Result rows

One published measurement is one row. A row says what ran, on what, and how it went. The fields below are the ones every row carries; lanes add their own metrics.

field	meaning
backend	the engine and arm that ran, for example an in-process arm or a server arm
lane	the lane (tables, graphs, vectors, time series, cross-model, lifecycle)
scale / tier	the corpus size
workload	the task within the lane
rep	the repetition number
engine_version	the exact engine version; container runs also record the image digest
durability_class	the durability class the run used, where the engine offers more than one
mem_cap	the memory cap the run had
io_*_mib	the bytes the cell's containers read from and wrote to disk, per phase and in total, from the container's io.stat. Rows written on the benchmark machine before the fix for issue #555 counted a stacked block device (an NVMe disk and the device-mapper device on top of it) twice, so those figures are about twice their value. The fixed harness counts the device once and stamps io_stat_dedup: 1 into the row; the frozen table halves the io figures of the earlier rows, which lack the stamp, and lists each corrected row with its raw values in generated/io_stat_correction.json
error / oom_killed	what went wrong, if anything; a failure is a result
metric fields	latencies (median and percentiles), throughput, peak memory and the like, per lane
*_censored / *_budget_s	whether a query hit its time budget, and the budget

Rows are never edited. A correction adds a row; the history keeps both. The one correction made after the fact is the io figures above, which is applied in the frozen copy the tables are built from, never in the recorded rows.