Skip to main content

Kubernetes RED metrics and OpenTelemetry metrics

Avuru Obs derives service RED metrics from ingested traces and accepts OTLP metrics from collectors and SDKs. These are different sources: a trace-derived rate describes the requests observed in spans, while an infrastructure metric describes a measurement collected from a node, container or application.

Read service rate, errors and duration​

Open a service's RED dashboard and choose a time window:

  • Rate: how much traced request traffic arrived over time.
  • Errors: failures in that traffic. Server-side 4xx refusals are shown separately; see trace outcomes.
  • Duration: latency distributions and percentiles, rather than an average alone.

If the service looks slow, open its traces in the same window and compare an individual request. A p95 spike identifies a period to investigate; it does not identify the cause. Sampling and missing spans affect what can be inferred from trace-derived metrics.

Add infrastructure and application metrics​

Enable the infrastructure metrics module and configure the relevant collection sources. The gateway accepts OTLP metrics over HTTP or gRPC; see OTLP setup and modules. Use a stable service.name and the correct project identity for application metrics.

OTel instruments include counters, gauges, histograms and up/down counters. Their aggregation and temporality matter: do not interpret a cumulative counter as an instantaneous rate. Verify the exporter and gateway configuration for the metric family you need. Not every ingested metric has a dedicated UI chart.

Compare CPU requests with usage​

The optional cost module combines Kubernetes reservations with observed usage. Compare requests with the observed peak, keeping the window in view. A historical peak is evidence, not a guarantee of future demand. An average alone is not a safe resize target.

Check an unexpected result​

Confirm the project, service, time window and signal source before concluding that a zero means inactivity. Missing collection is different from a measured zero. For a practical investigation, follow the slow checkout guide.