Skip to content

Observability — deployment

Target

This page describes intended behaviour or structure. No implementation is claimed unless a section is explicitly marked Current implementation or Evidence.

Rewritten 2026-10-09 (design plan P8)

The June-2026 page (collector, trace, log, metric and dashboard services as separate deployables) is archived under catalogs 2026-06.

Host Modules Topologies
ObservabilityHost 8 saas-default, hybrid, on-prem, hybrid-compute

Shape

flowchart TB
    subgraph host [ObservabilityHost · AKS · 2–8 replicas]
        D[DashboardService]
        S[SloService]
        A[AlertRuleService]
        Q[QualityScoreService]
        F[FeedbackService]
    end
    host --> SQL["(observability-sql)"]
    host --> BLOB["(artifact-blob · evidence, snapshots)"]
    host --> CACHE["(cache)"]
    host <--> OT["(OpenTelemetry backend)"]
    IN["(Orleans inbox · event-transport)"] --> host
    BFF[Studio BFF] --> host
    CP[Control Plane API] --> host
Hold "Alt" / "Option" to enable pan & zoom

One host, five modules, horizontal replicas behind the internal gateway; the inbox is partitioned by tenant so a noisy tenant's backlog does not delay another's projections.

Per topology

Topology Backend Store Notes
saas-default Azure Monitor (Application Insights, Log Analytics) in the tenant-selected region (D31) Azure SQL observability-sql; Blob for evidence default
hybrid Azure Monitor for the control side; the customer's collector for runtime as saas-default runtime spans arrive through the customer's OTLP exporter with the Factory attributes
on-prem Jaeger or Seq + Prometheus behind an OTLP collector SQL Server the host's TraceBackend option selects the query adapter
hybrid-compute as saas-default; model-serving spans exported from the compute pool as saas-default compute pool spans join on run_id and deployment_id

Data and recovery

  • Retention: timelines follow the run's retention (1 year default, tenant-configurable); overviews 13 months; alerts 13 months; evaluation results 2 years, evidence references 7 years with the evidence artifacts themselves under audit retention (D31).
  • Backup and restore: RPO 15 min, RTO 4 h (D30); this store is last in the restore order because every projection is rebuildable from the run journal and the producers' events; evidence and snapshots live in artifact-blob and are referenced.
  • Backend outage: projections keep serving; trace drill-down and metric-sourced objectives degrade with a visible note.

Ports and runtime

HTTPS 8100, gRPC 8102, silo 12100, gateway 31100 (messaging profile observability). Workload identity through the identity platform; secrets observability-sql, otel-backend, messaging-transport from Key Vault with 90-day rotation.