What it does Seldon Server was a Java-based machine learning platform for deploying models on Kubernetes. It wrapped prediction and recommendation workloads into containerized microservices, with REST and gRPC APIs, a CLI for management, and a Grafana dashboard fed by Kafka Streams.

The interesting bit The project tried to be the whole kitchen — model serving, recommendation engines, A/B testing, real-time analytics — before narrowing its ambition. In 2018 the team archived it and pivoted to Seldon Core, which does less (just model serving) but claims to do it more deliberately. The README essentially tells you to go use the new thing.

Key highlights

• Supports models from TensorFlow, Keras, XGBoost, Vowpal Wabbit, Gensim, and “any other model-building tool” (per the README’s generous wording)
• Built-in recommendation engine with out-of-the-box algorithms, cascading, and ensembles
• OAuth 2.0-secured APIs and a CLI for cluster management
• Real-time analytics pipeline: Kafka Streams → Fluentd → InfluxDB → Grafana
• Runs on GCP, AWS, Azure, or on-premise Kubernetes

Caveats

• Explicitly archived and unmaintained since January 2018; the README redirects to Seldon Core
• Documentation links point to an external site (docs.seldon.io) with unclear current status
• The “any other model-building tool” claim is vague on actual integration mechanics

Verdict Worth a historical look if you’re tracing the evolution of K8s ML infrastructure, or if you’re stuck maintaining a legacy Seldon Server deployment. Everyone else should head to Seldon Core or a modern alternative.