Data Engineering

This project implements a fully containerized data engineering workflow for downloading, ingesting, transforming, validating, and preparing NYC Yellow Taxi Trip Data. It uses Apache Spark for computation, Apache Airflow for orchestration, MinIO as Amazon S3-compatible object storage, and DuckDB as an analytics warehouse. The entire environment is reproducible using Docker Compose and set to run locally.

1. Project Overview

The workflow simulates a modern data engineering architecture with dedicated layers:

  1. Local Layer – Raw TLC data downloaded directly to the local filesystem.
  2. Landing Layer (MinIO/S3) – Central storage for raw but structured data.
  3. Prepared Layer (MinIO/S3) – Cleaned, filtered and transformed datasets generated via Spark Pipeline.
  4. Warehouse Layer (DuckDB) – Analytical storage loaded from the Prepared layer.
  5. EDA Layer (JupyterLab) – Data exploration and downstream modeling using notebooks.

2. Repository Structure

├── airflow
│   ├── dags
│   │   └── taxi_pipeline_dag.py
│   ├── Dockerfile
│   └── requirements.txt
├── config
│   └── storage_config.yaml
├── data
│   └── yellow_tripdata_2024-01.parquet
├── infra
│   └── docker
│       ├── notebook.Dockerfile
│       └── spark.Dockerfile
├── notebooks
│   ├── 01_raw_data_exploration.ipynb
│   └── 02_prepared_data_exploration.ipynb
├── spark_jobs
│   ├── ingest_landing.py
│   └── transform_prepared.py
├── warehouse
│   └── taxi.duckdb
├── README.md
├── docker-compose.yml
├── LICENSE
└── requirements.txt

3. Dataset

Source and availability

This project is configured and tested for years up to and including 2024 (e.g., 2024-01). The TLC continues to publish newer months/years, but schemas may change over time (for example, additional columns may appear), so treat 2025+ as “use at your own risk” unless you update your transformations accordingly.

Manual download [optional]

mkdir -p data

# Example month (2024-01)
curl -L "https://d37ci6vzurychx.cloudfront.net/trip-data/yellow_tripdata_2024-01.parquet"   -o data/yellow_tripdata_2024-01.parquet

# Taxi zones
curl -L "https://d37ci6vzurychx.cloudfront.net/misc/taxi_zone_lookup.csv"   -o data/taxi_zone_lookup.csv

Raw trip record columns

Column Type Description
VendorID int TPEP provider code (vendor).
tpep_pickup_datetime datetime Trip meter start timestamp.
tpep_dropoff_datetime datetime Trip meter stop timestamp.
passenger_count float/int Passenger count (driver-entered).
trip_distance float Trip distance in miles (taximeter).
RatecodeID float/int Final rate code for the trip.
store_and_fwd_flag string “Store and forward” indicator.
PULocationID int Pickup TLC Taxi Zone ID.
DOLocationID int Drop-off TLC Taxi Zone ID.
payment_type int Payment type code.
fare_amount float Fare amount.
extra float Extras/surcharges (varies).
mta_tax float MTA tax.
tip_amount float Tip amount.
tolls_amount float Tolls amount.
improvement_surcharge float Improvement surcharge.
total_amount float Total amount charged.
congestion_surcharge float Congestion surcharge.
Airport_fee float Airport fee (may be null / not present in older years).

Note: the TLC schema can vary slightly across years.

Taxi zone lookup columns

Column Type Description
LocationID int Taxi Zone ID used by TLC trip files.
Borough string Borough name.
Zone string Zone name.
service_zone string Service zone label (e.g., “Yellow Zone”, “Boro Zone”).

4. System Architecture

Component overview

Technologies

Component Responsibility
Airflow Orchestrates the ETL pipeline end-to-end.
Spark Performs ingestion and transformation into Landing and Prepared layers.
MinIO S3-compatible storage for Landing and Prepared zones.
DuckDB Analytical warehouse loaded from the Prepared layer.
JupyterLab Exploration and EDA environment.
Docker Compose Orchestrates and isolates the entire system.

High-level flow

Architecture

5. Data Layouts in MinIO

Landing zone

s3://<bucket>/landing/taxi/year=YYYY/month=MM/
    part-*.parquet

s3://<bucket>/landing/reference/taxi_zones/
    part-*.parquet

Prepared zone

Prepared trips (enriched with zone attributes; partitioned by day):

s3://<bucket>/prepared/taxi/year=YYYY/month=MM/
    pickup_date=YYYY-MM-DD/
        part-*.parquet

Both paths are dynamically parameterized with year and month, which are provided by Airflow at runtime.

6. Docker Services

Airflow

Spark

MinIO

JupyterLab

7. Airflow Pipeline

The DAG is defined in:

airflow/dags/taxi_pipeline_dag.py

Tasks

  1. download_taxi_data

    • Downloads the TLC Parquet file for a given year / month into data/.
    • URL pattern: https://d37ci6vzurychx.cloudfront.net/trip-data/yellow_tripdata_YYYY-MM.parquet
    • Downloads the Taxi Zone Lookup CSV into data/ (only once, not partitioned): https://d37ci6vzurychx.cloudfront.net/misc/taxi_zone_lookup.csv

  2. ingest_to_landing

    • Runs ingest_landing.py on Spark with --year and --month.
    • Reads local Parquet from data/.
    • Writes raw structured data into MinIO Landing:
      • Taxi Zones (only once, not partitioned): 
        s3a://<bucket>/landing/reference/taxi_zones/
      • Monthly Trips: 
        s3a://<bucket>/landing/taxi/year=YYYY/month=MM/

  3. transform_to_prepared

    • Runs transform_prepared.py on Spark with --year and --month.
    • Reads Landing data for the selected month.
    • Cleans and enriches the data.
    • Joins the Taxi Zone Lookup Table twice:
      • PULocationID → pickup zone attributes (PU_*)
      • DOLocationID → drop-off zone attributes (DO_*)

    • Writes the output to the Prepared layer (partitioned by pickup_date):

      s3a://<bucket>/prepared/taxi/year=YYYY/month=MM/

  4. load_to_duckdb

    • Uses DuckDB to read the Prepared Parquet for the selected month.

    • Creates schema and table if needed:

      CREATE SCHEMA IF NOT EXISTS taxi.taxi;
CREATE OR REPLACE TABLE taxi.taxi.trips_prepared AS
SELECT * FROM read_parquet('<prepared_glob>');

    • The taxi.taxi.trips_prepared table is overwritten for each pipeline run. After each run, it contains only the data for the triggered year/month.

Triggering the DAG

The DAG is manual only (schedule=None).

In the Airflow UI:

  1. Open the nyc_taxi_pipeline DAG.
  2. Click Trigger DAG.
  3. Provide the run configuration as JSON:
{
  "year": 2024,
  "month": 1
}

If year/month are omitted, the DAG falls back to its default params (year=2024, month=1).

Trigger DAG

8. Spark Jobs

spark_jobs/ingest_landing.py

Responsibilities:

spark_jobs/transform_prepared.py

Responsibilities:

9. Analytics Schema

The Prepared layer (and thus taxi.taxi.trips_prepared) contains:

Column Description
pickup_date Date derived from tpep_pickup_datetime.
pickup_hour Hour-of-day derived from tpep_pickup_datetime.
dropoff_date Date derived from tpep_dropoff_datetime.
trip_duration_min Trip duration in minutes.
avg_mph Average trip speed in mph (distance / duration).
PU_Borough Borough for pickup zone.
PU_Zone Zone name for pickup zone.
PU_service_zone Service zone label for pickup zone.
DO_Borough Borough for drop-off zone.
DO_Zone Zone name for drop-off zone.
DO_service_zone Service zone label for drop-off zone.

10. Notebooks

This project includes two Jupyter notebooks that support exploratory analysis of both the raw TLC data and the prepared data.

10.1 Raw Data Exploration

File: notebooks/01_raw_data_exploration.ipynb

10.2 Prepared Data Exploration

File: notebooks/02_prepared_data_exploration.ipynb

11. Environment Variables

All secrets and configuration are driven through a .env file in the project root.

The repository ships an .env.example file as a starting point. You must create your own .env and fill in values, including AIRFLOW_FERNET_KEY (which is not committed).

Required Variables

MINIO_ROOT_USER=...
MINIO_ROOT_PASSWORD=...
AWS_ACCESS_KEY_ID=...
AWS_SECRET_ACCESS_KEY=...
S3_BUCKET_NAME=...
S3_ENDPOINT_URL=http://minio:9000

AIRFLOW_ADMIN_USER=...
AIRFLOW_ADMIN_PASSWORD=...
AIRFLOW_ADMIN_EMAIL=...

AIRFLOW_FERNET_KEY=...   # must be generated locally (see below)

TZ=UTC

The Fernet key is critical for Airflow to securely store connections and variables. Please follow the next section to generate it and set it in your .env.

12. Generating AIRFLOW_FERNET_KEY

Airflow uses a Fernet key to encrypt sensitive values (e.g., connections, variables). You must generate a key and set AIRFLOW_FERNET_KEY in your .env.

From the project root:

docker compose run --rm airflow \
  python -c "from cryptography.fernet import Fernet; print(Fernet.generate_key().decode())"

Copy the printed value and set it in .env:

AIRFLOW_FERNET_KEY=<paste-generated-key-here>

Option 2: Generate using local Python

If your host Python has cryptography installed:

python -c "from cryptography.fernet import Fernet; print(Fernet.generate_key().decode())"

13. Setup and Execution

Step 1 — Prepare environment

git clone https://github.com/Yarroudh/dataengineering
cd dataengineering
cp .env.example .env

Edit .env and fill in:

Step 2 — Build services

docker compose build

Step 3 — Start services

docker compose up -d

Step 4 — Access UIs

Service URL
Airflow http://localhost:8080
MinIO Console http://localhost:9001
JupyterLab http://localhost:8888

Step 5 — Run the pipeline via Airflow

  1. Open Airflow at http://localhost:8080.

  2. Log in using the admin credentials from .env.

  3. Open the nyc_taxi_pipeline DAG.

  4. Click Trigger DAG.

  5. Provide year/month, for example:

    {
  "year": 2024,
  "month": 1
}

  6. Monitor tasks:

    • download_taxi_data
    • ingest_to_landing
    • transform_to_prepared
    • load_to_duckdb

After a successful run, DuckDB (warehouse/taxi.duckdb) contains taxi.taxi.trips_prepared for the selected month.

Step 6 — Explore data in JupyterLab

  1. Open http://localhost:8888.

  2. Run:

    • notebooks/01_raw_data_exploration.ipynb (raw profiling)
    • notebooks/02_prepared_data_exploration.ipynb (prepared data)

14. Troubleshooting

Containers not starting

docker compose logs <service>

Airflow cannot parse the DAG

Spark cannot access MinIO

15. Future Enhancements

16. License

This project is distributed under the terms defined in the LICENSE file.

17. Acknowledgements