// the find
SeldonIO/seldon-server
Machine Learning Platform and Recommendation Engine built on Kubernetes
Seldon Server is an archived (2018) ML model serving and recommendation platform built on Kubernetes, MySQL, Kafka, and Java. The team pivoted to Seldon Core, which is actively maintained and does the same job better. This repo is a historical artifact.
- The architecture decision to separate prediction (supervised model serving) from recommendation (collaborative filtering) into distinct API endpoints was sensible and ahead of its time for 2016-era ML tooling.
- Kafka Streams for real-time action ingestion feeding the recommender is a solid design — user events flow in asynchronously without blocking the serving path.
- The microservice model for wrapping arbitrary ML frameworks (TensorFlow, XGBoost, Vowpal Wabbit) via Docker containers predates KFServing/Torchserve by years; the idea was right even if the implementation aged poorly.
- A/B testing and algorithm ensembling baked into the routing layer, not bolted on afterward, is genuinely useful for recommendation systems.
- Archived since 2018. The README itself tells you to go use Seldon Core instead. There is no reason to adopt this.
- MySQL as the primary store for a recommendation engine is a fundamental mismatch — no vector search, no graph traversal, just relational tables being abused for item similarity lookups.
- The build system is a wall of per-component shell scripts with no coherent orchestration. Getting this running locally in 2024 would be a archaeology project.
- Java Spring for the core server means heavyweight startup, verbose configuration, and a dependency tree that was already showing its age when this was archived.