Snowpark ML FileSystem and FileSet

The Snowpark ML library includes FileSystem, an abstraction that is similar to a file system for an internal, server-side encrypted Snowflake stage. Specifically, it is an fsspec AbstractFileSystem implementation. The library also includes FileSet, a related class that allows you to move machine learning data from a Snowflake table to the stage, and from there to feed the data to PyTorch or TensorFlow (see Snowpark ML Framework Connectors).


See the Snowpark ML API Reference for complete details of these APIs.


This topic assumes that the Snowpark ML module is already installed. If it isn’t, see Installing Snowpark ML.

Creating and Using a File System

Creating a Snowpark ML file system requires either a Snowflake Connector for Python Connection object or a Snowpark Python Session. See Authenticating to Snowflake for instructions.

After you have either a connection or a session, you can create a Snowpark ML SFFileSystem instance through which you can access data in your internal stage.

If you have a Snowflake Connector for Python connection, pass it as the sf_connection argument:

import fsspec
from import sfcfs

sf_fs1 = sfcfs.SFFileSystem(sf_connection=sf_connection)

If you have a Snowpark Python session, pass it as the snowpark_session argument:

import fsspec
from import sfcfs

sf_fs2 = sfcfs.SFFileSystem(snowpark_session=sp_session)

SFFileSystem inherits many features from fsspec.FileSystem, such as local caching of files. You can enable this and other features by instantiating a Snowflake file system through the fsspec.filesystem factory function, passing target_protocol="sfc" to use the Snowflake FileSystem implementation:

local_cache_path = "/tmp/sf_files/"
cached_fs = fsspec.filesystem("cached", target_protocol="sfc",
                    target_options={"sf_connection": sf_connection,
                                    "cache_types": "bytes",
                                    "block_size": 32 * 2**20},

The Snowflake file system supports most read-only methods defined for a fsspec FileSystem, including find, info, isdir, isfile, and exists.

Specifying Files

To specify files in a stage, use a path in the form @database.schema.stage/file_path.

Listing Files

The file system’s ls method is used to get a list of the files in the stage:

print(*"@ML_DATASETS.public.my_models/sales_predict/"), end='\n')

Opening and Reading Files

You can open files in the stage by using the file system’s open method. You can then read the files by using the same methods you use with ordinary Python files. The file object is also a context manager that can be used with Python’s with statement, so it is automatically closed when it’s no longer needed.

path = '@ML_DATASETS.public.my_models/test/data_7_7_3.snappy.parquet'

with, mode='rb') as f:

You can also use the SFFileSystem instance with other components that accept fsspec file systems. Here, the Parquet data file mentioned in the previous code block is passed to PyArrow’s read_table method:

import pyarrow.parquet as pq

table = pq.read_table(path, filesystem=sf_fs1)
table.take([1, 3])

Python components that accept files (or file-like objects) can be passed a file object opened from the Snowflake file system. For example, if you have a gzip-compressed file in your stage, you can use it with Python’s gzip module by passing it to gzip.GzipFile as the fileobj parameter:

path = "sfc://@ML_DATASETS.public.my_models/dataset.csv.gz"

with, mode='rb', sf_connection=sf_connection) as f:
    g = gzip.GzipFile(fileobj=f)
    for i in range(3):

Creating and Using a FileSet

A Snowflake FileSet represents an immutable snapshot of the result of a SQL query in the form of files in an internal server-side encrypted stage. These files can be accessed through a FileSystem to feed data to tools such as PyTorch and TensorFlow so that you can train models at scale and within your existing data governance model. To create a FileSet, use the FileSet.make method.

You need a Snowflake Python connection or a Snowpark session to create a FileSet. See Authenticating to Snowflake for instructions. You must also provide the path to an existing internal server-side encrypted stage, or a subdirectory under such a stage, where the FileSet will be stored.

To create a FileSet from a Snowpark DataFrame, construct a DataFrame and pass it to FileSet.make as snowpark_dataframe; do not call the DataFrame’s collect method:

# Snowpark Python equivalent of "SELECT * FROM MYDATA LIMIT 5000000"
df = snowpark_session.table('mydata').limit(5000000)
fileset_df = fileset.FileSet.make(

To create a FileSet using a Snowflake Connector for Python connection, pass the connection to Fileset.make as sf_connection, and pass the SQL query as query:

fileset_sf = fileset.FileSet.make(
    query="SELECT * FROM MYDATA LIMIT 5000000",
    shuffle=True,           # see later section about shuffling


See Shuffling Data in FileSets for information about shuffling your data by using the shuffle parameter.

Use the files method to get a list of the files in the FileSet:


For information about feeding the data in the FileSet to PyTorch or TensorFlow, see Snowpark ML Framework Connectors.