Polars: Groupby and idxmin#
Welcome back to Cameron’s Corner! It’s the third week of January, and, instead of talking about graphs, I want to take a dive into Polars. I recently addressed a question on Polars’ Discord server, diving into the different ways to perform an “index minimum” operation across groups.
Sure, there’s a built-in Expression.idx_min(), but it operates a little differently than it does in pandas. Let’s take a look:
Data#
from polars import (
DataFrame as pl_DataFrame,
Categorical,
Config,
datetime_range,
col,
all as pl_all
)
from datetime import datetime, timedelta
from numpy.random import default_rng
from string import ascii_uppercase
def make_df(ngroups, group_size):
size = ngroups * group_size
rng = default_rng(0)
return pl_DataFrame({
'date' : datetime_range(
start := datetime(2000, 1, 1),
end=start + timedelta(hours=size-1),
interval='1H',
eager=True
),
'group': [*ascii_uppercase[:ngroups]] * group_size,
'value': rng.integers(0, 1_000, size=size),
}).cast({'group': Categorical})
pl_df = make_df(3, 10)
with Config(tbl_rows=6):
display(pl_df)
| date | group | value |
|---|---|---|
| datetime[μs] | cat | i64 |
| 2000-01-01 00:00:00 | "A" | 850 |
| 2000-01-01 01:00:00 | "B" | 636 |
| 2000-01-01 02:00:00 | "C" | 511 |
| … | … | … |
| 2000-01-02 03:00:00 | "A" | 33 |
| 2000-01-02 04:00:00 | "B" | 764 |
| 2000-01-02 05:00:00 | "C" | 729 |
Approaches#
pandas#
from pandas import DataFrame as pd_DataFrame
pd_df = pl_df.to_pandas()
pd_df.groupby('group', observed=True)['value'].idxmin()
group
A 27
B 7
C 23
Name: value, dtype: int64
Polars#
pl_df.group_by('group').agg(col('value').arg_min()).head()
| group | value |
|---|---|
| cat | u32 |
| "A" | 9 |
| "B" | 2 |
| "C" | 7 |
As you can see, the results are different! If you take your best guess as to what has happened here, then you might wonder, “Is pandas doing something with the index?”
Of course it is!
The difference in these results is that pandas returns index values in reference to the original DataFrame, or, more specifically, the original .index. Our index was a simple RangeIndex from
0 to the length of our data. The idxmin operation returns the index value, where
a minimum occurs in our Series ('value'). In contrast, Polars returns the index position
relative to the group.
While this difference might seem superficial—we all know that Polars does not
like the .index—it does lead to an interesting question:
“How do I find the row in each group where the minimum occurred?”
In pandas, we can simply wrap the result in a .loc
pd_df.loc[
# same operation as before!
pd_df.groupby('group', observed=True)['value'].idxmin()
]
| date | group | value | |
|---|---|---|---|
| 27 | 2000-01-02 03:00:00 | A | 33 |
| 7 | 2000-01-01 07:00:00 | B | 16 |
| 23 | 2000-01-01 23:00:00 | C | 2 |
However, in Polars, we don’t have an easy way to convert from these grouped idx_min
back to our original data. Instead, we need to alter our original expression. I have
noted a few ways to arrive at this result:
registry = []
def register(func):
registry.append(func)
return func
@register
def filter_over_argmin(df):
group_row_ids = col('group').cum_count().over(col('group')) - 1
group_arg_min = col('value').arg_min().over(col('group'))
return df.filter(group_row_ids == group_arg_min)
@register
def sort_keep_first(df):
return df.sort(['group', 'value']).unique('group', keep='first')
@register
def groupby_sort_first(df):
return df.group_by('group').agg(pl_all().sort_by('value').first())
@register
def groupby_get(df):
return (
df.group_by('group').agg(pl_all().get(col('value').arg_min()))
)
Now, let’s see these queries in action. I’ll use LazyFrames to ensure that any
available optimizations can be made in advance.
I’m also going to up the ante by working with a much larger data set:
pl_df = make_df(25, 10_000)
pd_df = pl_df.to_pandas()
with Config(tbl_rows=6):
display(pl_df)
| date | group | value |
|---|---|---|
| datetime[μs] | cat | i64 |
| 2000-01-01 00:00:00 | "A" | 850 |
| 2000-01-01 01:00:00 | "B" | 636 |
| 2000-01-01 02:00:00 | "C" | 511 |
| … | … | … |
| 2028-07-08 13:00:00 | "W" | 665 |
| 2028-07-08 14:00:00 | "X" | 392 |
| 2028-07-08 15:00:00 | "Y" | 992 |
results = {}
for func in registry:
res = func(pl_df.lazy())
print(f'{func.__name__:-^40}')
%timeit -o -n 10 -r 5 res.collect()
results[func] = res.collect()
print()
-----------filter_over_argmin-----------
4.28 ms ± 141 µs per loop (mean ± std. dev. of 5 runs, 10 loops each)
------------sort_keep_first-------------
10.3 ms ± 330 µs per loop (mean ± std. dev. of 5 runs, 10 loops each)
-----------groupby_sort_first-----------
4.04 ms ± 338 µs per loop (mean ± std. dev. of 5 runs, 10 loops each)
--------------groupby_get---------------
1.72 ms ± 91.8 µs per loop (mean ± std. dev. of 5 runs, 10 loops each)
And, for reference, here is how pandas performs:
%timeit -n 10 -r 5 pd_df.loc[pd_df.groupby('group', observed=True)['value'].idxmin()]
3.57 ms ± 131 µs per loop (mean ± std. dev. of 5 runs, 10 loops each)
Finally, let’s verify each of our Polars results arrived at the same answer:
from polars import align_frames
from itertools import pairwise
for left, right in pairwise(align_frames(*results.values(), on='group')):
left, right = left.sort('group'), right.sort('group')
assert left['value'].equals(right['value'])
Wrap-Up#
There we have it: how to answer a group_by...idxmin problem in both pandas and
Polars. As usual, when it comes to Polars, make sure you leverage the expression
syntax as best you can, and leave the optimizations to the engine. However, I will
note that you may see some variance in the speed of your operation, depending on
the approach you take, as seen above. Try to hit parallel-processing fastpaths like
groupby operations (depending on your cardinality) and working across columns.
What do you think? Let us know on the DUTC Discord server.
That’s all for today. Until next time!