airflow-ml : ML and Pre-Processing Automation

This is a workflow manager which uses Airflow to automate all the Machine Learning and Data Pre-processing pipelines in our system. Some of the most important steps in our pipeline are

• Fetching and indexing raw data and entities which are in the bucket
• Economic News Detection (To-be-done)
• Text Extraction - By hitting the news URL and fetching data
• Some cleaning, indexing and preprocessing - Probably storing it in a relational database
• Sentiment Analysis
• Fuzzy Matching
• Entity Extraction
• Custom Scoring Model (BERT)

Development

Airflow needs persistent storage to store the details of all the executions and pipelines. Right now, we run the postgres docker image as a temporary solution. We need to change it to production ready database.

When setting up airflow for the first time, we need to run the migrations which is in the docker-compose so that the tables can be setup.

• Set the $AIRFLOW_HOME to the current file root export AIRFLOW_HOME=$(pwd)
• airflow.cfg shouldn't be copied
• remove the default airflow.cfg while setting up locally

Testing Locally

• airflow webserver
• airflow backfill indexing -s 2019-12-30

airflow test indexing test_entities 2020-01-01 airflow test indexing test_entities 2020-01-01

Migrations: Run only the first time for setting up tables

docker-compose up initdb

docker-compose up --build

• delete .data and run docker-compose up --build postgres
• docker-compose up --build initdb
• and no airflow.cfg and unittest
• User Auth
• Disable auto-start docker update --restart=no rey_worker_1 rey_scheduler_1 rey_webserver_1 rey_flower_1 rey_redis_1 rey_postgres_1

AIRFLOW__CORE__SQL_ALCHEMY_CONN=postgresql+psycopg2://airflow:airflow@postgres:5432/airflow
AIRFLOW__CORE__LOAD_EXAMPLES=False
AIRFLOW__WEBSERVER__RBAC=True

• If you fuck up db

  • airflow resetdb
  • docker exec -it rey_scheduler_1 /entrypoint.sh bash

• Creating a New User

  airflow create_user -r Admin -u dev -e adarsh@alrt.ai -f username -l s -p password

• Configuring Fernet Key for Production

  Fetch the Fernet key and update initdb to make it work properly (using docker exec) export FERNET_KEY = python -c "from cryptography.fernet import Fernet; FERNET_KEY = Fernet.generate_key().decode(); print(FERNET_KEY)"

Distributing Workers

MetaStore: postgresql+psycopg2://airflow:airflow@postgres:5432/airflow BrokerURL: redis://redis:6379/1

For running workers on different nodes, connect to the metastore db and the redis queue by configuring the docker-compose with POSTGRES_HOST and REDIS_HOST. Also remove the dependency to the local scheduler

Nginx configuration

• delete /etc/nginx/sites-enabled/default
• create /etc/nginx/conf.d/reverse-proxies.conf
• add A Record airflow

•       listen 80 default_server;
        listen [::]:80 default_server;
        server_name airflow.alrt.ai;
  
        location / {
              proxy_pass http://localhost:8080/;
              proxy_http_version 1.1;
              proxy_set_header Upgrade $http_upgrade;
              proxy_set_header Connection 'upgrade';
              proxy_set_header Host $host;
              proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;
              proxy_cache_bypass $http_upgrade;
              proxy_set_header X-Forwarded-Proto $scheme;
        }
  

}

## References

* [Configuring Airflow in docker-compose](https://medium.com/@xnuinside/quick-guide-how-to-run-apache-airflow-cluster-in-docker-compose-615eb8abd67a)
* [Best Practices](https://gtoonstra.github.io/etl-with-airflow/principles.html)
* [Config](https://github.com/kjam/data-pipelines-course/issues/1)
* [Fernet Key and Stuff](https://medium.com/@itunpredictable/apache-airflow-on-docker-for-complete-beginners-cf76cf7b2c9a)