TopolearnPer-group totals, counts and averages are the most common analysis there is.
df.group_by("key").agg(...) splits the rows by key, evaluates each aggregation expression within every group, and returns one row per group. Group by several columns with group_by("a", "b").
Inside agg, any expression works:
pl.col("amount").sum(),.mean(),.median(),.min(),.max(),.std()pl.len()for the row count,pl.col("user").n_unique()for distinct valuespl.col("amount").filter(pl.col("type") == "refund").sum()to aggregate a subsetpl.col("item").first()orpl.col("item")alone, which collects each group's values into a list
Name results with .alias() or keyword arguments, e.g. agg(total=pl.col("amount").sum()). All aggregations in one agg run in parallel.
Order is not guaranteed. Group-by output comes back in whatever order the parallel engine finishes, and in Polars 2 the streaming engine is the default, so don't rely on first-seen order. Sort the result, or pass maintain_order=True (slower) when order matters.
Compared with pandas there is no index to reset afterwards: the keys are ordinary columns in the result. If you know SQL, group_by().agg() is exactly GROUP BY with the select list inside agg.
Example
import polars as pl
sales = pl.DataFrame({
"shop": ["A", "B", "A", "B", "A"],
"kind": ["sale", "sale", "refund", "sale", "sale"],
"amount": [10.0, 5.0, -3.0, 7.0, 4.0],
})
summary = (
sales.group_by("shop")
.agg(
n=pl.len(),
net=pl.col("amount").sum(),
refunds=pl.col("amount").filter(pl.col("kind") == "refund").sum(),
kinds=pl.col("kind").unique().sort(),
)
.sort("shop") # output order is not guaranteed
)
print(summary.rows())
# [('A', 3, 11.0, -3.0, ['refund', 'sale']), ('B', 2, 12.0, 0.0, ['sale'])]Remember
group_by(keys).agg(exprs)gives one row per group- Any expression can be an aggregation, including filtered ones
- Sort the result or use
maintain_order=True; order is not guaranteed
Watch out for
- Assuming groups come back in first-appearance order.