Kubernetes operator for ML models that speaks GitOps

What it does

Yatai is the Kubernetes deployment operator for BentoML, an ML model packaging framework. It installs into a cluster and adds three CRDs: BentoRequest, Bento, and BentoDeployment. You push a packaged model (a “Bento”) to Yatai’s registry, then apply a YAML manifest to deploy it with autoscaling, resource limits, and ingress — standard kubectl workflow, no custom CLI required for deployment.

The interesting bit

The project splits into three installable components — core Yatai, image builder, and deployment controller — which lets teams adopt only what they need. The BentoDeployment CRD is the hook: it treats an ML service like any other Kubernetes workload, meaning existing GitOps pipelines and RBAC policies apply without special-casing model deployments.

Key highlights

• CRD-native deployments: Define model services in YAML, version them in git, deploy with kubectl apply
• Per-component installation: Core registry, image builder, and deployment controller install separately via Helm/scripts
• Autoscaling and resource controls: HPA-style minReplicas/maxReplicas plus CPU/memory limits at both deployment and runner level
• Web UI + API tokens: Browser-based setup and token management for pushing Bentos from the CLI
• Elastic License 2.0: Not open source; source-available with usage restrictions

Caveats

• Under construction for BentoML 1.2: The README warns that Yatai for the current BentoML release is being rebuilt; a “Yatai 2.0 Proposal” is open for comment
• Quick-install scripts are dev-only: Production requires manual Helm installation; the one-liner scripts bundle PostgreSQL and MinIO for testing
• Three separate installs for full functionality: The quick start requires running three different bash scripts to get registry, builder, and deployment features

Verdict

Worth evaluating if you’re already using BentoML and want Kubernetes deployments without building a custom bridge. Skip if you need a fully open-source stack or aren’t committed to the BentoML packaging format — the ELv2 license and version-rewrite uncertainty add friction for new adopters.