Should we start with logs, metrics, or traces?

If you’re firefighting latency and timeouts, start with traces at the edge plus your top 5 flows. Metrics are your SLO backbone, logs tell stories, but traces explain where time went. Then stitch them together with exemplars and log links.

Is vendor X better than self-hosted?

Depends on your team. If you have K8s + S3 and Grafana already, Tempo is cost-effective and scalable. If you need strong UI and event-based analysis, Honeycomb shines. Datadog is convenient if you’re already in the ecosystem. The key is tail-based sampling and consistent propagation—those matter more than the logo.

Can we avoid code changes entirely?

You can go far with auto-instrumentation and mesh, but you’ll miss DB calls, queue hops, and business context. Plan for small, targeted code changes (2–3 manual spans and attributes) in critical services.

How do we keep costs sane?

Tail-based sampling (keep errors and slow, sample 1–5% of normal), batch processors in the Collector, and attribute scrubbing. Monitor `otelcol_exporter_sent_spans` and ingestion volume; set budgets per env (`deployment.environment`).

Guides · Nov 13, 2025 · 10 minute read

The Tracing Rollout That Finally Stuck: OpenTelemetry + Collector + Tail Sampling in K8s

You don’t need another tracing pilot. You need traces that answer pager questions in minutes, not hours—without bankrupting your ingest bill.

Back to all posts

The Tracing Rollout That Finally Stuck: OpenTelemetry + Collector + Tail Sampling in K8s

Key takeaways

Implementation checklist