Skip to content

aiops-dashboard.md

Latest commit

 

History

History
104 lines (78 loc) · 5.83 KB

aiops-dashboard.md

File metadata and controls

104 lines (78 loc) · 5.83 KB
Raw

AIOps Dashboard

Status: active

The in-app operational view of an AI agent fleet — live health, provider and agent performance, cost, reliability (alerts + circuit breakers), and short-window trends. Distinct from the infra logging/metrics stack in operations/observability.md and from the deep historical analytics surface.

Table of Contents

What this covers

The AIOps dashboard answers "how is my AI fleet doing right now, and over the last few hours?" It reads from Ai::Analytics::DashboardService (the AiopsMetrics concern) which aggregates Ai::AgentExecution, Ai::ProviderMetric, cost attribution, and circuit-breaker state — all account-scoped. It is read-only: it surfaces live operational data, it does not mutate fleet state.

AIOps lives on the dedicated Operations hub (OperationsPage, /app/ai/operations/*), inside an AiErrorBoundary. It is path-based (canonical PathTabs, one URL segment per tab):

  • AIOps tab — the operational core: KPIs + system health + an active-provider-alerts callout, trend charts, providers table, agents table. Body component: AiOpsContent (frontend/src/features/ai/aiops/components/AiOpsDashboard.tsx).
  • Alerts tab — the alert-management center plus provider reliability (circuit breakers + recent errors) via ReliabilitySection.
  • Execution Traces tab — the distributed-trace viewer (ExecutionTracesContent).

AIOps cost analysis is part of the Cost domain (/app/ai/cost), not Operations. The monitoring-only Observability hub (/app/ai/observability: Health · Systems · Conversations · Evaluation) is a separate sidebar item. See AI Navigation IA.

Backend: Api::V1::Ai::AiOpsController under scope :aiops — all reads gated by the ai.aiops.read permission (record_metrics is the only writer, gated by ai.aiops.write and used by the worker, not the UI).

Data fetching follows the platform standard: @tanstack/react-query hooks (features/ai/aiops/api/aiopsApi.ts) with a query-key factory. Sections self-fetch via the shared useAiOpsDashboard hook; react-query dedupes by query key, so all sections share a single fetch regardless of which tab renders them.

Where the data surfaces

The sections render across the Operations hub tabs:

Operations tab AIOps content Source
AIOps execution / latency / cost KPIs, system-health components, active-provider-alerts callout, hourly trend charts, providers table, agents table dashboard.overview, dashboard.health, dashboard.alerts[], /trends, dashboard.providers[], dashboard.agents[]
Alerts alert-management center + provider reliability: circuit-breaker status + recent execution errors dashboard.circuit_breakers[], /recent_errors

AIOps cost_analysis data surfaces in the Cost domain (/app/ai/cost), not Operations.

API contract

All responses use the standard { success, data, meta } envelope; the frontend BaseApiService unwraps data. Endpoints (under /api/v1/ai/aiops):

Endpoint Returns
GET /dashboard?time_range= { dashboard: { health, overview (incl. latency_aggregate), providers[], agents[], cost_analysis, alerts[], circuit_breakers[], real_time, generated_at }, time_range }
GET /real_time live snapshot: { current_requests_per_second, current_avg_latency_ms, current_error_rate (0–1), active_connections, queue_depth, timestamp }
GET /trends?time_range= { trends: { bucket: "hour", bucket_count, latency[], error_rate[], throughput[], cost[] }, time_range }
GET /latency_aggregate?time_range= { latency_aggregate: { avg_ms, p95_ms, p99_ms, max_ms, sample_provider_count }, time_range }
GET /recent_errors?limit= { recent_errors: [{ execution_id, agent_name, error, failed_at }], count, timestamp }

success_rate fields are percentages (0–100); current_error_rate and trend error_rate are fractions (0–1). The frontend types latency_aggregate, trends, and recent_errors as optional so the UI degrades gracefully if a deployment predates those endpoints.

Trend semantics

GET /trends returns four parallel time series sharing one x-axis. Key invariants:

  • Hourly buckets, capped at 168 (7 days). Bucket keys are ISO8601 UTC timestamps; the frontend localizes for display.
  • Zero-filled: every series has exactly bucket_count points — buckets with no activity are emitted as zeros, never omitted. This keeps the chart x-axes aligned.
  • Latency p95/p99 come from Ai::ProviderMetric (the only source with true percentiles). When an account has no provider metrics for a window, the series fall back to per-execution aggregates and report p95 = p99 = avg (a documented approximation).

Operational vs analytics boundary

AIOps is intentionally operational and live: health, providers, alerts, circuit breakers, real-time, short hourly trends, recent errors. It does not duplicate the deeper /api/v1/ai/analytics/* surface (ROI, forecasting, insights, recommendations, export, day-level 30-day trends). When you need historical/financial analysis, use that surface and cost-and-finops.md; when you need "is the fleet healthy now," use AIOps.

See also