Seldon Core

Background#

Seldon Technologies was founded in London in 2014 by Alex Housley and Clive Cox, originally as a recommendation engine and pivoting to ML infrastructure. Seldon Core launched in 2018 as one of the first Kubernetes-native model-serving frameworks — predating KFServing/KServe — and shaped much of the early thinking around inference graphs on Kubernetes.

The project follows an open-core commercial model. Seldon Core v1 is Apache 2.0; v2 (the current generation) moved to a Business Source License in 2023, with a paid Seldon Enterprise edition layered on top. The licence change was contentious but has not slowed adoption in their target enterprise market.

Inference Graphs#

Seldon's headline abstraction is the inference graph — a DAG of components (predictors, transformers, explainers, routers, combiners) that requests flow through. A typical production graph might be: input transformer (normalisation) → router (A/B split or contextual bandit) → predictor (the model) → output transformer (post-processing) → explainer (sidecar).

Each node is a containerised microservice, communicating via gRPC or REST. The runtime handles tracing across the graph, propagates request metadata, and exposes per-node metrics. The same pattern can be assembled in KServe with transformers and explainers, but Seldon's graph model is more expressive — multi-step routing, ensemble combiners, and shadow testing are first-class.

Seldon Core v1 vs v2#

A/B Testing and Multi-Armed Bandits#

Seldon ships built-in routers for weighted A/B splits and Thompson-sampling multi-armed bandits. This is the differentiator vs simpler servers — production teams running experiments on customer-facing models often pick Seldon for the routing layer alone, even when the model serving could be done in vLLM or Triton.

Explainability#

Seldon authored Alibi and Alibi Detect, the open-source libraries for model explanations (SHAP, anchors, counterfactuals) and outlier/drift detection respectively. Both integrate as graph nodes — an Explainer sidecar can be added to any SeldonDeployment to produce per-request feature attributions. This matters in regulated industries (financial services, healthcare) where every production prediction must be explainable.

When to Pick Seldon vs KServe#

Pick Seldon Core (v2) when your workload is fundamentally compositional — multi-step pipelines, A/B-tested ensembles, drift detection wired into the request path — and you accept a commercial licence. Pick KServe when you serve mostly single models behind an OpenAI-compatible endpoint, want CNCF-aligned governance, and prefer Apache 2.0 throughout.

Mid-2026, KServe has the wider open-source adoption — particularly for LLMs via vLLM — while Seldon retains a strong enterprise foothold in regulated industries where its routing and explainability features pay off.