[] vs .loc vs .iloc

This is the most-confused topic in pandas, and it isn't because it's hard — it's because three similar-looking things answer three different questions.

[] selects columns

orders["cups"]              # a Series — one column
orders[["city", "cups"]]    # a DataFrame — a list of columns

That's the odd one. For every other indexable object in Python, brackets pick an element; here they pick a column. It's a shortcut for the common case, and it has two exceptions — a boolean mask filters rows, and a slice takes rows. Those exceptions are exactly why bare brackets get confusing, so use them for columns and masks and nothing else.

.loc works by label

df.loc["tue"]              # the row LABELLED tue
df.loc["tue", "cups"]      # row first, then column
df.loc["mon":"wed"]        # three rows — the end IS included
df.loc[df.cups > 100, "cups"]   # filter and column, in one go

Two things to internalise. Rows come first, columns second. And a .loc slice includes its endpoint — unlike every other slice in Python. That's deliberate: with labels there's no "one past the end" to name, so excluding the end would make label slicing nearly useless.

.iloc works by position

df.iloc[1]        # the second row
df.iloc[1, 1]     # row 1, column 1
df.iloc[0:2]      # two rows — the end is EXCLUDED
df.iloc[-1]       # the last row

This one behaves like an ordinary Python index: zero-based, negatives from the end, end excluded.

Where it bites

As long as your index is words, mixing them up throws an error and you fix it. The danger is when the labels are integers — which is what filtering leaves behind, since the original 0, 1, 2… labels come along with their rows:

busy = orders[orders["cups"] > 100]
busy.index          # [0, 3, 5, 7, 8, 10] — gappy

busy.loc[3]         # the row LABELLED 3
busy.iloc[3]        # the FOURTH row
busy[3]             # KeyError — no column called 3

The first two both work and return different rows. No exception, no warning. That's the bug that survives code review, and it's why the advice is always to be explicit.

Which to reach for

• Labels mean something — dates, ids, names — use .loc.
• You genuinely mean position — "the first five rows" — use .iloc.
• Columns or a boolean mask — bare [] is fine and reads well.

And for a single cell in a loop, .at and .iat are the fast paths — same label-versus-position split, much less overhead, one cell only.