Observability in CNCF

In theory, non-functional requirements such as Security and Observability could also be fully described in prompts and specifications.

In practice, however, even observability alone requires many cross-cutting concerns such as distributed tracing, metrics, structured diagnostics, correlation IDs, payload protection, confidentiality masking, and integration with external observability platforms.

If these concerns are repeatedly delegated to individually generated implementations, implementation and operational complexity rapidly increase.

In SimpleModeling, CML and CNCF absorb these quality attributes so that AI can focus on developing domain models and domain logic.

Another major issue is the operational cost of generative AI itself.

If cross-cutting quality attributes such as Security and Observability must be described repeatedly in prompts and specifications, prompt and specification size rapidly grows.

As concerns such as traces, metrics, correlation, payload protection, redaction, export, retry handling, and backend integration begin interacting, implementation and operational complexity grow almost exponentially.

As a result, the amount of context required for generative AI also explodes, significantly increasing generation, review, and regeneration costs.

Moreover, when generative AI encounters requirements it cannot fully understand or requirements that become too complex, it tends to omit, simplify, or overlook them.

For this reason, approaches that attempt to push everything into prompts and specifications make quality assurance itself unstable.

CML&CNCF addresses this problem through structured absorption by the model compiler and execution framework.

CNCF automatically collects observability information within the runtime and provides an operational observability environment, including dashboards.

However, when considering observability across the entire system including CNCF, an operational management function is needed to aggregate observability information emitted by each system.

For this purpose, CNCF supports integration with OpenTelemetry, the industry-standard observability framework.

Inside CNCF, structured execution information is maintained for:

By integrating with OpenTelemetry, the observability information collected by CNCF can be viewed through dedicated tools such as Jaeger, Prometheus, and Grafana as part of the overall system observability data.

This section explains observability in CNCF using a tutorial example.

The tutorial package located below is used in this example.

After extraction, the sample code is located in the following directory.

This sample verifies the smallest possible trace-export configuration.

Configuration:

CNCF sends traces directly to Jaeger.

CNCF itself also provides native observability capabilities.

First, inspect the operation execution results using CNCF-native observability.

The recommended inspection order is as follows.

Next, verify the exported trace in Jaeger.

Search condition:

Verification points:

The previous example used only a minimal Jaeger-based observability configuration.

This example uses a full observability stack consisting of OpenTelemetry Collector, Jaeger, Prometheus, and Grafana.

This topology is also suitable for production-oriented deployments.

After extracting the ZIP archive, the sample code is located in the following directory.

This sample demonstrates a standard Collector-centered observability topology.

The observability system topology is shown below.

CNCF sends observability data to the OpenTelemetry Collector, which then propagates the information to Jaeger and Prometheus.

Grafana then retrieves information from Jaeger and Prometheus and visualizes it through dashboards.

Start the system as follows.

Generate observable runtime activity using the following commands.

The recommended inspection order is as follows.

In AI-driven development, it is not realistic to repeatedly describe cross-cutting quality attributes such as Security and Observability in exhaustive prompts and specifications.

As specifications grow, the amount of context required for generative AI, as well as generation, review, and regeneration costs, rapidly increase, destabilizing quality assurance itself.

Furthermore, when generative AI encounters requirements that are too complex or insufficiently understood, it tends to omit, simplify, or overlook them.

For this reason, CML&CNCF absorbs Security and Observability as runtime capabilities provided by the model compiler and execution framework rather than delegating them to application code.

CNCF also treats OpenTelemetry not as the primary observability system, but as an export boundary while maintaining CNCF Native Observability as the authoritative source of diagnostic truth.

The samples in this article demonstrated how the CNCF runtime integrates observability capabilities, from a minimal Jaeger-only setup to a full Collector-centered observability stack.