Skip to content

Latest commit

 

History

History
314 lines (259 loc) · 16.2 KB

README.md

File metadata and controls

314 lines (259 loc) · 16.2 KB

AI for Earth - Creating APIs

These images and examples are meant to illustrate how to build containers for use in the AI for Earth API system.

Notice

Additional to a running docker environment, GPU images require NVIDIA Docker package to support CUDA.

CUDA Toolkit

To view the license for the CUDA Toolkit included in the cuda base image, click here

CUDA Deep Neural Network library (cuDNN)

To view the license for cuDNN included in the cuda base image, click here

Contents

  1. Repo Layout
  2. Quickstart
  3. Grantee Onboarding Procedure
  4. Contributing

Repo Layout

  • Containers
    • base-py [Base AI for Earth Python image]
      • nvidia/cuda:9.2 ubuntu 16.04 base
      • supervisor
      • miniconda3 4.5.4
      • Python 3.6.6
      • uwsgi
      • flask and flask-restful
      • azure-storage-blob
      • applicationinsights
      • AI for Earth task management tools
      • AI for Earth Azure SAS Blob tools
    • base-r [Base AI for Earth R image]
      • nvidia/cuda:9.2 ubuntu 16.04 base
      • supervisor
      • Python 3
      • Microsoft R Open 3.4.3
      • sp
      • gdal
      • plumber
      • future
      • reticulate
      • azure-storage-blob
      • applicationinsights
      • AI for Earth task management tools
      • AI for Earth Azure SAS Blob tools
    • blob-py [Base AI for Earth Python image with Azure Blob mounting tools]
      • AI for Earth base-py base
      • AI for Earth Azure Blob mounting tools
    • blob-r [Base AI for Earth R image with Azure Blob mounting tools]
      • AI for Earth base-r base
      • AI for Earth Azure Blob mounting tools
  • Examples

Notes

  • Docker commands for the base and blob images must be run at the version level of the repo. Ex. docker build . -f base-py/Dockerfile. Example Docker commands can be run within the example codebase.

Quickstart Tutorial

A quickstart tutorial for converting a trained machine learning model to an API can be found at Quickstart.md.

Grantee Onboarding Procedure

AI for Earth Grantees may onboard their models/code to Azure and, in most cases, surface their solution as a private API or an AI for Earth API. This section describes the fastest path to obtaining this result.

Step 1: Identifying which container works with your technology stack

Currently, we offer two base image platform choices: Python and R and have an optional Azure Blob mounting capability available.

If you require an image with a different technology stack, we'd love to hear from you so we can expand our offerings.

Step 2: Build an example service

To demonstrate how easy it is to run an API service in Azure, we'll explore the base-py example.

Orientation

Under the examples/base-py directory, we find these objects:

  • my_api folder [Contains the API service files.]
  • Dockerfile [A Dockerfile is a text document that contains all the commands a user could call on the command line to assemble an image. This Dockerfile contains all the commands needed to run the example API service.]
  • LocalForwarder.config [Application Insights tracing configuration. Replace your_key_goes_here with your actual key.]
  • startup.sh [A simple bash script that is used as the service's entrypoint. We use a bash script as it allows us to easily run other required commands without having to expand the Dockerfile. See the blob-mount-py/startup.sh as an expanded example.]
  • supervisor.conf [We use supervisor as our process control system. If your API service crashes, etc., supervisor will restart it for you. supervisor.conf holds the configuration necessary to execute supervisord, in the context of your API service.]

We explore the API file components via an outside-in approach, starting with the Dockerfile.

If you are unfamiliar with Docker or container technologies, we encourage you to review the Docker Get Started documentation.

Dockerfile walkthrough:

FROM mcr.microsoft.com/aiforearth/base-py:latest
  • The very first line of our Dockerfile describes the base image, upon which we'll build our API service environment. In general, you may use any base image you would like, but if you wish to be considered as an AI For Earth API, you must use one of the official AI for Earth base images. If you find that the provided base images are not sufficient, please connect with us so we may expand our offerings.
RUN echo "source activate ai4e_py_api" >> ~/.bashrc \
    && conda install -c conda-forge -n ai4e_py_api numpy pandas
  • All of the default packages have been installed within the Anaconda virtual environment: ai4e_py_api. Any additional necessary packages should be installed within this environment. This example demonstrates how to install numpy and pandas from the conda-forge package repository.
COPY ./my_api /app/my_api/
COPY ./supervisord.conf /etc/supervisord.conf
COPY ./startup.sh /
RUN chmod +x /startup.sh

COPY ./LocalForwarder.config /lf/
  • Copy the API service directory into the container (your API service code will be placed in the /app/my_api directory).
  • Copy the supervisor configuration file to the proper location.
  • Copy the startup.sh bash script and add execute permission to it.
ENV APPINSIGHTS_INSTRUMENTATIONKEY ''
ENV APPINSIGHTS_LIVEMETRICSSTREAMAUTHENTICATIONAPIKEY  ''
ENV LOCALAPPDATA '/app_insights_data'
ENV OCAGENT_TRACE_EXPORTER_ENDPOINT 'localhost:55678'
ENV SERVICE_OWNER "AI4E_Test"
ENV SERVICE_CLUSTER "Local Docker"
ENV SERVICE_MODEL_NAME "base-py example"
ENV SERVICE_MODEL_FRAMEWORK "Python"
ENV SERVICE_MODEL_FRAMEOWRK_VERSION "3.6.6"
ENV SERVICE_MODEL_VERSION "1.0"
  • All logging and metric collection flows through Application Insights. Set the APPINSIGHTS_INSTRUMENTATIONKEY environment variable to your instrumentation key. This is required when offering an official AI for Earth API, but is highly suggested otherwise. If you wish not to use Application Insights, you can safely remove this line. All of the SERVICE_ prefixed environment variables offer logging filters within Application Insights.
HEALTHCHECK --interval=1m --timeout=3s --start-period=20s \
  CMD curl -f http://localhost/ || exit 1
  • The HEALTHCHECK will make a request to the / endpoint to ensure that the service is running. This is used by Docker to determine if the service is healthy.
EXPOSE 80
ENTRYPOINT [ "/startup.sh" ]
  • Expose the port that you wish to use to call your API.
  • Specify the entrypoint, or the file to execute when starting the container.
ENV API_PREFIX=/v1/my_api/tasker
  • Add a URL prefix that your API will use.

The supervisor.conf file contains a number of lines, but there are two that are important to execute your API service code.

directory=/app/my_api/

  • If you modify the Dockerfile to place your code in a different directory than /app/my_api/, you must change it here.

Next, we look at the "command" specification. This is a single line, but we'll break it apart for exploration. We example the important parts of the command, below: /usr/local/envs/ai4e_py_api/bin/uwsgi

  • uwsgi is used as our application server. We point to its executable as the command. If you change the Anaconda virtual environment, you must change this line.

--virtualenv /usr/local/envs/ai4e_py_api

  • Anaconda virtual environment declaration.

--callable app

  • This your application name. Default is "app".

--http 0.0.0.0:80

  • Specifices which port to use. Your API service will listen to this port. This port must be exposed in your Dockerfile.

--wsgi-file /app/my_api/runserver.py

  • This is the location file in your API service code that contains your 'app' definition: app = Flask(__name__)

This is an example entrypoint to an API service. We explore the important lines, below:

from ai4e_service import AI4EService
  • The AI for Earth api toolset contains service (AI4EService) wrapper that includes:
    • A task manager for use in long-running/async APIs services. The task manager stores state within the container, itself. Therefore, it is not a production-ready task manager. The task manager has been designed, however, to be directly pluggable into AI for Earth's distributed task manager. If you are designing your API for eventual adoption into the AI for Earth official APIs, utilizing this task manager will allow you to test your API and allow you to migrate to the AI for Earth distributed architecure without any code changes.
    • An integrated health check endpoint for Docker to determine the health of the service.
    • A request validation function that is called on each request, before your code is called. This function checks to ensure that the service is not being terminated, that the max connections has not been reached, that the content type in the request is expected, and that the request size doesn't exceed the max request size. If any of these conditions are discovered, a 503 will be sent to the caller.
    • Two aspect-oriented function wrappers: async and sync, which, when applied to your function, turns the function into a callable API.
from ai4e_app_insights_wrapper import AI4EAppInsights
  • We have built a wrapper around the Application Insights Python API that allows us to quickly adopt your APIs into our official API platform (without any code changes). If you wish to explore this option, please use this wrapper.
log = AI4EAppInsights()
  • Initialize the AI for Earth Application Insights class, which you will use to send log messages to Application Insights.
with app.app_context():
    ai4e_service = AI4EService(app, log)
  • Creates the AI4EService, which we defined above.
def process_request_data(request):
    return_values = {'data': None}
    try:
        # Attempt to load the body
        return_values['data'] = request.data
    except:
        log.log_error('Unable to load the request data')   # Log to Application Insights
    return return_values
  • This is a request data processing function, which is run before your model code. Use this function to validate any input and to assign the input data to a dictionary.
def run_model(taskId, body):
    # Update the task status, so the caller knows it has been accepted and is running.
    ai4e_service.api_task_manager.UpdateTaskStatus(taskId, 'running model')

    log.log_debug('Running model', taskId) # Log to Application Insights
    #INSERT_YOUR_MODEL_CALL_HERE
    sleep(10)  # replace with real code
  • This is the method that actually runs your service code. Notice that we wrap the "real code" section with status updates. This allows the user to request status updates from the task GET endpoint - included via AI for Earth task management.
@ai4e_service.api_async_func(
    api_path = '/', 
    methods = ['POST'], 
    request_processing_function = process_request_data, # This is the data process function that you created above.
    maximum_concurrent_requests = 5, # If the number of requests exceed this limit, a 503 is returned to the caller.
    content_types = ['application/json'],
    content_max_length = 1000, # In bytes
    trace_name = 'post:my_long_running_funct')
def default_post(*args, **kwargs):
    # Since this is an async function, we need to keep the task updated.
    taskId = kwargs.get('taskId')
    log.log_debug('Started task', taskId) # Log to Application Insights

    # Get the data from the dictionary key that you assigned in your process_request_data function.
    request_data = kwargs.get('data')

    if not request_data:
        ai4e_service.api_task_manager.FailTask(taskId, 'Task failed - Body was empty or could not be parsed.')
        return -1

    # Load the request data into JSON format.
    request_json = json.loads(request_data)

    # Run your model function
    run_model(taskId, request_json)

    # Once complete, ensure the status is updated.
    log.log_debug('Completed task', taskId) # Log to Application Insights
    # Update the task with a completion event.
    ai4e_service.api_task_manager.CompleteTask(taskId, 'completed')
  • Set up a POST, long-running, endpoint for your API.
@ai4e_service.api_sync_func(api_path = '/echo/<string:text>', methods = ['GET'], maximum_concurrent_requests = 1000, trace_name = 'get:echo', kwargs = {'text'})
def echo(*args, **kwargs):
    return 'Echo: ' + kwargs['text']
  • Set up a GET, sync endpoint.

Step 3: Run a service locally

Now that you have your container code completed, we can build your container and run it locally.

Build an image

From within the context of your service directory (eg. ./examples/base-py), execute the command to build your image.

docker build . -t <your_ACR_login_server>/<your_service_name>:<build_number>

So if we were to build our base-py example for use in the AI for Earth ACR:

docker build . -t ai4eregistry.azurecr.io/grantee-py-example:1

Run your container

docker run -p <host port>:<container port> "<your_ACR_login_server>/<your_service_name>:<build_number>"

base-py example:

docker run -p 8081:80 "ai4eregistry.azurecr.io/grantee-py-example:1"

Notice that, in the above example, we map our computer's port 8081 to the container's port 80. This reflects the container port we exposed in our Dockerfile and defined in the command in our supervisord.conf file.

If you need to log into the image while it's running to debug, etc.:

docker ps  #shows you the running instances, find the CONTAINER ID of your container.
docker exec -it <container id> bash

Step 4: Run a service in Azure

Publish your image

To properly pull your image from within Kubernetes or Azure Container Instances, we must publish the image to a container registry. To get started using Azure Container Registry, please read this quickstart.

docker push <your_ACR_login_server>/<your_service_name>:<build_number>

base-py example:

docker push ai4eregistry.azurecr.io/grantee-py-example:1

Create and run your container

There are several ways to run your container in Azure. The easiest and fastest way to get up-and-running is to utilize Azure Container Instances (ACI). ACI gives us a few ways to deploy our container:

  1. Via the Azure CLI
  2. Via the Azure Portal
  3. Via PowerShell

Contributing

This project welcomes contributions and suggestions. Most contributions require you to agree to a Contributor License Agreement (CLA) declaring that you have the right to, and actually do, grant us the rights to use your contribution. For details, visit https://cla.microsoft.com.

When you submit a pull request, a CLA-bot will automatically determine whether you need to provide a CLA and decorate the PR appropriately (e.g., label, comment). Simply follow the instructions provided by the bot. You will only need to do this once across all repos using our CLA.

This project has adopted the Microsoft Open Source Code of Conduct. For more information see the Code of Conduct FAQ or contact [email protected] with any additional questions or comments.