Skip to content

Unified Observability Architecture

Last reviewed: August 2026

In a multicloud environment, each vendor provides its own observability tools.

Vendor Logs Metrics Traces
AWS CloudWatch Logs CloudWatch Metrics X-Ray
Azure Azure Monitor Logs Azure Monitor Metrics Application Insights
Google Cloud Cloud Logging Cloud Monitoring Cloud Trace
OCI OCI Logging OCI Monitoring OCI APM

Operating these tools separately causes:

  • Silos — When a single request crosses multiple clouds, tracking the full flow is difficult
  • Duplicate cost — Licensing, storage, and training costs for each platform
  • Lack of consistency — Dashboards and alerts are scattered, confusing the operations team
  • Vendor lock-in — Deep reliance on a specific tool increases switching cost

OpenTelemetry is a CNCF project that provides a vendor-neutral observability standard. It collects logs/metrics/traces in a unified way.

graph LR
    A1[AWS workload] --> O[OpenTelemetry<br/>Collector]
    A2[Azure workload] --> O
    A3[Google Cloud workload] --> O
    A4[OCI workload] --> O
    O --> B[Central observability<br/>platform]
  • Language SDK — Instrumentation libraries for major languages (Java, Python, Go, JavaScript, .NET, etc.)
  • Collector — An agent that collects, processes, and forwards data
  • Semantic Conventions — Standard attribute names/formats (e.g., http.method, service.name)

Official vendor support:

Send data from each cloud’s workloads to a single central platform.

graph LR
    A[AWS] --> C[Central platform]
    B[Azure] --> C
    G[Google Cloud] --> C
    O[OCI] --> C
  • Advantages — A single dashboard, cross-cloud correlation analysis
  • Disadvantages — The central platform is a single point of failure; data movement cost
  • Use — Common in general multicloud operating organizations

Keep data in each cloud, and query multiple sources simultaneously at query time.

graph LR
    U[Operator·query] --> F[Query federation layer]
    F --> A[AWS logs]
    F --> B[Azure logs]
    F --> G[Google Cloud logs]
  • Advantages — No data movement, saves egress cost
  • Disadvantages — Query latency, requires a federation engine
  • Use — Cases with strict data sovereignty requirements (supported by tools like Grafana)

Only critical metrics are centrally aggregated; detailed logs stay in place.

Most organizations use a 3rd-party platform to unify observability across multiple clouds.

Platform Characteristics Notes
Datadog Unified dashboards, broad integrations, strong APM SaaS-focused
New Relic Full-stack APM, usage-based pricing SaaS
Dynatrace AI-based automatic anomaly detection (Davis AI) Enterprise
Splunk Strong log analysis, integrated security analytics (SIEM) Enterprise
Elastic Observability Open-source based, flexible deployment Self-hosting possible
Grafana Cloud Managed Prometheus/Loki/Tempo OpenTelemetry-friendly

If cloud portability and cost control matter, you can build an open-source stack yourself.

Area Open source
Metrics Prometheus, Thanos, VictoriaMetrics
Logs Elasticsearch/OpenSearch, Loki
Traces Jaeger, Tempo
Dashboards Grafana
Collectors OpenTelemetry Collector, Fluent Bit

CNCF’s Cloud Native Landscape — Observability organizes the entire ecosystem.

Observability cost is generally proportional to volume collected (GB) and retention period (days). In a multicloud setting, you also need to consider egress cost.

  • Log volume — Can vary by 10x depending on the application log level (DEBUG vs ERROR)
  • Metric cardinality — More tag combinations increase storage cost (e.g., per-user-ID metrics)
  • Trace sampling — Storing only 1-10% of all traces can still be sufficient for analysis
  • Cross-cloud egress — With the Fan-in pattern, several TB of data can move each month
  • Sampling — Keep only representative traces
  • Compression/tiering — Move older logs to cheaper storage
  • Filtering — Remove unnecessary logs at the collection stage
  • Aggregation — Send only aggregated metrics centrally, instead of raw logs
  • In-region processing — Aggregate within a region and forward only the metrics when possible

Multicloud Unified Monitoring (Single Pane of Glass)

Section titled “Multicloud Unified Monitoring (Single Pane of Glass)”

Looking at AWS CloudWatch, Azure Monitor, and Google Cloud Cloud Monitoring separately is inefficient. In a multicloud environment, you need a unified dashboard that shows the state of every cloud from one place.

Approach Description Tools
OpenTelemetry standardization Vendor-neutral instrumentation → collected into a single backend OTel Collector + Grafana/Datadog
3rd-party unified platform Metrics/logs from every vendor into a single SaaS Datadog, New Relic, Dynatrace, Splunk
Open-source stack Self-operated, no vendor lock-in Prometheus + Grafana + Loki + Tempo

Considerations when building unified monitoring:

  • Convert each vendor’s native metrics to OTel or Prometheus format
  • Route alerts to a single channel (PagerDuty, Opsgenie)
  • Standardize tags/labels so the dashboard can filter by vendor
  • Cost: 3rd-party SaaS bills by collection volume, so log volume needs to be managed

Items to check when adopting multicloud observability:

  • Use the OpenTelemetry standard to avoid vendor-locked instrumentation
  • Define common tagging conventions such as service.name, environment
  • Set a trace sampling policy (head sampling / tail sampling)
  • Establish collection/retention policy by log level (e.g., ERROR 90 days, INFO 7 days)
  • Keep cloud-native metrics (CPU, network) in the vendor tool
  • Consolidate application metrics/traces into the central platform
  • Define SLOs and build an error budget dashboard (see SLI/SLO and Error Budgets)
  • Standardize alert routing (single integration into PagerDuty, Opsgenie, etc.)
  • Monitor cost (the cost of the observability platform itself)
  • Regular dashboard/alert reviews — Check quarterly whether dashboards reflect the current architecture.
  • Remove alert noise — Remove ignored alerts or adjust thresholds. Alert fatigue causes real incidents to be missed.
  • SLO-based alert tuning — Switching to alerts based on error budget burn rate reduces noise.
  • Sending all logs to the central platform at DEBUG level — Collection cost explodes. Establish a log-level policy per environment (WARN and above in production).
  • Using high-cardinality values (user ID, request ID) in metric tags — Time-series explosion drastically worsens storage cost and query performance.
  • Running traces at 100% sampling — Most normal requests have low analytical value. Tail sampling (storing only errors/slow requests) can cut costs by 90%+.