What it does ModelDB tracks machine learning experiments: hyperparameters, metrics, artifacts, and model versions. You log runs from Python or Scala, browse results in a web UI, and query history when you need to reproduce something six months later. It runs on Docker or Kubernetes with a Java backend, PostgreSQL database, and a React frontend.

The interesting bit Born at MIT CSAIL and now maintained by Verta.ai, ModelDB is old enough to have seen the MLops tooling landscape explode around it. The architecture is pleasantly explicit: a Go proxy translates gRPC between frontend and Java backend, while HTTP handles binary artifacts directly. No magic, just plumbing you can trace.

Key highlights

• Python and Scala clients; log hyperparameters and metrics with a few lines of code
• Git-like model versioning that bundles code, data, config, and environment
• Pluggable storage with Hibernate/Liquibase for database flexibility
• Helm charts and Docker Compose for deployment; AWS AMI also available
• Claims integration with TensorFlow and PyTorch, though the README doesn’t detail how

Caveats

• The “5 minutes” quickstart requires manual file creation or repo cloning first
• “Battle-tested in production” and “beautiful dashboards” are claimed but not evidenced in the README
• Commit activity and maintenance cadence are unclear from the sources

Verdict Worth evaluating if you want self-hosted experiment tracking without SaaS lock-in, especially if your stack already leans Java/Scala. Skip if you need a vibrant plugin ecosystem or want something that feels native to Python notebooks.