Abstract
Serverless computing can favor the emergence of complex and error-prone applications. In order to gain observability in such applications, distributed tracing can be used. However, as serverless computing relies on the pay-per-use billing model, utilizing distributed tracing tools can have a noticeable impact on the resulting costs. Therefore, this paper investigates the impact of distributed tracing in serverless applications by exploring and comparing the efficiency characteristics of three selected distributed tracing tools - Zipkin, OpenTelemetry, and SkyWalking. In particular, the runtime, the memory usage, and the initialization duration were examined by benchmarking AWS Lambda function invocations. In the experiments, Zipkin imposed the lowest runtime overhead with an average of 10.73 %, while SkyWalking introduced the highest overhead with an average runtime overhead of 50.67 %. OpenTelemetry added 24.19 % additional runtime. Besides runtime overheads, significantly higher memory usage and initialization durations were detected for all tools. Therefore, the results suggest that distributed tracing can significantly impact the efficiency of serverless applications. Nevertheless, differences could be observed concerning tracing mechanisms and use cases. This helps developers to carefully select the most suitable tracing tool considering factors such as runtime overhead, memory usage, and initialization durations.