Performance and resource bounds#
Performance claims need a stated workload. The repository includes a synthetic benchmark for fleet accounting, separately from retrieval and model fitting:
uv run --no-sync python benchmarks/fleet_accounting.py \
--sites 787 --hours 2207 --repeats 3 \
--output artifacts/fleet-accounting-benchmark.json
The default performs accounting for 1,736,909 synthetic site-hours. It validates matched energy against an independent scalar calculation, then reports timings, throughput, Python/package versions and machine context. It reuses one site’s hourly frame and retains a summary row per site. It does not measure national downloads, disk I/O, geometric optimization, TDP or model training. This is not a benchmark of the historical production dataset.
The architecture uses NumPy/pandas operations within a site and bounded per-site processing for studies. Network retrieval has finite response sizes, timeouts, retries and request spacing. Shared public weather cells can be retrieved once and mapped to many sites; this does not make them independent on-site measurements. Native ERA5 access is optional and selects the relevant coordinates/time span rather than downloading the global archive.
LightGBM’s worker count is explicit and bounded. CI sets BLAS/OpenMP thread counts to one so multiple numerical libraries do not compete for the same cores. Set the same limits when launching multiple per-site workers; prefer one layer of parallelism. Large data and generated artifacts are not loaded as part of importing a package.
Timing thresholds are not a correctness gate on shared CI runners. Before changing a hot path, rerun the same benchmark on the same machine and inspect both timing and numerical equivalence. Record production workloads separately.