What is the difference between loc and iloc, and when does each fail?

The short version

loc uses the labels in the index. iloc uses the integer position, counting from zero. On a fresh DataFrame these look identical, because the labels happen to be 0, 1, 2. After a filter or a sort they diverge, and code that used the wrong one starts returning the wrong row silently.

The moment they diverge

import pandas as pd

df = pd.DataFrame({"state": ["Sokoto", "Kebbi", "Zamfara"], "schools": [412, 380, 265]})

print(df.loc[1, "state"])    # Kebbi
print(df.iloc[1]["state"])   # Kebbi     identical, so far

filtered = df[df["schools"] > 300]      # drops Zamfara, keeps index labels 0 and 1
print(filtered.loc[1, "state"])         # Kebbi     label 1
print(filtered.iloc[1]["state"])        # Kebbi     position 1

sorted_df = df.sort_values("schools")   # index is now 2, 1, 0
print(sorted_df.loc[1, "state"])        # Kebbi     still label 1
print(sorted_df.iloc[1]["state"])       # Kebbi     position 1 is label 1 here
print(sorted_df.iloc[0]["state"])       # Zamfara   position 0
print(sorted_df.loc[0, "state"])        # Sokoto    label 0

Read the last two lines again. Same number, different rows. That is the bug.

Slicing behaves differently too

This trips up people who assume Python slice rules apply everywhere.

df.iloc[0:2]   # two rows. The end is exclusive, like a Python list.
df.loc[0:2]    # three rows. The end is inclusive, because labels are not positions.

When each one fails

loc raises a KeyError when the label is absent. That is good: it is telling you the row you asked for does not exist.

iloc raises an IndexError when the position is out of range, and returns the wrong row when it is in range but you meant a label.

The dangerous failure is the silent one, and only iloc has it.

The rule worth adopting

Use loc unless you specifically want positional access, which in analysis code is rare. And after any filter, if you plan to use positions at all, reset the index so positions and labels agree again:

filtered = df[df["schools"] > 300].reset_index(drop=True)

Pass drop=True unless you actually want the old index kept as a column.

The one that catches everyone

df[df["schools"] > 300]["state"] = "unknown"    # does nothing, warns
df.loc[df["schools"] > 300, "state"] = "unknown"  # assigns

The first line operates on a temporary copy, which is then discarded. Pandas warns you with SettingWithCopyWarning. Do not silence that warning. Use .loc on the left hand side of an assignment, always.