From Data Source to Dashboard

I just finished a course in data engineering, and with it, my first DE project! Here’s an outline of what I did for the final project.

Huge thanks to the folks at Data Talks Club for putting the course together: Alexey Grigorev, Ankush Khanna, Sejal Vaidya, and Victoria Perez Mola.

My project in GitHub

Problem Description

For this project, I wanted to see what qualities of a song lead to popularity, and how that has changed over time. This knowledge is useful for people in the music industry, such as pop artists and producers. I used data from the Billboard Hot 100 (a weekly-updated list of top songs in the US), with records from 1964 to 2018.

One of the files, song_chart, lists the top 100 songs from each week. The other, acoustic_features, contains an entry for each individual song with fields describing its musicality. The following fields interested me the most:

1. key: A, A#, B, etc.
2. mode: Major or minor
3. danceability: “how suitable a track is for dancing based on a combination of musical elements including tempo, rhythm stability, beat strength, and overall regularity”
4. energy: “a perceptual measure of intensity and activity”
5. valence: “musical positiveness conveyed by a track”

Full descriptions of all fields are located in the ReadMe files, which are uploaded to GCS in the data ingestion step.

Ingestion to Data Lake

I used Terraform for IaC, and Airflow for workflow orchestration. My DAGs completed the following tasks:

1. Retrieve zip files from the data source
2. Extract desired CSVs and ReadMe files from the zip archives
3. Convert CSVs to parquet
4. Upload the files to GCS
5. Create external BigQuery tables from the parquet files (Note: song_chart became song_pop, and acoustic_features became song_feat)

At this point, I explored the data a bit with the BigQuery SQL workspace. I tried partitioning the song_pop table by week and the song_feat table by mode, but this did not improve query performance, so I decided not to move forward with that.

I made tables with the following SQL statements:

CREATE OR REPLACE TABLE `final-dtc-project.song_data.song_feature_data`
AS SELECT * FROM `final-dtc-project.song_data.external_table_song_feat`;

CREATE OR REPLACE TABLE `final-dtc-project.song_data.song_popularity_data`
AS SELECT * FROM `final-dtc-project.song_data.external_table_song_pop`;

Transformation in Data Warehouse

I was unable to fix the issues related to itegrating Spark and BigQuery, so I took the following steps to run a Spark job using Dataproc (ref: https://medium.com/google-cloud/apache-spark-and-jupyter-notebooks-made-easy-with-dataproc-component-gateway-fa91d48d6a5a):

1. Opened a terminal window in my GCP console, and enabled the required APIs with the following command:

gcloud services enable dataproc.googleapis.com \
  compute.googleapis.com \
  storage-component.googleapis.com \
  bigquery.googleapis.com \
  bigquerystorage.googleapis.com

2. Still in the terminal, I ran the following to commission a Dataproc cluster:

REGION=us-west1

CLUSTER_NAME=dproc-cluster

 gcloud beta dataproc clusters create ${CLUSTER_NAME} \
     --region=${REGION} \
     --master-machine-type e2-standard-2 \
     --worker-machine-type e2-standard-2 \
     --image-version 1.5-debian \
     --initialization-actions gs://dataproc-initialization-actions/python/pip-install.sh \
     --metadata 'PIP_PACKAGES=google-cloud-storage google-cloud-bigquery' \
     --optional-components=ANACONDA,JUPYTER \
     --enable-component-gateway

3. I then navigated to my Dataproc cluster in the console, went to “Web Interfaces,” clicked on Jupyter, and created a python3 notebook in the GCS folder.

In the notebook, I used Spark UDFs to transform the key and mode columns of the song feature dataframe to be more readable. I then used Spark SQL to filter the table of top 100 songs per week into just the number 1 song per week, and then left-joined this table with the table of song features.

Dashboard

Link to interactive dashboard: https://datastudio.google.com/reporting/b77b7528-3e73-48fa-8bee-dad34a9e8641

I used Google Data Studio to put together my dashboard. I chose to track average valence and danceability over time, because their relationship appears to have an interesting evolution. I find the most recent trend, from 2016-2018, to be particularly intruiging.

For the categorical variable, I chose key (note that the bar plot does not account for songs remaining in #1 for more than a week, so “Record Count” likely counts some songs multiple times). As someone with little knowledge of music theory, I’m curious as to why D# is so unpopular.