Productizing ML at Scale at Twitter with Yi Zhaung - TWIML Talk #271

Hyperparameter Tuning

Hyperparameter tuning at Twitter involves a sophisticated workflow that automates the tedious process of parameter optimization. Yi Zhaung explains that their system allows users to perform randomized and grid searches by launching numerous experiments on an internal cloud, automatically recording parameters and results. This is complemented by a service called Wet Lab, which employs Bayesian optimization to recommend the next set of hyperparameters to test, enhancing both exploration and exploitation.

Our ML workflow solution has completely automated solution for using wet lab.

--- Yi Zhaung

The core artifacts in this system are TensorFlow graphs and trained models, which are versioned by users to ensure transparency and efficiency in development workflows 1.

   

Orchestration Tools

Twitter's machine learning workflows are managed using orchestration tools built on Apache Airflow. Yi Zhaung notes that these tools primarily support offline training, while real-time predictions are handled by separate prediction servers. This setup ensures that when a user interacts with Twitter, such as viewing advertisements, the system can respond immediately with real-time scores.

Our predictions are like this. Like a user comes to twitter.com, for example, we need to present advertisement we immediately need to respond.

--- Yi Zhaung

This division between offline training and real-time prediction allows for efficient model development and deployment 2 3.