Umeå University's logo

Observability grants engineers the ability to comprehensively understand their systems by generating, exporting, and analyzing telemetry; logs, metrics, and traces. This is especially important for microservice-based systems as their distributed nature makes them notoriously difficult to debug. OpenTelemetry is an observability tool which has steadily risen in popularity since its inception in 2019, its purpose is to generate and export telemetry. This thesis investigates the performance impact of implementing OpenTelemetry in a microservice-based system as performance overhead can be a key concern for organizations. 

The experiments conducted in this thesis demonstrate that OpenTelemetry impacts performance, for instance, in terms of CPU overhead, which reached 42% under certain conditions. Compared to equivalent tools, OpenTelemetry performed better for tracing but slightly worse for metrics. The latter result might stem from the OpenTelemetry agent not being designed to be used solely for metric instrumentation. Furthermore, the results indicate that manually generating traces with OpenTelemetry incurs less overhead compared to using the agent for automatic trace instrumentation. This difference is most notable in terms of CPU overhead, where automatic instrumentation incurred roughly twice as much overhead. Finally, the experiments demonstrated that batching and head-based sampling can significantly reduce OpenTelemetry's performance impact. This reduction is most notable in terms of CPU and latency overhead, which decreased to 3.6% and 3.4% respectively.