the function getColumn(int index) that allows accessing to the table columns and you already used it in your example. Your example above is almost correct. Look at the documentation of Table.getColumn:

   * .....If you want to keep a reference to
   * the column around past the life time of the table, you will need to increment the reference
   * count on the column yourself.

That means, this is needed when calling getColumn:

if( i != 0 ){
        cols[i] = df.getColumn[i];
        cols[i].incRefCount();
    }

Add that second line to your example. That's all you need. When the original table is destroyed, the reused columns are not destroyed but moved to be managed by the new table because their reference counts have been increased.

[hiroki-nishimoto-fixstars]

Thank you for answering.

It's really helpful to know more about reference counting.

Your example above is almost correct.

So, are the following conclusions correct?

• If I want to update one of column of table, I must declare new table with exist columns and new update column, and assign to exist table.
• Exist table column is not able to be updated directly.

[ttnghia]

Yes these are correct. A table is just a manager for the columns, so creating a new table holding existing columns of another table is relatively cheap.