Much telemetry is bad: noisy, costly, and hides real signals. Auto-instrumentation can create high-cardinality floods, while over-instrumentation results in excessively verbose traces with countless near-useless internal spans. These common issues obscure insights and strain collection pipelines, leading to frustration and wasted resources.

This session dissects these "bad telemetry" patterns using real-world examples (like verbose traces and auto-instrumentation noise at global scale). We'll demonstrate concrete fixes both at the source, and using OpenTelemetry Collector configurations—like filter, attributes, span, and transform processors—to tame this data at the pipeline level, improving quality and utility.

Attendees will learn to identify bad telemetry, apply OTel best practices to create, filter, and reshape data, and cultivate higher-quality, cost-effective, actionable insights.