Snowpark Library for Python release notes for 2024

This article contains the release notes for the Snowpark Library for Python, including the following when applicable:

  • Behavior changes

  • New features

  • Customer-facing bug fixes

Snowflake uses semantic versioning for Snowpark Library for Python updates.

Version 1.12.1 (2024-02-08)

Version 1.12.1 of the Snowpark library introduces some new features.


  • Use split_blocks=True by default, during to_pandas conversion, for optimal memory allocation. This parameter is passed to pyarrow.Table.to_pandas, which enables PyArrow to split the memory allocation into smaller, more manageable blocks instead of allocating a single contiguous block. This results in better memory management when dealing with larger datasets.

Bug fixes

  • Fixed a bug in DataFrame.to_pandas that caused an error when evaluating on a Dataframe with an IntergerType column with null values.

Version 1.12.0 (2024-01-29)

Version 1.12.0 of the Snowpark library introduces some new features.

Behavior Changes (API Compatible)

  • When parsing data types during a to_pandas operation, we rely on GS precision value to fix precision issues for large integer values. This may affect users where a column that was earlier returned as int8 gets returned as int64. Users can fix this by explicitly specifying precision values for their return column.

  • Aligned behavior for in case of table stored procedures where running would not trigger a stored procedure unless a collect() operation was performed.

  • StoredProcedureRegistration now automatically adds snowflake-snowpark-python as a package dependency on the client’s local version of the library. An error is thrown if the server cannot support that version.

New features

  • Exposed statement_params in StoredProcedure.__call__.

  • Added two optional arguments to Session.add_import:

    • chunk_size: The number of bytes to hash per chunk of the uploaded files.

    • whole_file_hash: By default only the first chunk of the uploaded import is hashed to save time. When this is set to True each uploaded file is fully hashed instead.

  • Added parameters external_access_integrations and secrets when creating a UDAF from Snowpark Python to allow integration with external access.

  • Added a new method Session.append_query_tag, which allows an additional tag to be added to the current query tag by appending it as a comma separated value.

  • Added a new method Session.update_query_tag, which allows updates to a JSON encoded dictionary query tag.

  • SessionBuilder.getOrCreate will now attempt to replace the singleton it returns when token expiration has been detected.

  • Added the following functions in snowflake.snowpark.functions:

    • array_except

    • create_map

    • sign / signum

  • Added the following functions to

    • Added the moving_agg function in to enable moving aggregations like sums and averages with multiple window sizes.

    • Added the cummulative_agg function in to enable moving aggregations like sums and averages with multiple window sizes.

Bug fixes

  • Fixed a bug in that caused Boolean values to erroneously override integer values.

  • Fixed a bug in Session.create_dataframe where the Snowpark DataFrames created using pandas DataFrames were not inferring the type for timestamp columns correctly. The behavior is as follows:

    • Earlier timestamp columns without a timezone would be converted to nanosecond epochs and inferred as LongType(), but will now be correctly maintained as timestamp values and be inferred as TimestampType(TimestampTimeZone.NTZ).

    • Earlier timestamp columns with a timezone would be inferred as TimestampType(TimestampTimeZone.NTZ) and loose timezone information but will now be correctly inferred as TimestampType(TimestampTimeZone.LTZ) and timezone information is retained correctly.

    • Set session parameter PYTHON_SNOWPARK_USE_LOGICAL_TYPE_FOR_CREATE_DATAFRAME to revert back to old behavior. Snowflake recommends that you update your code to align with correct behavior because the parameter will be removed in the future.

  • Fixed a bug that DataFrame.to_pandas gets decimal type when scale is not 0, and creates an object dtype in pandas. Instead, we cast the value to a float64 type.

  • Fixed bugs that wrongly flattened the generated SQL when one of the following happens:

    • DataFrame.filter() is called after DataFrame.sort().limit().

    • DataFrame.sort() or filter() is called on a DataFrame that already has a window function or sequence-dependent data generator column. For instance,"a", seq1().alias("b")).select("a", "b").sort("a") won’t flatten the sort clause anymore.

    • A window or sequence-dependent data generator column is used after DataFrame.limit(). For instance, df.limit(10).select(row_number().over()) won’t flatten the limit and select in the generated SQL.

  • Fixed a bug where aliasing a DataFrame column raised an error when the DataFame was copied from another DataFrame with an aliased column. For instance,

    df ="a").alias("b"))
    df = copy(df)"b").alias("c"))  # Threw an error. Now it's fixed.
  • Fixed a bug in Session.create_dataframe that the non-nullable field in a schema is not respected for Boolean type. Note that this fix is only effective when the user has the privilege to create a temp table.

  • Fixed a bug in SQL simplifier where non-select statements in session.sql dropped a SQL query when used with limit().

  • Fixed a bug that raised an exception when session parameter ERROR_ON_NONDETERMINISTIC_UPDATE is true.