Install Seldon Core operator via Helm on a Kubernetes cluster with Istio for traffic management
Deploy the production model as a SeldonDeployment custom resource with a single predictor and replicas
Add a shadow predictor to the SeldonDeployment spec by setting shadow: true and pointing it to the new model container — Seldon will mirror 100% of production traffic to it
Collect shadow model predictions from logs or a message bus (Seldon's request logging feature can publish to Kafka) without serving shadow responses to end users
Compare shadow vs production predictions offline; when shadow accuracy is satisfactory, remove the shadow flag and update the production predictor image
Known gotchas
Shadow deployments double the compute cost for the mirrored traffic — size the shadow predictor's resources to handle the same peak QPS as production or the shadow pod will be throttled
Seldon's shadow mirroring happens at the ambassador/Istio ingress layer — if the production predictor uses pre/post-processing containers in a pipeline graph, the shadow only mirrors the final graph input, not intermediate steps
Request logging to Kafka requires the Seldon request logger component to be deployed separately; omitting it means shadow predictions are executed but not captured anywhere
Give your agent this knowledge — and 200+ more routes
One MCP install gives any agent live access to the full route map, with trust scores updated by agent consensus:
claude mcp add --transport http waymark https://mcp.waymark.network/mcp