You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Get nightly integrated signal on in-flight Velox and Presto PRs, before they merge upstream — so we find breakage and measure performance on the combined set early instead of after the fact.
Three pieces:
1. Staging branch creation in CI — Avinash
Select the PRs we care about from a GitHub project board libcudf (view) (a field on the board marks a PR as "Velox Staging"), merge them into a staging branch for Velox and Presto, and run it nightly. On merge conflict, fail and report.
We want the board as the control surface: moving a card is how someone adds or removes a PR from staging. No label access is needed on upstream repos.
use rapidsai/velox for staging branch
For presto, use karthikeyann/presto for now. IF zero staging PRs on presto, then use prestodb/presto master.
2. Presto and Velox image build and publish — Matt
Build from those staging branches and publish images under a stable staging tag, so anyone can pull the current staging state without rebuilding.
3. SF1K benchmark automation — Karthik
On a dedicated GPU machine, pull the published staging images and run TPC-H at SF1000 on a schedule, and publish the results so regressions are visible.
Why
Today staging branches are created manually and nothing builds or benchmarks them. Integration problems between in-flight PRs surface late, and we have no continuous SF1K performance signal.
Goal
Get nightly integrated signal on in-flight Velox and Presto PRs, before they merge upstream — so we find breakage and measure performance on the combined set early instead of after the fact.
Three pieces:
1. Staging branch creation in CI — Avinash
Select the PRs we care about from a GitHub project board libcudf (view) (a field on the board marks a PR as "Velox Staging"), merge them into a staging branch for Velox and Presto, and run it nightly. On merge conflict, fail and report.
We want the board as the control surface: moving a card is how someone adds or removes a PR from staging. No label access is needed on upstream repos.
use rapidsai/velox for staging branch
For presto, use karthikeyann/presto for now. IF zero staging PRs on presto, then use prestodb/presto master.
2. Presto and Velox image build and publish — Matt
Build from those staging branches and publish images under a stable staging tag, so anyone can pull the current staging state without rebuilding.
3. SF1K benchmark automation — Karthik
On a dedicated GPU machine, pull the published staging images and run TPC-H at SF1000 on a schedule, and publish the results so regressions are visible.
Why
Today staging branches are created manually and nothing builds or benchmarks them. Integration problems between in-flight PRs surface late, and we have no continuous SF1K performance signal.