3
sagemaker
303d

Airflow… airflow… I hate you so freaking much, you are a bloated piece of software that packages wayyy too much and increase the complexity of any solution we built on top of you. Plus unit-testing and integration tests are wayyy too difficult with you. But man… your recent UI changes are a massive welcome.

A year ago I was tasked with either upgrading airflow to version 2 or to migrate the code to another tool. Naively I thought “well might as well upgrade to avoid a rewrite”. Little did I know that the reason airflow didn’t scale out well for us, was due to people over the years not having a grasp on airflow primitives like “pools”, “workers” and “operators”. Ended up refactoring the entire codebase for both infra and DAGs anyhow AND upgrading the beast AND lower cost by a factor of 2 (from $100-$150 daily to $50-$70 daily).

But seriously feels like I could’ve solved the scheduling issue with literally any message queue+decent library (like celery or Faust) and I’d have half the headache.

Comments
Add Comment