zhuqi-lucas commented on code in PR #21956:
URL: https://github.com/apache/datafusion/pull/21956#discussion_r3246190490
##########
datafusion/datasource-parquet/src/access_plan.rs:
##########
@@ -16,10 +16,28 @@
// under the License.
use crate::sort::reverse_row_selection;
+use arrow::array::{Array, ArrayRef, BooleanArray};
+use arrow::datatypes::Schema;
use datafusion_common::{Result, assert_eq_or_internal_err};
+use datafusion_physical_expr::expressions::Column;
+use datafusion_physical_expr_common::sort_expr::LexOrdering;
+use log::debug;
+use parquet::arrow::arrow_reader::statistics::StatisticsConverter;
use parquet::arrow::arrow_reader::{RowSelection, RowSelector};
use parquet::file::metadata::{ParquetMetaData, RowGroupMetaData};
+/// Fraction of adjacent (in sorted-by-min order) row group pairs whose
+/// `[min, max]` ranges overlap above which `reorder_by_statistics` will
+/// bail out without reordering.
+///
+/// When stats overlap heavily (e.g. unsorted columns like ClickBench's
+/// `EventTime` on `hits_partitioned`), reordering by min cannot enable
+/// row-group-level pruning — every "later" RG still has values that
+/// could appear in TopK. The reorder cost (CPU sort + lost IO sequential
+/// locality + parallel scheduling pessimization across workers all
+/// pulling "best" RGs first) then dominates, producing a net regression.
Review Comment:
Yes, this logic was added because the clickbench benchmark result several
topk have regression, but i can remove those logic and trigger again, if it
works well, we don't need to add those logic.
--
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]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
