Pipeline Architecture

Pipeline Consists of various modules:


Data is captured in real time from the goodreads API using the Goodreads Python wrapper (View usage – Fetch Data Module). The data collected from the goodreads API is stored on local disk and is timely moved to the Landing Bucket on AWS S3. ETL jobs are written in spark and scheduled in airflow to run every 10 minutes.

ETL Flow

  • Data Collected from the API is moved to landing zone s3 buckets.
  • ETL job has s3 module which copies data from landing zone to working zone.
  • Once the data is moved to working zone, spark job is triggered which reads the data from working zone and apply transformation. Dataset is repartitioned and moved to the Processed Zone.
  • Warehouse module of ETL jobs picks up data from processed zone and stages it into the Redshift staging tables.
  • Using the Redshift staging tables and UPSERT operation is performed on the Data Warehouse tables to update the dataset.
  • ETL job execution is completed once the Data Warehouse is updated.
  • Airflow DAG runs the data quality check on all Warehouse tables once the ETL job execution is completed.
  • Airflow DAG has Analytics queries configured in a Custom Designed Operator. These queries are run and again a Data Quality Check is done on some selected Analytics Table.
  • Dag execution completes after these Data Quality check.

Environment Setup

Hardware Used

EMR – I used a 3 node cluster with below Instance Types:

4 vCore, 16 GiB memory, EBS only storage
EBS Storage:64 GiB

Redshift: For Redshift I used 2 Node cluster with Instace Types dc2.large

Setting Up Airflow

I have written detailed instruction on how to setup Airflow using AWS CloudFormation script. Check out – Airflow using AWS CloudFormation

NOTE: This setup uses EC2 instance and a Postgres RDS instance. Make sure to check out charges before running the CloudFromation Stack.

Project uses sshtunnel to submit spark jobs using a ssh connection from the EC2 instance. This setup does not automatically install sshtunnel for apache airflow. You can install by running below command:

pip install apache-airflow[sshtunnel]

Finally, copy the dag and plugin folder to EC2 inside airflow home directory. Also, checkout Airflow Connection for setting up connection to EMR and Redshift from Airflow.

Setting up EMR

Spinning up EMR cluster is preety straight forward. You can use AWS Guide available here.

ETL jobs in the project uses psycopg2 to connect to Redshift cluster to run staging and warehouse queries.
To install psycopg2 on EMR:

sudo pip-3.6 install psycopg2

psycopg2 uses postgresql-devel and postgresql-libs, and sometimes pscopg2 installation may fail if these dependencies are not available. To install run commands:

sudo yum install postgresql-libs
sudo yum install postgresql-devel

ETL jobs also use boto3 move files between s3 buckets. To install boto3 run:

pip-3.6 install boto3 --user

Finally, pyspark uses python2 as default setup on EMR. To change to python3, setup environment variables:

export PYSPARK_PYTHON=python3

Copy the ETL scripts to EMR and we have our EMR ready to run jobs.

Setting up Redshift

You can follow the AWS Guide to run a Redshift cluster or alternatively you can use Script to create cluster automatically.

How to run

Make sure Airflow webserver and scheduler is running.
Open the Airflow UI http://< ec2-instance-ip >:< configured-port >

GoodReads Pipeline DAG
Pipeline DAG

DAG View:
DAG View

DAG Tree View:
DAG Tree

DAG Gantt View:
DAG Gantt View

Testing the Limits

The goodreadsfaker module in this project generates Fake data which is used to test the ETL pipeline on heavy load.

To test the pipeline I used goodreadsfaker to generate 11.4 GB of data which is to be processed every 10 minutes (including ETL jobs + populating data into warehouse + running analytical queries) by the pipeline which equates to around 68 GB/hour and about 1.6 TB/day.

Source DataSet Count:
Source Dataset Count

DAG Run Results:
GoodReads DAG Run

Data Loaded to Warehouse:
GoodReads Warehouse Count


  • Data increase by 100x. read > write. write > read

    • Redshift: Analytical database, optimized for aggregation, also good performance for read-heavy workloads
    • Increase EMR cluster size to handle bigger volume of data
  • Pipelines would be run on 7am daily. how to update dashboard? would it still work?

    • DAG is scheduled to run every 10 minutes and can be configured to run every morning at 7 AM if required.
    • Data quality operators are used at appropriate position. In case of DAG failures email triggers can be configured to let the team know about pipeline failures.
  • Make it available to 100+ people

    • We can set the concurrency limit for your Amazon Redshift cluster. While the concurrency limit is 50 parallel queries for a single period of time, this is on a per cluster basis, meaning you can launch as many clusters as fit for you business.

Other News