Adding and changing columns

Assigning to a column name that doesn't exist creates it; assigning to one that does replaces it. The right-hand side is an ordinary vectorised expression — no loop:

orders["price"] = orders["revenue"] / orders["cups"]
orders["cups"] = orders["cups"] * 2

Conditional columns

orders["band"] = np.where(orders["cups"] > 100, "busy", "quiet")

orders["band"] = pd.cut(
    orders["cups"], bins=[0, 90, 120, 999], labels=["quiet", "steady", "busy"]
)

np.where is the two-outcome case you already know. pd.cut bands a continuous column into labelled ranges — and returns a category, so it sorts in band order rather than alphabetically. pd.qcut does the same by quantile, giving equal-sized groups.

assign returns a new frame

orders.assign(
    price=lambda d: d["revenue"] / d["cups"],
    busy=lambda d: d["cups"] > 100,
)

Same job, but it returns a copy rather than editing in place — which is what makes it chainable. Use a lambda when a new column depends on another one you're defining in the same call, since d refers to the frame as it is at that point.

The trap: chained assignment

This looks reasonable and does not work:

orders[orders["cups"] > 100]["price"] = 0     # writes nowhere

There are two bracket operations. The first produces a temporary frame; the second writes into that temporary; then it's discarded. Your original is untouched.

Historically pandas emitted a SettingWithCopyWarning here — a warning famous for being ignored, partly because it sometimes appeared when nothing was wrong, and the write sometimes worked anyway depending on memory layout.

The fix is to do it in one indexing operation:

orders.loc[orders["cups"] > 100, "price"] = 0      # one bracket — works

One bracket, not two. That's the whole rule, and it's the same instinct as the NumPy view-versus-copy lesson.

What changed in pandas 3

Copy-on-Write is now the default, and it makes the behaviour consistent: an operation never modifies its parent. So chained assignment now reliably does nothing, rather than working by accident on some layouts and not others.

That's a real improvement — a silent failure that happens every time is far easier to find than one that happens sometimes. The habit doesn't change: select once, then assign.

Renaming and dropping

orders.rename(columns={"cups": "cups_sold"})
orders.drop(columns=["rating"])
orders.insert(1, "price", values)      # at a specific position

These all return new frames by default. You'll see inplace=True in older code; it's discouraged now — it doesn't save memory, it breaks chaining, and it makes code harder to reason about.