§4.1 · Measurement windows
Three windows analysed, each ≥ 14 days of continuous operation:
| Window | Period | Conditions |
|---|---|---|
| W1 | TBD | Steady-state (no induced failures) |
| W2 | TBD | One scheduled WAN partition (≤ 30 min) |
| W3 | TBD | One simulated hub power loss (UPS-only) |
DRAFT — window dates land once §3.5 calibration is signed off.
§4.2 · Latency
- Edge path: publish → derived metric (Pi-local timestamp difference).
- Cloud path: publish →
recorded_atin Aurora minus publish ts. - Both bucketed in 100 ms bins, plotted as eCDFs (NOT means — the tail matters).
DRAFT — eCDFs pending. Hypothesis (§3.2 H₀): no meaningful tail divergence under steady state.
§4.3 · Energy
- Edge: Wh from a Pi wall-meter, sampled at 10 Hz.
- Cloud: Lambda invocation count × billed-MB-ms + upstream MQTT bytes converted to Wh via the AWS-published efficiency coefficient for eu-central-1 (footnote on coefficient sourcing).
DRAFT — calibration is the open item; need to validate that the AWS coefficient is applicable to one operator at this scale.
§4.4 · Monetary cost
Per-month, normalised to 1 sensor node:
- Edge: amortised hardware + electricity (PL Tauron 2026 prices, documented in dataset).
- Cloud: AWS bill — Aurora ACU-hours + Lambda + RDS Data API + CloudWatch Logs + Secrets Manager.
Free-tier conditions documented; over-tier excluded from the steady-state analysis (per §3.5).
DRAFT — pricing snapshots taken at window end; reproducible from the dataset.
§4.5 · Data integrity
- Gap count per window per path.
- Total gap duration per window per path.
- Samples lost per window per path.
The methodology decision in §3.6 means gaps are recorded as gaps, not hidden behind interpolation. Both paths treat gaps the same way; the benchmark is symmetric.
DRAFT — INC-001 (the 4-day gap) is in the dataset; W1/W2/W3 windows are chosen to exclude it (separately analysed in §5).
§4.6 · Steady-state results
[One table per variable. Both paths’ median + tail values + Wilcoxon signed-rank p (or appropriate non-parametric test) for paired samples.]
DRAFT — results table layout locked; values populate by 2026-08.
§4.7 · Failure-condition results
[W2 and W3 results. Where do paths diverge? By how much? Operationally, which is more recoverable?]
DRAFT — populated after W2 / W3 windows complete.
Status: structural draft v0.1, 2026-05-07. Citation:
https://plantir.garden/thesis/2026/benchmarkis locked per ADR-011. Related ADRs:docs/adr/010-public-sensor-data-policy.md.