hhhizzz commented on PR #10141:
URL: https://github.com/apache/arrow-rs/pull/10141#issuecomment-5011741531

   I reran one representative point from `row_selection_cursor` on the current
   head (`8c217d8`) against the merge-base (`ee30b61`), using five alternating
   current/base runs on a pinned CPU.
   
   For `uniform50/L32`:
   
   - forced `Selectors`: +7.67% paired geometric mean
   - forced `Mask`: +2.80%
   
   All five pairs moved in the same direction and Criterion produced no 
warnings.
   
   However, I want to be precise about the interpretation: this benchmark forces
   the two policies in order to map their crossover; it does not exercise 
`Auto`.
   Therefore these numbers are a signal that the underlying forced paths became
   slower at L32, not evidence that the Auto decision or threshold is wrong. A
   single point is also not enough to establish a broad regression.
   
   Could we rerun the complete L4–L40 curve in `row_selection_cursor` on the 
current head and compare:
   
   1. the Selector/Mask crossover against the merge-base;
   2. the strategy selected by Auto at each point; and
   3. Auto's runtime against the faster forced strategy?
   
   Since this PR adds a second input representation, it would be particularly
   useful to run the matrix for both selector-backed and mask-backed
   `RowSelection`s. That would show whether Auto should use a different 
threshold
   or cost model when choosing a strategy requires first converting the backing
   representation.


-- 
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]

Reply via email to