You are viewing documentation about an older version (1.4.0). View latest version

snowflake.snowpark.functions.udf

snowflake.snowpark.functions.udf(func: Callable | None = None, *, return_type: DataType | None = None, input_types: List[DataType] | None = None, name: str | Iterable[str] | None = None, is_permanent: bool = False, stage_location: str | None = None, imports: List[str | Tuple[str, str]] | None = None, packages: List[str | module] | None = None, replace: bool = False, if_not_exists: bool = False, session: Session | None = None, parallel: int = 4, max_batch_size: int | None = None, statement_params: Dict[str, str] | None = None, source_code_display: bool = True, strict: bool = False, secure: bool = False) → UserDefinedFunction | partial[source]

Registers a Python function as a Snowflake Python UDF and returns the UDF.

It can be used as either a function call or a decorator. In most cases you work with a single session. This function uses that session to register the UDF. If you have multiple sessions, you need to explicitly specify the session parameter of this function. If you have a function and would like to register it to multiple databases, use session.udf.register instead. See examples in UDFRegistration.

Parameters:

  • func – A Python function used for creating the UDF.

  • return_type – A DataType representing the return data type of the UDF. Optional if type hints are provided.

  • input_types – A list of DataType representing the input data types of the UDF. Optional if type hints are provided.

  • name – A string or list of strings that specify the name or fully-qualified object identifier (database name, schema name, and function name) for the UDF in Snowflake, which allows you to call this UDF in a SQL command or via call_udf(). If it is not provided, a name will be automatically generated for the UDF. A name must be specified when is_permanent is True.

  • is_permanent – Whether to create a permanent UDF. The default is False. If it is True, a valid stage_location must be provided.

  • stage_location – The stage location where the Python file for the UDF and its dependencies should be uploaded. The stage location must be specified when is_permanent is True, and it will be ignored when is_permanent is False. It can be any stage other than temporary stages and external stages.

  • imports – A list of imports that only apply to this UDF. You can use a string to represent a file path (similar to the path argument in add_import()) in this list, or a tuple of two strings to represent a file path and an import path (similar to the import_path argument in add_import()). These UDF-level imports will override the session-level imports added by add_import(). Note that an empty list means no import for this UDF, and None or not specifying this parameter means using session-level imports.

  • packages – A list of packages that only apply to this UDF. These UDF-level packages will override the session-level packages added by add_packages() and add_requirements(). Note that an empty list means no package for this UDF, and None or not specifying this parameter means using session-level packages.

  • replace – Whether to replace a UDF that already was registered. The default is False. If it is False, attempting to register a UDF with a name that already exists results in a SnowparkSQLException exception being thrown. If it is True, an existing UDF with the same name is overwritten.

  • if_not_exists – Whether to skip creation of a UDF when one with the same signature already exists. The default is False. if_not_exists and replace are mutually exclusive and a ValueError is raised when both are set. If it is True and a UDF with the same signature exists, the UDF creation is skipped.

  • session – Use this session to register the UDF. If it’s not specified, the session that you created before calling this function will be used. You need to specify this parameter if you have created multiple sessions before calling this method.

  • parallel – The number of threads to use for uploading UDF files with the PUT command. The default value is 4 and supported values are from 1 to 99. Increasing the number of threads can improve performance when uploading large UDF files.

  • max_batch_size – The maximum number of rows per input Pandas DataFrame or Pandas Series inside a vectorized UDF. Because a vectorized UDF will be executed within a time limit, which is 60 seconds, this optional argument can be used to reduce the running time of every batch by setting a smaller batch size. Note that setting a larger value does not guarantee that Snowflake will encode batches with the specified number of rows. It will be ignored when registering a non-vectorized UDF.

  • statement_params – Dictionary of statement level parameters to be set while executing this action.

  • source_code_display – Display the source code of the UDF func as comments in the generated script. The source code is dynamically generated therefore it may not be identical to how the func is originally defined. The default is True. If it is False, source code will not be generated or displayed.

  • strict – Whether the created UDF is strict. A strict UDF will not invoke the UDF if any input is null. Instead, a null value will always be returned for that row. Note that the UDF might still return null for non-null inputs.

  • secure – Whether the created UDF is secure. For more information about secure functions, see Secure UDFs.

Returns:

A UDF function that can be called with Column expressions.

Note

1. When type hints are provided and are complete for a function, return_type and input_types are optional and will be ignored. See details of supported data types for UDFs in UDFRegistration.

  • You can use use Variant to annotate a variant, and use Geography to annotate a geography when defining a UDF.

  • You can use use PandasSeries to annotate a Pandas Series, and use PandasDataFrame to annotate a Pandas DataFrame when defining a vectorized UDF. Note that they are generic types so you can specify the element type in a Pandas Series and DataFrame.

  • typing.Union is not a valid type annotation for UDFs, but typing.Optional can be used to indicate the optional type.

  • Type hints are not supported on functions decorated with decorators.

2. A temporary UDF (when is_permanent is False) is scoped to this session and all UDF related files will be uploaded to a temporary session stage (session.get_session_stage()). For a permanent UDF, these files will be uploaded to the stage that you provide.

3. By default, UDF registration fails if a function with the same name is already registered. Invoking udf() with replace set to True will overwrite the previously registered function.

4. When registering a vectorized UDF, pandas library will be added as a package automatically, with the latest version on the Snowflake server. If you don’t want to use this version, you can overwrite it by adding pandas with specific version requirement using package argument or add_packages().

See also

UDFRegistration

UDFs can be created as anonymous UDFs

Example:

>>> from snowflake.snowpark.types import IntegerType
>>> add_one = udf(lambda x: x+1, return_type=IntegerType(), input_types=[IntegerType()])
>>> df = session.create_dataframe([1, 2, 3], schema=["a"])
>>> df.select(add_one(col("a")).as_("ans")).collect()
[Row(ANS=2), Row(ANS=3), Row(ANS=4)]
Copy

or as named UDFs that are accessible in the same session. Instead of calling udf as function, it can be also used as a decorator:

Example:

>>> @udf(name="minus_one", replace=True)
... def minus_one(x: int) -> int:
...     return x - 1
>>> df.select(minus_one(col("a")).as_("ans")).collect()
[Row(ANS=0), Row(ANS=1), Row(ANS=2)]
>>> session.sql("SELECT minus_one(10)").collect()
[Row(MINUS_ONE(10)=9)]
Copy
Langage: Français