pandas

Index, columns, values, dtypes — the four parts every later method moves around.

Labels take part in the arithmetic — the one idea NumPy doesn’t have.

One type per column — plus the nullable and categorical ones pandas had to add.

read_csv, and the four calls to run before you trust a single row.

Three ways in, three different questions — and the one that silently returns the wrong row.

Why df[df["cups"] > 100] works, and why the index comes out full of holes.

The readable ways to say a filter, once masks stop scaling.

Creating columns, and the chained assignment that silently does nothing.

Dropping and filling are different claims — and they give different answers.

“Duplicate” is a decision you make, and subset= is where you make it.

One capital letter, one wrong number, no error message.

Housekeeping — and the two footguns hiding in it.

describe() is a checklist, not a report — run it on everything.

The three motions hiding inside every .groupby() call.

Told apart by one thing: the shape of what comes back.

The same aggregation, with the second key across the top instead of down the side.

No keys, no matching — and two traps that come free with alignment.

The named side is kept in full — and duplicate keys multiply your rows.

Compute in long form, present in wide. Most awkward pandas is a shape problem.

An index whose labels are tuples — and two methods for moving levels sideways.

A date read from a file is a string, and strings sort in the wrong order.

One changes how many points there are. The other changes what each one says.

Slide the column down a row, and every “compared to last time” question falls out.

Store each value once — and get meaningful sorting for free.

One expression, no half-modified frames, and the same answer every time you run it.