Support for Azure Blob Storage and Azure Data Lake

[upendrarv]

Hi,

Any plans to add support for Azure Blob storage and/or Azure Data Lake ?

Thanks for the great work !

[jsgoller1]

Hi there @upendrarv, thanks for opening this issue. Can you clarify specifically what you're asking?

[selitvin]

This is not something we were planning to work on at the moment.

For this to work, we would need a proper Azure blob storage support from pyarrow. I am not sure what is arrow's progress on having a proper support of these Azure filesystems.The only thing I see in their documentation is this. It might be possible to do something using that technique in petastorm, but might get hairy as we do use pyarrow's ParquetDataset class.

[upendrarv]

@jsgoller1 : Currently I am looking to refer the PetaStorm datasets stored in Azure Data Lake Storage and Azure Blob Storage.. Currently there is a support for S3 and GCS (Case 5 and Case 6 in [fs_util.py].(https://github.com/uber/petastorm/blob/c370bac366e86ca07c051cbd0daffacf866ecde1/petastorm/fs_utils.py)). Similarly I am looking for support for Azure Blob File System.

@selitvin : Thanks for the reply. In Python, adlfs package available. We can use adlfs to access the Azure Blob File System instead of taking dependency on pyarrow. I have created a prototype code to access the petastorm datasets present in Azure Blob File System. With below piece of code, I am able to read the dataset and train my ML model.

Below is the sample code:

import os
import posixpath

from adlfs import AzureBlobFileSystem
from petastorm.reader import Reader
from pyarrow.filesystem import FileSystem, DaskFileSystem
from pyarrow.util import implements, _stringify_path

class ABFSWrapper(DaskFileSystem):
    @implements(FileSystem.isdir)
    def isdir(self, path):
        path = _sanitize_azure_path(_stringify_path(path))
        try:
            contents = self.fs.ls(path)
            if len(contents) == 1 and contents[0] == path:
                return False
            else:
                if not any(ftype in path for ftype in
                           ['parquet', 'parq', 'metadata']):
                    raise ValueError('Directory is not a partition of'
                                     ' *.parquet. Try passing a globstring.')
                return True
        except OSError:
            return False

    @implements(FileSystem.isfile)
    def isfile(self, path):
        path = _sanitize_azure_path(_stringify_path(path))
        try:
            contents = self.fs.ls(path)
            return len(contents) == 1 and contents[0] == path
        except OSError:
            return False

    def walk(self, path, invalidate_cache=True, refresh=False):
        """
        Directory tree generator, like os.walk
        Generator version of what is in ABFSClient, which yields a flattened list of files
        """
        path = _sanitize_azure_path(_stringify_path(path))
        directories = set()
        files = set()

        for key in list(self.fs.ls(path, invalidate_cache=invalidate_cache)):
            path = key
            if self.isdir(path):
                directories.add(path)
            elif self.isfile(path):
                files.add(path)
            else:
                pass

        files = sorted([posixpath.split(f)[1] for f in files
                        if f not in directories])
        directories = sorted([posixpath.split(x)[1]
                              for x in directories])

        yield path, directories, files

        for directory in directories:
            for tup in self.walk(directory, refresh=refresh):
                yield tup
def _sanitize_azure_path(path):
    if path.startswith('adl://'):
        return path.replace('adl://', '')
    else:
        return path

abfs = AzureBlobFileSystem(account_name="xxxxxxxxxxxxxxxxxx", account_key="xxxxxxxxxxxxxxxxx")
wrapped_fs = ABFSWrapper(abfs)

dataset_url_without_protocol_prefix = "/petastorm/parquet"
with Reader(dataset_path=dataset_url_without_protocol_prefix, pyarrow_filesystem=wrapped_fs) as reader:
    .....................................................

The same piece of code can be added to Petastorm Library - similar to gcsfs_helper folder. Let me know your thoughts on the same.

[selitvin]

This is nice! Would you like to take a shot at adding a PR that would add this To Petastorm?

[sam-h-bean]

@upendrarv Do it! I am facing this right now trying to instantiate a batch reader from a parquet table in Azure Blob Storage