Skip to content


An unofficial generic client stack for AzureML web services, working with both python 2 and 3.

Python versions Build Status Tests Status codecov

Documentation PyPI Downloads Downloads per week GitHub stars

New AzureMLClient base class to create high-level clients is here, check it out

azmlclient helps you consume web services deployed on the AzureML platform easily. It provides you with a low-level API to call web services in request-response or batch mode. It also offers optional tools if you wish to provide high-level applicative APIs on top of these web services.

As opposed to AzureML client library,

  • this library is much simpler and is only focused on consuming web services.
  • It is compliant with all services deployed from AzureML experiments (using the AzureML studio UI), and should also work with python and R "dataframe" web services (not checked though).
  • It does not require your AzureML workspace id and API key, only the deployed services' URL and API key.

You may use it for example

  • to show to your customers how to consume your AzureML cloud services.
  • to make simple 'edge' devices consume your AzureML cloud services (if they support python :) ).


> pip install azmlclient

1. Low level API

This API is the python equivalent of the "execute" generic AzureML operation. It supports both request-response and batch mode, as well as swagger and non-swagger format.

First examples

First create variables holding the endpoint information provided by AzureML

base_url = 'https://<geo><wId>/services/<sId>'
api_key = '<apiKey>'

Then create

  • the inputs - a dictionary containing all you inputs as pandas.DataFrame objects
  • the parameters - a dictionary
  • and optionally define a list of expected output names
inputs = {"trainDataset": training_df, "input2": input2_df}
params = {"param1": "val1", "param2": "val2"}
output_names = ["my_out1","my_out2"]

Finally call in Request-Response mode:

from azmlclient import execute_rr
outputs = execute_rr(api_key, base_url, 
                     inputs=inputs, params=params, output_names=output_names)

Or in Batch mode. In this case you also need to configure the Blob storage to be used:

from azmlclient import execute_bes

# Define the blob storage to use for storing inputs and outputs
blob_account = '<account_id>'       # 'myblobs'
blob_apikey = '<api_key>'           # 'mi3Qxcd5rwuM9r5k7h2ipXNww2T0Bw=='
blob_container = '<container>'      # 'rootcontainer'
blob_path_prefix = '<path_prefix>'  # 'folder/path'

# Perform the call (polling is done by default every 5s until job end)
outputs = execute_bes(api_key, base_url,
                      blob_account, blob_apikey, blob_container, 
                      inputs=inputs, params=params, output_names=output_names)

Formatting options

execute_bes provides several options to control how dataframes are converted to json in the request payloads:

  • swagger_format is a boolean (default False) enabling the more verbose "swagger" (= json objects) format
  • replace_NaN_with and replace_NaT_with control how NaN and NaT are converted

Debug and proxies

Users may wish to create a requests session object using the helper method provided, in order to override environment variable settings for HTTP requests. For example to use Fiddler as a proxy to debug the web service calls:

from azmlclient import create_session_for_proxy
session = create_session_for_proxy(http_proxyhost='localhost', 

Then you may use that object in the requests_session parameter of the methods:

outputsRR = execute_rr(..., requests_session=session)
outputsB = execute_bes(..., requests_session=session)

Note that the session object will be passed to the underlying azure blob storage client to ensure consistency.

Advanced usage

Advanced users may with to create BatchClient or RequestResponseClient classes to better control what's happening.

from azmlclient import RequestResponseClient

# 0- Create the client
rr_client = RequestResponseClient(requests_session=requests_session)

# 1- Create the query body
request_body = rr_client.create_request_body(inputs, params)

# 2- Execute the query and receive the response body
response_body = rr_client.execute_rr(base_url, api_key, request_body)
# 3- parse the response body into a dictionary of dataframes
result_dfs = rr_client.read_response_json_body(response_body, output_names)

2. Providing high-level APIs

Even though the above API is enough to consume your AzureML web services, it is still very low-level:

  • the services are not mapped to python methods with friendly names
  • their inputs, outputs and parameters have to be created by hand from python structures
  • changing the call mode between request-response and batch requires you to change your code
  • there is no easy way to switch between remote and local call, for example for hybrid implementations (computationally intensive operations in the cloud, computationally cheap operations executed locally)

For all these reasons, azmlclient offers tools to help you provide higher-level APIs.

Creating the main client class

Let's imagine that we have two AzureML services deployed: one for adding dataframe columns and another for subtracting them. We wish to provide our users with a more pythonic way to call them than the low-level api that we saw previously.

A nice way to do this is to create a "client class", that will hide away the AzureML specific syntax. We will name our class MathsProvider, it will offer one pythonic method mapped on each AzureML service: add_columns(a_name, b_name, df) and subtract_columns(a_name, b_name, df) respectively.

It is extremely easy to create such a class, by inheriting from AzureMLClient. This helper base class provide a bunch of mechanisms to automate both configuration and support for alternate call modes (local, request-response, batch) as we'll see below.

For each service that we want to offer, we create a method. That method should

  • be decorated with @azureml_service,
  • transform the received arguments (python objects) into azureml inputs and parameters dictionaries, in the same format that presented previously in the low-level api,
  • use the self.call_azureml(...) helper function to perform the AzureML call. Note that this helper function handles the call mode (request response or batch) for you as we'll see below.
  • unpack the various results and create the appropriate outputs (python objects) from them.

For example:

from azmlclient import AzureMLClient, azureml_service

class MathsProvider(AzureMLClient):
    A client for the `add_columns` and `subtract_columns` AzureML web services
    def add_columns(self, a_name, b_name, df):
        Offers a pythonic API around the `add_columns` azureML service

        :param a_name: name of the first column to add (a string)
        :param b_name: name of the second column to add (a string)
        :param df: the input dataframe, that should at least contain the 2 columns selected
        # (1) create the web service inputs and parameters from provided data.
        ws_inputs = {'input': df}
        ws_params = {'a_name': a_name, 'b_name': b_name}

        # (2) call the azureml web service
        result_dfs = self.call_azureml(self.add_columns,
                                       ws_output_names=['output']  # optional

        # (3) unpack the results
        return result_dfs['output']

    def subtract_columns(self, a_name, b_name, df):
        # (similar contents than `add_columns` here) 

Using it

Using your new client is extremely easy: simply instantiate it with a ClientConfig configuration object describing the AzureML services endpoints and you're set:

from azmlclient import ClientConfig, GlobalConfig, ServiceConfig
import pandas as pd

# create a configuration indicating the endpoints for each service id
cfg = ClientConfig(add_columns=ServiceConfig(base_url="https://.....", 

# instantiate the client
client = MathsProvider(cfg)

# use it
df = pd.DataFrame({'x': [1, 2, 3], 'y': [0, 5, 10]})
result_df = client.add_columns('x', 'y', df)

The configuration object can alternately be loaded from a .yaml file such as this one:

cfg = ClientConfig.load_yaml(yaml_file_path)

or from a configparser-compliant .ini/.cfg file such as this one:

cfg = ClientConfig.load_config(cfg_file_path)

Note that the service names in the configuration are by default the method names in your client class. If you wish to use different names, simply provide the service name to the @azureml_service decorator, for example:

    def minus_columns(self, a_name, b_name, df):


Finally, note that the yaml and ini/cfg configuration files can be templates, using the jinja2 syntax:


base_url = https://localhost:4443/a_plus_b
api_key = {{ api_key }}


This way, you can for example specify api keys at configuration loading time without storing them in the configuration file. Simply provide the variables and their values as keyword arguments in any of the load_config or load_yaml function:

cfg = ClientConfig.load_config(cfg_file_path, api_key="abc25d4789e=o")


If you wish to debug the calls made by your client, there are two things that you can do:

  • (recommended) use a tool to capture network traffic such as Fiddler or (more complex) Wireshark. Some tools such as Fiddler require you to change to http(s) proxy. This can be done by setting the http(s)_proxy configuration option as shown below. Note that this can also be done programatically by passing a GlobalConfig(https_proxy=...,) in the ClientConfig constructor.

# (only for debug) use fiddler proxy and skip ssl verification
# http_proxy = ...
https_proxy = http://localhost:8888
ssl_verify = false

   # (only for debug) use fiddler proxy and skip ssl verification 
   # http_proxy = ...
   https_proxy = http://localhost:8888
   ssl_verify: false
  • alternatively you can use the with client.debug_requests() context manager on your client. This will print the http requests contents on stdout:
with client.debug_requests():
    result_df = client.add_columns('x', 'y', df)

Alternate call modes: local, batch..

In the example above, client.add_columns calls the web service in request-response mode. This call mode can be changed temporarily thanks to the context managers provided:

# change to BATCH mode
with client.batch_calls(polling_period_seconds=20):
    result_df = client.add_columns('x', 'y', df)

# change to RR mode (useless since that's already the default)
with client.rr_calls():
    result_df = client.add_columns('x', 'y', df)

# change to LOCAL mode
with client.local_calls():
    result_df = client.add_columns('x', 'y', df)

For the local calls by default it does not work and yields:

NotImplementedError: Local execution is not available for this client. 
Please override `__init_local_impl__` or set a non-none `self._local_impl` 
if you wish local calls to be made available

But if you override the __init_local_impl__ method and return an object on which the methods are available, it works:

class MathsProviderLocal(object):
    A local implementation of the same services
    def add_columns(self, a_name, b_name, df):
        return pd.DataFrame({'sum': df[a_name] + df[b_name]})

    def subtract_columns(self, a_name, b_name, df):
        return pd.DataFrame({'diff': df[a_name] - df[b_name]})

class MathsProvider(AzureMLClient):
    def __init_local_impl__(self):
        """ Use our local implementation in 'local' call mode"""
        return MathsProviderLocal()

    def add_columns(self, a_name, b_name, df):

    def subtract_columns(self, a_name, b_name, df):

we can test it :

>>> with client.local_calls():
>>>    result_df = client.add_columns('x', 'y', df)
>>> print(result_df)

0    1
1    7
2   13

Note that the default call mode can also be changed permanentlyby specifying another mode in the AzureMLClient constructor arguments, or by changing the client._current_call_mode attribute.

3. Payload conversion goodies

You can use the static methods available on the classes to convert data between json and python easily.

From json to python

You can convert the body from the HTTP web service calls into python objects using RequestResponseClient static methods:

  • for RR requests using RequestResponseClient.decode_request_json_body(body)
  • for RR responses using RequestResponseClient.read_response_json_body(body)

For example for requests:

from azmlclient import RequestResponseClient

# -- read from file
# with open('./request_payload.json') as f:
#     request_payload =

# -- read from variable
request_payload = """
"GlobalParameters": {"p1": "p1_val", "p2": "p2_val"}, 
"Inputs": {
      "trainDataset": {
         "ColumnNames": ["a", "b"], 
         "Values": [[0, 0], [1, 1], [2, 999]]}

# parse the payload from a web service request to create the python equivalents
dfs, params = RequestResponseClient.decode_request_json_body(request_payload)

# display results:
for input_name, df in dfs.items():
    print("Input %r:" % input_name)

""" % "\n   ".join("%r: %r" % (k, v) for k, v in params.items()))


Input 'trainDataset':
   a    b
0  0    0
1  1    1
2  2  999

   'p1': 'p1_val'
   'p2': 'p2_val'

From python to json

You can create json requests and even fake responses from a RequestResponseClient instance:

  • the json request body using rr_client.create_request_body(dfs, params)
  • the json response body using rr_client.create_response_body(dfs)

For example for the request:

import pandas as pd
from azmlclient import RequestResponseClient

# some dataframe
train_df = pd.DataFrame(data={'a': [0, 1, 2],
                              'b': [1, 2, 3]})
# give it a name
dfs = {'trainDataset': train_df}

# parameters
params = {'foo': 1, 'bar': 2}

# convert to json payload and print
json_payload = RequestResponseClient().create_request_body(dfs, params)


{"Inputs": {"trainDataset": {"ColumnNames": ["a", "b"], "Values": [[0, 1], [1, 2], [2, 3]]}}, "GlobalParameters": {"bar": 2, "foo": 1}}

Main features

  • Creates the Web Services requests from dataframe inputs and dataframe/dictionary parameters, and maps the responses to dataframes too
  • Maps the errors to more friendly python exceptions
  • Supports both Request/Response and Batch mode
  • In Batch mode, performs all the Blob storage and retrieval for you.
  • Properly handles file encoding in both modes (utf-8 is used by default as the pivot encoding)
  • Supports global requests.Session configuration to configure the HTTP clients behaviour (including the underlying blob storage client).
  • Provides tools to create higher-level clients supporting both remote and local call modes.

See Also


Do you like this library ? You might also like my other python libraries

Want to contribute ?

Details on the github page: