Controlling Data Overhead with the OpenTelemetry Collector
summary
Shows concrete Collector configs and choices: filtering PII, clamping uncontrolled metrics, and using processors for normalization, sampling and metric transforms (e.g. marking error counts as cumulative). Describes batching, memory limits and an alert pipeline that emits logs from host network metrics. Notes high-cardinality user-path explosions and that very few full traces are inspected. Gives a production lesson: Doordash rewrites Collector batching/queueing in four variants and cuts CPU overhead dramatically, demonstrating when to push logic into the Collector vs SDK.
AI-written summary · about
chapters (10)
- 0:00Introduction
- 1:07Talk overview
- 3:14OpenTelemetry Collector overview
- 8:49Instrumentation tips
- 10:14Collector config example
- 12:03Network usage monitoring
- 14:10Cardinality explained
- 18:21Tracing costs
- 21:50Baggage and future uses
- 22:54Q&A
related talks
28mControlling data overhead from the OpenTelmetry Collector
16mUsing the OpenTelemetry Collector to Empower End Users
01h 03mGet a Grip on your Telemetry using the OpenTelemetry Collector
25mOpenTelemetry Collector Deployment Patterns
27mTricked Out Traces
31mTroubleshooting and Monitoring Production Applications using OpenTelemetry
29mMastering OpenTelemetry Collector Configuration
13mManaging Observability Data at the Edge with the OpenTelemetry Collector and OTTL