diff --git a/datafusion/physical-plan/src/aggregates/group_values/multi_group_by/clustered.rs b/datafusion/physical-plan/src/aggregates/group_values/multi_group_by/clustered.rs index efb4e24e6cff9..c6b293349da99 100644 --- a/datafusion/physical-plan/src/aggregates/group_values/multi_group_by/clustered.rs +++ b/datafusion/physical-plan/src/aggregates/group_values/multi_group_by/clustered.rs @@ -188,23 +188,29 @@ mod tests { use arrow::array::UInt32Array; use arrow::array::{ - Array, AsArray, BooleanArray, Decimal32Array, Decimal64Array, Decimal128Array, - Decimal256Array, FixedSizeBinaryArray, Int32Array, Int64Array, - IntervalDayTimeArray, IntervalMonthDayNanoArray, IntervalYearMonthArray, + Array, AsArray, BooleanArray, FixedSizeBinaryArray, Int32Array, Int64Array, StringArray, }; use arrow::buffer::{Buffer, NullBuffer}; - use arrow::compute::{cast, take}; - use arrow::datatypes::{ - Field, IntervalDayTime, IntervalMonthDayNano, IntervalUnit, i256, - }; + use arrow::datatypes::Field; use datafusion_expr::GroupSelection; + #[cfg(not(feature = "force_hash_collisions"))] + use arrow::array::{ + Decimal32Array, Decimal64Array, Decimal128Array, Decimal256Array, + IntervalDayTimeArray, IntervalMonthDayNanoArray, IntervalYearMonthArray, + }; + #[cfg(not(feature = "force_hash_collisions"))] + use arrow::compute::{cast, take}; + #[cfg(not(feature = "force_hash_collisions"))] + use arrow::datatypes::{IntervalDayTime, IntervalMonthDayNano, IntervalUnit, i256}; + use super::*; // The values repeat in seven-row runs, with four keys per value of the // first column. Construct logical types directly where casting would // rescale decimals or cannot represent all interval components. + #[cfg(not(feature = "force_hash_collisions"))] fn clustered_key_array( key_type: &DataType, rows: usize, @@ -338,15 +344,21 @@ mod tests { }) } - // Compare with the existing hash implementation, changing actual input - // boundaries and removing completed groups between batches. This checks the - // semantic contract rather than duplicating the adjacent-run algorithm. + // Compare with the existing hash implementation across input boundaries and + // emits, without duplicating the adjacent-run algorithm. The reference hashes + // while the clustered implementation does not, so forced collisions make the + // reference quadratic without testing collision handling in this implementation. + #[cfg(not(feature = "force_hash_collisions"))] #[rstest::rstest] #[case::sorted(false)] #[case::unsorted(true)] fn clustered_keys_match_hash_grouping_across_batches_and_emits( #[case] unsorted: bool, + // Keep these values in sync with KEY_TYPE_SHARD_COUNT below. + #[values(0, 1, 2, 3)] shard: usize, ) -> Result<()> { + const KEY_TYPE_SHARD_COUNT: usize = 4; + let mut key_types = vec![ DataType::Boolean, DataType::Int8, @@ -394,7 +406,11 @@ mod tests { key_types.push(DataType::Timestamp(unit, timezone.map(Into::into))); } } - for key_type in key_types { + for key_type in key_types + .into_iter() + .skip(shard) + .step_by(KEY_TYPE_SHARD_COUNT) + { for nullable in [false, true] { for batch_size in [1, 2, 7, 8, 63, 1024] { let rows = 129.max(batch_size + 1);