Issue #16250: Window Range Performance #16320
Conversation
- Use <= for previous start checks now that they are properly vectorised.
- This reduces the benchmark time by 5x to 3.858s.
* Use <= for previous start checks now that they are properly vectorised. * This reduces the benchmark time by 5x to 3.858s.
There was a problem hiding this comment.
Thanks for the PR! I think it's all good, although I'm not sure if I follow how a 1-line change makes a benchmark 5x faster. Care to explain?
I don't think the regression failure is related, that benchmark query shouldn't have anything to do with the window operator, although we might want to re-run CI just to be sure.
@lnkuiper I may have made a mistake on which query I looked at, but from stepping through the code, we were in a situation where the previous bound got reused a lot when it was the first value in the partition: there are only 83 dates for ~10MR, and if you go back a year, you often end up at the start of the partition. The check specifically excluded this situation for no good reason (we know the previous bound is valid) and since it came up a lot in this data set we ended up searching instead of reusing the value. |
|
Thanks! |
Issue duckdb/duckdb#16250: Window Range Performance (duckdb/duckdb#16320)
Issue duckdb/duckdb#16250: Window Range Performance (duckdb/duckdb#16320)
Issue duckdb/duckdb#16250: Window Range Performance (duckdb/duckdb#16320)
Issue duckdb/duckdb#16250: Window Range Performance (duckdb/duckdb#16320)
Issue duckdb/duckdb#16250: Window Range Performance (duckdb/duckdb#16320)
Issue duckdb/duckdb#16250: Window Range Performance (duckdb/duckdb#16320)