Yelp Unifies ML Model Training with Training Orchestrator

Yelp has launched Training Orchestrator, an internal framework that replaces the separate Spark training scripts individual teams maintained. The company is consolidating machine learning model training onto one system instead of leaving each team to own its own training path. Coverage on InfoQ by Claudio Masolo frames the release as a unification of how Yelp runs training work.

Technically, Training Orchestrator is configuration-driven and runs training as a DAG-based execution model. Teams describe training through configuration rather than custom Spark scripts, and the framework executes the work as a directed acyclic graph of steps. That shifts ownership from ad hoc scripts toward a shared orchestrator that controls how training jobs are defined and run.

For engineers and builders, the change cuts down on duplicated Spark training code across teams. A single configuration surface and a shared DAG execution path make training jobs easier to reuse, inspect, and operate. New models and experiments can follow the same entry point instead of forking another team-specific script.

In competitive and market terms, this matches a common pattern at larger ML organizations: move from team-owned training scripts to a platform layer that standardizes how models are trained. Yelp is applying that pattern inside its own stack by putting Spark-era training under one orchestrator rather than leaving training mechanics fragmented by team.

The practical takeaway is that configuration-driven, DAG-based training orchestration is how Yelp is standardizing ML training. Watch how much of the remaining team-level Spark training work moves onto Training Orchestrator, and whether the same configuration model becomes the default path for new training jobs.