Practical introduction to OpenTelemetry tracing by Nicolas Frankel

OpenTelemetry is a standardized approach for distributed tracing across different technologies and components, merging OpenTracing and OpenCensus standards

Two main approaches to implement OpenTelemetry:

• Auto-instrumentation (recommended for beginners) - uses runtime agents and requires no code changes
• Manual instrumentation - requires adding dependencies and code but offers more control

Tracing concepts:

• Traces follow requests across distributed system components
• Spans represent units of work within components
• Spans can have multiple children but only one parent
• Context propagation happens via HTTP headers (W3C trace context)

For asynchronous messaging:

• Need to pass trace context as metadata if supported
• Manually extract/inject trace context if metadata not supported
• Requires more manual work compared to HTTP

Key considerations for production:

• Sampling is crucial - don’t trace every request
• Start with low sampling rate (~10%) and increase based on confidence
• Consider performance impact and storage costs
• Configure proper centralized collection and storage

Compatible with multiple backends:

• Can use Jaeger, Zipkin or any OpenTelemetry-compatible collector
• Allows switching backends without changing application code
• OpenTelemetry Collector can forward to multiple destinations

Structured logging is preferred over unstructured:

• Produces data that can be directly indexed
• Avoids complex parsing and transformation
• More efficient for centralized logging systems