Back to Dask

Indexing into Dask DataFrames

docs/source/dataframe-indexing.rst

2026.3.06.0 KB
Original Source

.. _dataframe.indexing:

Indexing into Dask DataFrames

Dask DataFrame supports some of Pandas' indexing behavior.

.. currentmodule:: dask.dataframe

.. autosummary::

DataFrame.iloc DataFrame.loc

Label-based Indexing

Just like Pandas, Dask DataFrame supports label-based indexing with the .loc accessor for selecting rows or columns, and __getitem__ (square brackets) for selecting just columns.

.. note::

To select rows, the DataFrame's divisions must be known (see :ref:dataframe.design and :ref:dataframe.performance for more information.)

.. code-block:: python

import dask.dataframe as dd import pandas as pd df = pd.DataFrame({"A": [1, 2, 3], "B": [3, 4, 5]}, ... index=['a', 'b', 'c']) ddf = dd.from_pandas(df, npartitions=2) ddf Dask DataFrame Structure: A B npartitions=1 a int64 int64 c ... ... Dask Name: from_pandas, 1 tasks

Selecting columns:

.. code-block:: python

ddf[['B', 'A']] Dask DataFrame Structure: B A npartitions=1 a int64 int64 c ... ... Dask Name: getitem, 2 tasks

Selecting a single column reduces to a Dask Series:

.. code-block:: python

ddf['A'] Dask Series Structure: npartitions=1 a int64 c ... Name: A, dtype: int64 Dask Name: getitem, 2 tasks

Slicing rows and (optionally) columns with .loc:

.. code-block:: python

ddf.loc[['b', 'c'], ['A']] Dask DataFrame Structure: A npartitions=1 b int64 c ... Dask Name: loc, 2 tasks

ddf.loc[df["A"] > 1, ["B"]] Dask DataFrame Structure: B npartitions=1 a int64 c ... Dask Name: try_loc, 2 tasks

ddf.loc[lambda df: df["A"] > 1, ["B"]] Dask DataFrame Structure: B npartitions=1 a int64 c ... Dask Name: try_loc, 2 tasks

Dask DataFrame supports Pandas' partial-string indexing <https://pandas.pydata.org/pandas-docs/stable/user_guide/timeseries.html#partial-string-indexing>_:

.. code-block:: python

ts = dd.demo.make_timeseries() ts Dask DataFrame Structure: id name x y npartitions=11 2000-01-31 int64 object float64 float64 2000-02-29 ... ... ... ... ... ... ... ... ... 2000-11-30 ... ... ... ... 2000-12-31 ... ... ... ... Dask Name: make-timeseries, 11 tasks

ts.loc['2000-02-12'] Dask DataFrame Structure: id name x y npartitions=1 2000-02-12 00:00:00.000000000 int64 object float64 float64 2000-02-12 23:59:59.999999999 ... ... ... ... Dask Name: loc, 12 tasks

Positional Indexing

Dask DataFrame does not track the length of partitions, making positional indexing with .iloc inefficient for selecting rows. :meth:DataFrame.iloc only supports indexers where the row indexer is slice(None) (which : is a shorthand for.)

.. code-block:: python

ddf.iloc[:, [1, 0]] Dask DataFrame Structure: B A npartitions=1 a int64 int64 c ... ... Dask Name: iloc, 2 tasks

Trying to select specific rows with iloc will raise an exception:

.. code-block:: python

ddf.iloc[[0, 2], [1]] Traceback (most recent call last) File "<stdin>", line 1, in <module> ValueError: 'DataFrame.iloc' does not support slicing rows. The indexer must be a 2-tuple whose first item is 'slice(None)'.

Partition Indexing

In addition to pandas-style indexing, Dask DataFrame also supports indexing at a partition level with :meth:DataFrame.get_partition and :attr:DataFrame.partitions. These can be used to select subsets of the data by partition, rather than by position in the entire DataFrame or index label.

Use :meth:DataFrame.get_partition to select a single partition by position.

.. code-block:: python

import dask ddf = dask.datasets.timeseries(start="2021-01-01", end="2021-01-07", freq="1h") ddf.get_partition(0) Dask DataFrame Structure: name id x y npartitions=1 2021-01-01 object int64 float64 float64 2021-01-02 ... ... ... ... Dask Name: get-partition, 2 graph layers

Note that the result is also a Dask DatFrame.

Index into :attr:DataFrame.partitions to select one or more partitions. For example, you can select every other partition with a slice:

.. code-block:: python

ddf.partitions[::2] Dask DataFrame Structure: name id x y npartitions=3 2021-01-01 object int64 float64 float64 2021-01-03 ... ... ... ... 2021-01-05 ... ... ... ... 2021-01-06 ... ... ... ... Dask Name: blocks, 2 graph layers

Or even more complicated selections based on the data in the partitions themselves (at the cost of computing the DataFrame up until that point). For example, we can create a boolean mask with the partitions that have more than some number of unique IDs using :meth:DataFrame.map_partitions:

.. code-block:: python

mask = ddf.id.map_partitions(lambda p: len(p.unique()) > 20).compute() ddf.partitions[mask] Dask DataFrame Structure: name id x y npartitions=5 2021-01-01 object int64 float64 float64 2021-01-02 ... ... ... ... ... ... ... ... ... 2021-01-06 ... ... ... ... 2021-01-07 ... ... ... ... Dask Name: blocks, 2 graph layers