Temperature-Prediction

Weather Temperature prediction

This repository contains the final project for the MLOps Zoomcamp course provided by DataTalks.Club (Cohort 2024). The project consists of a Machine Learning Pipeline with Experiment Tracking, Workflow Orchestration, Model Deployment and Monitoring.

Project Description
Technologies
Project Architecture
Reproducibility
- 1- Local Setup
- 2- Cloud Setup
Further Improvements

Project description

For the Mlops_zoomcamp final project, i choose to implement a simple solution to forecast weather temperature. The main goal is to apply all the skills learnt from the course. So, the emphasis will not be made on the model and the accuracy. We use the data provided by Environment Canada's weather data https://climate.weather.gc.ca/. We only focus on the weather stations of the province of Ontario (canada). There is a jupyter notebook (Extract_station_data.ipynb) explaining how to extract the data related to a station (eg. station id, station name). There is also a python script (download_stations_id.py) to download the station data for all canadian provinces and save them in the folder data/stations/. The notebook Data_Extraction_and_Cleaning.ipynb shows how to download weather data for a given station.

Technologies

Name	Scope
Jupyter Notebook	Exploratory data analysis
Docker	Application containerization
Docker-Compose	Multi-container Docker applications definition and running
Mage.ai	Workflow orchestration and pipeline notebooks
MLFlow	Experiment tracking and model registry
PostgreSQL RDS	MLFLow backend entity storage and Terraform backend
Flask	Web server
EvidentlyAI	ML models evaluation and monitoring
pytest	Python unit testing suite
pylint	Python static code analysis
black	Python code formatting
isort	Python import sorting
Pre-Commit Hooks	Code issue identification before submission
GitHub Actions	CI/CD pipelines

Project architecture

1 Data preparation pipeline

We only used two features namely Date/Time (LST) and stationID and the goal is to predict Temp (°C).

2 Training pipeline

We used Xgboost as regression model. We track the experiment with mlfow and register the model having the best score. In the test_and_promote block, we compared the registred model with the previous and promote it to production stage if it is the best model.

3 Monitoring-pipleine

When a new data is available in the system, we run evidently to check any "Column Drift" or "Dataset Drift". When a drift is observed, we retraining the model with the old and new data.

4 Mage orchestration

We orchestrate our pipelines with Mage.ai

5 Model deployment

We choose a web service based on flask to deploy our model. Firstly, the code will fetch the model available in mlflow production stage and make a prediction. We containerize the code with docker and it can be deploy on AWS ECS by running in the terminal the following command: cd model_web_service followed by make scratch_deploy.

Note

We don't use the s3 as backend for terraform for the web service deployment. So the terraform state file will be stored locally

6 MLops best practices

We make for the web service:

unit test and integration test
pylint, black and isort were used to enhance code quality
make file is written to automatically run the code and deploy it.
pre-commit is also used

For the ocherstration in mage:

In the terminal of mage we run pylint, black and isort were used to enhance code quality.
make file is written to automatically run the code and deploy it.
CI/CD is written.

Reproducibility

Before anything:

change the name of dev.env in the main folder and in model_web_service/integration-tests/ by .env. In these files, change the values of the environment variables after the comment # ======= To change ========.
Put your AWS_SECRET in dev.secret and change the filename to .secret.
Remove the _ from the name of the file ._docker_password.txt and write in the file your docker password.

1 Local setup

Set up mlflow server on Aws EC2 by following this instruction with S3 bucket as artifact storage and RDS postgres as backend.

You can set up mlflow locally by:

installing first mlflow: pip install mlflow or conda install -c conda-forge mlflow.
running mlflow ui --backend-store-uri sqlite:///mlflow.db
Change in the file temperature_prediction/utils/logging.py:

DEFAULT_TRACKING_URI set to http://127.0.0.1:5000

In the main project directory, run make run-local to launch Mage.
You can leave the data present in data/Training/old/ and run the training pipeline. You can also delete these data and start from scratch by running the data preparation pipeline following by the training pipeline. For the ochestration, i suggest to follow my setting:
- run the data preparation pipeline eah day at any hour of your choice
- run the retrain pipeline 30 min after the time you set for the data preparation pipeline.
In the terminal cd model_web_service and make run-local to launch the web-service.
To get a new prediction run:

curl --header "Content-Type: application/json" \
  --request POST \
  --data '{"Station ID": 50840, "Date/Time (LST)": "2024-08-09 23:00:00"}' \
  http://localhost:9696/predict`

2 Cloud setup

Set up mlflow server on Aws EC2 by following this instruction with S3 bucket as artifact storage and RDS postgres as backend.
In the main project directory, run make run-local to launch Mage and run the pipeline deploying_to_production. This will create two users MageDeployer and MageContinuousIntegrationDeployer. The first will have the grants to deploy Mage the second one will be used in CI/CD configuration. Running the pipeline will deploy mage on aws ECS but it will not contain your project. By committing the project on github it'll run the github actions and put the new image Aws ECR. Normally, all the pipelines should be there.

Warning

Unfortunately, the pipelines don't appear in the mage (for me). But we can clone the repository in the control version panel.

Another way of publishing Mage, is to run the block permissions of the pipeline deploying_to_production. It will create the user MageDeployer. Use its credentials in the .env and .secret and then run make scratch_deploy. Use the load balancer DNS given to load mage.

To deploy the web service, in the terminal, cd in cd model_web_service and run make scratch_deploy.
To get a new prediction run:

curl --header "Content-Type: application/json" \
  --request POST \
  --data '{"Station ID": 50840, "Date/Time (LST)": "2024-08-09 23:00:00"}' \
  http://the_balancer_output/predict`