I have the following DF on which i am trying to construct a new column which is will be cum_sum.
import polars as pl
pl.Config(tbl_rows=20)
df= pl.LazyFrame({
"a" : [0,1,2,3,4,5,6,7,8,9],
"b" : [33,35,33,32,33,37,33,29,34,36],
"c" : [33,35,35,35,35,37,37,37,37,37],
})
df = df.with_columns(pl.when(pl.col('b') == pl.col('c')).then(0).otherwise(1).alias('temp'))
df = df.with_columns(pl.when(pl.col('temp') == 0).then(0).otherwise(pl.col('temp').cum_sum()).alias('c_sum_n'))
Currently i am getting the following output ,
>>> df.collect()
shape: (10, 5)
┌─────┬─────┬─────┬──────┬─────────┐
│ a ┆ b ┆ c ┆ temp ┆ c_sum_n │
│ --- ┆ --- ┆ --- ┆ --- ┆ --- │
│ i64 ┆ i64 ┆ i64 ┆ i32 ┆ i32 │
╞═════╪═════╪═════╪══════╪═════════╡
│ 0 ┆ 33 ┆ 33 ┆ 0 ┆ 0 │
│ 1 ┆ 35 ┆ 35 ┆ 0 ┆ 0 │
│ 2 ┆ 33 ┆ 35 ┆ 1 ┆ 1 │
│ 3 ┆ 32 ┆ 35 ┆ 1 ┆ 2 │
│ 4 ┆ 33 ┆ 35 ┆ 1 ┆ 3 │
│ 5 ┆ 37 ┆ 37 ┆ 0 ┆ 0 │
│ 6 ┆ 33 ┆ 37 ┆ 1 ┆ 4 │
│ 7 ┆ 29 ┆ 37 ┆ 1 ┆ 5 │
│ 8 ┆ 34 ┆ 37 ┆ 1 ┆ 6 │
│ 9 ┆ 36 ┆ 37 ┆ 1 ┆ 7 │
└─────┴─────┴─────┴──────┴─────────┘
but the desired output i am looking is
>>> df.collect()
shape: (10, 5)
┌─────┬─────┬─────┬──────┬─────────┐
│ a ┆ b ┆ c ┆ temp ┆ c_sum_n │
│ --- ┆ --- ┆ --- ┆ --- ┆ --- │
│ i64 ┆ i64 ┆ i64 ┆ i32 ┆ i32 │
╞═════╪═════╪═════╪══════╪═════════╡
│ 0 ┆ 33 ┆ 33 ┆ 0 ┆ 0 │
│ 1 ┆ 35 ┆ 35 ┆ 0 ┆ 0 │
│ 2 ┆ 33 ┆ 35 ┆ 1 ┆ 1 │
│ 3 ┆ 32 ┆ 35 ┆ 1 ┆ 2 │
│ 4 ┆ 33 ┆ 35 ┆ 1 ┆ 3 │
│ 5 ┆ 37 ┆ 37 ┆ 0 ┆ 0 │
│ 6 ┆ 33 ┆ 37 ┆ 1 ┆ 1 │
│ 7 ┆ 29 ┆ 37 ┆ 1 ┆ 2 │
│ 8 ┆ 34 ┆ 37 ┆ 1 ┆ 3 │
│ 9 ┆ 36 ┆ 37 ┆ 1 ┆ 4 │
└─────┴─────┴─────┴──────┴─────────┘
i,e when col b == col c, i would like to have cum_sum value reset to zero and start all over again. Any help, either in python polars or rust polars will be helpful.
Here's one way to do this using over:
We extract the predicate col("b") == col("c"), then we do a cum_sum of ~predicate over the cum_sum of predicate.
import polars as pl
df = pl.DataFrame(
{
"a": [0, 1, 2, 3, 4, 5, 6, 7, 8, 9],
"b": [33, 35, 33, 32, 33, 37, 33, 29, 34, 36],
"c": [33, 35, 35, 35, 35, 37, 37, 37, 37, 37],
}
)
predicate = pl.col("b") == pl.col("c")
df = df.with_columns((~predicate).cum_sum().over(predicate.cum_sum()).alias("c_sum_n"))
print(df)
Output:
shape: (10, 4)
┌─────┬─────┬─────┬─────────┐
│ a ┆ b ┆ c ┆ c_sum_n │
│ --- ┆ --- ┆ --- ┆ --- │
│ i64 ┆ i64 ┆ i64 ┆ u32 │
╞═════╪═════╪═════╪═════════╡
│ 0 ┆ 33 ┆ 33 ┆ 0 │
│ 1 ┆ 35 ┆ 35 ┆ 0 │
│ 2 ┆ 33 ┆ 35 ┆ 1 │
│ 3 ┆ 32 ┆ 35 ┆ 2 │
│ 4 ┆ 33 ┆ 35 ┆ 3 │
│ 5 ┆ 37 ┆ 37 ┆ 0 │
│ 6 ┆ 33 ┆ 37 ┆ 1 │
│ 7 ┆ 29 ┆ 37 ┆ 2 │
│ 8 ┆ 34 ┆ 37 ┆ 3 │
│ 9 ┆ 36 ┆ 37 ┆ 4 │
└─────┴─────┴─────┴─────────┘
How this works:
Doing a cum_sum on the predicate gives us different values for each group which is exactly what we need to split the cum_sum over ~predicate:
print(df.with_columns(group=predicate.cum_sum()))
Output:
┌─────┬─────┬─────┬───────┐
│ a ┆ b ┆ c ┆ group │
│ --- ┆ --- ┆ --- ┆ --- │
│ i64 ┆ i64 ┆ i64 ┆ u32 │
╞═════╪═════╪═════╪═══════╡
│ 0 ┆ 33 ┆ 33 ┆ 1 │
│ 1 ┆ 35 ┆ 35 ┆ 2 │
│ 2 ┆ 33 ┆ 35 ┆ 2 │
│ 3 ┆ 32 ┆ 35 ┆ 2 │
│ 4 ┆ 33 ┆ 35 ┆ 2 │
│ 5 ┆ 37 ┆ 37 ┆ 3 │
│ 6 ┆ 33 ┆ 37 ┆ 3 │
│ 7 ┆ 29 ┆ 37 ┆ 3 │
│ 8 ┆ 34 ┆ 37 ┆ 3 │
│ 9 ┆ 36 ┆ 37 ┆ 3 │
└─────┴─────┴─────┴───────┘