Skip to main content
Engineering blog

Pandas user-defined functions (UDFs) are one of the most significant enhancements in Apache SparkTM for data science. They bring many benefits, such as enabling users to use Pandas APIs and improving performance.

However, Pandas UDFs have evolved organically over time, which has led to some inconsistencies and is creating confusion among users. The full release of Apache Spark 3.0, expected soon, will introduce a new interface for Pandas UDFs that leverages Python type hints to address the proliferation of Pandas UDF types and help them become more Pythonic and self-descriptive.

This blog post introduces new Pandas UDFs with Python type hints, and the new Pandas Function APIs including grouped map, map, and co-grouped map.

Pandas UDFs

Pandas UDFs were introduced in Spark 2.3, see also Introducing Pandas UDF for PySpark. Pandas is well known to data scientists and has seamless integrations with many Python libraries and packages such as NumPy, statsmodel, and scikit-learn, and Pandas UDFs allow data scientists not only to scale out their workloads, but also to leverage the Pandas APIs in Apache Spark.

The user-defined functions are executed by:

  • Apache Arrow, to exchange data directly between JVM and Python driver/executors with near-zero (de)serialization cost.
  • Pandas inside the function, to work with Pandas instances and APIs.

The Pandas UDFs work with Pandas APIs inside the function and Apache Arrow for exchanging data. It allows vectorized operations that can increase performance up to 100x, compared to row-at-a-time Python UDFs.

The example below shows a Pandas UDF to simply add one to each value, in which it is defined with the function called pandas_plus_one decorated by pandas_udf with the Pandas UDF type specified as PandasUDFType.SCALAR.

from pyspark.sql.functions import pandas_udf, PandasUDFType

@pandas_udf('double', PandasUDFType.SCALAR)
def pandas_plus_one(v):
    # `v` is a pandas Series
    return v.add(1)  # outputs a pandas Series

spark.range(10).select(pandas_plus_one("id")).show()

The Python function takes and outputs a Pandas Series. You can perform a vectorized operation for adding one to each value by using the rich set of Pandas APIs within this function. (De)serialization is also automatically vectorized by leveraging Apache Arrow under the hood.

Python Type Hints

Python type hints were officially introduced in PEP 484 with Python 3.5. Type hinting is an official way to statically indicate the type of a value in Python. See the example below.

def greeting(name: str) -> str:
    return 'Hello ' + name

The name: strindicates the name argument is of str type and the -> syntax indicates the greeting() function returns a string.

Python type hints bring two significant benefits to the PySpark and Pandas UDF context.

  • It gives a clear definition of what the function is supposed to do, making it easier for users to understand the code. For example, unless it is documented, users cannot know if greeting can take None or not if there is no type hint. It can avoid the need to document such subtle cases with a bunch of test cases and/or for users to test and figure out by themselves.
  • It can make it easier to perform static analysis. IDEs such as PyCharm and Visual Studio Code can leverage type annotations to provide code completion, show errors, and support better go-to-definition functionality.

Proliferation of Pandas UDF Types

Since the release of Apache Spark 2.3, a number of new Pandas UDFs have been implemented, making it difficult for users to learn about the new specifications and how to use them. For example, here are three Pandas UDFs that output virtually the same results:

from pyspark.sql.functions import pandas_udf, PandasUDFType                                                                                                                                                                                                                                                                                                                                                                                                                              

@pandas_udf('long', PandasUDFType.SCALAR)
def pandas_plus_one(v):
    # `v` is a pandas Series
    return v + 1  # outputs a pandas Series

spark.range(10).select(pandas_plus_one("id")).show()
from pyspark.sql.functions import pandas_udf, PandasUDFType


# New type of Pandas UDF in Spark 3.0.
@pandas_udf('long', PandasUDFType.SCALAR_ITER)
def pandas_plus_one(itr):
    # `iterator` is an iterator of pandas Series.
    return map(lambda v: v + 1, itr)  # outputs an iterator of pandas Series.

spark.range(10).select(pandas_plus_one("id")).show()
from pyspark.sql.functions import pandas_udf, PandasUDFType

@pandas_udf("id long", PandasUDFType.GROUPED_MAP)
def pandas_plus_one(pdf):
# `pdf` is a pandas DataFrame
return pdf + 1  # outputs a pandas DataFrame

# `pandas_plus_one` can _only_ be used with `groupby(...).apply(...)`
spark.range(10).groupby('id').apply(pandas_plus_one).show()

Although each of these UDF types has a distinct purpose, several can be applicable. In this simple case, you could use any of the three. However, each of the Pandas UDFs expects different input and output types, and works in a different way with a distinct semantic and different performance. It confuses users about which one to use and learn, and how each works.

Furthermore, pandas_plus_one in the first and second cases can be used where the regular PySpark columns are used. Consider the argument of withColumn or the function with the combinations of other expressions such as pandas_plus_one("id") + 1. However, the last pandas_plus_one can only be used with groupby(...).apply(pandas_plus_one).

This level of complexity has triggered numerous discussions with Spark developers, and drove the effort to introduce the new Pandas APIs with Python type hints via an official proposal. The goal is to enable users to naturally express their pandas UDFs using Python type hints without confusion as in the problematic cases above. For example, the cases above can be written as below:

def pandas_plus_one(v: pd.Series) -> pd.Series:
    return v + 1
def pandas_plus_one(itr: Iterator[pd.Series]) -> Iterator[pd.Series]:
    return map(lambda v: v + 1, itr)
def pandas_plus_one(pdf: pd.DataFrame) -> pd.DataFrame:
    return pdf + 1

 

New Pandas APIs with Python Type Hints

To address the complexity in the old Pandas UDFs, from Apache Spark 3.0 with Python 3.6 and above, Python type hints such as pandas.Series, pandas.DataFrame, Tuple, and Iterator can be used to express the new Pandas UDF types.

In addition, the old Pandas UDFs were split into two API categories: Pandas UDFs and Pandas Function APIs. Although they work internally in a similar way, there are distinct differences.

You can treat Pandas UDFs in the same way that you use other PySpark column instances. However, you cannot use the Pandas Function APIs with these column instances. Here are these two examples:

# Pandas UDF
import pandas as pd
from pyspark.sql.functions import pandas_udf, log2, col

@pandas_udf('long')
def pandas_plus_one(s: pd.Series) -> pd.Series:
    return s + 1

# pandas_plus_one("id") is identically treated as _a SQL expression_ internally.
# Namely, you can combine with other columns, functions and expressions.
spark.range(10).select(
    pandas_plus_one(col("id") - 1) + log2("id") + 1).show()
# Pandas Function API
from typing import Iterator
import pandas as pd


def pandas_plus_one(iterator: Iterator[pd.DataFrame]) -> Iterator[pd.DataFrame]:
    return map(lambda v: v + 1, iterator)


# pandas_plus_one is just a regular Python function, and mapInPandas is
# logically treated as _a separate SQL query plan_ instead of a SQL expression. 
# Therefore, direct interactions with other expressions are impossible.
spark.range(10).mapInPandas(pandas_plus_one, schema="id long").show()

Also, note that Pandas UDFs require Python type hints whereas the type hints in Pandas Function APIs are currently optional. Type hints are planned for Pandas Function APIs and may be required at some point in the future.

New Pandas UDFs

Instead of defining and specifying each Pandas UDF type manually, the new Pandas UDFs infer the Pandas UDF type from the given Python type hints at the Python function. There are currently four supported cases of the Python type hints in Pandas UDFs:

  • Series to Series
  • Iterator of Series to Iterator of Series
  • Iterator of Multiple Series to Iterator of Series
  • Series to Scalar (a single value)

Before we do a deep dive into each case, let’s look at three key points about working with the new Pandas UDFs.

  • Although Python type hints are optional in the Python world in general, you must specify Python type hints for the input and output in order to use the new Pandas UDFs.
  • Users can still use the old way by manually specifying the Pandas UDF type. However, using Python type hints is encouraged.
  • The type hint should use pandas.Series in all cases. However, there is one variant in which pandas.DataFrame should be used for its input or output type hint instead: when the input or output column is of StructType.Take a look at the example below:
    import pandas as pd
    from pyspark.sql.functions import pandas_udf
    
    
    df = spark.createDataFrame(
        [[1, "a string", ("a nested string",)]],
        "long_col long, string_col string, struct_col struct<col1 string>")
    
    @pandas_udf("col1 string, col2 long")
    def pandas_plus_len(
            s1: pd.Series, s2: pd.Series, pdf: pd.DataFrame) -> pd.DataFrame:
        # Regular columns are series and the struct column is a DataFrame.
        pdf['col2'] = s1 + s2.str.len() 
        return pdf  # the struct column expects a DataFrame to return
    
    df.select(pandas_plus_len("long_col", "string_col", "struct_col")).show()

Series to Series

Series to Series is mapped to scalar Pandas UDF introduced in Apache Spark 2.3. The type hints can be expressed as pandas.Series, ... -> pandas.Series. It expects the given function to take one or more pandas.Series and outputs one pandas.Series. The output length is expected to be the same as the input.

import pandas as pd
from pyspark.sql.functions import pandas_udf       

@pandas_udf('long')
def pandas_plus_one(s: pd.Series) -> pd.Series:
    return s + 1

spark.range(10).select(pandas_plus_one("id")).show()

The example above can be mapped to the old style with scalar Pandas UDF, as below.

from pyspark.sql.functions import pandas_udf, PandasUDFType                                                                                                                                                                                                                                                                                                                                                                                                                   

@pandas_udf('long', PandasUDFType.SCALAR)
def pandas_plus_one(v):
    return v + 1

spark.range(10).select(pandas_plus_one("id")).show()

Iterator of Series to Iterator of Series

This is a new type of Pandas UDF coming in Apache Spark 3.0. It is a variant of Series to Series, and the type hints can be expressed as Iterator[pd.Series] -> Iterator[pd.Series]. The function takes and outputs an iterator of pandas.Series.

The length of the whole output must be the same length of the whole input. Therefore, it can prefetch the data from the input iterator as long as the lengths of entire input and output are the same. The given function should take a single column as input.

from typing import Iterator
import pandas as pd
from pyspark.sql.functions import pandas_udf       

@pandas_udf('long')
def pandas_plus_one(iterator: Iterator[pd.Series]) -> Iterator[pd.Series]:
    return map(lambda s: s + 1, iterator)

spark.range(10).select(pandas_plus_one("id")).show()

It is also useful when the UDF execution requires expensive initialization of some state. The pseudocode below illustrates the case.

@pandas_udf("long")
def calculate(iterator: Iterator[pd.Series]) -> Iterator[pd.Series]:
    # Do some expensive initialization with a state
    state = very_expensive_initialization()
    for x in iterator:
        # Use that state for the whole iterator.
        yield calculate_with_state(x, state)

df.select(calculate("value")).show()

Iterator of Series to Iterator of Series can be also mapped to the old Pandas UDF style. See the example below.

from pyspark.sql.functions import pandas_udf, PandasUDFType                                                                                                                                                                                                                   

@pandas_udf('long', PandasUDFType.SCALAR_ITER)
def pandas_plus_one(iterator):
    return map(lambda s: s + 1, iterator)

spark.range(10).select(pandas_plus_one("id")).show()

Iterator of Multiple Series to Iterator of Series

This type of Pandas UDF will be also introduced in Apache Spark 3.0, together with Iterator of Series to Iterator of Series. The type hints can be expressed as Iterator[Tuple[pandas.Series, ...]] -> Iterator[pandas.Series].

It has the similar characteristics and restrictions with Iterator of Series to Iterator of Series. The given function takes an iterator of a tuple of pandas.Series and outputs an iterator of pandas.Series. It is also useful when to use some states and when to prefetch the input data. The length of the entire output should also be the same as the length of the entire input. However, the given function should take multiple columns as input, unlike Iterator of Series to Iterator of Series.

from typing import Iterator, Tuple
import pandas as pd
from pyspark.sql.functions import pandas_udf       

@pandas_udf("long")
def multiply_two(
        iterator: Iterator[Tuple[pd.Series, pd.Series]]) -> Iterator[pd.Series]:
    return (a * b for a, b in iterator)

spark.range(10).select(multiply_two("id", "id")).show()

This can also be mapped to the old Pandas UDF style as below.

from pyspark.sql.functions import pandas_udf, PandasUDFType                                                                                                                                                                                                                                                                                                                                                                                                                              

@pandas_udf('long', PandasUDFType.SCALAR_ITER)
def multiply_two(iterator):
    return (a * b for a, b in iterator)

spark.range(10).select(multiply_two("id", "id")).show()

Series to Scalar

Series to Scalar is mapped to the grouped aggregate Pandas UDF introduced in Apache Spark 2.4. The type hints are expressed as pandas.Series, ... -> Any. The function takes one or more pandas.Series and outputs a primitive data type. The returned scalar can be either a Python primitive type, e.g., int, float, or a NumPy data type such as numpy.int64, numpy.float64, etc. Any should ideally be a specific scalar type accordingly.

import pandas as pd
from pyspark.sql.functions import pandas_udf
from pyspark.sql import Window

df = spark.createDataFrame(
    [(1, 1.0), (1, 2.0), (2, 3.0), (2, 5.0), (2, 10.0)], ("id", "v"))

@pandas_udf("double")
def pandas_mean(v: pd.Series) -> float:
    return v.sum()

df.select(pandas_mean(df['v'])).show()
df.groupby("id").agg(pandas_mean(df['v'])).show()
df.select(pandas_mean(df['v']).over(Window.partitionBy('id'))).show()

The example above can be converted to the example with the grouped aggregate Pandas UDF as you can see here:

import pandas as pd
from pyspark.sql.functions import pandas_udf, PandasUDFType
from pyspark.sql import Window

df = spark.createDataFrame(
    [(1, 1.0), (1, 2.0), (2, 3.0), (2, 5.0), (2, 10.0)], ("id", "v"))

@pandas_udf("double", PandasUDFType.GROUPED_AGG)
def pandas_mean(v):
    return v.sum()

df.select(pandas_mean(df['v'])).show()
df.groupby("id").agg(pandas_mean(df['v'])).show()
df.select(pandas_mean(df['v']).over(Window.partitionBy('id'))).show()

New Pandas Function APIs

This new category in Apache Spark 3.0 enables you to directly apply a Python native function, which takes and outputs Pandas instances against a PySpark DataFrame. Pandas Functions APIs supported in Apache Spark 3.0 are: grouped map, map, and co-grouped map.

Note that the grouped map Pandas UDF is now categorized as a group map Pandas Function API. As mentioned earlier, the Python type hints in Pandas Function APIs are optional currently.

Grouped Map

Grouped map in the Pandas Function API is applyInPandas at a grouped DataFrame, e.g., df.groupby(...). This is mapped to the grouped map Pandas UDF in the old Pandas UDF types. It maps each group to each pandas.DataFrame in the function. Note that it does not require for the output to be the same length of the input.

import pandas as pd

df = spark.createDataFrame(
    [(1, 1.0), (1, 2.0), (2, 3.0), (2, 5.0), (2, 10.0)], ("id", "v"))

def subtract_mean(pdf: pd.DataFrame) -> pd.DataFrame:
    v = pdf.v
    return pdf.assign(v=v - v.mean())

df.groupby("id").applyInPandas(subtract_mean, schema=df.schema).show()

Grouped map type is mapped to grouped map Pandas UDF supported from Spark 2.3, as below:

import pandas as pd
from pyspark.sql.functions import pandas_udf, PandasUDFType

df = spark.createDataFrame(
    [(1, 1.0), (1, 2.0), (2, 3.0), (2, 5.0), (2, 10.0)], ("id", "v"))

@pandas_udf(df.schema, PandasUDFType.GROUPED_MAP)
def subtract_mean(pdf):
    v = pdf.v
    return pdf.assign(v=v - v.mean())

df.groupby("id").apply(subtract_mean).show()

 

Map

Map Pandas Function API is mapInPandas in a DataFrame. It is new in Apache Spark 3.0. It maps every batch in each partition and transforms each. The function takes an iterator of pandas.DataFrame and outputs an iterator of pandas.DataFrame. The output length does not need to match the input size.

from typing import Iterator
import pandas as pd

df = spark.createDataFrame([(1, 21), (2, 30)], ("id", "age"))

def pandas_filter(iterator: Iterator[pd.DataFrame]) -> Iterator[pd.DataFrame]:
    for pdf in iterator:
        yield pdf[pdf.id == 1]

df.mapInPandas(pandas_filter, schema=df.schema).show()

Co-grouped Map

Co-grouped map, applyInPandas in a co-grouped DataFrame such as df.groupby(...).cogroup(df.groupby(...)), will also be introduced in Apache Spark 3.0. Similar to the grouped map, it maps each group to each pandas.DataFrame in the function but it groups with another DataFrame by common key(s) and then the function is applied to each cogroup. Likewise, there is no restriction on the output length.

import pandas as pd

df1 = spark.createDataFrame(
    [(1201, 1, 1.0), (1201, 2, 2.0), (1202, 1, 3.0), (1202, 2, 4.0)],
    ("time", "id", "v1"))
df2 = spark.createDataFrame(
    [(1201, 1, "x"), (1201, 2, "y")], ("time", "id", "v2"))

def asof_join(left: pd.DataFrame, right: pd.DataFrame) -> pd.DataFrame:
    return pd.merge_asof(left, right, on="time", by="id")

df1.groupby("id").cogroup(
    df2.groupby("id")
).applyInPandas(asof_join, "time int, id int, v1 double, v2 string").show()

 

Conclusion and Future Work

The upcoming release of Apache Spark 3.0 (read our preview blog for details). will offer Python type hints to make it simpler for users to express Pandas UDFs and Pandas Function APIs. In the future, we should consider adding support for other type hint combinations in both Pandas UDFs and Pandas Function APIs. Currently, the supported cases are only few of many possible combinations of Python type hints. There are also other ongoing discussions in the Apache Spark community. Visit Side Discussions and Future Improvement to learn more.

Learn more about Spark 3.0 in our preview webinar.  Try out these new capabilities today for free on Databricks as part of the Databricks Runtime 7.0 Beta.

Try Databricks for free

Related posts

Engineering blog

New Pandas UDFs and Python Type Hints in the Upcoming Release of Apache Spark 3.0

May 20, 2020 by Hyukjin Kwon in Engineering Blog
Pandas user-defined functions (UDFs) are one of the most significant enhancements in Apache Spark TM for data science. They bring many benefits, such...
See all Engineering Blog posts