Window function
Window function — windowFunction
Compute a window/analytics value into a new column, optionally scoped to a partition and ordered within it. Row order is preserved; the result is added as a new column.
Use cases
- Rank rows within each group (top product per region).
- Running totals, cumulative max/min over an ordered key.
- Compare a row to the previous/next one with
lag/lead.
What it does
A window function adds a new column computed from a window of rows — scoped to a partition (group) and ordered within it. The original row order is preserved; the calculation happens internally and the result is added alongside the existing columns.
Below: cumsum partitioned by region and ordered by date adds a per-region
running total without collapsing rows.
Before
| region | date | amount |
|---|---|---|
| North | 2024-01-01 | 100 |
| North | 2024-01-02 | 150 |
| South | 2024-01-01 | 80 |
| South | 2024-01-02 | 200 |
4 rows · 3 cols
Window: cumsum (partition_by=region, order_by=date, target=amount) → running_total
After
| region | date | amount | running_totalnew |
|---|---|---|---|
| North | 2024-01-01 | 100 | 100 |
| North | 2024-01-02 | 150 | 250 |
| South | 2024-01-01 | 80 | 80 |
| South | 2024-01-02 | 200 | 280 |
4 rows · 4 cols
Configuration
| Config key | Type | Required | Description |
|---|---|---|---|
function | string | Yes | row_number, rank, dense_rank, cumcount, cumsum, cummax, cummin, lag, lead |
new_column | string | Yes | Name of the column to add |
partition_by | string[] | No | Restart the window within each group (empty = whole table) |
order_by | string[] | Conditional | Row order within the window; required for rank/dense_rank |
target | string | Conditional | Value column; required for cumsum/cummax/cummin/lag/lead |
offset | int | No | Shift distance for lag/lead (default 1) |
descending | bool | No | Order descending (default false) |
Generated Python code
# function: cumsum, partition_by: ['region'], order_by: ['date'], target: 'amount'
df_2 = df_1.assign(running_total=lambda _d: _d.sort_values('date', kind='stable').groupby('region', sort=False)['amount'].cumsum())
Tips & common mistakes
- Each function needs its own inputs: ranking needs
order_by; value functions (cumsum,cummax,cummin,lag,lead) need atarget. rank/dense_rankrank by the firstorder_bycolumn.- For
lag/lead, rows at the window edge with no neighbor are null. - Use
partition_byto restart the calculation per group; leave it empty to run across the whole table.
Row order in the output is preserved — the window sorts internally and restores the original order, so this node is safe to place anywhere.