Read csv from Google Cloud storage to pandas dataframe

UPDATE

As of version 0.24 of pandas, read_csv supports reading directly from Google Cloud Storage. Simply provide link to the bucket like this:

df = pd.read_csv('gs://bucket/your_path.csv')

The read_csv will then use gcsfs module to read the Dataframe, which means it had to be installed (or you will get an exception pointing at missing dependency).

I leave three other options for the sake of completeness.

• Home-made code
• gcsfs
• dask

I will cover them below.

The hard way: do-it-yourself code

I have written some convenience functions to read from Google Storage. To make it more readable I added type annotations. If you happen to be on Python 2, simply remove these and code will work all the same.

It works equally on public and private data sets, assuming you are authorised. In this approach you don't need to download first the data to your local drive.

How to use it:

fileobj = get_byte_fileobj('my-project', 'my-bucket', 'my-path')
df = pd.read_csv(fileobj)

The code:

from io import BytesIO, StringIO
from google.cloud import storage
from google.oauth2 import service_account

def get_byte_fileobj(project: str,
                     bucket: str,
                     path: str,
                     service_account_credentials_path: str = None) -> BytesIO:
    """
    Retrieve data from a given blob on Google Storage and pass it as a file object.
    :param path: path within the bucket
    :param project: name of the project
    :param bucket_name: name of the bucket
    :param service_account_credentials_path: path to credentials.
           TIP: can be stored as env variable, e.g. os.getenv('GOOGLE_APPLICATION_CREDENTIALS_DSPLATFORM')
    :return: file object (BytesIO)
    """
    blob = _get_blob(bucket, path, project, service_account_credentials_path)
    byte_stream = BytesIO()
    blob.download_to_file(byte_stream)
    byte_stream.seek(0)
    return byte_stream

def get_bytestring(project: str,
                   bucket: str,
                   path: str,
                   service_account_credentials_path: str = None) -> bytes:
    """
    Retrieve data from a given blob on Google Storage and pass it as a byte-string.
    :param path: path within the bucket
    :param project: name of the project
    :param bucket_name: name of the bucket
    :param service_account_credentials_path: path to credentials.
           TIP: can be stored as env variable, e.g. os.getenv('GOOGLE_APPLICATION_CREDENTIALS_DSPLATFORM')
    :return: byte-string (needs to be decoded)
    """
    blob = _get_blob(bucket, path, project, service_account_credentials_path)
    s = blob.download_as_string()
    return s


def _get_blob(bucket_name, path, project, service_account_credentials_path):
    credentials = service_account.Credentials.from_service_account_file(
        service_account_credentials_path) if service_account_credentials_path else None
    storage_client = storage.Client(project=project, credentials=credentials)
    bucket = storage_client.get_bucket(bucket_name)
    blob = bucket.blob(path)
    return blob

gcsfs

gcsfs is a "Pythonic file-system for Google Cloud Storage".

How to use it:

import pandas as pd
import gcsfs

fs = gcsfs.GCSFileSystem(project='my-project')
with fs.open('bucket/path.csv') as f:
    df = pd.read_csv(f)

dask

Dask "provides advanced parallelism for analytics, enabling performance at scale for the tools you love". It's great when you need to deal with large volumes of data in Python. Dask tries to mimic much of the pandas API, making it easy to use for newcomers.

Here is the read_csv

How to use it:

import dask.dataframe as dd

df = dd.read_csv('gs://bucket/data.csv')
df2 = dd.read_csv('gs://bucket/path/*.csv') # nice!

# df is now Dask dataframe, ready for distributed processing
# If you want to have the pandas version, simply:
df_pd = df.compute()

Another option is to use TensorFlow which comes with the ability to do a streaming read from Google Cloud Storage:

from tensorflow.python.lib.io import file_io
with file_io.FileIO('gs://bucket/file.csv', 'r') as f:
  df = pd.read_csv(f)

Using tensorflow also gives you a convenient way to handle wildcards in the filename. For example:

Reading wildcard CSV into Pandas

Here is code that will read all CSVs that match a specific pattern (e.g: gs://bucket/some/dir/train-*) into a Pandas dataframe:

import tensorflow as tf
from tensorflow.python.lib.io import file_io
import pandas as pd

def read_csv_file(filename):
  with file_io.FileIO(filename, 'r') as f:
    df = pd.read_csv(f, header=None, names=['col1', 'col2'])
    return df

def read_csv_files(filename_pattern):
  filenames = tf.gfile.Glob(filename_pattern)
  dataframes = [read_csv_file(filename) for filename in filenames]
  return pd.concat(dataframes)

usage

DATADIR='gs://my-bucket/some/dir'
traindf = read_csv_files(os.path.join(DATADIR, 'train-*'))
evaldf = read_csv_files(os.path.join(DATADIR, 'eval-*'))

As of pandas==0.24.0 this is supported natively if you have gcsfs installed: https://github.com/pandas-dev/pandas/pull/22704.

Until the official release you can try it out with pip install pandas==0.24.0rc1.

I was taking a look at this question and didn't want to have to go through the hassle of installing another library, gcsfs, which literally says in the documentation, This software is beta, use at your own risk... but I found a great workaround that I wanted to post here in case this is helpful to anyone else, using just the google.cloud storage library and some native python libraries. Here's the function:

import pandas as pd
from google.cloud import storage
import os
import io
os.environ['GOOGLE_APPLICATION_CREDENTIALS'] = 'path/to/creds.json'


def gcp_csv_to_df(bucket_name, source_file_name):
    storage_client = storage.Client()
    bucket = storage_client.bucket(bucket_name)
    blob = bucket.blob(source_blob_name)
    data = blob.download_as_string()
    df = pd.read_csv(io.BytesIO(data))
    print(f'Pulled down file from bucket {bucket_name}, file name: {source_file_name}')
    return df

Further, although it is outside of the scope of this question, if you would like to upload a pandas dataframe to GCP using a similar function, here is the code to do so:

def df_to_gcp_csv(df, dest_bucket_name, dest_file_name):
    storage_client = storage.Client()
    bucket = storage_client.bucket(dest_bucket_name)
    blob = bucket.blob(dest_file_name)
    blob.upload_from_string(df.to_csv(), 'text/csv')
    print(f'DataFrame uploaded to bucket {dest_bucket_name}, file name: {dest_file_name}')

Hope this is helpful! I know I'll be using these functions for sure.