Save Dataframe to csv directly to s3 Python

If you pass None as the first argument to to_csv() the data will be returned as a string. From there it’s an easy step to upload that to S3 in one go.

It should also be possible to pass a StringIO object to to_csv(), but using a string will be easier.

I like s3fs which lets you use s3 (almost) like a local filesystem.

You can do this:

import s3fs

bytes_to_write = df.to_csv(None).encode()
fs = s3fs.S3FileSystem(key=key, secret=secret)
with fs.open('s3://bucket/path/to/file.csv', 'wb') as f:
    f.write(bytes_to_write)

s3fs supports only rb and wb modes of opening the file, that’s why I did this bytes_to_write stuff.

I read a csv with two columns from bucket s3, and the content of the file csv i put in pandas dataframe.

Example:

config.json

{
  "credential": {
    "access_key":"xxxxxx",
    "secret_key":"xxxxxx"
}
,
"s3":{
       "bucket":"mybucket",
       "key":"csv/user.csv"
   }
}

cls_config.json

#!/usr/bin/env python
# -*- coding: utf-8 -*-

import os
import json

class cls_config(object):

    def __init__(self,filename):

        self.filename = filename


    def getConfig(self):

        fileName = os.path.join(os.path.dirname(__file__), self.filename)
        with open(fileName) as f:
        config = json.load(f)
        return config

cls_pandas.py

#!/usr/bin/env python
# -*- coding: utf-8 -*-

import pandas as pd
import io

class cls_pandas(object):

    def __init__(self):
        pass

    def read(self,stream):

        df = pd.read_csv(io.StringIO(stream), sep = ",")
        return df

cls_s3.py

#!/usr/bin/env python
# -*- coding: utf-8 -*-

import boto3
import json

class cls_s3(object):

    def  __init__(self,access_key,secret_key):

        self.s3 = boto3.client('s3', aws_access_key_id=access_key, aws_secret_access_key=secret_key)

    def getObject(self,bucket,key):

        read_file = self.s3.get_object(Bucket=bucket, Key=key)
        body = read_file['Body'].read().decode('utf-8')
        return body

test.py

#!/usr/bin/env python
# -*- coding: utf-8 -*-

from cls_config import *
from cls_s3 import *
from cls_pandas import *

class test(object):

    def __init__(self):
        self.conf = cls_config('config.json')

    def process(self):

        conf = self.conf.getConfig()

        bucket = conf['s3']['bucket']
        key = conf['s3']['key']

        access_key = conf['credential']['access_key']
        secret_key = conf['credential']['secret_key']

        s3 = cls_s3(access_key,secret_key)
        ob = s3.getObject(bucket,key)

        pa = cls_pandas()
        df = pa.read(ob)

        print df

if __name__ == '__main__':
    test = test()
    test.process()

This is a more up to date answer:

import s3fs

s3 = s3fs.S3FileSystem(anon=False)

# Use 'w' for py3, 'wb' for py2
with s3.open('<bucket-name>/<filename>.csv','w') as f:
    df.to_csv(f)

The problem with StringIO is that it will eat away at your memory. With this method, you are streaming the file to s3, rather than converting it to string, then writing it into s3. Holding the pandas dataframe and its string copy in memory seems very inefficient.

If you are working in an ec2 instant, you can give it an IAM role to enable writing it to s3, thus you dont need to pass in credentials directly. However, you can also connect to a bucket by passing credentials to the S3FileSystem() function. See documention:https://s3fs.readthedocs.io/en/latest/

You can directly use the S3 path. I am using Pandas 0.24.1

In [1]: import pandas as pd

In [2]: df = pd.DataFrame( [ [1, 1, 1], [2, 2, 2] ], columns=['a', 'b', 'c'])

In [3]: df
Out[3]:
   a  b  c
0  1  1  1
1  2  2  2

In [4]: df.to_csv('s3://experimental/playground/temp_csv/dummy.csv', index=False)

In [5]: pd.__version__
Out[5]: '0.24.1'

In [6]: new_df = pd.read_csv('s3://experimental/playground/temp_csv/dummy.csv')

In [7]: new_df
Out[7]:
   a  b  c
0  1  1  1
1  2  2  2

Release Note:

S3 File Handling

pandas now uses s3fs for handling S3 connections. This shouldn’t break any code. However, since s3fs is not a required dependency, you will need to install it separately, like boto in prior versions of pandas. GH11915.

since you are using boto3.client(), try:

import boto3
from io import StringIO #python3 
s3 = boto3.client('s3', aws_access_key_id='key', aws_secret_access_key='secret_key')
def copy_to_s3(client, df, bucket, filepath):
    csv_buf = StringIO()
    df.to_csv(csv_buf, header=True, index=False)
    csv_buf.seek(0)
    client.put_object(Bucket=bucket, Body=csv_buf.getvalue(), Key=filepath)
    print(f'Copy {df.shape[0]} rows to S3 Bucket {bucket} at {filepath}, Done!')

copy_to_s3(client=s3, df=df_to_upload, bucket='abc', filepath='def/test.csv')

You can also use the AWS Data Wrangler:

import awswrangler as wr
    
wr.s3.to_csv(
    df=df,
    path="s3://...",
)

Note that it will handle multipart upload for you to make the upload faster.

I found this can be done using client also and not just resource.

from io import StringIO
import boto3
s3 = boto3.client("s3",
                  region_name=region_name,
                  aws_access_key_id=aws_access_key_id,
                  aws_secret_access_key=aws_secret_access_key)
csv_buf = StringIO()
df.to_csv(csv_buf, header=True, index=False)
csv_buf.seek(0)
s3.put_object(Bucket=bucket, Body=csv_buf.getvalue(), Key='path/test.csv')

I use AWS Data Wrangler. For example:

import awswrangler as wr
import pandas as pd

# read a local dataframe
df = pd.read_parquet('my_local_file.gz')

# upload to S3 bucket
wr.s3.to_parquet(df=df, path='s3://mys3bucket/file_name.gz')

The same applies to csv files. Instead of read_parquet and to_parquet, use read_csv and to_csv with the proper file extension.

You can use

• pandas
• boto3
• s3fs (version ≤0.4)

I use to_csv with s3:// in path and storage_options

key = "folder/file.csv"

df.to_csv(
    f"s3://{YOUR_S3_BUCKET}/{key}",
    index=False,
    storage_options={
        "key": AWS_ACCESS_KEY_ID,
        "secret": AWS_SECRET_ACCESS_KEY,
        "token": AWS_SESSION_TOKEN,
    },

To handle large files efficiently you can also use an open-source S3-compatible MinIO, with its minio python client package, like in this function of mine:

import minio
import os
import pandas as pd

minio_client = minio.Minio(..)

def write_df_to_minio(df, 
                    minio_client, 
                    bucket_name, 
                    file_name="new-file.csv",
                    local_temp_folder="/tmp/", 
                    content_type="application/csv",
                    sep=",",
                    save_row_index=False):

    df.to_csv(os.path.join(local_temp_folder, file_name), sep=sep, index=save_row_index)
    
    minio_results = minio_client.fput_object(bucket_name=bucket_name,
                                             object_name=file_name,
                                             file_path=os.path.join(local_temp_folder, file_name),
                                             content_type=content_type)

    assert minio_results.object_name == file_name

Another option is to do this with cloudpathlib, which supports S3 and also Google Cloud Storage and Azure Blob Storage. See example below.

import pandas as pd
from cloudpathlib import CloudPath

# read data from S3
df = pd.read_csv(CloudPath("s3://covid19-lake/rearc-covid-19-testing-data/csv/states_daily/states_daily.csv"))

# look at some of the data
df.head(1).T.iloc[:10]
#>                                       0
#> date                           20210307
#> state                                AK
#> positive                        56886.0
#> probableCases                       NaN
#> negative                            NaN
#> pending                             NaN
#> totalTestResultsSource  totalTestsViral
#> totalTestResults              1731628.0
#> hospitalizedCurrently              33.0
#> hospitalizedCumulative           1293.0

# writing to S3
with CloudPath("s3://bucket-you-can-write-to/data.csv").open("w") as f:
    df.to_csv(f)

CloudPath("s3://bucket-you-can-write-to/data.csv").exists()
#> True

Note, that you can’t call df.to_csv(CloudPath("s3://drivendata-public-assets/test-asdf2.csv")) directly because of the way pandas handles paths/handles passed to it. Instead you need to open the file for writing and pass that handle directly to to_csv.

This comes with a few added benefits in terms of setting particular options or different authentication mechanisms or keeping a persistent cache so you don’t always need to redownload from S3.

from io import StringIO
import boto3
#Creating Session With Boto3.
session = boto3.Session(
aws_access_key_id='<your_access_key_id>',
aws_secret_access_key='<your_secret_access_key>'
)
#Creating S3 Resource From the Session.
s3_res = session.resource('s3')
csv_buffer = StringIO()
df.to_csv(csv_buffer)
bucket_name = 'stackvidhya'
s3_object_name = 'df.csv'
s3_res.Object(bucket_name, s3_object_name).put(Body=csv_buffer.getvalue())
print("Dataframe is saved as CSV in S3 bucket.")

For those who might have problems with S3FS or fsspec using Lambda:

You have to create a layer for each libary and insert them in your Lambda.

You can find how to crate a layer here.