A comprehensive guide to ML App Deployment using Flask
Machine Learning (ML) models typically leverage the capability to learn patterns from previously seen (training) data and apply them to predict the outcome for new (test) data. Deploying ML models as web apps can help test the efficacy of trained ML models by allowing access to multiple users and testing environments, thereby gathering test performance metrics. However, ML web app deployment in production can be a highly complex process that may need to ensure minimal down-time for users in the event of app update processes. Cloud-based deployment solutions such as Google Cloud Platform (GCP) have highly simplified the process of continuous integration and continuous deployment (CI/CD) through pipelines and triggers that can be designed to ensure sanity checks as well as the integrity of the integrated code base for updated application runs.
In this tutorial, we will review a detailed example wherein we deploy a ML model web app on GCP through a CD pipeline. To follow the steps from data analysis to final app deployment, fork the github repository at  and follow the steps below.
Step 1: Data and ML Model Description
The first step is understanding the data followed by ML model creation and subsequent deployment. For this step we refer to the file model.py under the folder app_files.
For this tutorial, we will be using the 50_Startup.csv data from  as shown in Fig. 1, where the goal is to train a regression model (f(X)) using the three features (x₁,x₂,x₃) as (R&D Spend, Administration Cost, Marketing Cost), and predict the company ‘Profit’ (Y).
Our goal is to estimate Profit as a function of its features as f(x)=g(x₁,x₂,x₃), where g(.) represents linear or non-linear combinations of the features x₁, x₂, x₃, respectively, that can be estimated from the training data.
While Linear Regression is the simplest solution to the above data set, one of the most scalable data models for large data sets continues to be ensemble models that are based on Decision Trees, since they enable data aggregation and expendability. In this work, we implement a non-linear Decision Tree-based Regression model as shown in Fig 2. To train and visualize the regression model, run the file model.py with code snippet shown below.
A Decision Tree Regressor proceeds training by partitioning sub-regions or blocks in the feature subspace such that similar samples flock together to minimize the residual sum of squares (RSS) for the estimate . At test time, each sample is projected onto the partitioned feature subspace followed by reading out the mean sample value from the respective sub-region as the estimated outcome . The non-linear nature of the trained decision tree regression model with a maximum depth of 5 is shown in Fig. 2. The code to visualize the trained model across each feature is shown below.
Step 1 Command at terminal: python model.py
Step 2: ML Web App using Flask
Having built the ML model, the next step is to package it and serve as a Web App. Here, we refer to the file app.py and html files under the templates folder contained in app_files.
We utilize Flask-based RESTful api from . REpresentational State Transfer (REST) is a standard architectural design for web APIs and Flask has been the choice for python web API serves since its inception in 2010. The two major components for designing the Flask-API are :
- Flask library and POST HTTP method to serve the predictions.
- Front end HTML files that are useful for users to interact with the ML model.
One significant change in the app.py file from general Flask-based python deployments in [4–5] is that the running app.py launches an API that runs at http://0.0.0:8000 instead of its default settings.
This change is made to enable a Google Cloud Run (GCR) that listens at http://0.0.0, port 8080. For details on ways to debug a GCR app see .
Step 2 Command at terminal: python app.py
Step 3: Create a Docker Container for Google Cloud
Now that we have a functional web app, we need to package it for a Google Cloud Run. For this purpose, we will create a Docker container, as in  that is a standalone software and can be installed on any physical or virtual machine to run containerized applications. The Docker file for this tutorial, as shown below, must include 4 key components.
- The operating system needed by the application;
- List of dependencies such as libraries and packages in the requirements.txt file;
- Application files;
- Command to launch the application.
Step 4: Create a YAML file for the Cloud Run
The final component required to run the CD pipeline is a YAML Ain’t Markup Language file based on the application in . YAML files follow the data serialization standard to build configuration settings for a CD pipeline. Details on building a YAML file for your application can be found at . The YAML file for this tutorial shown below contains 4 steps.
- If the docker pull request fails, then exit 0 will be executed, i.e. no error message will be displayed and build process will continue with the last build.
- Create a build if none exists so far, otherwise use the cache build.
- Push the docker image to GCR
- Deploy the application to GCR.
Now that we have all the components to run and deploy the ML web app, lets create a trigger to deploy the app on Google Cloud.
Step 5: Create the CD Pipeline and Run
Now that we have an app that runs locally, the next task is to set up a project in GCP that will generate triggers and the cloud run. Follow the steps below to ensure you have a viable Google cloud project to begin the CD process.
- Sign up at Google Console: https://console.cloud.google.com/home/
- Create a new Project as shown below:
3. Ensure ‘Billing’ is activated for the project.
4. Start from Github code to create a CD pipeline. Use the forked Github repository from your Github account for this purpose.
5. Go to the page: https://console.cloud.google.com/cloud-build/triggers. Ensure your project is selected on the top panel. Next click ‘Create Trigger’ button at the bottom as shown below.
6. This will take you to the following page where you will need to create a ‘Push to a branch’ event for your forked Github repository. To deploy the app, select the ‘Cloud Build’ configuration file option as shown below and hit ‘Create’ at the bottom.
7. Now a trigger gets created. Hit RUN option next to the trigger and look at the logs that get generated. This would take a few minutes to run and deploy. At the end you should have something that looks like below.
8. Now go to the GCR console at: https://console.cloud.google.com/run and notice your app (flaskappml) running as shown below:
Clicking on ‘flaskappml’ takes you to the GCR page with the URL displayed on top right. This is the URL to your deployed webapp that you can now test on any device.
Don’t forget to clean up the triggers and the GCR before your leave. This is crucial to ensure appropriate resource usage.
Building ML-models and deploying them on cloud platforms not only enhances usability, but it also helps gather insights into the limiting conditions for the model. To ensure robust deployed web apps, CI pipelines enable code to be committed to a repository in a systematic manner while CD pipelines enable code sanity triggers and checks to ensure successful app deployment. Extending this tutorial to other web apps is a relatively simple process. This will require the following 4 major changes:
- Updating the app_files to reflect your application of interest.
- Updating the Dockerfile and requirements.txt files to represent the libraries needed for your application.
- Updating the YAML file to include more DevOps test cases.
- Committing all updated code as a continuous integration pipeline to a Github or Git repo.
Now you have the resources needed to build CI/CD pipelines for your own ML-based use cases.
 S. Roychowdhury. Flask ML Model CD Pipeline Tutorial [Online]: https://github.com/sohiniroych/Flask-ML-Pipeline_GCP-Tutorial/
 K. Veerakumar. Startup — Multiple Linear Regression. [Online]:https://www.kaggle.com/karthickveerakumar/startup-logistic-regression
 L. Li. Classification and Regression Analysis with Decision Trees [Online]: Classification and Regression Analysis with Decision Trees | by Lorraine Li | Towards Data Science
 A. Jaleel. Startup:REST API with Flask [Online]: https://www.kaggle.com/ahammedjaleel/startup-rest-api-with-flask
 M. Grinberg. Designing a RESTful API with Python and Flask [Online]: https://blog.miguelgrinberg.com/post/designing-a-restful-api-with-python-and-flask
 Google Cloud. Trouble shooting. [Online] https://cloud.google.com/run/docs/troubleshooting
 J. Araujo. Machine Learning deployment pipeline on Google Cloud Run [Online]: https://github.com/jgvaraujo/ml-deployment-on-gcloud
Google Cloud. app.yaml Reference [Online]: https://cloud.google.com/appengine/docs/standard/python/config/appref
A Hands-on Tutorial to Continuous Deployment Pipelines for ML-based Web Apps on Google Cloud was originally published in Towards Data Science on Medium, where people are continuing the conversation by highlighting and responding to this story.