Red Hat turns telemetry correlation into a Granite-assisted SRE quickstart

Red Hat has documented a deployable reference pipeline that correlates application logs and traces, then asks a Granite model to produce a root-cause analysis for an operator. The AI quickstart is an integration blueprint as much as an AI demo: OpenTelemetry supplies the signals, Kafka carries them, Apache Camel applications route and process them, and Infinispan groups events by trace ID before the model sees the incident context.

What the pipeline does

The design uses three Camel applications on OpenShift. A correlator consumes logs and traces from Kafka and groups related data in Infinispan. An analyzer sends that correlated context to a Granite LLM. A UI stores the resulting analyses and lets an operator inspect a trace or ask follow-up questions with a custom prompt, according to the Red Hat walkthrough.

That separation matters. Instead of presenting a model with an undifferentiated stream of telemetry, the pipeline performs deterministic collection and trace-level correlation first. The LLM is used after the system has narrowed the evidence to one transaction path. The source describes the result as actionable remediation guidance, but teams evaluating the pattern should still treat the generated explanation as a diagnostic lead rather than proof.

What practitioners can run

The quickstart targets the Red Hat Developer Sandbox and lists OpenShift 4.21 or later with Pipelines, a shared Granite model from OpenShift AI, Helm 3, oc, tkn and Git as its requirements. It says a regular sandbox user can deploy the components inside that user’s namespace without cluster-admin access. A create.sh script installs the infrastructure, builds the applications through Tekton and deploys them with Helm; a supplied log generator then produces simulated order failures for the analysis flow.

The linked repository snapshot also includes manual deployment, monitoring, troubleshooting and technical-design documents. Red Hat pins the documentation page to commit 2d0fb88; GitHub dates that commit to June 25, 2026. The public quickstart page itself does not display a publication date, so this desk is not treating the underlying code as an August release.

Where to be cautious

Red Hat labels the content as authored by its experts but not tested on every supported configuration. The demo’s own test traffic uses a 30% simulated failure rate, and the documented stack assumes the sandbox’s shared model service. Before adapting it to production, platform teams would need to define access controls for telemetry and prompts, validate the model’s output against incident evidence, and replace the demo’s deployment and test assumptions with their own operating controls.

The useful contribution is therefore not a claim that an LLM can diagnose every outage. It is a concrete example of where generative analysis fits in an observability pipeline: after trace correlation, with the source signals still available for an engineer to inspect.