berkaysynnada commented on PR #7793: URL: https://github.com/apache/arrow-datafusion/pull/7793#issuecomment-1764521549
> Which selectivity estimator provides an inexact point instead of a lower/upper bound? I can give an example of how `FilterExec::statistics()` currently works on a table having one column. Let's assume that the number of rows from input statistics() is exactly 1000, and the range of the column values is [1,100]. The predicate is `column_a>90`. After the analysis, the new range of the column would become [91,100], which estimates the selectivity as 0.1 (intervals are assumed to be uniform, they cannot be treated any differently yet). https://github.com/apache/arrow-datafusion/blob/fa2bb6c4e80d80e3d1d26ce85f7c21232f036dd5/datafusion/physical-expr/src/analysis.rs#L218-L223 Now, when we update the number of rows, we use that selectivity value and find the new row count as 1000*0.1=100. https://github.com/apache/arrow-datafusion/blob/fa2bb6c4e80d80e3d1d26ce85f7c21232f036dd5/datafusion/physical-plan/src/filter.rs#L233-L235 These lines become: https://github.com/synnada-ai/arrow-datafusion/blob/23c564b093f8a60b95bbf9f774a1afe9b2908bc7/datafusion/physical-plan/src/filter.rs#L221-L224 Though the input row count is exact, we cannot take that exactness information further. At that point, how can we define a range for 100 rows? That is why we needed such Inexact implementation. -- 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]
