kumarUjjawal commented on code in PR #24589: URL: https://github.com/apache/datafusion/pull/24589#discussion_r3874118598
########## datafusion/physical-plan/benches/window_filter.rs: ########## @@ -0,0 +1,263 @@ +// 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. + +//! Microbenchmark for window aggregates with `FILTER`. The benchmark uses +//! pre-ordered input to exclude sorting and query planning, but executes the +//! physical window plan so both stateful and whole-partition evaluation paths +//! are represented. + +use std::hint::black_box; +use std::sync::Arc; + +use arrow::array::{BooleanArray, Float64Array, UInt64Array}; +use arrow::datatypes::{DataType, Field, Schema, SchemaRef}; +use arrow::record_batch::RecordBatch; +use criterion::{BenchmarkId, Criterion, criterion_group, criterion_main}; +use datafusion_common::{ScalarValue, config::ConfigOptions}; +use datafusion_execution::TaskContext; +use datafusion_expr::{ + Operator, WindowFrame, WindowFrameBound, WindowFrameUnits, WindowFunctionDefinition, +}; +use datafusion_functions::math::power; +use datafusion_functions_aggregate::sum::sum_udaf; +use datafusion_physical_expr::expressions::{BinaryExpr, col, lit}; +use datafusion_physical_expr::{ + LexOrdering, PhysicalExpr, PhysicalSortExpr, ScalarFunctionExpr, +}; +use datafusion_physical_plan::test::TestMemoryExec; +use datafusion_physical_plan::windows::{ + BoundedWindowAggExec, WindowAggExec, create_window_expr, +}; +use datafusion_physical_plan::{ExecutionPlan, InputOrderMode, collect}; + +const BATCH_SIZE: usize = 8192; +const NUM_BATCHES: usize = 4; + +#[derive(Clone, Copy)] +enum ArgumentKind { + Column, + Divide, + Power, +} + +impl ArgumentKind { + fn name(self) -> &'static str { + match self { + Self::Column => "column", + Self::Divide => "divide", + Self::Power => "power_udf", + } + } +} + +fn schema() -> SchemaRef { + Arc::new(Schema::new(vec![ + Field::new("id", DataType::UInt64, false), + Field::new("value", DataType::Float64, false), + Field::new("include", DataType::Boolean, false), + ])) +} + +fn make_batches(filter_percent: usize) -> Vec<RecordBatch> { + (0..NUM_BATCHES) + .map(|batch_index| { + let start = batch_index * BATCH_SIZE; + let end = start + BATCH_SIZE; + let id = UInt64Array::from_iter_values((start..end).map(|i| i as u64)); + let value = Float64Array::from_iter_values((start..end).map(|i| i as f64)); + let include = BooleanArray::from( + (start..end) + .map(|i| (i * filter_percent) % 100 < filter_percent) + .collect::<Vec<_>>(), + ); + + RecordBatch::try_new( + schema(), + vec![Arc::new(id), Arc::new(value), Arc::new(include)], + ) + .unwrap() + }) + .collect() +} + +fn window_argument(kind: ArgumentKind, schema: &Schema) -> Arc<dyn PhysicalExpr> { + let column = col("value", schema).unwrap(); + match kind { + ArgumentKind::Column => column, + ArgumentKind::Divide => { + Arc::new(BinaryExpr::new(column, Operator::Divide, lit(10.0_f64))) + } + ArgumentKind::Power => Arc::new( + ScalarFunctionExpr::try_new( + power(), + vec![column, lit(1.5_f64)], + schema, + Arc::new(ConfigOptions::default()), + ) + .unwrap(), + ), + } +} + +fn cumulative_frame() -> WindowFrame { + WindowFrame::new_bounds( + WindowFrameUnits::Rows, + WindowFrameBound::Preceding(ScalarValue::UInt64(None)), + WindowFrameBound::CurrentRow, + ) +} + +fn sliding_frame() -> WindowFrame { + WindowFrame::new_bounds( + WindowFrameUnits::Rows, + WindowFrameBound::Preceding(ScalarValue::UInt64(Some(10))), + WindowFrameBound::CurrentRow, + ) +} + +fn whole_partition_frame() -> WindowFrame { + WindowFrame::new_bounds( + WindowFrameUnits::Rows, + WindowFrameBound::Preceding(ScalarValue::UInt64(None)), + WindowFrameBound::Following(ScalarValue::UInt64(None)), + ) +} + +fn make_window_plan( + filter_percent: usize, + argument_kind: ArgumentKind, + window_frame: WindowFrame, +) -> Arc<dyn ExecutionPlan> { + let schema = schema(); + let order_by = vec![PhysicalSortExpr { + expr: col("id", &schema).unwrap(), + options: Default::default(), + }]; + let window_expr = create_window_expr( + &WindowFunctionDefinition::AggregateUDF(sum_udaf()), + format!("sum({}) FILTER (WHERE include)", argument_kind.name()), + &[window_argument(argument_kind, &schema)], + &[], + &order_by, + Arc::new(window_frame), + Arc::clone(&schema), + false, + false, + Some(col("include", &schema).unwrap()), Review Comment: Every case here passes a filter, so the run has no baseline without FILTER at the same shape. That makes it hard to read the cost of the filter machinery apart from the base window cost. Would it work to add one case per group that passes `None` here? ########## datafusion/physical-plan/benches/window_filter.rs: ########## @@ -0,0 +1,263 @@ +// 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. + +//! Microbenchmark for window aggregates with `FILTER`. The benchmark uses +//! pre-ordered input to exclude sorting and query planning, but executes the +//! physical window plan so both stateful and whole-partition evaluation paths +//! are represented. + +use std::hint::black_box; +use std::sync::Arc; + +use arrow::array::{BooleanArray, Float64Array, UInt64Array}; +use arrow::datatypes::{DataType, Field, Schema, SchemaRef}; +use arrow::record_batch::RecordBatch; +use criterion::{BenchmarkId, Criterion, criterion_group, criterion_main}; +use datafusion_common::{ScalarValue, config::ConfigOptions}; +use datafusion_execution::TaskContext; +use datafusion_expr::{ + Operator, WindowFrame, WindowFrameBound, WindowFrameUnits, WindowFunctionDefinition, +}; +use datafusion_functions::math::power; +use datafusion_functions_aggregate::sum::sum_udaf; +use datafusion_physical_expr::expressions::{BinaryExpr, col, lit}; +use datafusion_physical_expr::{ + LexOrdering, PhysicalExpr, PhysicalSortExpr, ScalarFunctionExpr, +}; +use datafusion_physical_plan::test::TestMemoryExec; +use datafusion_physical_plan::windows::{ + BoundedWindowAggExec, WindowAggExec, create_window_expr, +}; +use datafusion_physical_plan::{ExecutionPlan, InputOrderMode, collect}; + +const BATCH_SIZE: usize = 8192; +const NUM_BATCHES: usize = 4; + +#[derive(Clone, Copy)] +enum ArgumentKind { + Column, + Divide, + Power, +} + +impl ArgumentKind { + fn name(self) -> &'static str { + match self { + Self::Column => "column", + Self::Divide => "divide", + Self::Power => "power_udf", + } + } +} + +fn schema() -> SchemaRef { + Arc::new(Schema::new(vec![ + Field::new("id", DataType::UInt64, false), + Field::new("value", DataType::Float64, false), + Field::new("include", DataType::Boolean, false), + ])) +} + +fn make_batches(filter_percent: usize) -> Vec<RecordBatch> { + (0..NUM_BATCHES) + .map(|batch_index| { + let start = batch_index * BATCH_SIZE; + let end = start + BATCH_SIZE; + let id = UInt64Array::from_iter_values((start..end).map(|i| i as u64)); + let value = Float64Array::from_iter_values((start..end).map(|i| i as f64)); + let include = BooleanArray::from( + (start..end) + .map(|i| (i * filter_percent) % 100 < filter_percent) Review Comment: At 10, 30 and 50 percent this expression never selects two rows in a row, so every selected run is one row long. The whole-partition case filters the full batch, and that shape is the most expensive one for `arrow::compute::filter`. `i % 100 < filter_percent` gives the same selectivity with clustered runs. Would it be worth a second pattern with clustered runs, so the numbers show the range rather than one end of it? ########## datafusion/physical-plan/benches/window_filter.rs: ########## @@ -0,0 +1,263 @@ +// 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. + +//! Microbenchmark for window aggregates with `FILTER`. The benchmark uses +//! pre-ordered input to exclude sorting and query planning, but executes the +//! physical window plan so both stateful and whole-partition evaluation paths +//! are represented. + +use std::hint::black_box; +use std::sync::Arc; + +use arrow::array::{BooleanArray, Float64Array, UInt64Array}; +use arrow::datatypes::{DataType, Field, Schema, SchemaRef}; +use arrow::record_batch::RecordBatch; +use criterion::{BenchmarkId, Criterion, criterion_group, criterion_main}; +use datafusion_common::{ScalarValue, config::ConfigOptions}; +use datafusion_execution::TaskContext; +use datafusion_expr::{ + Operator, WindowFrame, WindowFrameBound, WindowFrameUnits, WindowFunctionDefinition, +}; +use datafusion_functions::math::power; +use datafusion_functions_aggregate::sum::sum_udaf; +use datafusion_physical_expr::expressions::{BinaryExpr, col, lit}; +use datafusion_physical_expr::{ + LexOrdering, PhysicalExpr, PhysicalSortExpr, ScalarFunctionExpr, +}; +use datafusion_physical_plan::test::TestMemoryExec; +use datafusion_physical_plan::windows::{ + BoundedWindowAggExec, WindowAggExec, create_window_expr, +}; +use datafusion_physical_plan::{ExecutionPlan, InputOrderMode, collect}; + +const BATCH_SIZE: usize = 8192; +const NUM_BATCHES: usize = 4; + +#[derive(Clone, Copy)] +enum ArgumentKind { + Column, + Divide, + Power, +} + +impl ArgumentKind { + fn name(self) -> &'static str { + match self { + Self::Column => "column", + Self::Divide => "divide", + Self::Power => "power_udf", + } + } +} + +fn schema() -> SchemaRef { + Arc::new(Schema::new(vec![ + Field::new("id", DataType::UInt64, false), + Field::new("value", DataType::Float64, false), + Field::new("include", DataType::Boolean, false), + ])) +} + +fn make_batches(filter_percent: usize) -> Vec<RecordBatch> { + (0..NUM_BATCHES) + .map(|batch_index| { + let start = batch_index * BATCH_SIZE; + let end = start + BATCH_SIZE; + let id = UInt64Array::from_iter_values((start..end).map(|i| i as u64)); + let value = Float64Array::from_iter_values((start..end).map(|i| i as f64)); + let include = BooleanArray::from( + (start..end) + .map(|i| (i * filter_percent) % 100 < filter_percent) + .collect::<Vec<_>>(), + ); + + RecordBatch::try_new( + schema(), + vec![Arc::new(id), Arc::new(value), Arc::new(include)], + ) + .unwrap() + }) + .collect() +} + +fn window_argument(kind: ArgumentKind, schema: &Schema) -> Arc<dyn PhysicalExpr> { + let column = col("value", schema).unwrap(); + match kind { + ArgumentKind::Column => column, + ArgumentKind::Divide => { + Arc::new(BinaryExpr::new(column, Operator::Divide, lit(10.0_f64))) + } + ArgumentKind::Power => Arc::new( + ScalarFunctionExpr::try_new( + power(), + vec![column, lit(1.5_f64)], + schema, + Arc::new(ConfigOptions::default()), + ) + .unwrap(), + ), + } +} + +fn cumulative_frame() -> WindowFrame { + WindowFrame::new_bounds( + WindowFrameUnits::Rows, + WindowFrameBound::Preceding(ScalarValue::UInt64(None)), + WindowFrameBound::CurrentRow, + ) +} + +fn sliding_frame() -> WindowFrame { + WindowFrame::new_bounds( + WindowFrameUnits::Rows, + WindowFrameBound::Preceding(ScalarValue::UInt64(Some(10))), + WindowFrameBound::CurrentRow, + ) +} + +fn whole_partition_frame() -> WindowFrame { + WindowFrame::new_bounds( + WindowFrameUnits::Rows, + WindowFrameBound::Preceding(ScalarValue::UInt64(None)), + WindowFrameBound::Following(ScalarValue::UInt64(None)), + ) +} + +fn make_window_plan( + filter_percent: usize, + argument_kind: ArgumentKind, + window_frame: WindowFrame, +) -> Arc<dyn ExecutionPlan> { + let schema = schema(); + let order_by = vec![PhysicalSortExpr { + expr: col("id", &schema).unwrap(), + options: Default::default(), + }]; + let window_expr = create_window_expr( + &WindowFunctionDefinition::AggregateUDF(sum_udaf()), + format!("sum({}) FILTER (WHERE include)", argument_kind.name()), + &[window_argument(argument_kind, &schema)], + &[], + &order_by, + Arc::new(window_frame), + Arc::clone(&schema), + false, + false, + Some(col("include", &schema).unwrap()), + ) + .unwrap(); + + let source = TestMemoryExec::try_new(&[make_batches(filter_percent)], schema, None) + .unwrap() + .try_with_sort_information(LexOrdering::new(order_by).into_iter().collect()) + .unwrap(); + let input: Arc<dyn ExecutionPlan> = + Arc::new(TestMemoryExec::update_cache(&Arc::new(source))); + + if window_expr.uses_bounded_memory() { + Arc::new( + BoundedWindowAggExec::try_new( + vec![window_expr], + input, + InputOrderMode::Sorted, + false, + ) + .unwrap(), + ) + } else { + Arc::new(WindowAggExec::try_new(vec![window_expr], input, false).unwrap()) + } +} + +fn benchmark_window_case( + c: &mut Criterion, + runtime: &tokio::runtime::Runtime, + name: &str, + window_frame: &WindowFrame, + argument_kinds: &[ArgumentKind], + filter_percents: &[usize], + sample_size: usize, +) { + let mut group = c.benchmark_group(format!("window_aggregate_filter/{name}")); + group.sample_size(sample_size); + + for &filter_percent in filter_percents { + for &argument_kind in argument_kinds { + let plan = + make_window_plan(filter_percent, argument_kind, window_frame.clone()); + let task_ctx = Arc::new(TaskContext::default()); + group.bench_function( + BenchmarkId::new( + argument_kind.name(), + format!("{filter_percent}_percent"), + ), + |b| { + b.iter(|| { + let batches = runtime + .block_on(collect(Arc::clone(&plan), Arc::clone(&task_ctx))) + .unwrap(); + black_box(batches); + }) + }, + ); + } + } + + group.finish(); +} + +fn window_filter_benchmark(c: &mut Criterion) { + let runtime = tokio::runtime::Runtime::new().unwrap(); + benchmark_window_case( + c, + &runtime, + "bounded_cumulative", + &cumulative_frame(), + &[ + ArgumentKind::Column, + ArgumentKind::Divide, + ArgumentKind::Power, + ], + &[10, 30, 50], + 10, + ); + benchmark_window_case( + c, + &runtime, + "bounded_sliding_10_rows", + &sliding_frame(), Review Comment: The sliding frame filters twice per row, once for the update slice and once for the retract slice. That makes it the heaviest filter path of the three groups, but it skips `divide`. Would you add it here too? -- 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]
