Model Evaluation Service  ·  View 16 of 21  ·  Operations

Deployment Architecture

Where it runs, and the two projects the execution plane deliberately cannot write to.

Editable source SVG draw.io All views
Google Cloud · project mes-prod · europe-west4 GKE Autopilot · control plane namespace Evaluation API 3 pods · 2 zones Run Planner 3 pods Gate Engine 4 pods GKE · execution node pool · Spot with on-demand floor Harness Workers 0-600 pods Sandbox Pods gVisor · egress deny Scoring Workers 0-200 pods Managed data services Cloud SQL HA · PITR BigQuery EU multi-region Cloud Storage dual-region · versioned Memorystore cache only Perimeter Private Service Connect VPC Service Controls data perimeter Egress Proxy allowlist Separate projects · blast-radius boundary mes-corpus Dataset Versions no delete role Canary Subset restricted mes-labels Human Labels rater programme Model Evaluation Service — Deployment Architecture Interface / broker Application we own Decision point Security / platform Data store The corpus and label projects are separate so that no identity in the execution plane can reach them for write. The harness reads datasets and writes traces; it cannot touch a label or delete a dataset version. v 1.0 · owner Data & AI Global Practice · date 2026-09

Decisions

  • Corpus and labels live in separate Google Cloud projects. No execution-plane identity can delete a dataset version or touch a label.
  • Execution runs on Spot capacity with an on-demand floor: a re-baseline is interruptible, a release-window gate run is not.
  • VPC Service Controls draw a data perimeter around production-derived content.

Availability

  • Gate decision API ≥ 99.9% monthly during release windows · score store read ≥ 99.9% · harness control plane ≥ 99.5% (it is on no user path).
  • Artefact store ≥ 99.99% for reads — a verdict that cannot cite its evidence is not a verdict.

Recovery

  • Score store RPO 0 / RTO ≤ 4 h · artefacts RPO 0 / RTO ≤ 2 h · run state RPO ≤ 5 min / RTO ≤ 1 h, resuming from checkpoint (stated assumptions).
◀  Frame Change and Re-Baseline All views Release Pipeline  ▶
Solution Architecture v1.0 · Google Cloud · Data & AI Global Practice · 2026-09 · 21 views · 18 architecture decision records16-deployment-architecture