TPC-H benchmark

The tpch-smoke job in cores.yml generates TPC-H at scale factor 10 with tpchgen-cli, ingests all 8 tables from Parquet into Delta through duckrun's write path (conn.read.parquet(...).write.saveAsTable(...)), then runs the 22 TPC-H queries through conn.sql over delta_scan — timing each. It runs on a standard GitHub-hosted runner on every push, and the scorecard below is rendered live and committed on every push to main.

It is a coverage + cost check, not a speed contest: the ingestion time is duckrun's write path, but the 22 query times are DuckDB reading Delta with no second engine to compare against — so read them as "the whole schema loads and all 22 queries run at SF=10", not a "duckrun is fast" claim.

🐤 TPC-H benchmark — duckrun on Delta Lake

What this checks: duckrun ingests the full TPC-H schema (8 tables) from Parquet into Delta through its write path (conn.read.parquet(...).write.saveAsTable(...)), then runs the 22 TPC-H queries through conn.sql over delta_scan. The ingestion time is duckrun's write path; the query times are DuckDB reading Delta — there is no second engine to race here, so read them as "the whole schema loads and all 22 queries run at this scale", not a duckrun is fast claim.

Ingest 8 tables in 45.6s · run 22 queries in 24.8s · SF 10 · 86.6M rows · 4 cores

Setup

Engine	duckrun · DuckDB 1.5.4 · delta_rs 1.5.0
Scale factor	10
Runner	GitHub-hosted · 4 cores

Ingestion — Parquet → Delta (duckrun write path)

Table	Rows	Write (s)
nation	25	0.97
region	5	0.01
customer	1,500,000	1.22
supplier	100,000	0.10
lineitem	59,986,052	30.07
orders	15,000,000	7.16
partsupp	8,000,000	4.85
part	2,000,000	1.23
Total	86,586,082	45.60

Queries — 22 TPC-H over delta_scan

Query	Duration (s)
Q01	1.425
Q02	0.360
Q03	0.900
Q04	0.630
Q05	1.550
Q06	0.440
Q07	1.034
Q08	1.253
Q09	2.055
Q10	1.363
Q11	0.159
Q12	0.595
Q13	1.336
Q14	0.783
Q15	0.513
Q16	0.316
Q17	2.322
Q18	1.526
Q19	1.149
Q20	1.168
Q21	3.422
Q22	0.534
Total	24.83