A controlled, experiment-first testbed for measuring the real performance cost and debugging value of observability in a containerized NestJS microservice system. The project is designed with publication-grade statistical rigor and delivered as a fully reproducible research artifact.

Highlights

• Five additive observability modes, from none to metrics, logs, traces, and the full OpenTelemetry pipeline, controlled through a single OBS_MODE switch and measured against a true uninstrumented baseline.

• Primary RQ1 result: a realistic open-model load campaign with N=10 per mode, analyzed using non-parametric methods: bootstrap 95% confidence intervals, Kruskal–Wallis, Mann–Whitney U, and Cliff’s δ.

• Metrics are statistically indistinguishable from the baseline. Structured logging is the dominant cost, with CPU usage increasing by 164% and median latency by 177%. The batched OTLP pipeline remains smooth despite carrying the most telemetry.

• Objective RQ2 detection result, measured through MTTD: a fault is automatically undetectable without a metrics pipeline and detected within approximately one scrape interval when metrics are enabled.

• 1M-row PostgreSQL and MongoDB indexing experiments with bootstrap confidence intervals, including a SQL-vs-NoSQL structural comparison.

• Orchestration comparison across Docker Compose, Docker Swarm, and Kubernetes.

• Reproducibility package including environment capture with image digests, a one-command-per-result reproduction guide, MIT and CC-BY-4.0 licensing, CITATION.cff, and Zenodo metadata.

Stack

pnpm monorepo, strict TypeScript, 52 tests, ESLint, Prettier, dependency-free SVG charts, and Docker Compose profiles.