Trace-Affinity for Scalable Observability Pipelines in Kubernetes Benchmarking

Modern microservice systems require scalability evaluation under controlled load and resource variations, yet trace-based observability is difficult to scale because telemetry data are often heterogeneous and trace reconstruction must correlate multiple events that describe the same request. In Kubernetes deployments, naive horizontal scaling of a trace-processing component can distribute trace fragments across replicas, leading to incomplete traces and fewer valid exported spans. This thesis proposes an affinity-preserving ingestion architecture that enforces trace locality by partitioning monitoring events using the trace identifier as key. The approach routes all events of a trace to the same processing instance while still allowing horizontal scaling across partitions, and exports reconstructed executions as OpenTelemetry spans for downstream analysis. Experiments with a microservice benchmark application vary workload and replica counts under an SLO-driven benchmarking methodology to assess throughput and trace correctness, showing that trace-affinity prevents reconstruction failures under horizontal scaling and enables reproducible, trace-informed scalability analysis.