Why does loc return several rows for a single label in pandas?

The short version

If the index has duplicate labels, .loc[label] returns every row carrying that label, which is a DataFrame, not a single row. Code expecting one row then breaks in confusing ways. Check that the index is unique before relying on label lookups.

Seeing it

import pandas as pd

df = pd.DataFrame({"value": [1, 2]}, index=["a", "a"])

len(df.loc["a"])   # 2 , not 1

Both rows are labelled a, so both are returned. With a unique index, .loc["a"] would give a Series representing one row. With duplicates, it gives a DataFrame. The type of the result changes depending on the data, which is what makes this so awkward: code that works on one dataset breaks on another.

Check for uniqueness

df.index.is_unique     # False here
df.index.duplicated().sum()   # how many duplicate labels

Run this before setting an index you intend to look up by. It takes a moment and prevents a whole class of bug.

Where duplicate indexes come from

Most often, concat without ignore_index=True. Each source frame carries its own 0, 1, 2 index, and stacking them produces repeated labels.

pd.concat([a, b])                      # duplicate labels
pd.concat([a, b], ignore_index=True)   # clean 0..n index

They also arise from set_index on a column that is not actually unique, which usually means your assumption about the data was wrong and is worth investigating rather than papering over.

The knock-on effects

A duplicated index does more than confuse .loc. It breaks alignment in arithmetic between Series, can make join multiply rows, and produces ValueError: cannot reindex on an axis with duplicate labels in operations that need a unique index. The error messages appear far from the cause, which is why checking early pays off.

Fixing it

If the labels are meaningless leftovers, discard them.

df = df.reset_index(drop=True)

If the index is meant to be a key and is not unique, the duplicates are a data problem. Look at them before deciding.

df[df.index.duplicated(keep=False)]

They may be genuine repeated records to aggregate, or accidental duplicates to remove.

Avoid meaningful indexes unless they earn it

The deeper lesson is that a meaningful index is a commitment: it must be unique and sorted to behave well. Unless you need label lookups, alignment or time-series tools, keeping the data in ordinary columns and filtering with boolean conditions avoids the entire category of index problems, at no cost.