pythonpython-polars

Python Polars join on column with greater or equal


I have two polars dataframe, one dataframe df_1 with two columns start and end the other dataframe df_2 one with a column dates and I want to do a left join on df_2 under the condition that the dates column is in between the start and end column.

To make it more obvious what I want to do here is an example

DATA

import polars as pl
from datetime import date

df_1 = pl.DataFrame(
    {
        "id": ["abc", "abc", "456"],
        "start": [date(2022, 1, 1), date(2022, 3, 4), date(2022, 5, 11)],
        "end": [date(2022, 2, 4), date(2022, 3, 10), date(2022, 5, 16)],
        "value": [10, 3, 4]

    }
)

df_2 = pl.DataFrame(
    {
        "id": ["abc", "abc", "456", "abc", "abc", "456"],
        "dates": [date(2022, 1, 2), date(2022, 3, 4), date(2022, 5, 11), date(2022, 1, 4), date(2022, 3, 7), date(2022, 5, 13)],
    }
)

So now I would join on id and that dates is in between start and end and the result should look like that

RESULT

shape: (6, 3)
┌─────┬────────────┬───────┐
│ id  ┆ dates      ┆ value │
│ --- ┆ ---        ┆ ---   │
│ str ┆ date       ┆ i64   │
╞═════╪════════════╪═══════╡
│ abc ┆ 2022-01-02 ┆ 10    │
│ abc ┆ 2022-03-04 ┆ 3     │
│ 456 ┆ 2022-05-11 ┆ 4     │
│ abc ┆ 2022-01-04 ┆ 10    │
│ abc ┆ 2022-03-07 ┆ 3     │
│ 456 ┆ 2022-05-13 ┆ 4     │
└─────┴────────────┴───────┘

Solution

  • (I'm going to assume that your intervals in df_1 do not overlap for a particular id - otherwise, there may not be a unique value that we can assign to the id/dates combinations in df_2.)

    One way to do this is with join_asof.

    The Algorithm

    (
        df_2
        .sort("dates")
        .join_asof(
            df_1.sort("start"),
            by="id",
            left_on="dates",
            right_on="start",
            strategy="backward",
        )
        .with_columns(
            pl.when(pl.col('dates') <= pl.col('end'))
            .then(pl.col('value'))
        )
        .select('id', 'dates', 'value')
    )
    
    
    shape: (6, 3)
    ┌─────┬────────────┬───────┐
    │ id  ┆ dates      ┆ value │
    │ --- ┆ ---        ┆ ---   │
    │ str ┆ date       ┆ i64   │
    ╞═════╪════════════╪═══════╡
    │ abc ┆ 2022-01-02 ┆ 10    │
    │ abc ┆ 2022-01-04 ┆ 10    │
    │ abc ┆ 2022-03-04 ┆ 3     │
    │ abc ┆ 2022-03-07 ┆ 3     │
    │ 456 ┆ 2022-05-11 ┆ 4     │
    │ 456 ┆ 2022-05-13 ┆ 4     │
    └─────┴────────────┴───────┘
    

    In Steps

    First, let's append some additional rows to df_2, to show what will happen if a particular row is not contained in an interval in df_1. I'll also add a row number, for easier inspection.

    df_2 = pl.DataFrame(
        {
            "id": ["abc", "abc", "456", "abc", "abc", "456", "abc", "abc", "abc"],
            "dates": [
                date(2022, 1, 2),
                date(2022, 3, 4),
                date(2022, 5, 11),
                date(2022, 1, 4),
                date(2022, 3, 7),
                date(2022, 5, 13),
                date(2021, 12, 31),
                date(2022, 3, 1),
                date(2023, 1, 1),
            ],
        }
    ).with_row_index()
    df_2
    
    shape: (9, 3)
    ┌───────┬─────┬────────────┐
    │ index ┆ id  ┆ dates      │
    │ ---   ┆ --- ┆ ---        │
    │ u32   ┆ str ┆ date       │
    ╞═══════╪═════╪════════════╡
    │ 0     ┆ abc ┆ 2022-01-02 │
    │ 1     ┆ abc ┆ 2022-03-04 │
    │ 2     ┆ 456 ┆ 2022-05-11 │
    │ 3     ┆ abc ┆ 2022-01-04 │
    │ 4     ┆ abc ┆ 2022-03-07 │
    │ 5     ┆ 456 ┆ 2022-05-13 │
    │ 6     ┆ abc ┆ 2021-12-31 │
    │ 7     ┆ abc ┆ 2022-03-01 │
    │ 8     ┆ abc ┆ 2023-01-01 │
    └───────┴─────┴────────────┘
    

    The join_asof step finds the latest start date that is on or before the dates date. Since intervals do not overlap, this is the only interval that might contain the dates date.

    For our purposes, I'll make a copy of the start column so that we can inspect the results. (The start column will not be in the results of the join_asof.)

    Note that for a join_asof, both DataFrames must be sorted by the asof columns (dates and start in this case).

    (
        df_2
        .sort("dates")
        .join_asof(
            df_1.sort("start").with_columns(pl.col("start").alias("start_df1")),
            by="id",
            left_on="dates",
            right_on="start",
            strategy="backward",
        )
        .sort("index")
    )
    
    shape: (9, 7)
    ┌───────┬─────┬────────────┬────────────┬────────────┬───────┬────────────┐
    │ index ┆ id  ┆ dates      ┆ start      ┆ end        ┆ value ┆ start_df1  │
    │ ---   ┆ --- ┆ ---        ┆ ---        ┆ ---        ┆ ---   ┆ ---        │
    │ u32   ┆ str ┆ date       ┆ date       ┆ date       ┆ i64   ┆ date       │
    ╞═══════╪═════╪════════════╪════════════╪════════════╪═══════╪════════════╡
    │ 0     ┆ abc ┆ 2022-01-02 ┆ 2022-01-01 ┆ 2022-02-04 ┆ 10    ┆ 2022-01-01 │
    │ 1     ┆ abc ┆ 2022-03-04 ┆ 2022-03-04 ┆ 2022-03-10 ┆ 3     ┆ 2022-03-04 │
    │ 2     ┆ 456 ┆ 2022-05-11 ┆ 2022-05-11 ┆ 2022-05-16 ┆ 4     ┆ 2022-05-11 │
    │ 3     ┆ abc ┆ 2022-01-04 ┆ 2022-01-01 ┆ 2022-02-04 ┆ 10    ┆ 2022-01-01 │
    │ 4     ┆ abc ┆ 2022-03-07 ┆ 2022-03-04 ┆ 2022-03-10 ┆ 3     ┆ 2022-03-04 │
    │ 5     ┆ 456 ┆ 2022-05-13 ┆ 2022-05-11 ┆ 2022-05-16 ┆ 4     ┆ 2022-05-11 │
    │ 6     ┆ abc ┆ 2021-12-31 ┆ null       ┆ null       ┆ null  ┆ null       │
    │ 7     ┆ abc ┆ 2022-03-01 ┆ 2022-01-01 ┆ 2022-02-04 ┆ 10    ┆ 2022-01-01 │
    │ 8     ┆ abc ┆ 2023-01-01 ┆ 2022-03-04 ┆ 2022-03-10 ┆ 3     ┆ 2022-03-04 │
    └───────┴─────┴────────────┴────────────┴────────────┴───────┴────────────┘
    

    The last three rows are the ones that I added.

    In the last step, we'll inspect the end date, and null out any values where dates is beyond end.

    (
        df_2
        .sort("dates")
        .join_asof(
            df_1.sort("start").with_columns(pl.col("start").alias("start_df1")),
            by="id",
            left_on="dates",
            right_on="start",
            strategy="backward",
        )
        .with_columns(
            pl.when(pl.col('dates') <= pl.col('end'))
            .then(pl.col('value'))
        )
        .sort("index")
    )
    
    shape: (9, 7)
    ┌───────┬─────┬────────────┬────────────┬────────────┬───────┬────────────┐
    │ index ┆ id  ┆ dates      ┆ start      ┆ end        ┆ value ┆ start_df1  │
    │ ---   ┆ --- ┆ ---        ┆ ---        ┆ ---        ┆ ---   ┆ ---        │
    │ u32   ┆ str ┆ date       ┆ date       ┆ date       ┆ i64   ┆ date       │
    ╞═══════╪═════╪════════════╪════════════╪════════════╪═══════╪════════════╡
    │ 0     ┆ abc ┆ 2022-01-02 ┆ 2022-01-01 ┆ 2022-02-04 ┆ 10    ┆ 2022-01-01 │
    │ 1     ┆ abc ┆ 2022-03-04 ┆ 2022-03-04 ┆ 2022-03-10 ┆ 3     ┆ 2022-03-04 │
    │ 2     ┆ 456 ┆ 2022-05-11 ┆ 2022-05-11 ┆ 2022-05-16 ┆ 4     ┆ 2022-05-11 │
    │ 3     ┆ abc ┆ 2022-01-04 ┆ 2022-01-01 ┆ 2022-02-04 ┆ 10    ┆ 2022-01-01 │
    │ 4     ┆ abc ┆ 2022-03-07 ┆ 2022-03-04 ┆ 2022-03-10 ┆ 3     ┆ 2022-03-04 │
    │ 5     ┆ 456 ┆ 2022-05-13 ┆ 2022-05-11 ┆ 2022-05-16 ┆ 4     ┆ 2022-05-11 │
    │ 6     ┆ abc ┆ 2021-12-31 ┆ null       ┆ null       ┆ null  ┆ null       │
    │ 7     ┆ abc ┆ 2022-03-01 ┆ 2022-01-01 ┆ 2022-02-04 ┆ null  ┆ 2022-01-01 │
    │ 8     ┆ abc ┆ 2023-01-01 ┆ 2022-03-04 ┆ 2022-03-10 ┆ null  ┆ 2022-03-04 │
    └───────┴─────┴────────────┴────────────┴────────────┴───────┴────────────┘
    

    You can see that the last three rows that I added (which purposely don't match any intervals in df_1) have null as value.

    Instead of using when/then/otherwise to set value to null, you can filter these out, if that's what you need.