zhengruifeng commented on PR #56200: URL: https://github.com/apache/spark/pull/56200#issuecomment-4581143719
### CI performance: before vs after **Samples:** BEFORE = 3 `build_uds` scheduled runs (2026-05-04, 2026-05-19, 2026-05-25 — `precompile` job was already running but `tpcds-1g` not yet consuming it). AFTER = 1 PR run (2026-05-29, this PR). | Step | BEFORE avg (n=3) | AFTER | Savings | |---|---:|---:|---:| | **Generate TPC-DS data** (SBT compile + data gen) | **9m28s** | **2m44s** | **~6m44s** | | Total job wall-time | 86m56s | 81m06s | **~5m50s (~7%)** | Samples: - BEFORE-1: [apache/spark run 26379069496](https://github.com/apache/spark/actions/runs/26379069496) (2026-05-25, `build_uds` scheduled) - BEFORE-2: [apache/spark run 26071135105](https://github.com/apache/spark/actions/runs/26071135105) (2026-05-19, `build_uds` scheduled) - BEFORE-3: [apache/spark run 25297013619](https://github.com/apache/spark/actions/runs/25297013619) (2026-05-04, `build_uds` scheduled) - AFTER: [zhengruifeng/spark run 26632958957](https://github.com/zhengruifeng/spark/actions/runs/26632958957) (2026-05-29, this PR) ### Where the savings come from Before this PR, the `tpcds-1g` job's first SBT invocation is `sql/Test/runMain GenTPCDSData`, but it also drives a full `sql/core` compile from scratch before the data generator can run. That is why the "Generate TPC-DS data" step took ~9.5 min even when no new data is needed -- roughly 7 of those minutes was SBT compile. With the precompile artifact, SBT sees the pre-built `target/` and goes straight to data generation: the step drops to 2m44s, a ~6m44s reduction that is very consistent across the 3 BEFORE samples (9m24s--9m35s). ### Test phase -- unaffected (as expected) | Phase | BEFORE avg | AFTER | Delta | |---|---:|---:|---:| | Sort merge join | 18m29s | 18m31s | +0m02s | | Broadcast hash join | 7m04s | 7m05s | +0m01s | | Shuffled hash join | 18m47s | 17m54s | −0m53s | | Collated data | 31m26s | 31m04s | −0m22s | All within CI noise -- the precompile artifact has no effect on SQL query execution speed. ### Bottom line ~7% wall-clock reduction per `tpcds-1g` run (~5m50s), driven by eliminating the duplicate `sql/core` compile that the shared precompile job already produced. Like the docker/k8s jobs in [SPARK-57069](https://issues.apache.org/jira/browse/SPARK-57069), the savings are real and stable but partially masked by the long (~75min) test-dominated job. -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected] --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
