다음을 통해 공유

Asynchronous web service consumption via batch processing in Python

  • 아티클

Important

This content is being retired and may not be updated in the future. The support for Machine Learning Server will end on July 1, 2022. For more information, see What's happening to Machine Learning Server?

Applies to: Machine Learning Server

In this article, you can learn how to consume a web service asynchronously, which is especially useful with large input data sets and long-running computations. The typical approach to consuming web services in Python, "Request Response" consumption, involves a single API call to execute the code in that web service once. The "Asynchronous Batch" approach involves the execution of code without manual intervention using multiple asynchronous API calls on a specific web service sent as a single request to Machine Learning Server. Then, Machine Learning Server immediately executes those operations once for every row of data provided.

Asynchronous batch workflow

Generally speaking, the process for asynchronous batch consumption of a web service involves the following:

  1. Call the web service on which the batch execution should be run
  2. Define the data records for the batch execution task
  3. Start (or cancel) the batch execution task
  4. Monitor task and interact with results

Use these following public API functions to define, start, and interact with your batch executions.

Asynchronous batch scoring in Python with Machine Learning Server

Example code

The batch example code is stored in a Jupyter notebook. This notebook format allows you to not only see the code alongside detailed explanations, but also allows you to try out the code.

This example walks through the asynchronous consumption of a Python model as a web service hosted in Machine Learning Server. Consume a simple linear model built using the rx_lin_mod function from the revoscalepy package.

► Download the Jupyter notebook to try it out.

Public functions for batch

You can use the following supported public functions to consume a service asynchronously.

Batch functions performed on the service object

Once you get the service object, use these public functions on that object.

Function Usage Description
batch view Define the data records to be batched and the thread count
get_batch_executions view Get the list of batch execution identifiers
get_batch view Get batch object using its unique execution identifier

Batch functions performed on the batch object

Once you have the batch object, use these public functions to interact with it.

Function Description Usage
start view Starts the execution of a batch scoring operation
cancel view Cancel the named batch execution
execution_id view Get the execution identifier for the named batch process
state view Poll for the state of the batch execution (failed, complete, ...)
results view Poll for batch execution results, partial or full results as defined
execution Get results for a given index row returned as an array
list_artifacts view List of every artifact files that was generated by this execution index
artifact view Print the contents of the named artifact file generated by the batch execution
download view Download one or all artifact files from execution index

1. Get the web service

Once you have authenticated, retrieve the web service from the server, assign it to a variable, and define the inputs to it as record data in a data frame, CSV, or TSV.

Batching begins by retrieving the web service containing the code against which you score the data records you define next. You can get a service using its name and version with the get_service function from azureml-model-management-sdk. The result is a service object.
The get_service function is covered in detail in the article "How to interact with and consume web services in R."

2. Define the data records to be batched

Next, use the public api function batch to define the input record data for the batch and set the number of concurrent threads for processing.

Syntax: batch(inputs, parallelCount = 10)

Argument Description
inputs Specify the data.frame name directly
parallelCount Default value is 10. Specify the number of concurrent threads that can be dedicated to processing records in the batch. Take care not to set a number so high that it negatively impacts performance.

Returns: The batch object

Example:

# -- Import the dataset from the microsoftml package
from microsoftml.datasets.datasets import get_dataset
mtcars = get_dataset('mtcars')

# -- Represent the dataset as a dataframe.
mtcars = mtcars.as_df()

# -- Define the data for the execution.
records = mtcars[['hp', 'wt']]

# -- Assign the record data to the batch service and set the thread count.
batch = svc.batch(records, parallelCount = 10)

3. Start, find, or cancel the batch execution

Next, use the public api functions to start the asynchronous batch execution on the batch object, monitor the execution, or even cancel it.

Start batch execution task

Start the batch task with start(). Machine Learning Server starts the batch execution and assigns an ID to the execution and returns the batch object.

Syntax: start()

No arguments.

Returns: The batch object

Example: batch = batch.start()

Note

We recommend you always use the id function after you start the execution so that you can find it easily with this id later such as: id = batch.execution_id print("Batch execution_id is: {}".format(id))


Get batch ID

Get the batch task's identifier from the service object so you can reference it during or after its execution using id().

Syntax: execution_id

No arguments.

Returns: The ID for the named batch object.

Example: id = batch.execution_id


### Get batch by ID

Retrieve the Batch based on an execution id.

Syntax: get_batch(execution_id)

Argument Description
execution_id The batch execution identifiers

Returns: The Batch object


List the Batch execution identifiers

Gets all batch executions currently queued for this service.

Syntax: list_batch_executions()

No arguments

Returns: List of batch execution identifiers

Example: list_batch_executions()


Cancel execution

Cancel the batch execution using cancel().

Syntax: cancel()

No arguments.

Returns: The batch object

Example: batch = batch.cancel()

4. Monitor, retrieve, and interact with results

While the batch task is running, you can monitor and poll the results. Once the batch task has completed, you can get the web service consumption output by index from the batch results object, including:

  • Monitor execution results and status
  • Get results for a given index row returned as an array
  • Get a list of every file that was generated by this execution index
  • Print the contents of a specific artifact or all artifacts returned
  • Download artifacts from execution index

Monitor execution results and status

There are several public functions you can use to get the results and status of a batch execution.

– Monitor or get the batch execution results

  • Syntax: results(showPartialResults = TRUE)

    Argument Description
    showPartialResults This argument returns the already processed results of the batch execution even if it has not been fully completed. If showPartialResults = FALSE, then returns only the results if the execution has completed.

  • Returns: A batch result object is returned, which in our example is called batchRes.
     

– Get the status of the batch execution.

  • Syntax: state
      no arguments
  • Returns: The status of the batch execution.
     

Example: In this example, we return partial results every three seconds until the batch execution fails or completes. Then, we return results for a given index row returned as an array.

batchRes = None
while(True):
    batchRes = batch.results()
    print(batchRes)
    if batchRes.state == "Failed":
        print("Batch execution failed")  
        break
    if batchRes.state == "Complete": 
        print("Batch execution succeeded")  
        break
    print("Polling for asynchronous batch to complete...")
    time.sleep(1)

Get list of generated artifacts

Retrieve a list of every artifact that was generated during the batch execution for a given data record, or index, with listArtifacts(). This function can be made part of a loop to get the list of the artifacts for every data record (see workflow example for a loop).

Syntax: list_artifacts(index)

Argument Description
index Index value for a given batch data record

Returns: A list of every artifact that was generated during the batch execution for a given data record

Example:

# List every artifact generated by this execution index for a specific row.
lst_artifact = batch.list_artifacts(1)
print(lst_artifact)

Display artifact contents

Display the contents of a named artifact returned in the preceding list with artifact(). Machine Learning Server returns the ID for the named batch object.

Syntax: artifact(index, fileName)

Argument Description
index Index value for a given batch data record
fileName Name of file artifact created during batch execution

Returns: The ID for the named batch object

Example:

# Then, get the contents of each artifact returned in the previous list.
# The result is a byte string of the corresponding object.
for obj in lst_artifact:
    content = batch.artifact(1, obj)

Download generated artifacts

Download any artifacts from a specific execution index using download(). By default, artifacts are downloaded to the current working directory getwd() unless a different dest = "<path>" is specified. You can choose to download a specific artifact or all artifacts.

Syntax: download(index, fileName, dest = "<path>")

Argument Description
index Index value for a given batch data record
fileName Name of specific artifact generated during batch execution. If omitted, all artifacts are downloaded for that index.
dest The download directory on your local machine. The default is the current working directory. The directory must already exist on the local machine.

Returns: The path to each downloaded artifact.

Example:

In this example, we download a named artifact for the first index to the current working directory.

batch.download(1, "answer.csv")

See also