Pandas — Merging DataFrames

• Use pd.merge(left, right, on="key") for an inner join.
• Switch join style with how="left" / "right" / "outer".
• Join on multi-column keys, or on differently-named keys (left_on/right_on).
• Stack frames vertically with pd.concat.

import pandas as pd

pupils = pd.DataFrame({
    "id": [1, 2, 3, 4],
    "name": ["Mia", "Leo", "Victor", "Mei"],
    "class_id": [1, 1, 2, 2],
})

classes = pd.DataFrame({
    "class_id": [1, 2, 3],
    "name": ["Year 7A", "Year 7B", "Year 7C"],
})

print(pupils.merge(classes, on="class_id", suffixes=("", "_class")))

   id     name  class_id name_class
0   1   Mia         1    Year 7A
1   2  Leo         1    Year 7A
2   3   Victor         2    Year 7B
3   4      Mei         2    Year 7B

Same column name on both sides (name) → use suffixes so pandas knows what to call them in the result.

Four join styles

how="inner"  keep only matches in BOTH (default)
how="left"   keep all left rows; NaN where right misses
how="right"  keep all right rows; NaN where left misses
how="outer"  keep everything from both

pupils.merge(classes, on="class_id", how="left")
pupils.merge(classes, on="class_id", how="outer", suffixes=("_p", "_c"))

Multi-column join

a.merge(b, on=["year", "subject"])

Differently-named keys

orders.merge(customers, left_on="customer_id", right_on="id")
# drops the duplicate id column afterwards
.drop(columns="id")

indicator=True — see who matched

merged = pupils.merge(classes, on="class_id", how="outer", indicator=True)
print(merged["_merge"].value_counts())

_merge
both          4
right_only    1   ← Year 7C had no pupil
left_only     0
Name: count, dtype: int64

Super useful for debugging unexpected join results.

concat — stack frames

# Same columns, more rows
all_pupils = pd.concat([pupils_2024, pupils_2025, pupils_2026], ignore_index=True)

# Different columns, side-by-side
combined = pd.concat([df_a, df_b], axis=1)

concat is for stacking — when both frames have the same shape and you want to add more rows (or more columns side-by-side).

import pandas as pd

sales = pd.read_csv("sales.csv")          # date, customer_id, product_id, qty
customers = pd.read_csv("customers.csv")  # id, name, city
products  = pd.read_csv("products.csv")   # id, name, price

# Two merges to enrich every sale with customer name and product name + price
joined = (sales
          .merge(customers, left_on="customer_id", right_on="id",
                 suffixes=("", "_c"))
          .drop(columns="id")
          .merge(products, left_on="product_id", right_on="id",
                 suffixes=("", "_p"))
          .drop(columns="id"))

joined["revenue"] = joined["qty"] * joined["price"]

# Top-spending customers
top = (joined.groupby("name")["revenue"].sum()
              .sort_values(ascending=False)
              .head(5))
print(top)

Read the diff

Chained merges enrich the sales frame with two lookup tables, the same way a SQL query would join three tables. Then groupby + sum + head gives you the top customers — what would have been a 30-line script in plain Python becomes 10 lines, vectorised.

merge = JOIN; concat = stack. Four join styles with how=. left_on/right_on for differently-named keys; suffixes for collisions. indicator=True when in doubt. Tomorrow we sort + slice for top-N reports.