udacity-nd027-data_lake

Udacity Data Engeneering Nanodegree Program - My Submission of Project: Data Lake

Summary

The goal of this project is to define fact and dimension tables for a star schema for a particular analytic focus, and write an ETL pipeline that transfers data from sets in Amazon S3 into five S3 parquet files with Apache Spark (via pyspark).

Raw Data

The raw data is in Amazon S3 and contains

1. *song_data*.jsons ('artist_id', 'artist_latitude', 'artist_location', 'artist_longitude', 'artist_name', 'duration', 'num_songs', 'song_id', 'title', 'year') - a subset of real data from the Million Song Dataset.
2. *log_data*.jsons ('artist', 'auth', 'firstName', 'gender', 'itemInSession', 'lastName', 'length', 'level', 'location', 'method', 'page', 'registration', 'sessionId', 'song', 'status', 'ts', 'userAgent', 'userId') - simulated activity logs from a music streaming app based on specified configurations.

preliminaries

Make sure you have an AWS secret and access key and an own S3-Bucket as well. Then edit the dl.cfg file and replace the following XXX entries:

[AWS]
KEY=YOUR_AWS_KEY
SECRET=YOUR_AWS_SECRET

Afterwards edit etl.py and let output_data point to your S3 Bucket.

Run ETL Pipeline - Define Fact and Dimension Tables for the Raw Data

Run python3 etl.py in a terminal to import the raw sets from Udacity's S3 above and store data in the following star schema:

A Fact Table

1. songplays (songplay_id, start_time, user_id, level, song_id, artist_id, session_id, location, user_agent) - records in log data associated with song plays

Four Dimension Tables

1. users (user_id, first_name, last_name, gender, level) - users in the app
2. songs (song_id, title, artist_id, year, duration) - songs in music database
3. artists (artist_id, name, location, latitude, longitude) - artists in music database
4. time (start_time, hour, day, week, month, year, weekday) - timestamps of records in songplays broken down into specific units

1. Discuss the Purpose of this Database in the Context of the Startup, Sparkify, and their Analytical Goals.

The startup Sparkify is an audio streaming services provider. "As a freemium service, basic features are free with advertisements and limited control, while additional features, such as offline listening and commercial-free listening, are offered via paid subscriptions. Users can search for music based on artist, album, or genre, and can create, edit, and share playlists." (Taken from https://en.wikipedia.org/wiki/Spotify)

So the commercial goal is to get as many users as long as possible to use sparkify. In order to achieve this, sparkify must meet/exceed the users expectations and satisfy them. Ways of doing this are

• providing a huge database of artists and songs
• a fancy and usable (Browser, Mobile-App, Desktop-App) GUI
• a good search engine (Analytical goal!)
• a good recommendation system (Analytical goal!)

This database serves the needs of a good recommendation system: The favourite songs of user X can easily be extracted of our fact_table given their number_of_plays (by X). So we can put users in clusters based on their favourite songs. And if X likes the songs a, b, & c and there are other Users in (one of) his cluster(-s) which likes the songs a, b, c, & d he also might like song d. Let's recommend this song to him. He supposably enjoys that and thus enjoys sparkify.

2. State and Justify your Database Schema Design and ETL Pipeline.

Database Schema Design

The denormalized fact table songplays provides most information sparkify needs for its basic analytical goals. Reading and aggregation is very fast and if we need additional information from the dimension tables

• users
• artists
• songs (missing genre here 'tho)
• time

the joins are very simple -> high readability.

ETL Pipeline

The ETL Pipeline is very simple,

1. Raw data from S3 is loaded in staging tables staging_events and staging_songs in Spark.
2. Data from staging_songs is directly written into the dimension tables songs and artists and written to parquet files.
3. Data from staging_events is directly written into the dimension tables users and time and written to parquet files.
4. For the fact table songplays the staging_events are filtered by page == 'Nextsong'. Then both parquet files for songs and artists are read and joined by artist id, to get the artist name. After that, this dataframe is joined by staging_events on song title, artist name and song duration. The is joined by time table on start_time to append year and month before it eventually is written to a parquet file as well.