Skip to content

app/recommendations/ — v2 recommendations API (KServe wrapper)

A small Go service introduced in ch.13.08. POSTs to the KServe predictor (which serves the model from MLflow); publishes each prediction to Kafka ml.predictions for the drift detector.

What it does

  1. Reads the verified x-jwt-payload header (Istio sets it post JWT verification; ch.13.04) and extracts the tenant claim.
  2. POSTs the request body to the KServe v1 protocol endpoint /v1/models/recommender:predict.
  3. Returns the predictions to the caller.
  4. Async-publishes the prediction event to Kafka ml.predictions for Alibi-Detect to consume.

What it does not do (out of scope)

  • Embedding cache for hot users.
  • Request batching to amortise KServe RPCs.
  • A/B test traffic split via feature-flag SDK.
  • Streaming response (predictions can be large).

The chapter calls these out; this is a 150-line demonstration.

Build + run

cd examples/bookstore-platform/app/recommendations
go vet ./...
go build ./...
docker build -t bookstore-platform/recommendations:dev .

Cross-references

  • Ch.13.08 — the chapter that introduces this service.
  • ../../ml/inferenceservice.yaml — the KServe predictor this service calls.
  • ../../ml/alibi-detect-drift.yaml — the drift detector that consumes ml.predictions.
  • ../../kafka/topics.yaml — ml.predictions is implicit (the topic is created on first write; production should add it to topics.yaml).