Querying results

Every run persists a 0.2 record under results/<domain>/<platform>/. The analytics layer flattens those records at read time into four SQL-queryable views, so measurements, findings, cost and time series from any cloud or product are queried the same way — domain and platform are just columns. csbench query (SQL) and any parquet/series read need the [store] extra:

pip install 'clousight-bench[store]'

csbench export to CSV/JSONL (for records / measurements / findings) works without it.

The four views

Only records whose record_digest verifies are included (tamper-evident); run plan aggregates are skipped. All four share run_id / domain / task_id / platform, so any join works. records — one row per run:

run_id, domain, task_id, platform, task_revision, scorer_revision,
status, started_at, finished_at,
benchmark_fp, environment_fp, implementation_fp, record_digest,
region, mode, execution, cost_usd, list_cost_usd, discount_usd

execution is simulated (a simulated runtime — mock cloud / local-sim / local-process) or live (real cloud), or unknown (migrated records). Simulated and live records are fingerprint-isolated — they get different environment fingerprints and are never pooled/compared. Filter real-cloud data with WHERE execution = 'live'. Cost comes from the pricing enricher (extensions["pricing"], NULL when the run was not priced). It is reported across three dimensions: list_cost_usd (vendor list price), discount_usd (what your discount layer saved), and cost_usd — the net, what finance actually pays.

Pricing: list / discount / net

The pricing enricher is the single cost authority. It multiplies a run’s usage by two independent, pluggable layers:

CLOUSIGHT_PRICING_DATA → the public list-price feed (vendor list prices; a small seed ships in the core).
CLOUSIGHT_PRICING_DISCOUNTS → your private discount layer (per-provider, optionally per-provider+service, pct, plus a default_pct). net = list × (1 − discount); absent → net == list.

See all three at once:

csbench query "SELECT platform, list_cost_usd, discount_usd, cost_usd AS net
               FROM records WHERE cost_usd IS NOT NULL"

measurements — one row per scored measurement:

run_id, domain, task_id, platform, benchmark_fp, environment_fp,
name, value_num, value_str, unit, evidence, aggregation, sample_count

Numeric values go to value_num; label/text values (e.g. recovery_mode = "auto-retry") go to value_str. findings — one row per finding:

run_id, domain, task_id, platform, code, severity, summary, evidence

series — the high-frequency time series (from each run’s series.parquet):

run_id, domain, task_id, platform, benchmark_fingerprint, series, t, value, unit

Examples

Compare a dimension across clouds:

csbench query "SELECT platform, avg(value_num) avg_cold_ms
               FROM measurements WHERE name='cold_start_ms'
               GROUP BY platform ORDER BY avg_cold_ms"

Cost rollup by product and platform (avoid cost as an alias — it is a DuckDB keyword):

csbench query "SELECT domain, platform, count(*) runs, sum(cost_usd) total_cost
               FROM records GROUP BY 1, 2"

Join a scalar to its time series:

csbench query "SELECT r.platform, s.t, s.value
               FROM series s JOIN records r USING (run_id)
               WHERE s.series='warm_start_ms'"

Shortcut for a whole table, no SQL:

csbench query --table measurements --where "task_id='T1.1'" --format csv

Exporting for external tools

Write a long table to parquet / csv / jsonl for a notebook or BI tool:

csbench export measurements --out ./measurements.parquet
csbench export records --out ./records.csv --format csv

Get started

Guides

Reference

The four views

Pricing: list / discount / net

Examples

Exporting for external tools

​The four views

​Pricing: list / discount / net

​Examples

​Exporting for external tools

The four views

Pricing: list / discount / net

Examples

Exporting for external tools