eejbyfeldt commented on code in PR #12269: URL: https://github.com/apache/datafusion/pull/12269#discussion_r1771340393
########## datafusion/physical-plan/src/aggregates/group_values/column_wise.rs: ########## @@ -0,0 +1,315 @@ +// Licensed to the Apache Software Foundation (ASF) under one +// or more contributor license agreements. See the NOTICE file +// distributed with this work for additional information +// regarding copyright ownership. The ASF licenses this file +// to you under the Apache License, Version 2.0 (the +// "License"); you may not use this file except in compliance +// with the License. You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, +// software distributed under the License is distributed on an +// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +// KIND, either express or implied. See the License for the +// specific language governing permissions and limitations +// under the License. + +use crate::aggregates::group_values::group_value_row::{ + ArrayRowEq, ByteGroupValueBuilder, PrimitiveGroupValueBuilder, +}; +use crate::aggregates::group_values::GroupValues; +use ahash::RandomState; +use arrow::compute::cast; +use arrow::datatypes::{ + Date32Type, Date64Type, Float32Type, Float64Type, Int16Type, Int32Type, Int64Type, + Int8Type, UInt16Type, UInt32Type, UInt64Type, UInt8Type, +}; +use arrow::record_batch::RecordBatch; +use arrow_array::{Array, ArrayRef}; +use arrow_schema::{DataType, SchemaRef}; +use datafusion_common::hash_utils::create_hashes; +use datafusion_common::{DataFusionError, Result}; +use datafusion_execution::memory_pool::proxy::{RawTableAllocExt, VecAllocExt}; +use datafusion_expr::EmitTo; +use datafusion_physical_expr::binary_map::OutputType; + +use hashbrown::raw::RawTable; + +/// Compare GroupValue Rows column by column +pub struct GroupValuesColumn { + /// The output schema + schema: SchemaRef, + + /// Logically maps group values to a group_index in + /// [`Self::group_values`] and in each accumulator + /// + /// Uses the raw API of hashbrown to avoid actually storing the + /// keys (group values) in the table + /// + /// keys: u64 hashes of the GroupValue + /// values: (hash, group_index) + map: RawTable<(u64, usize)>, + + /// The size of `map` in bytes + map_size: usize, + + /// The actual group by values, stored column-wise. Compare from + /// the left to right, each column is stored as `ArrayRowEq`. + /// This is shown faster than the row format + group_values: Vec<Box<dyn ArrayRowEq>>, + + /// reused buffer to store hashes + hashes_buffer: Vec<u64>, + + /// Random state for creating hashes + random_state: RandomState, +} + +impl GroupValuesColumn { + pub fn try_new(schema: SchemaRef) -> Result<Self> { + let map = RawTable::with_capacity(0); + Ok(Self { + schema, + map, + map_size: 0, + group_values: vec![], + hashes_buffer: Default::default(), + random_state: Default::default(), + }) + } +} + +impl GroupValues for GroupValuesColumn { + fn intern(&mut self, cols: &[ArrayRef], groups: &mut Vec<usize>) -> Result<()> { + let n_rows = cols[0].len(); + + if self.group_values.is_empty() { + let len = cols.len(); + let mut v = Vec::with_capacity(len); Review Comment: I think we should give `len` a better name like `n_cols` or just "inline" it like ```suggestion let mut v = Vec::with_capacity(cols.len()); ``` because for me it currently just hurts readability. ########## datafusion/physical-plan/src/aggregates/group_values/group_value_row.rs: ########## @@ -0,0 +1,446 @@ +// Licensed to the Apache Software Foundation (ASF) under one +// or more contributor license agreements. See the NOTICE file +// distributed with this work for additional information +// regarding copyright ownership. The ASF licenses this file +// to you under the Apache License, Version 2.0 (the +// "License"); you may not use this file except in compliance +// with the License. You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, +// software distributed under the License is distributed on an +// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +// KIND, either express or implied. See the License for the +// specific language governing permissions and limitations +// under the License. + +use arrow::array::BooleanBufferBuilder; +use arrow::array::BufferBuilder; +use arrow::array::GenericBinaryArray; +use arrow::array::GenericStringArray; +use arrow::array::OffsetSizeTrait; +use arrow::array::PrimitiveArray; +use arrow::array::{Array, ArrayRef, ArrowPrimitiveType, AsArray}; +use arrow::buffer::NullBuffer; +use arrow::buffer::OffsetBuffer; +use arrow::buffer::ScalarBuffer; +use arrow::datatypes::ArrowNativeType; +use arrow::datatypes::ByteArrayType; +use arrow::datatypes::DataType; +use arrow::datatypes::GenericBinaryType; +use arrow::datatypes::GenericStringType; + +use std::sync::Arc; +use std::vec; + +use datafusion_physical_expr_common::binary_map::{OutputType, INITIAL_BUFFER_CAPACITY}; + +/// Trait for group values column-wise row comparison +/// +/// Implementations of this trait store a in-progress collection of group values +/// (similar to various builders in Arrow-rs) that allow for quick comparison to +/// incoming rows. +/// +pub trait ArrayRowEq: Send + Sync { + /// Returns equal if the row stored in this builder at `lhs_row` is equal to + /// the row in `array` at `rhs_row` + fn equal_to(&self, lhs_row: usize, array: &ArrayRef, rhs_row: usize) -> bool; + /// Appends the row at `row` in `array` to this builder + fn append_val(&mut self, array: &ArrayRef, row: usize); + /// Returns the number of rows stored in this builder + fn len(&self) -> usize; + /// Builds a new array from all of the stored rows + fn build(self: Box<Self>) -> ArrayRef; + /// Builds a new array from the first `n` stored rows, shifting the + /// remaining rows to the start of the builder + fn take_n(&mut self, n: usize) -> ArrayRef; +} + +pub struct PrimitiveGroupValueBuilder<T: ArrowPrimitiveType> { + group_values: Vec<T::Native>, + nulls: Vec<bool>, + // whether the array contains at least one null, for fast non-null path + has_null: bool, + nullable: bool, +} + +impl<T> PrimitiveGroupValueBuilder<T> +where + T: ArrowPrimitiveType, +{ + pub fn new(nullable: bool) -> Self { + Self { + group_values: vec![], + nulls: vec![], + has_null: false, + nullable, + } + } +} + +impl<T: ArrowPrimitiveType> ArrayRowEq for PrimitiveGroupValueBuilder<T> { + fn equal_to(&self, lhs_row: usize, array: &ArrayRef, rhs_row: usize) -> bool { + // non-null fast path + // both non-null + if !self.nullable { + return self.group_values[lhs_row] + == array.as_primitive::<T>().value(rhs_row); + } + + // lhs is non-null + if self.nulls[lhs_row] { + if array.is_null(rhs_row) { + return false; + } + + return self.group_values[lhs_row] + == array.as_primitive::<T>().value(rhs_row); + } + + array.is_null(rhs_row) + } + + fn append_val(&mut self, array: &ArrayRef, row: usize) { + // non-null fast path + if !self.nullable || !array.is_null(row) { + let elem = array.as_primitive::<T>().value(row); + self.group_values.push(elem); + self.nulls.push(true); + } else { + self.group_values.push(T::default_value()); + self.nulls.push(false); + self.has_null = true; + } + } + + fn len(&self) -> usize { + self.group_values.len() + } + + fn build(self: Box<Self>) -> ArrayRef { + if self.has_null { + Arc::new(PrimitiveArray::<T>::new( + ScalarBuffer::from(self.group_values), + Some(NullBuffer::from(self.nulls)), + )) + } else { + Arc::new(PrimitiveArray::<T>::new( + ScalarBuffer::from(self.group_values), + None, + )) + } + } + + fn take_n(&mut self, n: usize) -> ArrayRef { + if self.has_null { + let first_n = self.group_values.drain(0..n).collect::<Vec<_>>(); + let first_n_nulls = self.nulls.drain(0..n).collect::<Vec<_>>(); + Arc::new(PrimitiveArray::<T>::new( + ScalarBuffer::from(first_n), + Some(NullBuffer::from(first_n_nulls)), + )) + } else { + let first_n = self.group_values.drain(0..n).collect::<Vec<_>>(); + self.nulls.truncate(self.nulls.len() - n); + Arc::new(PrimitiveArray::<T>::new(ScalarBuffer::from(first_n), None)) + } + } +} + +pub struct ByteGroupValueBuilder<O> +where + O: OffsetSizeTrait, +{ + output_type: OutputType, + buffer: BufferBuilder<u8>, + /// Offsets into `buffer` for each distinct value. These offsets as used + /// directly to create the final `GenericBinaryArray`. The `i`th string is + /// stored in the range `offsets[i]..offsets[i+1]` in `buffer`. Null values + /// are stored as a zero length string. + offsets: Vec<O>, + /// Null indexes in offsets, if `i` is in nulls, `offsets[i]` should be equals to `offsets[i+1]` + nulls: Vec<usize>, +} + +impl<O> ByteGroupValueBuilder<O> +where + O: OffsetSizeTrait, +{ + pub fn new(output_type: OutputType) -> Self { + Self { + output_type, + buffer: BufferBuilder::new(INITIAL_BUFFER_CAPACITY), + offsets: vec![O::default()], + nulls: vec![], + } + } + + fn append_val_inner<B>(&mut self, array: &ArrayRef, row: usize) + where + B: ByteArrayType, + { + let arr = array.as_bytes::<B>(); + if arr.is_null(row) { + self.nulls.push(self.len()); + // nulls need a zero length in the offset buffer + let offset = self.buffer.len(); + + self.offsets.push(O::usize_as(offset)); + return; + } + + let value: &[u8] = arr.value(row).as_ref(); + self.buffer.append_slice(value); + self.offsets.push(O::usize_as(self.buffer.len())); + } + + fn equal_to_inner<B>(&self, lhs_row: usize, array: &ArrayRef, rhs_row: usize) -> bool + where + B: ByteArrayType, + { + // Handle nulls + let is_lhs_null = self.nulls.iter().any(|null_idx| *null_idx == lhs_row); + let arr = array.as_bytes::<B>(); + if is_lhs_null { + return arr.is_null(rhs_row); + } else if arr.is_null(rhs_row) { + return false; + } + + let arr = array.as_bytes::<B>(); + let rhs_elem: &[u8] = arr.value(rhs_row).as_ref(); + let rhs_elem_len = arr.value_length(rhs_row).as_usize(); + debug_assert_eq!(rhs_elem_len, rhs_elem.len()); + let l = self.offsets[lhs_row].as_usize(); + let r = self.offsets[lhs_row + 1].as_usize(); + let existing_elem = unsafe { self.buffer.as_slice().get_unchecked(l..r) }; + existing_elem.len() == rhs_elem.len() && rhs_elem == existing_elem Review Comment: Do we really need to check the length manually? Will that not just be a part of the `==` for slices. (https://doc.rust-lang.org/1.81.0/src/core/slice/cmp.rs.html#59) ```suggestion rhs_elem == existing_elem ``` ########## datafusion/physical-plan/src/aggregates/group_values/column_wise.rs: ########## @@ -0,0 +1,315 @@ +// Licensed to the Apache Software Foundation (ASF) under one +// or more contributor license agreements. See the NOTICE file +// distributed with this work for additional information +// regarding copyright ownership. The ASF licenses this file +// to you under the Apache License, Version 2.0 (the +// "License"); you may not use this file except in compliance +// with the License. You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, +// software distributed under the License is distributed on an +// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +// KIND, either express or implied. See the License for the +// specific language governing permissions and limitations +// under the License. + +use crate::aggregates::group_values::group_value_row::{ + ArrayRowEq, ByteGroupValueBuilder, PrimitiveGroupValueBuilder, +}; +use crate::aggregates::group_values::GroupValues; +use ahash::RandomState; +use arrow::compute::cast; +use arrow::datatypes::{ + Date32Type, Date64Type, Float32Type, Float64Type, Int16Type, Int32Type, Int64Type, + Int8Type, UInt16Type, UInt32Type, UInt64Type, UInt8Type, +}; +use arrow::record_batch::RecordBatch; +use arrow_array::{Array, ArrayRef}; +use arrow_schema::{DataType, SchemaRef}; +use datafusion_common::hash_utils::create_hashes; +use datafusion_common::{DataFusionError, Result}; +use datafusion_execution::memory_pool::proxy::{RawTableAllocExt, VecAllocExt}; +use datafusion_expr::EmitTo; +use datafusion_physical_expr::binary_map::OutputType; + +use hashbrown::raw::RawTable; + +/// Compare GroupValue Rows column by column +pub struct GroupValuesColumn { + /// The output schema + schema: SchemaRef, + + /// Logically maps group values to a group_index in + /// [`Self::group_values`] and in each accumulator + /// + /// Uses the raw API of hashbrown to avoid actually storing the + /// keys (group values) in the table + /// + /// keys: u64 hashes of the GroupValue + /// values: (hash, group_index) + map: RawTable<(u64, usize)>, + + /// The size of `map` in bytes + map_size: usize, + + /// The actual group by values, stored column-wise. Compare from + /// the left to right, each column is stored as `ArrayRowEq`. + /// This is shown faster than the row format + group_values: Vec<Box<dyn ArrayRowEq>>, + + /// reused buffer to store hashes + hashes_buffer: Vec<u64>, + + /// Random state for creating hashes + random_state: RandomState, +} + +impl GroupValuesColumn { + pub fn try_new(schema: SchemaRef) -> Result<Self> { + let map = RawTable::with_capacity(0); + Ok(Self { + schema, + map, + map_size: 0, + group_values: vec![], + hashes_buffer: Default::default(), + random_state: Default::default(), + }) + } +} + +impl GroupValues for GroupValuesColumn { + fn intern(&mut self, cols: &[ArrayRef], groups: &mut Vec<usize>) -> Result<()> { + let n_rows = cols[0].len(); + + if self.group_values.is_empty() { + let len = cols.len(); + let mut v = Vec::with_capacity(len); + + for f in self.schema.fields().iter() { + let nullable = f.is_nullable(); + match f.data_type() { + &DataType::Int8 => { + let b = PrimitiveGroupValueBuilder::<Int8Type>::new(nullable); + v.push(Box::new(b) as _) + } + &DataType::Int16 => { + let b = PrimitiveGroupValueBuilder::<Int16Type>::new(nullable); + v.push(Box::new(b) as _) + } + &DataType::Int32 => { + let b = PrimitiveGroupValueBuilder::<Int32Type>::new(nullable); + v.push(Box::new(b) as _) + } + &DataType::Int64 => { + let b = PrimitiveGroupValueBuilder::<Int64Type>::new(nullable); + v.push(Box::new(b) as _) + } + &DataType::UInt8 => { + let b = PrimitiveGroupValueBuilder::<UInt8Type>::new(nullable); + v.push(Box::new(b) as _) + } + &DataType::UInt16 => { + let b = PrimitiveGroupValueBuilder::<UInt16Type>::new(nullable); + v.push(Box::new(b) as _) + } + &DataType::UInt32 => { + let b = PrimitiveGroupValueBuilder::<UInt32Type>::new(nullable); + v.push(Box::new(b) as _) + } + &DataType::UInt64 => { + let b = PrimitiveGroupValueBuilder::<UInt64Type>::new(nullable); + v.push(Box::new(b) as _) + } + &DataType::Float32 => { + let b = PrimitiveGroupValueBuilder::<Float32Type>::new(nullable); + v.push(Box::new(b) as _) + } + &DataType::Float64 => { + let b = PrimitiveGroupValueBuilder::<Float64Type>::new(nullable); + v.push(Box::new(b) as _) + } + &DataType::Date32 => { + let b = PrimitiveGroupValueBuilder::<Date32Type>::new(nullable); + v.push(Box::new(b) as _) + } + &DataType::Date64 => { + let b = PrimitiveGroupValueBuilder::<Date64Type>::new(nullable); + v.push(Box::new(b) as _) + } + &DataType::Utf8 => { + let b = ByteGroupValueBuilder::<i32>::new(OutputType::Utf8); + v.push(Box::new(b) as _) + } + &DataType::LargeUtf8 => { + let b = ByteGroupValueBuilder::<i64>::new(OutputType::Utf8); + v.push(Box::new(b) as _) + } + &DataType::Binary => { + let b = ByteGroupValueBuilder::<i32>::new(OutputType::Binary); + v.push(Box::new(b) as _) + } + &DataType::LargeBinary => { + let b = ByteGroupValueBuilder::<i64>::new(OutputType::Binary); + v.push(Box::new(b) as _) + } + dt => todo!("{dt} not impl"), + } + } + self.group_values = v; + } + + // tracks to which group each of the input rows belongs + groups.clear(); + + // 1.1 Calculate the group keys for the group values + let batch_hashes = &mut self.hashes_buffer; + batch_hashes.clear(); + batch_hashes.resize(n_rows, 0); + create_hashes(cols, &self.random_state, batch_hashes)?; + + for (row, &target_hash) in batch_hashes.iter().enumerate() { + let entry = self.map.get_mut(target_hash, |(exist_hash, group_idx)| { + // Somewhat surprisingly, this closure can be called even if the + // hash doesn't match, so check the hash first with an integer + // comparison first avoid the more expensive comparison with + // group value. https://github.com/apache/datafusion/pull/11718 + if target_hash != *exist_hash { + return false; + } + + fn check_row_equal( + array_row: &dyn ArrayRowEq, + lhs_row: usize, + array: &ArrayRef, + rhs_row: usize, + ) -> bool { + array_row.equal_to(lhs_row, array, rhs_row) + } + + for (i, group_val) in self.group_values.iter().enumerate() { + if !check_row_equal(group_val.as_ref(), *group_idx, &cols[i], row) { + return false; + } + } + + true + }); + + let group_idx = match entry { + // Existing group_index for this group value + Some((_hash, group_idx)) => *group_idx, + // 1.2 Need to create new entry for the group + None => { + // Add new entry to aggr_state and save newly created index + // let group_idx = group_values.num_rows(); + // group_values.push(group_rows.row(row)); + + let mut checklen = 0; + let group_idx = self.group_values[0].len(); + for (i, group_value) in self.group_values.iter_mut().enumerate() { + group_value.append_val(&cols[i], row); + let len = group_value.len(); + if i == 0 { + checklen = len; + } else { + debug_assert_eq!(checklen, len); + } + } + + // for hasher function, use precomputed hash value + self.map.insert_accounted( + (target_hash, group_idx), + |(hash, _group_index)| *hash, + &mut self.map_size, + ); + group_idx + } + }; + groups.push(group_idx); + } + + Ok(()) + } + + fn size(&self) -> usize { + let group_values_size = self.group_values.len(); Review Comment: My understanding is that size should return allocated size in bytes, so this does not look correct for `self.group_values`. I would expect us need to do something like ``` let group_values_size = self.group_values.iter().map(|column| column.size()).sum() ``` and add a size method to `ArrayRowEq` so it can do some size calculation based on the field. ########## datafusion/physical-expr-common/src/group_value_row.rs: ########## @@ -0,0 +1,393 @@ +// Licensed to the Apache Software Foundation (ASF) under one +// or more contributor license agreements. See the NOTICE file +// distributed with this work for additional information +// regarding copyright ownership. The ASF licenses this file +// to you under the Apache License, Version 2.0 (the +// "License"); you may not use this file except in compliance +// with the License. You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, +// software distributed under the License is distributed on an +// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +// KIND, either express or implied. See the License for the +// specific language governing permissions and limitations +// under the License. + +use arrow::array::BooleanBufferBuilder; +use arrow::array::BufferBuilder; +use arrow::array::GenericBinaryArray; +use arrow::array::GenericStringArray; +use arrow::array::OffsetSizeTrait; +use arrow::array::PrimitiveArray; +use arrow::array::{Array, ArrayRef, ArrowPrimitiveType, AsArray}; +use arrow::buffer::NullBuffer; +use arrow::buffer::OffsetBuffer; +use arrow::buffer::ScalarBuffer; +use arrow::datatypes::ArrowNativeType; +use arrow::datatypes::ByteArrayType; +use arrow::datatypes::DataType; +use arrow::datatypes::GenericBinaryType; +use arrow::datatypes::GenericStringType; + +use std::sync::Arc; +use std::vec; + +use crate::binary_map::OutputType; +use crate::binary_map::INITIAL_BUFFER_CAPACITY; + +/// Trait for group values column-wise row comparison +pub trait ArrayRowEq: Send + Sync { + fn equal_to(&self, lhs_row: usize, array: &ArrayRef, rhs_row: usize) -> bool; + fn append_val(&mut self, array: &ArrayRef, row: usize); + fn len(&self) -> usize; + fn is_empty(&self) -> bool; + fn build(self: Box<Self>) -> ArrayRef; + fn take_n(&mut self, n: usize) -> ArrayRef; +} + +pub struct PrimitiveGroupValueBuilder<T: ArrowPrimitiveType> { + group_values: Vec<T::Native>, + nulls: Vec<bool>, + // whether the array contains at least one null, for fast non-null path + has_null: bool, + nullable: bool, +} + +impl<T> PrimitiveGroupValueBuilder<T> +where + T: ArrowPrimitiveType, +{ + pub fn new(nullable: bool) -> Self { + Self { + group_values: vec![], + nulls: vec![], + has_null: false, + nullable, + } + } +} + +impl<T: ArrowPrimitiveType> ArrayRowEq for PrimitiveGroupValueBuilder<T> { + fn equal_to(&self, lhs_row: usize, array: &ArrayRef, rhs_row: usize) -> bool { + // non-null fast path + // both non-null + if !self.nullable { + return self.group_values[lhs_row] + == array.as_primitive::<T>().value(rhs_row); + } + + // lhs is non-null + if self.nulls[lhs_row] { + if array.is_null(rhs_row) { + return false; + } + + return self.group_values[lhs_row] + == array.as_primitive::<T>().value(rhs_row); + } + + array.is_null(rhs_row) + } + + fn append_val(&mut self, array: &ArrayRef, row: usize) { + // non-null fast path + if !self.nullable || !array.is_null(row) { Review Comment: For me it also took sometime when reading to convince my self that `||` was correct. Maybe it would be easier to read if it we swaped the two cases and removed the `!`? e.g ``` if self.nullable && array.is_null(row) { self.group_values.push(T::default_value()); self.nulls.push(false); self.has_null = true; } else { let elem = array.as_primitive::<T>().value(row); self.group_values.push(elem); self.nulls.push(true); } ``` ########## datafusion/sqllogictest/test_files/group_by.slt: ########## @@ -5148,3 +5148,42 @@ NULL statement ok drop table test_case_expr + +statement ok +drop table t; + +# test multi group by for binary type with nulls +statement ok +create table t(a int, b bytea) as values (1, 0xa), (1, 0xa), (2, null), (null, 0xb), (null, 0xb); + +query I?I +select a, b, count(*) from t group by grouping sets ((a, b), (a), (b)); Review Comment: Do we want to add a comment here to note that the result here is actually incorrect when compraing to other query engines. Due to https://github.com/apache/datafusion/issues/12570 ########## datafusion/physical-plan/src/aggregates/group_values/group_value_row.rs: ########## @@ -0,0 +1,446 @@ +// Licensed to the Apache Software Foundation (ASF) under one +// or more contributor license agreements. See the NOTICE file +// distributed with this work for additional information +// regarding copyright ownership. The ASF licenses this file +// to you under the Apache License, Version 2.0 (the +// "License"); you may not use this file except in compliance +// with the License. You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, +// software distributed under the License is distributed on an +// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +// KIND, either express or implied. See the License for the +// specific language governing permissions and limitations +// under the License. + +use arrow::array::BooleanBufferBuilder; +use arrow::array::BufferBuilder; +use arrow::array::GenericBinaryArray; +use arrow::array::GenericStringArray; +use arrow::array::OffsetSizeTrait; +use arrow::array::PrimitiveArray; +use arrow::array::{Array, ArrayRef, ArrowPrimitiveType, AsArray}; +use arrow::buffer::NullBuffer; +use arrow::buffer::OffsetBuffer; +use arrow::buffer::ScalarBuffer; +use arrow::datatypes::ArrowNativeType; +use arrow::datatypes::ByteArrayType; +use arrow::datatypes::DataType; +use arrow::datatypes::GenericBinaryType; +use arrow::datatypes::GenericStringType; + +use std::sync::Arc; +use std::vec; + +use datafusion_physical_expr_common::binary_map::{OutputType, INITIAL_BUFFER_CAPACITY}; + +/// Trait for group values column-wise row comparison +/// +/// Implementations of this trait store a in-progress collection of group values +/// (similar to various builders in Arrow-rs) that allow for quick comparison to +/// incoming rows. +/// +pub trait ArrayRowEq: Send + Sync { + /// Returns equal if the row stored in this builder at `lhs_row` is equal to + /// the row in `array` at `rhs_row` + fn equal_to(&self, lhs_row: usize, array: &ArrayRef, rhs_row: usize) -> bool; + /// Appends the row at `row` in `array` to this builder + fn append_val(&mut self, array: &ArrayRef, row: usize); + /// Returns the number of rows stored in this builder + fn len(&self) -> usize; + /// Builds a new array from all of the stored rows + fn build(self: Box<Self>) -> ArrayRef; + /// Builds a new array from the first `n` stored rows, shifting the + /// remaining rows to the start of the builder + fn take_n(&mut self, n: usize) -> ArrayRef; +} + +pub struct PrimitiveGroupValueBuilder<T: ArrowPrimitiveType> { + group_values: Vec<T::Native>, + nulls: Vec<bool>, + // whether the array contains at least one null, for fast non-null path + has_null: bool, + nullable: bool, +} + +impl<T> PrimitiveGroupValueBuilder<T> +where + T: ArrowPrimitiveType, +{ + pub fn new(nullable: bool) -> Self { + Self { + group_values: vec![], + nulls: vec![], + has_null: false, + nullable, + } + } +} + +impl<T: ArrowPrimitiveType> ArrayRowEq for PrimitiveGroupValueBuilder<T> { + fn equal_to(&self, lhs_row: usize, array: &ArrayRef, rhs_row: usize) -> bool { + // non-null fast path + // both non-null + if !self.nullable { + return self.group_values[lhs_row] + == array.as_primitive::<T>().value(rhs_row); + } + + // lhs is non-null + if self.nulls[lhs_row] { + if array.is_null(rhs_row) { + return false; + } + + return self.group_values[lhs_row] + == array.as_primitive::<T>().value(rhs_row); + } + + array.is_null(rhs_row) + } + + fn append_val(&mut self, array: &ArrayRef, row: usize) { + // non-null fast path + if !self.nullable || !array.is_null(row) { + let elem = array.as_primitive::<T>().value(row); + self.group_values.push(elem); + self.nulls.push(true); + } else { + self.group_values.push(T::default_value()); + self.nulls.push(false); + self.has_null = true; + } + } + + fn len(&self) -> usize { + self.group_values.len() + } + + fn build(self: Box<Self>) -> ArrayRef { + if self.has_null { + Arc::new(PrimitiveArray::<T>::new( + ScalarBuffer::from(self.group_values), + Some(NullBuffer::from(self.nulls)), + )) + } else { + Arc::new(PrimitiveArray::<T>::new( + ScalarBuffer::from(self.group_values), + None, + )) + } + } + + fn take_n(&mut self, n: usize) -> ArrayRef { + if self.has_null { + let first_n = self.group_values.drain(0..n).collect::<Vec<_>>(); + let first_n_nulls = self.nulls.drain(0..n).collect::<Vec<_>>(); + Arc::new(PrimitiveArray::<T>::new( + ScalarBuffer::from(first_n), + Some(NullBuffer::from(first_n_nulls)), + )) + } else { + let first_n = self.group_values.drain(0..n).collect::<Vec<_>>(); + self.nulls.truncate(self.nulls.len() - n); + Arc::new(PrimitiveArray::<T>::new(ScalarBuffer::from(first_n), None)) + } + } +} + +pub struct ByteGroupValueBuilder<O> +where + O: OffsetSizeTrait, +{ + output_type: OutputType, + buffer: BufferBuilder<u8>, + /// Offsets into `buffer` for each distinct value. These offsets as used + /// directly to create the final `GenericBinaryArray`. The `i`th string is + /// stored in the range `offsets[i]..offsets[i+1]` in `buffer`. Null values + /// are stored as a zero length string. + offsets: Vec<O>, + /// Null indexes in offsets, if `i` is in nulls, `offsets[i]` should be equals to `offsets[i+1]` + nulls: Vec<usize>, +} + +impl<O> ByteGroupValueBuilder<O> +where + O: OffsetSizeTrait, +{ + pub fn new(output_type: OutputType) -> Self { + Self { + output_type, + buffer: BufferBuilder::new(INITIAL_BUFFER_CAPACITY), + offsets: vec![O::default()], + nulls: vec![], + } + } + + fn append_val_inner<B>(&mut self, array: &ArrayRef, row: usize) + where + B: ByteArrayType, + { + let arr = array.as_bytes::<B>(); + if arr.is_null(row) { + self.nulls.push(self.len()); + // nulls need a zero length in the offset buffer + let offset = self.buffer.len(); + + self.offsets.push(O::usize_as(offset)); + return; + } + + let value: &[u8] = arr.value(row).as_ref(); + self.buffer.append_slice(value); + self.offsets.push(O::usize_as(self.buffer.len())); + } + + fn equal_to_inner<B>(&self, lhs_row: usize, array: &ArrayRef, rhs_row: usize) -> bool + where + B: ByteArrayType, + { + // Handle nulls + let is_lhs_null = self.nulls.iter().any(|null_idx| *null_idx == lhs_row); + let arr = array.as_bytes::<B>(); + if is_lhs_null { + return arr.is_null(rhs_row); + } else if arr.is_null(rhs_row) { + return false; + } + + let arr = array.as_bytes::<B>(); + let rhs_elem: &[u8] = arr.value(rhs_row).as_ref(); + let rhs_elem_len = arr.value_length(rhs_row).as_usize(); + debug_assert_eq!(rhs_elem_len, rhs_elem.len()); + let l = self.offsets[lhs_row].as_usize(); + let r = self.offsets[lhs_row + 1].as_usize(); + let existing_elem = unsafe { self.buffer.as_slice().get_unchecked(l..r) }; + existing_elem.len() == rhs_elem.len() && rhs_elem == existing_elem + } +} + +impl<O> ArrayRowEq for ByteGroupValueBuilder<O> +where + O: OffsetSizeTrait, +{ + fn equal_to(&self, lhs_row: usize, column: &ArrayRef, rhs_row: usize) -> bool { + // Sanity array type + match self.output_type { + OutputType::Binary => { + debug_assert!(matches!( + column.data_type(), + DataType::Binary | DataType::LargeBinary + )); + self.equal_to_inner::<GenericBinaryType<O>>(lhs_row, column, rhs_row) + } + OutputType::Utf8 => { + debug_assert!(matches!( + column.data_type(), + DataType::Utf8 | DataType::LargeUtf8 + )); + self.equal_to_inner::<GenericStringType<O>>(lhs_row, column, rhs_row) + } + _ => unreachable!("View types should use `ArrowBytesViewMap`"), + } + } + + fn append_val(&mut self, column: &ArrayRef, row: usize) { + // Sanity array type + match self.output_type { + OutputType::Binary => { + debug_assert!(matches!( + column.data_type(), + DataType::Binary | DataType::LargeBinary + )); + self.append_val_inner::<GenericBinaryType<O>>(column, row) + } + OutputType::Utf8 => { + debug_assert!(matches!( + column.data_type(), + DataType::Utf8 | DataType::LargeUtf8 + )); + self.append_val_inner::<GenericStringType<O>>(column, row) + } + _ => unreachable!("View types should use `ArrowBytesViewMap`"), + }; + } + + fn len(&self) -> usize { + self.offsets.len() - 1 + } + + fn build(self: Box<Self>) -> ArrayRef { + let Self { + output_type, + mut buffer, + offsets, + nulls, + } = *self; + + let null_buffer = if nulls.is_empty() { + None + } else { + // Only make a `NullBuffer` if there was a null value + let num_values = offsets.len() - 1; + let mut bool_builder = BooleanBufferBuilder::new(num_values); + bool_builder.append_n(num_values, true); + nulls.into_iter().for_each(|null_index| { + bool_builder.set_bit(null_index, false); + }); + Some(NullBuffer::from(bool_builder.finish())) + }; + + // SAFETY: the offsets were constructed correctly in `insert_if_new` -- + // monotonically increasing, overflows were checked. + let offsets = unsafe { OffsetBuffer::new_unchecked(ScalarBuffer::from(offsets)) }; + let values = buffer.finish(); + match output_type { + OutputType::Binary => { + // SAFETY: the offsets were constructed correctly + Arc::new(unsafe { + GenericBinaryArray::new_unchecked(offsets, values, null_buffer) + }) + } + OutputType::Utf8 => { + // SAFETY: + // 1. the offsets were constructed safely + // + // 2. we asserted the input arrays were all the correct type and + // thus since all the values that went in were valid (e.g. utf8) + // so are all the values that come out + Arc::new(unsafe { + GenericStringArray::new_unchecked(offsets, values, null_buffer) + }) + } + _ => unreachable!("View types should use `ArrowBytesViewMap`"), + } + } + + fn take_n(&mut self, n: usize) -> ArrayRef { + debug_assert!(self.len() >= n); + + let mut nulls_count = 0; Review Comment: Seems like we never use the result in `nulls_count`? Could it just be removed? -- 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]
