ms-ai-architect/skills/ms-ai-governance/references/monitoring-observability/real-time-streaming-monitoring.md

Real-Time Streaming and Live Monitoring Dashboards

Last updated: 2026-02 Status: GA Category: Monitoring & Observability

---

Introduksjon

Real-time streaming og live monitoring er kritiske kapabiliteter for operasjonell overvåking av AI-applikasjoner i produksjon. Mens tradisjonell logging og metrics aggregeres over tid (typisk 1-5 minutter), tilbyr real-time løsninger innsikt med under ett sekunds latency, noe som er essensielt for debugging, incident response og operasjonell overvåking av AI-tjenester.

Microsoft-stakken tilbyr to primære løsninger: Application Insights Live Metrics for sanntidsovervåking av applikasjonsytelse med 1-sekunds latency, og Fabric Real-Time Intelligence med Real-Time Dashboards for streaming analytics på tvers av multiple datakilder. Live Metrics er optimalisert for utviklere som trenger umiddelbar tilbakemelding under deployment og debugging, mens Real-Time Intelligence er bygget for operasjonelle scenarier som krever kontinuerlig overvåking av store datavolumer.

Begge løsningene har ulike arkitekturer: Live Metrics bruker en dedikert push-basert control channel som streamer data direkte fra applikasjonen til portalen uten persistering, mens Real-Time Dashboards benytter Kusto Query Language (KQL) mot eventhouses eller Azure Data Explorer med refresh-intervaller ned til 10 sekunder.

---

Kjernekomponenter

Application Insights Live Metrics

Komponent	Beskrivelse	Latency
Live Metrics Stream	Push-basert streaming av telemetri fra app til portal	< 1 sekund
Custom Filters	Real-time filtering på URL, duration, telemetry type	Real-time
Server Instance Filtering	Isoler spesifikke server-instanser for debugging	Real-time
Performance Counters	Windows performance counters (CPU, memory, requests)	< 1 sekund
Exception Stack Traces	Full stack traces for exceptions når de skjer	Real-time
Control Channel	Secure channel for filter-signaler (krever Entra ID auth)	N/A

Støttede plattformer:

// ASP.NET Core - OpenTelemetry (anbefalt, enabled by default)
builder.Services.AddOpenTelemetry().UseAzureMonitor(options => {
    options.EnableLiveMetrics = true; // Default: true
});

// ASP.NET Core - Classic API (manual config)
using Microsoft.ApplicationInsights.Extensibility.PerfCounterCollector.QuickPulse;

var quickPulseProcessor = null;
config.DefaultTelemetrySink.TelemetryProcessorChainBuilder
    .Use((next) => {
        quickPulseProcessor = new QuickPulseTelemetryProcessor(next);
        return quickPulseProcessor;
    })
    .Build();

var quickPulseModule = new QuickPulseTelemetryModule();
quickPulseModule.Initialize(config);
quickPulseModule.RegisterTelemetryProcessor(quickPulseProcessor);

Nøkkelegenskaper:

• On-demand streaming: Data sendes KUN når Live Metrics-panen er åpen (spart kostnader)
• Ingen persistering: Data vises kun i real-time, ikke lagret for historisk analyse
• Sampling: Stack traces og exceptions samples, men alle metrics og counters sendes
• Gratis: Ingen ekstra kostnad for Live Metrics data (kun standard ingestion)

Fabric Real-Time Intelligence & Real-Time Dashboards

Komponent	Beskrivelse	Refresh Rate
Real-Time Dashboard	No-code dashboard med KQL queries	10 sekunder - manuell
Eventhouse	Time-series optimalisert database for streaming data	Subsecond ingestion
Eventstream	No-code streaming pipelines med transformasjoner	Near real-time
Data Activator	Event detection med subsecond latency, trigger actions	< 1 sekund
Copilot for Dashboards	AI-generert dashboard fra natural language prompts	N/A

Støttede datakilder:

Datakilde	Use Case	Latency
Eventhouse	Fabric-native streaming events	Subsecond
Azure Data Explorer	Log og telemetry analytics	< 5 sekunder
Application Insights	App performance metrics	1-5 minutter*
Log Analytics	Azure resource logs	1-5 minutter*

*Application Insights og Log Analytics har inherent ingestion latency (1-5 min), men kan queries via Real-Time Dashboard.

KQL Query Eksempel (streaming dashboard):

// Real-time monitoring av AI request latency
AIRequests
| where timestamp > ago(5m)
| summarize
    AvgDuration = avg(duration),
    P95Duration = percentile(duration, 95),
    RequestCount = count()
    by bin(timestamp, 10s), operation_Name
| render timechart

Auto-refresh konfigurasjon:

• Minimum refresh interval: 10 sekunder
• Anbefalt for high-volume scenarier: 30-60 sekunder (reduserer compute cost)
• Manual refresh: Alltid tilgjengelig for ad-hoc analyse

---

Arkitekturmønstre

Mønster 1: Live Metrics for CI/CD Validation

Scenario: Validere deployment i sanntid under release pipeline.

Arkitektur:

Developer → Deploy til Azure → App starter → Live Metrics åpnes i portal
                                             ↓
                                    Monitor exceptions, latency, dependencies
                                             ↓
                                    Validate fix → Close Live Metrics (stop streaming)

Fordeler:

• Umiddelbar feedback (< 1 sekund latency)
• Ingen setup utover Application Insights instrumentation
• Gratis (no-cost streaming)
• Filter på spesifikk server instance under rolling deployment

Ulemper:

• Krever manuell observasjon (ingen alerting)
• Data persisteres ikke (kun for live debugging)
• Control channel må sikres med Entra ID for production (API keys deprecated Sept 2025)

Når bruke:

• Debugging av nye deployments
• Load testing validation
• Exception tracking under release
• Server-specific performance issues

Mønster 2: Real-Time Dashboard for Operations

Scenario: Kontinuerlig overvåking av AI-tjenester i produksjon med alerting.

Arkitektur:

AI App → Application Insights → Eventhouse (via Eventstream)
                                      ↓
                              Real-Time Dashboard (10s refresh)
                                      ↓
                              Data Activator → Teams/Email Alert

Fordeler:

• Persistent storage i Eventhouse (long-term analytics)
• Proactive alerting via Data Activator
• Multi-source dashboards (combine App Insights + Azure Monitor)
• No-code dashboard creation med Copilot

Ulemper:

• Higher latency enn Live Metrics (10s min refresh)
• Krever Fabric capacity (kostnader)
• Mer kompleks setup (Eventstream pipeline)

Når bruke:

• Production operations med SLA requirements
• Multi-service monitoring (microservices)
• Compliance requirements (data retention)
• Business stakeholder visibility (shareable dashboards)

Mønster 3: Hybrid - Live Metrics + Real-Time Dashboard

Scenario: Kombinere ad-hoc debugging (Live Metrics) med persistent monitoring (Real-Time Dashboard).

Arkitektur:

AI App → Application Insights ─┬→ Live Metrics (on-demand)
                               └→ Eventhouse → Real-Time Dashboard
                                                      ↓
                                               Data Activator Alerts

Fordeler:

• Best of both worlds: Live debugging + persistent monitoring
• Cost-efficient (Live Metrics kun når åpen, dashboard alltid tilgjengelig)
• Enkel escalation fra dashboard til Live Metrics for deep-dive

Ulemper:

• Dobbel ingestion (App Insights + Eventhouse) hvis begge brukes samtidig
• Mer kompleks setup

Når bruke:

• Enterprise AI-løsninger med både dev og ops teams
• Critical workloads som krever både proactive alerting og reactive debugging

---

Beslutningsveiledning

Når bruke Live Metrics

Kriterium	Anbefaling
Use Case	Debugging, deployment validation, load testing
Latency krav	< 1 sekund
Data retention	Ikke nødvendig (kun live view)
Alerting	Ikke påkrevd (manuell observasjon OK)
Cost sensitivity	Høy (gratis streaming)
Team	Developers, DevOps

Når bruke Real-Time Dashboard

Kriterium	Anbefaling
Use Case	Operations monitoring, SLA tracking, compliance
Latency krav	10 sekunder - 1 minutt OK
Data retention	Påkrevd (historisk analyse)
Alerting	Kritisk (proactive incident response)
Multi-source	Ja (combine App Insights + Azure Monitor + custom events)
Team	Operations, business stakeholders

Vanlige feil

❌ Bruke Live Metrics for production alerting Live Metrics har ingen alerting-kapabilitet. Bruk Real-Time Dashboard med Data Activator.

❌ Åpne Live Metrics kontinuerlig i produksjon Live Metrics streamer data kun når panen er åpen, men slutter ikke automatisk. Lukk etter debugging for å stoppe streaming overhead på app.

❌ Forvente persistering i Live Metrics Data i Live Metrics discarderes når du lukker panen. Bruk Logs eller Metrics Explorer for historiske queries.

❌ Sette Real-Time Dashboard refresh til 10s for alle scenarier Høyere refresh rate = høyere compute cost. Bruk 30-60s for most production dashboards, 10s kun for critical metrics.

❌ Bruke usecured control channel i production API keys for Live Metrics retired Sept 2025. Migrer til Entra ID authentication.

Røde flagg

🚩 Live Metrics viser ingen data

• Sjekk firewall: Live Metrics bruker separate endpoints (live.applicationinsights.azure.com) enn vanlig telemetri
• Verifiser TLS 1.2 support (Live Metrics krever TLS 1.2+)
• Bekreftet at OpenTelemetry Distro er nyeste versjon (live metrics enabled by default)

🚩 Real-Time Dashboard viser gammel data (> 1 min latency)

• Application Insights har inherent ingestion latency (1-5 min). For true real-time, stream direkt til Eventhouse via Eventstream.

🚩 Data Activator trigger for ofte (false positives)

• Bruk anomaly detection functions i KQL (series_decompose_anomalies) for å detektere avvik fra baseline istedenfor statiske thresholds.

---

Integrasjon med Microsoft-stakken

Azure OpenAI + Live Metrics

// Track Azure OpenAI calls i Live Metrics
var activity = new Activity("AzureOpenAI.ChatCompletion");
activity.Start();

try {
    var response = await openAIClient.GetChatCompletionsAsync(deploymentName, options);

    telemetryClient.TrackDependency(
        "AzureOpenAI",
        "ChatCompletion",
        deploymentName,
        activity.StartTimeUtc,
        activity.Duration,
        success: true);
} catch (Exception ex) {
    telemetryClient.TrackException(ex);
    throw;
} finally {
    activity.Stop();
}

Live Metrics vil vise:

• Dependency latency i real-time
• Exception stack traces hvis Azure OpenAI API feiler
• CPU/memory impact av token processing

Copilot Studio + Real-Time Dashboard

Scenario: Monitor Copilot Studio agent performance med Real-Time Dashboard.

1. Enable Application Insights for Copilot Studio (via Power Platform admin)
2. Create Eventstream som subscriber til App Insights metrics
3. Build Real-Time Dashboard med KQL queries:

// Copilot Studio conversation success rate (10s buckets)
customEvents
| where name == "ConversationCompleted"
| extend success = tostring(customDimensions.Success)
| summarize
    SuccessRate = countif(success == "true") * 100.0 / count()
    by bin(timestamp, 10s)
| render timechart

4. Setup Data Activator til å trigger alert hvis SuccessRate < 95% i 1 minutt.

Azure AI Foundry + Eventhouse

Azure AI Foundry Observability dashboard støtter native integration med Application Insights, som kan streams til Eventhouse for real-time dashboards:

1. Enable Application Insights for AI Foundry project
2. Use Eventstream til å route App Insights logs til Eventhouse
3. Create Real-Time Dashboard med queries for:
   • Token usage per minute
   • Model latency (P50, P95, P99)
   • Content safety violations
   • Groundedness scores

---

Offentlig sektor (Norge)

GDPR og datasuverenitet

Live Metrics:

• Ingen persistering: Data i Live Metrics lagres ikke, kun streames til browser. GDPR Article 6(1)(f) "legitimate interests" for debugging.
• PII i custom filters: Bruk IKKE custom filters med potensielt sensitive data (customer ID, email) før Entra ID authentication er aktivert (påkrevd fra Sept 2025).

Real-Time Dashboard:

• Eventhouse data residency: Velg Fabric capacity i Norway East/West for datasuverenitet.
• Data retention policies: Eventhouse støtter granular retention policies per table (påkrevd for Forvaltningsloven § 10 journalføring).

Schrems II og dataoverføring

• Live Metrics endpoints: live.applicationinsights.azure.com er hosted i Azure public cloud. For Schrems II compliance, bruk Azure Government (ingen Live Metrics) eller on-prem Azure Stack HCI med Azure Arc-enabled Application Insights.
• Real-Time Dashboard: Fabric eventhouses kan deployes til Norway regions med Microsoft EU Data Boundary compliance.

AI Act Article 72 - Logging

Real-Time Dashboards dekker AI Act Article 72 krav for "automatic recording of events (logs)":

• High-risk AI systems (Article 6): Real-Time Dashboard + Eventhouse retention ≥ 6 måneder.
• Audit trail: KQL queries mot Eventhouse gir immutable audit log for AI decisions.
• Incident response: Data Activator alerts sikrer rask respons på AI failures (Article 9 risk management).

Forvaltningsloven § 11 - Begrunnelse

Real-Time Dashboards kan kombineres med AI Foundry tracing til å bygge "begrunnelse for vedtak":

// Retrieve AI decision trace for specific case
AITraces
| where timestamp between (datetime(2026-02-05T10:00) .. datetime(2026-02-05T10:05))
| where customDimensions.caseId == "CASE-12345"
| project timestamp, operation_Name, promptTokens, completionTokens, groundednessScore
| order by timestamp asc

Dette gir sporbarhet for AI-baserte vedtak (påkrevd av Forvaltningsloven § 11).

---

Kostnad og lisensiering

Application Insights Live Metrics

Komponent	Kostnad	Merknad
Live Metrics streaming	Gratis	Ingen ekstra ingestion cost
Underlying telemetry	Standard App Insights pricing	$2.88/GB (pay-as-you-go)
Entra ID authentication	Inkludert i Entra ID P1/P2	Påkrevd fra Sept 2025

Optimaliseringstips:

• Lukk Live Metrics etter debugging (stopper streaming overhead på app)
• Bruk sampling for underlying telemetry (påvirker ikke Live Metrics, men reduserer ingestion cost)

Fabric Real-Time Intelligence

Komponent	Kostnad (estimat)	Merknad
Real-Time Dashboard	Inkludert i Fabric capacity	Ingen ekstra kostnad
Eventhouse storage	~$0.10/GB/måned	Time-series compressed
Eventstream compute	Inkludert i Fabric capacity	Avhenger av throughput
Data Activator	Separate SKU (preview pricing TBA)	Per reflex/trigger
Minimum Fabric capacity	F2 SKU (~$262/måned)	Kan skales opp/ned

Kostnadsmodell for eventhouse:

• Ingestion: Ingen ekstra kostnad (dekket av capacity)
• Storage: Compressed time-series (~10:1 compression ratio for telemetry)
• Query compute: CU usage avhenger av dashboard refresh rate og query complexity

Optimaliseringstips:

• Bruk table update policies for pre-aggregation (reduserer query compute)
• Set dashboard refresh til 30-60s for non-critical metrics (reduserer CU usage)
• Enable caching for ofte-brukte queries (cache retention 5 min - 1 time)
• Bruk materialized views for expensive aggregations (calculate once, query mange)

Total Cost of Ownership (TCO) Eksempel

Scenario: AI-tjeneste med 1M requests/dag, 5 GB telemetry/dag.

Option 1: Live Metrics only

• App Insights ingestion: 5 GB/dag × 30 dager × $2.88/GB = $432/måned
• Live Metrics: $0/måned
• Total: $432/måned

Option 2: Real-Time Dashboard + Eventhouse

• App Insights ingestion: $432/måned
• Eventhouse storage: 150 GB × $0.10 = $15/måned
• Fabric F2 capacity: $262/måned
• Total: $709/måned

Trade-off:

• +65% cost for Real-Time Dashboard, MEN får persistent storage, alerting, multi-source dashboards.
• For enterprise workloads med SLA requirements: Real-Time Dashboard ROI gjennom redusert downtime.

---

For arkitekten (Cosmo)

Spørsmål å stille

1. Hva er primary use case for real-time monitoring?

   • Debugging/deployment validation → Live Metrics
   • Production operations med SLA → Real-Time Dashboard
   • Begge → Hybrid approach

2. Hva er akseptabel latency for monitoring?

   • < 1 sekund → Live Metrics
   • 10 sekunder - 1 minutt → Real-Time Dashboard
   • Varierer per metric → Kombiner begge

3. Er data retention påkrevd (compliance, audit)?

   • Nei → Live Metrics sufficient
   • Ja → Real-Time Dashboard med Eventhouse

4. Har dere eksisterende Fabric capacity?

   • Ja → Legg til Real-Time Dashboard (no extra infra cost)
   • Nei → Vurder cost/benefit mot managed App Insights only

5. Trenger dere alerting basert på real-time metrics?

   • Ja → Data Activator (krever Real-Time Dashboard)
   • Nei → Live Metrics eller standard Azure Monitor alerts (1-5 min latency)

6. Hvor mange datakilder skal monitors?

   • Kun én app → Live Metrics
   • Multiple apps/services → Real-Time Dashboard (unified view)

7. Hvem er primary audience for dashboards?

   • Developers → Live Metrics (ad-hoc debugging)
   • Operations/business → Real-Time Dashboard (shareable, no-code)

8. Er Schrems II compliance påkrevd?

   • Ja → Fabric eventhouse i Norway regions
   • Nei → Standard Application Insights OK

Fallgruver

⚠️ Over-reliance på Live Metrics for production Live Metrics er designet for debugging, ikke production monitoring. Mangler alerting og persistering.

⚠️ Underestimere Fabric capacity krav Real-Time Dashboard krever minimum F2 SKU. Start med F2, skaler opp hvis CU throttling.

⚠️ Ignorere API key deprecation (Sept 2025) Migrer til Entra ID authentication for Live Metrics control channel NÅ, ikke vent til deadline.

⚠️ Sette for aggressive refresh rates 10s refresh på alle dashboards gir høy CU cost. Bruk 30-60s for de fleste metrics.

⚠️ Blande real-time streaming med batch ETL Real-Time Dashboard er IKKE erstatning for data warehouse. Bruk for operational monitoring, ikke business analytics.

Anbefalinger per modenhetsnivå

Nivå 1 - Proof of Concept:

• Start med Live Metrics (gratis, zero-config)
• Enable for ASP.NET Core/Java/Python apps (enabled by default med OpenTelemetry)
• Bruk under deployment validation

Nivå 2 - Pilot (produksjon med begrenset scope):

• Introduser Real-Time Dashboard for critical services
• Deploy Eventhouse i Norway region (GDPR compliance)
• Setup Data Activator for 2-3 critical alerts (error rate, latency)
• Start med F2 capacity, monitor CU usage

Nivå 3 - Production (enterprise-scale):

• Hybrid approach: Live Metrics for developers + Real-Time Dashboard for ops
• Multi-source dashboards (App Insights + Azure Monitor + custom events)
• Materialized views for expensive aggregations
• Entra ID authentication for Live Metrics control channel
• KQL alert queries med anomaly detection (series_decompose_anomalies)

Nivå 4 - Optimalisert:

• Custom Eventstream pipelines med pre-aggregation
• Dedicated Fabric capacity (F8+) for high-throughput
• Automated dashboard generation med Copilot
• Integration med Power BI for business stakeholder reporting
• Cross-region replication av Eventhouse (disaster recovery)

---

Kilder og verifisering

Microsoft Learn (Verified via MCP)

1. Live Metrics: Monitor and diagnose with 1-second latency https://learn.microsoft.com/en-us/azure/azure-monitor/app/live-stream Confidence: Verified (Feb 2026) - Authoritative doc for Live Metrics

2. What is Real-Time Dashboard? https://learn.microsoft.com/en-us/fabric/real-time-intelligence/real-time-dashboards-overview Confidence: Verified (Feb 2026) - Fabric Real-Time Intelligence GA features

3. Configure Azure Monitor OpenTelemetry https://learn.microsoft.com/en-us/azure/azure-monitor/app/opentelemetry-configuration#live-metrics Confidence: Verified (Feb 2026) - OpenTelemetry Live Metrics config

4. Troubleshoot Live Metrics issues https://learn.microsoft.com/en-us/troubleshoot/azure/azure-monitor/app-insights/troubleshoot-live-metrics Confidence: Verified (Feb 2026) - Firewall, TLS requirements

5. Monitor .NET and Node.js applications with Application Insights (Classic API) https://learn.microsoft.com/en-us/azure/azure-monitor/app/classic-api#collecting-telemetry-data Confidence: Verified (Feb 2026) - Manual Live Metrics setup (legacy)

6. Build real-time monitoring and observable systems for media https://learn.microsoft.com/en-us/azure/architecture/example-scenario/monitoring/monitoring-observable-systems-media Confidence: Verified (Feb 2026) - Real-time architecture patterns

7. Observability in generative AI https://learn.microsoft.com/en-us/azure/ai-foundry/concepts/observability Confidence: Verified (Feb 2026) - AI Foundry monitoring integration

8. Implement advanced monitoring for Azure OpenAI through a gateway https://learn.microsoft.com/en-us/azure/architecture/ai-ml/guide/azure-openai-gateway-monitoring#near-real-time-monitoring Confidence: Verified (Feb 2026) - Near real-time vs batch monitoring trade-offs

Confidence per seksjon

Seksjon	Confidence	Kilde
Application Insights Live Metrics	Verified	MCP fetch: live-stream doc
Fabric Real-Time Dashboard	Verified	MCP fetch: real-time-dashboards-overview
Arkitekturmønstre	Baseline	Synthesized fra multiple MCP sources
Code samples	Verified	MCP code search: C# Live Metrics setup
Kostnad og lisensiering	Baseline	Pricing calculated fra public Azure pricing (Feb 2026)
Offentlig sektor compliance	Baseline	Applied GDPR/AI Act principles til verified features

MCP calls: 6 (3 × search, 2 × fetch, 1 × code search) Unique sources: 8 Microsoft Learn URLs Last verified: 2026-02-05