What a DataFrame actually is

Most people meet pandas through a spreadsheet metaphor, and the metaphor holds for about ten minutes. Then something like filtered[1] throws a KeyError and the metaphor quietly stops working.

Here's the version that keeps working. A DataFrame is not a grid of cells with numbered rows. It's a set of columns, each one a typed array, glued to a shared set of row labels called the index. The index is a real object with real values. When you create a DataFrame without specifying one, pandas invents 0, 1, 2, … — and because those look exactly like positions, almost everyone assumes they are positions. They aren't, and the next lesson shows what happens when the two disagree.

One type per column

The other half of the picture is dtypes. Every column has exactly one type for all of its values. That's not a limitation, it's the trade that makes pandas fast: a column of int64 is one contiguous block of memory that numpy can rip through, not six million separate Python integers.

It also explains a wart you'll hit eventually — put a string into an integer column and the whole column changes type, because there's no such thing as one odd cell out.

Selecting: one bracket or two

df["cups"] gives you a Series — a single column, still carrying the index. df[["cups"]] gives you a DataFrame with one column. Same data, different container, and the difference matters constantly because Series and DataFrames have different methods. When you pass a list, you get a table back; when you pass a single label, you get a column back.