sparkpolars 0.1.1rc17

Conversion between PySpark and Polars DataFrames

sparkpolars

sparkpolars is a lightweight library designed for seamless conversions between Apache Spark and Polars without unnecessary dependencies. (Dependencies are only required when explicitly requested.)

Installation

pip install sparkpolars
# or
conda install skandev::sparkpolars

Requirements

• Python ≥ 3.10
• Apache Spark ≥ 3.3.0 (must be pre-installed)
• Polars ≥ 1.0 (must be pre-installed)
• Pyspark must also be installed if you plan to use this library

Why Does This Library Exist?

The Problem

Typical conversions between Spark and Polars often involve an intermediate Pandas step:

# Traditional approach:
# Spark -> Pandas -> Polars
# or
# Polars -> Pandas -> Spark

The Solution

sparkpolars eliminates unnecessary dependencies like pandas and pyarrow by leveraging native functions such as .collect() and schema interpretation.

Key Benefits

• 🚀 No extra dependencies – No need for Pandas or PyArrow
• ✅ Reliable handling of complex types – Provides better consistency for MapType, StructType, and nested ArrayType, where existing conversion methods can be unreliable

Features

• Convert a Spark DataFrame to a Polars DataFrame or LazyFrame
• Ensures schema consistency: preserves LongType as Int64 instead of mistakenly converting to Int32
• Three conversion modes: NATIVE, ARROW, PANDAS
• NATIVE mode properly converts MapType, StructType, and nested ArrayType
• ARROW and PANDAS modes may have limitations with complex types
• Configurable conversion settings for Polars list(struct) to Spark MapType
• Timezone and time unit customization for Polars Datetime

Usage

0. Supercharge Polars and Spark DataFrame

In your __init__.py file at the root project you can do the following for ease of use

from sparkpolars import toPolars, to_spark
from pyspark.sql import DataFrame as SparkDataFrame
from polars import DataFrame as PolarsDataFrame, LazyFrame as PolarsLazyFrame

__all__ = [
    "toPolars",
    "to_spark",
]

SparkDataFrame.toPolars = toPolars
PolarsDataFrame.to_spark = to_spark
PolarsLazyFrame.to_spark = to_spark

1. From Spark to Polars DataFrame

from pyspark.sql import SparkSession

spark = SparkSession.builder.appName("example").getOrCreate()

df = spark.createDataFrame([(1, 2)], ["a", "b"])

polars_df = df.toPolars()

2. From Spark to Polars LazyFrame

from pyspark.sql import SparkSession

spark = SparkSession.builder.appName("example").getOrCreate()

df = spark.createDataFrame([(1, 2)], ["a", "b"])

polars_df = df.toPolars(lazy=True)

3. From Polars DataFrame to Spark

from pyspark.sql import SparkSession
from polars import DataFrame

spark = SparkSession.builder.appName("example").getOrCreate()

df = DataFrame({"a": [1], "b": [2]})  # It can also be a LazyDataFrame

spark_df = df.to_spark(spark=spark)
# or
spark_df = df.to_spark()  # It will try to get the Spark ActiveSession

4. Using Specific Mode

from sparkpolars import ModeMethod

spark_df = df.to_spark(mode=ModeMethod.NATIVE)
spark_df = df.to_spark(mode=ModeMethod.PANDAS)
spark_df = df.to_spark(mode=ModeMethod.ARROW)

polars_df = df.toPolars(mode=ModeMethod.NATIVE)
polars_df = df.toPolars(mode=ModeMethod.PANDAS)
polars_df = df.toPolars(mode=ModeMethod.ARROW)

5. Using Config

from sparkpolars import Config

conf = Config(
    map_elements=["column_should_be_converted_to_map_type", ...],  # Specify columns to convert to MapType
    time_unit="ms",  # Literal["ns", "us", "ms"], defaults to "us"
)
spark_df = df.to_spark(config=conf)

polars_df = df.toPolars(config=conf)

Known Limitations

JVM Timezone Discrepancy

Spark timestamps are collected via the JVM, which may differ from Spark’s timezone settings. If issues arise, verify the JVM timezone.

Memory Constraints

Collecting large datasets into memory can exceed available driver memory, leading to failures. (as for pandas/arrow)

Handling MapType:

From Spark to Polars

If you have in Spark:

Type: StructField("example", MapType(StringType(), IntegerType()))

Data: {"a": 1, "b": 2}

Then it will become in Polars:

Type: {"example": List(Struct("key": String, "value": Int32))}

Data: [{"key": "a", "value": 1}, {"key": "b", "value": 2}]

From Polars to Spark

If you have in Polars:

Type: {"example": List(Struct("key": String, "value": Int32))}

Data: [{"key": "a", "value": 1}, {"key": "b", "value": 2}]

Then it will become in Spark without specifying any config (Default Behavior):

Type: StructField("example", ArrayType(StructType(StructField("key", StringType())), StructField("value", IntegerType())))

Data: [{"key": "a", "value": 1}, {"key": "b", "value": 2}]

If you want this data to be converted to MapType:

from sparkpolars import Config
conf = Config(
    map_elements=["example"]
)

Type: StructField("example", MapType(StringType(), IntegerType()))

Data: {"a": 1, "b": 2}

License

• MIT License

Contribution

• Create an associated issue, or assign yourself to an existing issue
• Fork the project
• Install all the dependencies pip install ".[dev,lint,test]
• Install pre-commit file pre-commit install
• Develop your feature
• Unit-test your feature
• Create a Pull request