Selecting rows and columns

Picking columns, the difference between loc and iloc, the index, and how copy-on-write changes what happens when you modify a selection.

Selecting columns

What: Square brackets with one name return a Series; with a list of names they return a DataFrame with those columns in that order.
Why: Most analysis starts by narrowing a wide table to the handful of columns a question needs.

Python

df["amount"]
df[["date", "region", "amount"]]
df.filter(like="price")

Watch out

Using attribute access like df.amount: it breaks for names with spaces or names that clash with methods such as count.

loc: select by label

Also called .loc

What: df.loc[rows, cols] selects by index labels and column names, and accepts boolean masks. Label slices include both ends.
Why: It is the one clear way to select and assign in a single step, which matters now that chained assignment is gone.

Python

df.loc[df["amount"] > 1000, ["region", "amount"]]
df.loc["2026-01-01":"2026-01-31"]   # dated index, both ends
df.loc[df["region"] == "North", "flag"] = True

iloc: select by position

Also called .iloc

What: df.iloc[rows, cols] selects by integer position, like Python lists: the end of a slice is excluded.
Why: Useful when position is what you mean, such as the first ten rows or the last column, whatever their labels are.

Python

df.iloc[:10]
df.iloc[:, -1]
df.iloc[[0, 5], [1, 2]]

Watch out

Mixing up loc and iloc on an integer index: loc[0:5] returns six rows, iloc[0:5] returns five.

The index

Also called set_index, reset_index, row labels

What: The index is the set of row labels. set_index turns a column into the index and reset_index turns the index back into a column.
Why: A meaningful index, such as a date or an ID, makes lookups, alignment and time series resampling simple.

Python

df = df.set_index("order_id")
df.loc[10432]
df = df.reset_index()

Watch out

Forgetting that arithmetic between two Series aligns on the index, so mismatched labels silently produce NaN.

Copy-on-write

Also called CoW, SettingWithCopyWarning, chained assignment

What: Since pandas 3.0 every selection or method result behaves as a new object: changing it never changes the original. pandas only copies data when you actually modify it.
Why: It ends the old confusion over views and copies and removes SettingWithCopyWarning, but chained assignment such as df["a"][0] = 1 no longer modifies df.

Python

# does nothing to df under copy-on-write
df[df["amount"] < 0]["amount"] = 0
# do it in one step instead
df.loc[df["amount"] < 0, "amount"] = 0

Watch out

Copying old tutorials that assign through two sets of brackets and wondering why the data never changes.

Sampling rows

Also called sample, nlargest, nsmallest

What: sample(n) or sample(frac=0.1) picks random rows; nlargest and nsmallest return the top or bottom rows by a column.
Why: Sampling makes exploring a big dataset fast, and nlargest answers top-10 questions without sorting everything.

Python

df.sample(frac=0.05, random_state=42)
df.nlargest(10, "amount")

Go deeper