Repository navigation
The benchmarks publish what a re-solve costs through Model.update, so the next re-solve work is decided on numbers #1899
Description
Prompt: "Do new issues for all 3!"
Note
The following content was generated by AI. Carried over from #700 (closed), whose "What remains" list held this item.
#700's slices have all landed: Model.update, a warm start across a rebuild, and start= (#1882, #1884). Its one open item is a measurement: what a sweep and an iterating driver cost on the update path today. #1840 added the bench cells for update and sweep (bench/arms/specsolve.py). No numbers from an idle machine have been published.
What to measure, and what it decides
- Run the
updateandsweepcells on an idle machine, A/B against the same base, and publish them with their method and base commit. - Split each re-solve into what was rebuilt, what was pushed to the solver, and the solve itself, so the number says where the time goes.
- The numbers decide whether Re-solving a model should cost the change, not the model #700's deliberately unfiled follow-up is worth filing: a partial rebuild that rebuilds only the model frames a changed parameter can touch. That is decidable from the plan with no data. Re-solving a model should cost the change, not the model #614 split a full build at 240k columns (base
119504a) as 8% bind, 24% variables, 41% constraints and 26% objective, but that measured a path that no longer runs.
Out of scope: any change to src/. This is bench work (chore).
Activity