Skip to contents

Import data from Parquet file(s)

Source: R/read_scan.R
from_parquet.Rd

read_parquet_polars() imports the data as a Polars DataFrame.

scan_parquet_polars() imports the data as a Polars LazyFrame.

Usage

read_parquet_polars(
  source,
  ...,
  n_rows = NULL,
  row_index_name = NULL,
  row_index_offset = 0L,
  parallel = "auto",
  hive_partitioning = NULL,
  hive_schema = NULL,
  try_parse_hive_dates = TRUE,
  glob = TRUE,
  rechunk = TRUE,
  low_memory = FALSE,
  storage_options = NULL,
  use_statistics = TRUE,
  cache = TRUE,
  include_file_paths = NULL
)

scan_parquet_polars(
  source,
  ...,
  n_rows = NULL,
  row_index_name = NULL,
  row_index_offset = 0L,
  parallel = "auto",
  hive_partitioning = NULL,
  hive_schema = NULL,
  try_parse_hive_dates = TRUE,
  glob = TRUE,
  rechunk = FALSE,
  low_memory = FALSE,
  storage_options = NULL,
  use_statistics = TRUE,
  cache = TRUE,
  include_file_paths = NULL
)

Arguments

source

Path to a file. You can use globbing with * to scan/read multiple files in the same directory (see examples).

...

Ignored.

n_rows

Maximum number of rows to read.

row_index_name

If not NULL, this will insert a row index column with the given name into the DataFrame.

row_index_offset

Offset to start the row index column (only used if the name is set).

parallel

This determines the direction of parallelism. "auto" will try to determine the optimal direction. Can be "auto", "columns", "row_groups", "prefiltered", or "none". See 'Details'.

hive_partitioning

Infer statistics and schema from Hive partitioned URL and use them to prune reads. If NULL (default), it is automatically enabled when a single directory is passed, and otherwise disabled.

hive_schema

A list containing the column names and data types of the columns by which the data is partitioned, e.g. list(a = pl$String, b = pl$Float32). If NULL (default), the schema of the Hive partitions is inferred.

try_parse_hive_dates

Whether to try parsing hive values as date/datetime types.

glob

Expand path given via globbing rules.

rechunk

In case of reading multiple files via a glob pattern, rechunk the final DataFrame into contiguous memory chunks.

low_memory

Reduce memory usage (will yield a lower performance).

storage_options

Experimental. List of options necessary to scan parquet files from different cloud storage providers (GCP, AWS, Azure, HuggingFace). See the 'Details' section.

use_statistics

Use statistics in the parquet file to determine if pages can be skipped from reading.

cache

Cache the result after reading.

include_file_paths

Include the path of the source file(s) as a column with this name.

Details

On parallel strategies

The prefiltered strategy first evaluates the pushed-down predicates in parallel and determines a mask of which rows to read. Then, it parallelizes over both the columns and the row groups while filtering out rows that do not need to be read. This can provide significant speedups for large files (i.e. many row-groups) with a predicate that filters clustered rows or filters heavily. In other cases, prefiltered may slow down the scan compared other strategies.

The prefiltered settings falls back to auto if no predicate is given.

Connecting to cloud providers

Polars supports scanning parquet files from different cloud providers. The cloud providers currently supported are AWS, GCP, and Azure. The supported keys to pass to the storage_options argument can be found here:

Currently it is impossible to scan public parquet files from GCP without a valid service account. Be sure to always include a service account in the storage_options argument.

Scanning from HuggingFace

It is possible to scan data stored on HuggingFace using a path starting with hf://. The hf:// path format is defined as hf://BUCKET/REPOSITORY@REVISION/PATH, where:

  • BUCKET is one of datasets or spaces

  • REPOSITORY is the location of the repository. this is usually in the format of username/repo_name. A branch can also be optionally specified by appending @branch.

  • REVISION is the name of the branch (or commit) to use. This is optional and defaults to main if not given.

  • PATH is a file or directory path, or a glob pattern from the repository root.

A Hugging Face API key can be passed to access private locations using either of the following methods:

  • Passing a token in storage_options to the scan function, e.g. scan_parquet(..., storage_options = list(token = <your HF token>))

  • Setting the HF_TOKEN environment variable, e.g. Sys.setenv(HF_TOKEN = <your HF token>).