Public corpus refresh¶
The public corpus on published-results goes stale when nobody submits,
not when CI stops running. Nightly tests do not publish bundles. This page
is the publication calendar.
Cadence¶
Monthly, on the 1st at 07:00 UTC, via .github/workflows/seed-corpus.yml
(schedule plus the existing workflow_dispatch). Each run opens a PR
against develop. Merging that PR triggers
sync-results-data-to-published.yml, which opens a draft mirror onto
published-results for a maintainer to accept.
Do not add a second nightly UAT gate for this. Generation already lives in
seed-corpus.yml; the schedule is what was missing.
Recurring matrix¶
This is the seed-corpus.yml matrix. It is DuckDB / DataFusion / Polars
(and optional ClickHouse Cloud when secrets exist). Cloud warehouses
(Snowflake, Databricks, BigQuery) are not on this calendar until live
credentials are confirmed.
Benchmark |
Scales |
Platforms |
Notes |
|---|---|---|---|
TPC-H |
0.01, 0.1, 1.0 |
DuckDB, DataFusion, Polars-DF (SF 0.01/0.1); DuckDB, DataFusion, ClickHouse Cloud optional (SF 1.0) |
Power phase only |
TPC-DS |
1 |
DuckDB, DataFusion |
Integer SF only. Runs with |
SSB |
0.01, 0.1 |
DuckDB, DataFusion, Polars-DF |
Power phase only |
TPC-H SF10 / SF100 and extra TPC-DS platforms stay operator-run (too large
for a GitHub-hosted monthly job). They still publish through
benchbox submit onto published-results, as in PR #1786. Those direct
operator submissions remain archive-only unless a separate reviewed PR
promotes them into develop’s curated release-preview corpus.
Operator trigger¶
gh workflow run seed-corpus.yml
# optional: -f benchmark=tpch
After the develop PR merges, confirm the mirror draft against
published-results and merge it. corpus-drift-check.yml remains the
loud canary if a push-triggered mirror is dropped.