User guide

Petastorm is an open source data access library developed at Uber ATG. This library enables single machine or distributed training and evaluation of deep learning models directly from datasets in Apache Parquet format. Petastorm supports popular Python-based machine learning (ML) frameworks such as Tensorflow, PyTorch, and PySpark. It can also be used from pure Python code.

Documentation web site:


pip install petastorm

There are several extra dependencies that are defined by the petastorm package that are not installed automatically. The extras are: tf, tf_gpu, torch, opencv, docs, test.

For example to trigger installation of GPU version of tensorflow and opencv, use the following pip command:

pip install petastorm[opencv,tf_gpu]

Generating a dataset

A dataset created using Petastorm is stored in Apache Parquet format. On top of a Parquet schema, petastorm also stores higher-level schema information that makes multidimensional arrays into a native part of a petastorm dataset.

Petastorm supports extensible data codecs. These enable a user to use one of the standard data compressions (jpeg, png) or implement her own.

Generating a dataset is done using PySpark. PySpark natively supports Parquet format, making it easy to run on a single machine or on a Spark compute cluster. Here is a minimalistic example writing out a table with some random data.

import numpy as np
from pyspark.sql import SparkSession
from pyspark.sql.types import IntegerType

from petastorm.codecs import ScalarCodec, CompressedImageCodec, NdarrayCodec
from petastorm.etl.dataset_metadata import materialize_dataset
from petastorm.unischema import dict_to_spark_row, Unischema, UnischemaField

# The schema defines how the dataset schema looks like
HelloWorldSchema = Unischema('HelloWorldSchema', [
    UnischemaField('id', np.int32, (), ScalarCodec(IntegerType()), False),
    UnischemaField('image1', np.uint8, (128, 256, 3), CompressedImageCodec('png'), False),
    UnischemaField('array_4d', np.uint8, (None, 128, 30, None), NdarrayCodec(), False),

def row_generator(x):
    """Returns a single entry in the generated dataset. Return a bunch of random values as an example."""
    return {'id': x,
            'image1': np.random.randint(0, 255, dtype=np.uint8, size=(128, 256, 3)),
            'array_4d': np.random.randint(0, 255, dtype=np.uint8, size=(4, 128, 30, 3))}

def generate_petastorm_dataset(output_url='file:///tmp/hello_world_dataset'):
    rowgroup_size_mb = 256

    spark = SparkSession.builder.config('spark.driver.memory', '2g').master('local[2]').getOrCreate()
    sc = spark.sparkContext

    # Wrap dataset materialization portion. Will take care of setting up spark environment variables as
    # well as save petastorm specific metadata
    rows_count = 10
    with materialize_dataset(spark, output_url, HelloWorldSchema, rowgroup_size_mb):

        rows_rdd = sc.parallelize(range(rows_count))\
            .map(lambda x: dict_to_spark_row(HelloWorldSchema, x))

        spark.createDataFrame(rows_rdd, HelloWorldSchema.as_spark_schema()) \
            .coalesce(10) \
            .write \
            .mode('overwrite') \
  • HelloWorldSchema is an instance of a Unischema object. Unischema is capable of rendering types of its fields into different framework specific formats, such as: Spark StructType, Tensorflow tf.DType and numpy numpy.dtype.
  • To define a dataset field, you need to specify a type, shape, a codec instance and whether the field is nullable for each field of the Unischema.
  • We use PySpark for writing output Parquet files. In this example, we launch PySpark on a local box (.master('local[2]')). Of course for a larger scale dataset generation we would need a real compute cluster.
  • We wrap spark dataset generation code with the materialize_dataset context manager. The context manager is responsible for configuring row group size at the beginning and write out petastorm specific metadata at the end.
  • The row generating code is expected to return a Python dictionary indexed by a field name. We use row_generator function for that.
  • dict_to_spark_row converts the dictionary into a pyspark.Row object while ensuring schema HelloWorldSchema compliance (shape, type and is-nullable condition are tested).
  • Once we have a pyspark.DataFrame we write it out to a parquet storage. The parquet schema is automatically derived from HelloWorldSchema.

Plain Python API

The petastorm.reader.Reader class is the main entry point for user code that accesses the data from an ML framework such as Tensorflow or Pytorch. The reader has multiple features such as:

  • Selective column readout
  • Multiple parallelism strategies: thread, process, single-threaded (for debug)
  • N-grams readout support
  • Row filtering (row predicates)
  • Shuffling
  • Partitioning for multi-GPU training
  • Local caching

Reading a dataset is simple using the petastorm.reader.Reader class which can be created using the petastorm.make_reader factory method:

from petastorm import make_reader

 with make_reader('hdfs://myhadoop/some_dataset') as reader:
    for row in reader:

hdfs://... and file://... are supported URL protocols.

Once a Reader is instantiated, you can use it as an iterator.

Tensorflow API

To hookup the reader into a tensorflow graph, you can use the tf_tensors function:

from petastorm.tf_utils import tf_tensors

with make_reader('file:///some/localpath/a_dataset') as reader:
   row_tensors = tf_tensors(reader)
   with tf.Session() as session:
       for _ in range(3):

Alternatively, you can use new API;

from petastorm.tf_utils import make_petastorm_dataset

with make_reader('file:///some/localpath/a_dataset') as reader:
    dataset = make_petastorm_dataset(reader)
    iterator = dataset.make_one_shot_iterator()
    tensor = iterator.get_next()
    with tf.Session() as sess:
        sample =

Pytorch API

As illustrated in, reading a petastorm dataset from pytorch can be done via the adapter class petastorm.pytorch.DataLoader, which allows custom pytorch collating function and transforms to be supplied.

Be sure you have torch and torchvision installed:

pip install torchvision

The minimalist example below assumes the definition of a Net class and train and test functions, included in pytorch_example:

import torch
from petastorm.pytorch import DataLoader

device = torch.device('cpu')
model = Net().to(device)
optimizer = torch.optim.SGD(model.parameters(), lr=0.01, momentum=0.5)

def _transform_row(mnist_row):
    transform = transforms.Compose([
        transforms.Normalize((0.1307,), (0.3081,))
    return (transform(mnist_row['image']), mnist_row['digit'])

transform = TransformSpec(_transform_row, removed_fields=['idx'])

with DataLoader(make_reader('file:///localpath/mnist/train', num_epochs=10,
                            transform_spec=transform, seed=1, shuffle_rows=True), batch_size=64) as train_loader:
    train(model, device, train_loader, 10, optimizer, 1)
with DataLoader(make_reader('file:///localpath/mnist/test', num_epochs=10,
                            transform_spec=transform), batch_size=1000) as test_loader:
    test(model, device, test_loader)

If you are working with very large batch sizes and do not need support for Decimal/strings we provide a petastorm.pytorch.BatchedDataLoader that can buffer using Torch tensors (cpu or cuda) with a signficantly higher throughput.

If the size of your dataset can fit into system memory, you can use an in-memory version dataloader petastorm.pytorch.InMemBatchedDataLoader. This dataloader only reades the dataset once, and caches data in memory to avoid additional I/O for multiple epochs.

Spark Dataset Converter API

Spark converter API simplifies the data conversion from Spark to TensorFlow or PyTorch. The input Spark DataFrame is first materialized in the parquet format and then loaded as a or

The minimalist example below assumes the definition of a compiled tf.keras model and a Spark DataFrame containing a feature column followed by a label column.

from petastorm.spark import SparkDatasetConverter, make_spark_converter
import tensorflow.compat.v1 as tf  # pylint: disable=import-error

# specify a cache dir first.
# the dir is used to save materialized spark dataframe files
spark.conf.set(SparkDatasetConverter.PARENT_CACHE_DIR_URL_CONF, 'hdfs:/...')

df = ... # `df` is a spark dataframe

# create a converter from `df`
# it will materialize `df` to cache dir.
converter = make_spark_converter(df)

# make a tensorflow dataset from `converter`
with converter.make_tf_dataset() as dataset:
    # the `dataset` is `` object
    # dataset transformation can be done if needed
    dataset =
    # we can train/evaluate model on the `dataset`
    # when exiting the context, the reader of the dataset will be closed

# delete the cached files of the dataframe.

The minimalist example below assumes the definition of a Net class and train and test functions, included in, and a Spark DataFrame containing a feature column followed by a label column.

from petastorm.spark import SparkDatasetConverter, make_spark_converter

# specify a cache dir first.
# the dir is used to save materialized spark dataframe files
spark.conf.set(SparkDatasetConverter.PARENT_CACHE_DIR_URL_CONF, 'hdfs:/...')

df_train, df_test = ... # `df_train` and `df_test` are spark dataframes
model = Net()

# create a converter_train from `df_train`
# it will materialize `df_train` to cache dir. (the same for df_test)
converter_train = make_spark_converter(df_train)
converter_test = make_spark_converter(df_test)

# make a pytorch dataloader from `converter_train`
with converter_train.make_torch_dataloader() as dataloader_train:
    # the `dataloader_train` is `` object
    # we can train model using the `dataloader_train`
    train(model, dataloader_train, ...)
    # when exiting the context, the reader of the dataset will be closed

# the same for `converter_test`
with converter_test.make_torch_dataloader() as dataloader_test:
    test(model, dataloader_test, ...)

# delete the cached files of the dataframes.

Analyzing petastorm datasets using PySpark and SQL

A Petastorm dataset can be read into a Spark DataFrame using PySpark, where you can use a wide range of Spark tools to analyze and manipulate the dataset.

# Create a dataframe object from a parquet file
dataframe =

# Show a schema

# Count all

# Show a single column'id').show()

SQL can be used to query a Petastorm dataset:

   'SELECT count(id) '
   'from parquet.`file:///tmp/hello_world_dataset`').collect()

You can find a full code sample here:,

Non Petastorm Parquet Stores

Petastorm can also be used to read data directly from Apache Parquet stores. To achieve that, use make_batch_reader (and not make_reader). The following table summarizes the differences make_batch_reader and make_reader functions.

make_reader make_batch_reader
Only Petastorm datasets (created using materializes_dataset) Any Parquet store (some native Parquet column types are not supported yet.
The reader returns one record at a time. The reader returns batches of records. The size of the batch is not fixed and defined by Parquet row-group size.
Predicates passed to make_reader are evaluated per single row. Predicates passed to make_batch_reader are evaluated per batch.
Can filter parquet file based on the filters argument. Can filter parquet file based on the filters argument


See the Troubleshooting page and please submit a ticket if you can’t find an answer.

See also

  1. Gruener, R., Cheng, O., and Litvin, Y. (2018) Introducing Petastorm: Uber ATG’s Data Access Library for Deep Learning. URL:
  2. 2019: “Petastorm: A Light-Weight Approach to Building ML Pipelines”.

How to Contribute

We prefer to receive contributions in the form of GitHub pull requests. Please send pull requests against the repository.

  • If you are looking for some ideas on what to contribute, check out github issues and comment on the issue.
  • If you have an idea for an improvement, or you’d like to report a bug but don’t have time to fix it please a create a github issue.

To contribute a patch:

  • Break your work into small, single-purpose patches if possible. It’s much harder to merge in a large change with a lot of disjoint features.
  • Submit the patch as a GitHub pull request against the master branch. For a tutorial, see the GitHub guides on forking a repo and sending a pull request.
  • Include a detailed describtion of the proposed change in the pull request.
  • Make sure that your code passes the unit tests. You can find instructions how to run the unit tests here.
  • Add new unit tests for your code.

Thank you in advance for your contributions!

See the Development for development related information.