![Sachin G.](/assets/transparent-ad5be28fbcd25b7b08d2cebe1d957125437fb5407d75ee717965ad22c8808791.gif "Sachin G.") SG Sachin G. Machine Learning Engineer Mid-Market (51-1000 emp.) 6/8/2026 "Powerful for complex ML pipelines, but comes with a steep infrastructure learning curve" 5/5 What do you like best about Apache Airflow? We use Apache Airflow as the central orchestrator for our entire machine learning and data engineering lifecycle. Specifically, it manages the schedules and dependencies for pulling raw data from our production databases, orchestrating daily feature engineering jobs, and kicking off automated model retraining pipelines. Once the models are evaluated, Airflow also triggers the deployment scripts to push updated model artifacts to our staging environment. It basically acts as the glue holding our data workflows together, ensuring that everything runs in the exact sequence required.What I appreciate most about Airflow is its "configuration as code" philosophy. Because every workflow is defined as a Python script, it fits perfectly into our existing development practices. We can version control our DAGs in Git, run code reviews on them, and easily write custom operators when the built-in ones don't quite fit our needs. The UI is also incredibly detailed; when a complex pipeline fails at 3:00 AM, the tree view and the ability to dig directly into the logs of a specific failed task save us an immense amount of troubleshooting time. The open-source community is massive, so if you are trying to connect Airflow to a database or a cloud service, chances are a robust provider package already exists. Review collected by and hosted on G2.com. What do you dislike about Apache Airflow? The biggest headache with Airflow is definitely the operational overhead and the steep learning curve required to keep it running smoothly. Managing the scheduler, web server, and workers—especially when handling resource-intensive machine learning tasks—requires a lot of infrastructure tuning. We struggled initially with the scheduler getting bogged down and tasks getting stuck in a queued state. Additionally, local development can be painful to set up realistically, and the fact that the scheduler constantly parses Python files means you have to be extremely careful about how you write your code to avoid severe performance degradation. It is not a tool you can just turn on and forget about; it requires dedicated DevOps attention. Review collected by and hosted on G2.com. What problems is Apache Airflow solving and how is that benefiting you? Before implementing Airflow, our data syncing and model training tasks were managed by a fragmented mix of cron jobs and custom shell scripts. If a single step failed, the subsequent scripts would either run with corrupt data or the whole system would silently halt without anyone knowing. Airflow completely solved this visibility and dependency issue. For example, we automated a massive customer churn prediction pipeline that requires seamless execution across data extraction, transformation, and model inference. Now, if the initial data extraction step fails due to a network glitch, Airflow automatically retries it a few times before alerting us on Slack, ensuring our downstream models always train on fresh, accurate data without manual intervention. Review collected by and hosted on G2.com. Show More Rating Updated (6/10/2026) Validated ReviewerIncentivizedSource: G2 invite See what 128 reviewers think of Apache Airflow 4.4 out of 5 · Verified reviews from real users [ Read all reviews ](https://www.g2.com/products/apache-airflow/reviews)