Рударење на масивни податоци: Домашна работа 3

This project is designed to process health data using Spark Streaming, Kafka, and machine learning. The main pipeline reads health indicators from Kafka, makes predictions using a trained model, and sends the results back to Kafka.

Project Structure

.
├── checkpoints                           # Directory for storing Spark checkpoint data
├── data                                  # Data-related files and scripts
│   ├── get_data_script.sh                # Script to fetch data
│   └── split_data.py                     # Script to split data into train/test sets
├── infrastructure                        # Infrastructure configuration files
│   ├── kafka-docker-compose.yml          # Docker Compose for Kafka
│   └── spark-docker-compose.yml          # Docker Compose for Spark
├── model                                 # Model-related files and scripts
│   ├── best_model_script.py              # Script to find the best model
├── poetry.lock                           # Poetry lockfile for dependency management
├── pyproject.toml                        # Poetry configuration file
├── README.md                             # Main project documentation
├── main_shell_script.sh                  # Main project script, run this to build everything needed
└── scripts                               # Main scripts
    ├── consume_kafka_msgs.py             # Script to consume messages from Kafka
    ├── make_predictions_spark_script.py  # Main Spark script for predictions
    ├── produce_kafka_msgs.py             # Script to produce messages to Kafka
    └── start_stream.sh                   # Shell script to start the streaming pipeline

Usage Instructions

1. Setting up the Environment

• Use the provided Docker Compose files to set up Kafka and Spark:

  docker-compose -f infrastructure/kafka-docker-compose.yml up -d
  docker-compose -f infrastructure/spark-docker-compose.yml up -d

2. Preparing the Data

• Use data/get_data_script.sh to fetch the raw data.
• Split the data into training and testing data for the model (offline data) and data for making the predictions (online data) using data/split_data.py.

3. Training the Model

• Run model/best_model_script.py to train models and save the best one as model/model.pkl.
• This script trains 3 models (Logistic Regression, XGBoost and Random Forest) using 3 fold CV for finding the best model.

4. Making the predictions using Spark Structured Streaming

• Use scripts/make_predictions_spark_script.py script as a spark streaming job
• This script creates a Spark Session, reads the steaming data, adds a prediction column to each row and sends the resultant rows to another Kafka topic.

5. The whole streaming pipeline

• Produce messages to Kafka using scripts/produce_kafka_msgs.py.
• Run the Spark streaming script scripts/make_predictions_spark_script.py to process the data and produce predictions.
• Consume the prediction results from Kafka using scripts/consume_kafka_msgs.py.

Dependencies

• Python 3.10 or higher
• Poetry for dependency management
• Docker and Docker Compose
  • Apache Spark
  • Apache Kafka

Commands

• Generate the offline.csv and online.csv files:

  source data/get_data_script.sh

• Find the best model and save it as a pkl file:

  python model/best_model_script.py

• Produce the needed Kafka messages for making predictions:

  python scripts/consume_kafka_msgs.py

• Start the pipeline inside the spark container:

  source scripts/start_stream.sh

• View Kafka messages:

  python scripts/consume_kafka_msgs.py

Notes

• There is one script that runs everything needed to get the data, find the best model and make the streaming pipelines, as well as start the docker containers.

  source main_shell_script.sh

• Install the needed python packages using poetry
• Make sure all containers are running on the same Docker network for communication.
• Update Kafka broker addresses in the scripts as necessary.