Skip to content

An End-to-End ETL data pipeline that leverages pyspark parallel processing to process about 25 million rows of data coming from a SaaS application using Apache Airflow as an orchestration tool and various data warehouse technologies and finally using Apache Superset to connect to DWH for generating BI dashboards for weekly reports

License

judeleonard/Prescriber-ETL-data-pipeline

Repository files navigation

Prescriber-ETL-data-pipeline showing End-to-End implementation using Apache Airflow, pyspark and Apache superset to build BI dashboards

Project Overview

The dataset for this project contains the information about US Medical prescribers, their cities, years of experience, the cost for each prescribed medicine etc. The data that houses these information came in two format having a relation.The city dimensions table in parquet format and the fact table that comes in csv format containing the prescribers' information. The goal is to load these data using the spec format into spark and utilizing spark rdd to process and transform by extracting insight off this large dataset about 4GB in size.

Architecture - Prescriber pipeline

Technologies and Functions

  • Pyspark: Data preprocessing, cleaning & transformation
  • Python: data extraction & loading
  • Amazon s3: cloud storage as DWH
  • Azure blob: Cloud storage as DWH
  • Postgres: DWH on premise
  • Docker: Containerization
  • Apache Airflow: Orchestration tool and for running Cron jobs
  • Apache Superset: BI dashboard

Project structure

    ├── airflow_data
    │   └── airflow.cfg
    ├── dags
    │   ├── credentials
    │   │   └── variables.json
    │   ├── presc_pipeline.py
    │   └── __pycache__
    │       └── presc_pipeline.cpython-37.pyc
    ├── data
    │   ├── USA_Presc_Medicare_Data_2021.csv
    │   └── us_cities_dimension.parquet
    ├── docker-compose.yaml
    ├── Dockerfile
    ├── images
    │   ├── city-report-dashboard-2022-12-05T04-26-37.676Z.jpg
    │   ├── db1.png
    │   ├── db2.png
    │   ├── db3.png
    │   ├── pipeline_status.png
    │   ├── prescriber-dashboard-2022-12-04T21-49-29.517Z.jpg
    │   └── spark-architecture.jpeg
    ├── logs
    │   └── dag_processor_manager
    │       └── dag_processor_manager.log
    ├── plugins
    ├── requirements.txt
    ├── scripts
    │   ├── __init__.py
    │   ├── __pycache__
    │   │   └── validate.cpython-37.pyc
    │   ├── spark_etl.py
    │   └── validate.py
    ├── staging
    └── superset
        ├── city_report_dashboard.gif
        └── prescriber_report_dashboard.gif

ERD

This is my proposed data model prior to data cleaning and transformation to help understand the structure of the data. Also note that some of the fields are not in the right datatypes the ERD is to help figure out how to map and cast each fields to suit our requirement purposes.

Spark ETL components/job

  • City transformation / report:

    • filter records by exempting cities which do not have any subscriber assigned
    • calculate the number of zips in each city
    • calculate the total transaction in each city
    • calculate the number distinct prescribers assigned to each city
  • Prescriber transformation/ report:

    • apply filter to consider prescribers only from 20 - 50 years of experience
    • rank the prescribers based on their transaction count for each state
    • Select top 5 prescribers from each state

Pipeline runtime status -> executed successfully

pipeline is scheduled to run on weekly basis

Tables in postgres

Querying report tables

city_report

prescriber_report

Visualizing city report dashboard in Apache superset

Visualizing prescribers report in Apache superset

About

An End-to-End ETL data pipeline that leverages pyspark parallel processing to process about 25 million rows of data coming from a SaaS application using Apache Airflow as an orchestration tool and various data warehouse technologies and finally using Apache Superset to connect to DWH for generating BI dashboards for weekly reports

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published