Pandas · Function

pd.read_parquet

Reads a Parquet object into a pandas DataFrame. It supports column selection, engine and storage configuration, dtype backends, and optional row filtering.

How it works

pandas.read_parquet loads a Parquet file, directory, URL, or binary file-like object through a supported Parquet engine and returns the data as a DataFrame. You can select columns, pass filesystem or storage settings, filter partitioned data, and control dtype conversion.

Syntax

pandas.read_parquet(path, engine='auto', columns=None, storage_options=None, dtype_backend=<no_default>, filesystem=None, filters=None, to_pandas_kwargs=None, **kwargs)

Parameters

Name	Type	Description
`path`	`str, path object or file-like object`	Parquet path, URL, path-like object, or binary file-like object.
`engine`	`{‘auto’, ‘pyarrow’, ‘fastparquet’}, default ‘auto’`	Parquet implementation to use; auto, pyarrow, or fastparquet.
`columns`	`list, default None`	Names of columns to load; None reads all columns.
`storage_options`	`dict, optional`	Options for the relevant storage connection.
`dtype_backend`	`{‘numpy_nullable’, ‘pyarrow’}, default ‘numpy_nullable’`	Nullable dtype backend selection for the resulting DataFrame.
`filesystem`	`fsspec or pyarrow filesystem, optional`	Filesystem object used to access the path.
`filters`	`List[Tuple] or List[List[Tuple]], optional`	Row-group or partition predicates passed to the engine.
`to_pandas_kwargs`	`dict, optional`	Additional keyword arguments forwarded to the underlying reader.
`**kwargs`	`dict`	Additional engine-specific keyword arguments.
`path`	`str, path object or file-like object`	Object to load.

Returns

DataFrame

Key points

The function loads Parquet data and produces a pandas DataFrame.
path may be a local path or a URL, in addition to a path-like or binary file-like object.
Use columns to restrict the read to named fields and avoid loading unneeded columns.
engine chooses the Parquet implementation: auto, pyarrow, or fastparquet.
storage_options carries connection-specific settings such as credentials or endpoint-related options.

Common mistakes

With the default engine selection, pandas consults the io.parquet.engine option; configure it when automatic engine choice is not suitable.

Example: Read selected Parquet columns

import pandas as pd

# Requires the optional pyarrow dependency.
df = pd.read_parquet("data.parquet", columns=["user_id", "score"])
print(df)

Reads the selected columns from a Parquet file into a pandas DataFrame.

When to use it

Local analytics

Load a local .parquet file into a DataFrame for analysis.

Column projection

Read only selected fields from a large Parquet dataset.

Remote or partitioned storage

Access Parquet data stored behind a supported URL or filesystem.

Interview questions

QWhat does read_parquet return and what can path represent?

It accepts a path, URL, or binary file-like object and returns a DataFrame. The engine can be selected explicitly or determined automatically.

Follow-up: How would you reduce I/O? Pass columns to read only the required fields, and use filters when supported by the Parquet engine.

QHow do you read only a subset of columns?

The columns argument is a list of column names. When provided, pandas reads only those columns instead of the complete Parquet dataset.

Follow-up: What additional optimization is available for partitioned data? Use filters to restrict rows during the Parquet read when the engine supports them.

QHow is the Parquet engine selected?

The function supports auto, pyarrow, and fastparquet engine choices. With auto, pandas uses the io.parquet.engine option to determine the engine.

Follow-up: What is the practical consequence of passing filesystem? The documented behavior uses pyarrow for that access path, so engine compatibility must be checked.

DataFrame.to_parquet
pandas.read_csv
pandas.read_feather
pandas.read_orc
pandas.read_pickle

Source: pandas.read_parquet — pandas 3.0.6 documentation. Checked 2026-10-15.