Skip to main content

ADR-0004: ML Serving for MCP (Explainable, Versioned, gRPC-first)

  • Status: Proposed
  • Date: 2025-10-21

Context​

  • Predictive flakiness and root-cause categorization require serving ML models with explainability (SHAP/LIME), versioning, and rollback.

Decision​

  • MCP server exposes a gRPC inference API with standard request/response contracts.
  • Initial serving: custom gRPC wrapper around model runtime (e.g., Python FastAPI sidecar or Rust/Go bindings); capture SHAP values when enabled.
  • Mandatory: model versioning, canary rollout, rollback triggers.

Consequences​

  • Consistent interface across models; polyglot model support.
  • Explainability overhead must be bounded; can be toggleable.

Alternatives considered​

  • TorchServe/TensorFlow Serving directly (less flexible for XAI hooks).
  • REST-only inference (less efficient for high-throughput).

Revisit criteria​

  • Latency/throughput constraints; operational overhead; model lifecycle maturity.

References​

  • PRD §5.2, §5.8; Algorithm doc; Research §3.2.