polars.LazyFrame.groupby_dynamic#

LazyFrame.groupby_dynamic(
index_column: IntoExpr,
*,
every: str | timedelta,
period: str | timedelta | None = None,
offset: str | timedelta | None = None,
truncate: bool = True,
include_boundaries: bool = False,
closed: ClosedInterval = 'left',
by: IntoExpr | Iterable[IntoExpr] | None = None,
start_by: StartBy = 'window',
check_sorted: bool = True,
) → LazyGroupBy[source]#

Group based on a time value (or index value of type Int32, Int64).

Time windows are calculated and rows are assigned to windows. Different from a normal groupby is that a row can be member of multiple groups. The time/index window could be seen as a rolling window, with a window size determined by dates/times/values instead of slots in the DataFrame.

A window is defined by:

  • every: interval of the window

  • period: length of the window

  • offset: offset of the window

The every, period and offset arguments are created with the following string language:

  • 1ns (1 nanosecond)

  • 1us (1 microsecond)

  • 1ms (1 millisecond)

  • 1s (1 second)

  • 1m (1 minute)

  • 1h (1 hour)

  • 1d (1 calendar day)

  • 1w (1 calendar week)

  • 1mo (1 calendar month)

  • 1q (1 calendar quarter)

  • 1y (1 calendar year)

  • 1i (1 index count)

Or combine them: “3d12h4m25s” # 3 days, 12 hours, 4 minutes, and 25 seconds

Suffix with “_saturating” to indicate that dates too large for their month should saturate at the largest date (e.g. 2022-02-29 -> 2022-02-28) instead of erroring.

By “calendar day”, we mean the corresponding time on the next day (which may not be 24 hours, due to daylight savings). Similarly for “calendar week”, “calendar month”, “calendar quarter”, and “calendar year”.

In case of a groupby_dynamic on an integer column, the windows are defined by:

  • “1i” # length 1

  • “10i” # length 10

Warning

The index column must be sorted in ascending order. If by is passed, then the index column must be sorted in ascending order within each group.

Parameters:
index_column

Column used to group based on the time window. Often of type Date/Datetime. This column must be sorted in ascending order (or, if by is specified, then it must be sorted in ascending order within each group).

In case of a dynamic groupby on indices, dtype needs to be one of {Int32, Int64}. Note that Int32 gets temporarily cast to Int64, so if performance matters use an Int64 column.

every

interval of the window

period

length of the window, if None it is equal to ‘every’

offset

offset of the window if None and period is None it will be equal to negative every

truncate

truncate the time value to the window lower bound

include_boundaries

Add the lower and upper bound of the window to the “_lower_bound” and “_upper_bound” columns. This will impact performance because it’s harder to parallelize

closed{‘right’, ‘left’, ‘both’, ‘none’}

Define which sides of the temporal interval are closed (inclusive).

by

Also group by this column/these columns

start_by{‘window’, ‘datapoint’, ‘monday’, ‘tuesday’, ‘wednesday’, ‘thursday’, ‘friday’, ‘saturday’, ‘sunday’}

The strategy to determine the start of the first window by.

  • ‘window’: Truncate the start of the window with the ‘every’ argument. Note that weekly windows start on Monday.

  • ‘datapoint’: Start from the first encountered data point.

  • a day of the week (only takes effect if every contains 'w'):

    • ‘monday’: Start the window on the Monday before the first data point.

    • ‘tuesday’: Start the window on the Tuesday before the first data point.

    • …

    • ‘sunday’: Start the window on the Sunday before the first data point.

check_sorted

When the by argument is given, polars can not check sortedness by the metadata and has to do a full scan on the index column to verify data is sorted. This is expensive. If you are sure the data within the by groups is sorted, you can set this to False. Doing so incorrectly will lead to incorrect output

Returns:
LazyGroupBy

Object you can call .agg on to aggregate by groups, the result of which will be sorted by index_column (but note that if by columns are passed, it will only be sorted within each by group).

See also

groupby_rolling

Notes

If you’re coming from pandas, then

# polars
df.groupby_dynamic("ts", every="1d").agg(pl.col("value").sum())

is equivalent to

# pandas
df.set_index("ts").resample("D")["value"].sum().reset_index()

though note that, unlike pandas, polars doesn’t add extra rows for empty windows. If you need index_column to be evenly spaced, then please combine with DataFrame.upsample().

Examples

>>> from datetime import datetime
>>> # create an example dataframe
>>> lf = pl.LazyFrame(
...     {
...         "time": pl.date_range(
...             start=datetime(2021, 12, 16),
...             end=datetime(2021, 12, 16, 3),
...             interval="30m",
...             eager=True,
...         ),
...         "n": range(7),
...     }
... )
>>> lf.collect()
shape: (7, 2)
┌─────────────────────┬─────┐
│ time                ┆ n   │
│ ---                 ┆ --- │
│ datetime[μs]        ┆ i64 │
╞═════════════════════╪═════╡
│ 2021-12-16 00:00:00 ┆ 0   │
│ 2021-12-16 00:30:00 ┆ 1   │
│ 2021-12-16 01:00:00 ┆ 2   │
│ 2021-12-16 01:30:00 ┆ 3   │
│ 2021-12-16 02:00:00 ┆ 4   │
│ 2021-12-16 02:30:00 ┆ 5   │
│ 2021-12-16 03:00:00 ┆ 6   │
└─────────────────────┴─────┘

Group by windows of 1 hour starting at 2021-12-16 00:00:00.

>>> lf.groupby_dynamic("time", every="1h", closed="right").agg(
...     [
...         pl.col("time").min().alias("time_min"),
...         pl.col("time").max().alias("time_max"),
...     ]
... ).collect()
shape: (4, 3)
┌─────────────────────┬─────────────────────┬─────────────────────┐
│ time                ┆ time_min            ┆ time_max            │
│ ---                 ┆ ---                 ┆ ---                 │
│ datetime[μs]        ┆ datetime[μs]        ┆ datetime[μs]        │
╞═════════════════════╪═════════════════════╪═════════════════════╡
│ 2021-12-15 23:00:00 ┆ 2021-12-16 00:00:00 ┆ 2021-12-16 00:00:00 │
│ 2021-12-16 00:00:00 ┆ 2021-12-16 00:30:00 ┆ 2021-12-16 01:00:00 │
│ 2021-12-16 01:00:00 ┆ 2021-12-16 01:30:00 ┆ 2021-12-16 02:00:00 │
│ 2021-12-16 02:00:00 ┆ 2021-12-16 02:30:00 ┆ 2021-12-16 03:00:00 │
└─────────────────────┴─────────────────────┴─────────────────────┘

The window boundaries can also be added to the aggregation result

>>> lf.groupby_dynamic(
...     "time", every="1h", include_boundaries=True, closed="right"
... ).agg([pl.col("time").count().alias("time_count")]).collect()
shape: (4, 4)
┌─────────────────────┬─────────────────────┬─────────────────────┬────────────┐
│ _lower_boundary     ┆ _upper_boundary     ┆ time                ┆ time_count │
│ ---                 ┆ ---                 ┆ ---                 ┆ ---        │
│ datetime[μs]        ┆ datetime[μs]        ┆ datetime[μs]        ┆ u32        │
╞═════════════════════╪═════════════════════╪═════════════════════╪════════════╡
│ 2021-12-15 23:00:00 ┆ 2021-12-16 00:00:00 ┆ 2021-12-15 23:00:00 ┆ 1          │
│ 2021-12-16 00:00:00 ┆ 2021-12-16 01:00:00 ┆ 2021-12-16 00:00:00 ┆ 2          │
│ 2021-12-16 01:00:00 ┆ 2021-12-16 02:00:00 ┆ 2021-12-16 01:00:00 ┆ 2          │
│ 2021-12-16 02:00:00 ┆ 2021-12-16 03:00:00 ┆ 2021-12-16 02:00:00 ┆ 2          │
└─────────────────────┴─────────────────────┴─────────────────────┴────────────┘

When closed=”left”, should not include right end of interval [lower_bound, upper_bound)

>>> lf.groupby_dynamic("time", every="1h", closed="left").agg(
...     [
...         pl.col("time").count().alias("time_count"),
...         pl.col("time").alias("time_agg_list"),
...     ]
... ).collect()
shape: (4, 3)
┌─────────────────────┬────────────┬───────────────────────────────────┐
│ time                ┆ time_count ┆ time_agg_list                     │
│ ---                 ┆ ---        ┆ ---                               │
│ datetime[μs]        ┆ u32        ┆ list[datetime[μs]]                │
╞═════════════════════╪════════════╪═══════════════════════════════════╡
│ 2021-12-16 00:00:00 ┆ 2          ┆ [2021-12-16 00:00:00, 2021-12-16… │
│ 2021-12-16 01:00:00 ┆ 2          ┆ [2021-12-16 01:00:00, 2021-12-16… │
│ 2021-12-16 02:00:00 ┆ 2          ┆ [2021-12-16 02:00:00, 2021-12-16… │
│ 2021-12-16 03:00:00 ┆ 1          ┆ [2021-12-16 03:00:00]             │
└─────────────────────┴────────────┴───────────────────────────────────┘

When closed=”both” the time values at the window boundaries belong to 2 groups.

>>> lf.groupby_dynamic("time", every="1h", closed="both").agg(
...     pl.col("time").count().alias("time_count")
... ).collect()
shape: (5, 2)
┌─────────────────────┬────────────┐
│ time                ┆ time_count │
│ ---                 ┆ ---        │
│ datetime[μs]        ┆ u32        │
╞═════════════════════╪════════════╡
│ 2021-12-15 23:00:00 ┆ 1          │
│ 2021-12-16 00:00:00 ┆ 3          │
│ 2021-12-16 01:00:00 ┆ 3          │
│ 2021-12-16 02:00:00 ┆ 3          │
│ 2021-12-16 03:00:00 ┆ 1          │
└─────────────────────┴────────────┘

Dynamic groupbys can also be combined with grouping on normal keys

>>> lf = pl.LazyFrame(
...     {
...         "time": pl.date_range(
...             start=datetime(2021, 12, 16),
...             end=datetime(2021, 12, 16, 3),
...             interval="30m",
...             eager=True,
...         ),
...         "groups": ["a", "a", "a", "b", "b", "a", "a"],
...     }
... )
>>> lf.collect()
shape: (7, 2)
┌─────────────────────┬────────┐
│ time                ┆ groups │
│ ---                 ┆ ---    │
│ datetime[μs]        ┆ str    │
╞═════════════════════╪════════╡
│ 2021-12-16 00:00:00 ┆ a      │
│ 2021-12-16 00:30:00 ┆ a      │
│ 2021-12-16 01:00:00 ┆ a      │
│ 2021-12-16 01:30:00 ┆ b      │
│ 2021-12-16 02:00:00 ┆ b      │
│ 2021-12-16 02:30:00 ┆ a      │
│ 2021-12-16 03:00:00 ┆ a      │
└─────────────────────┴────────┘
>>> (
...     lf.groupby_dynamic(
...         "time",
...         every="1h",
...         closed="both",
...         by="groups",
...         include_boundaries=True,
...     )
... ).agg([pl.col("time").count().alias("time_count")]).collect()
shape: (7, 5)
┌────────┬─────────────────────┬─────────────────────┬─────────────────────┬────────────┐
│ groups ┆ _lower_boundary     ┆ _upper_boundary     ┆ time                ┆ time_count │
│ ---    ┆ ---                 ┆ ---                 ┆ ---                 ┆ ---        │
│ str    ┆ datetime[μs]        ┆ datetime[μs]        ┆ datetime[μs]        ┆ u32        │
╞════════╪═════════════════════╪═════════════════════╪═════════════════════╪════════════╡
│ a      ┆ 2021-12-15 23:00:00 ┆ 2021-12-16 00:00:00 ┆ 2021-12-15 23:00:00 ┆ 1          │
│ a      ┆ 2021-12-16 00:00:00 ┆ 2021-12-16 01:00:00 ┆ 2021-12-16 00:00:00 ┆ 3          │
│ a      ┆ 2021-12-16 01:00:00 ┆ 2021-12-16 02:00:00 ┆ 2021-12-16 01:00:00 ┆ 1          │
│ a      ┆ 2021-12-16 02:00:00 ┆ 2021-12-16 03:00:00 ┆ 2021-12-16 02:00:00 ┆ 2          │
│ a      ┆ 2021-12-16 03:00:00 ┆ 2021-12-16 04:00:00 ┆ 2021-12-16 03:00:00 ┆ 1          │
│ b      ┆ 2021-12-16 01:00:00 ┆ 2021-12-16 02:00:00 ┆ 2021-12-16 01:00:00 ┆ 2          │
│ b      ┆ 2021-12-16 02:00:00 ┆ 2021-12-16 03:00:00 ┆ 2021-12-16 02:00:00 ┆ 1          │
└────────┴─────────────────────┴─────────────────────┴─────────────────────┴────────────┘

Dynamic groupby on an index column

>>> lf = pl.LazyFrame(
...     {
...         "idx": pl.int_range(0, 6, eager=True),
...         "A": ["A", "A", "B", "B", "B", "C"],
...     }
... )
>>> lf.groupby_dynamic(
...     "idx",
...     every="2i",
...     period="3i",
...     include_boundaries=True,
...     closed="right",
... ).agg(pl.col("A").alias("A_agg_list")).collect()
shape: (3, 4)
┌─────────────────┬─────────────────┬─────┬─────────────────┐
│ _lower_boundary ┆ _upper_boundary ┆ idx ┆ A_agg_list      │
│ ---             ┆ ---             ┆ --- ┆ ---             │
│ i64             ┆ i64             ┆ i64 ┆ list[str]       │
╞═════════════════╪═════════════════╪═════╪═════════════════╡
│ 0               ┆ 3               ┆ 0   ┆ ["A", "B", "B"] │
│ 2               ┆ 5               ┆ 2   ┆ ["B", "B", "C"] │
│ 4               ┆ 7               ┆ 4   ┆ ["C"]           │
└─────────────────┴─────────────────┴─────┴─────────────────┘