datafusion_physical_expr/window/
standard.rs

1// Licensed to the Apache Software Foundation (ASF) under one
2// or more contributor license agreements.  See the NOTICE file
3// distributed with this work for additional information
4// regarding copyright ownership.  The ASF licenses this file
5// to you under the Apache License, Version 2.0 (the
6// "License"); you may not use this file except in compliance
7// with the License.  You may obtain a copy of the License at
8//
9//   http://www.apache.org/licenses/LICENSE-2.0
10//
11// Unless required by applicable law or agreed to in writing,
12// software distributed under the License is distributed on an
13// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
14// KIND, either express or implied.  See the License for the
15// specific language governing permissions and limitations
16// under the License.
17
18//! Physical exec for standard window function expressions.
19
20use std::any::Any;
21use std::ops::Range;
22use std::sync::Arc;
23
24use super::{StandardWindowFunctionExpr, WindowExpr};
25use crate::window::window_expr::{get_orderby_values, WindowFn};
26use crate::window::{PartitionBatches, PartitionWindowAggStates, WindowState};
27use crate::{reverse_order_bys, EquivalenceProperties, PhysicalExpr};
28use arrow::array::{new_empty_array, ArrayRef};
29use arrow::compute::SortOptions;
30use arrow::datatypes::Field;
31use arrow::record_batch::RecordBatch;
32use datafusion_common::utils::evaluate_partition_ranges;
33use datafusion_common::{Result, ScalarValue};
34use datafusion_expr::window_state::{WindowAggState, WindowFrameContext};
35use datafusion_expr::WindowFrame;
36use datafusion_physical_expr_common::sort_expr::{LexOrdering, PhysicalSortExpr};
37
38/// A window expr that takes the form of a [`StandardWindowFunctionExpr`].
39#[derive(Debug)]
40pub struct StandardWindowExpr {
41    expr: Arc<dyn StandardWindowFunctionExpr>,
42    partition_by: Vec<Arc<dyn PhysicalExpr>>,
43    order_by: LexOrdering,
44    window_frame: Arc<WindowFrame>,
45}
46
47impl StandardWindowExpr {
48    /// create a new standard window function expression
49    pub fn new(
50        expr: Arc<dyn StandardWindowFunctionExpr>,
51        partition_by: &[Arc<dyn PhysicalExpr>],
52        order_by: &LexOrdering,
53        window_frame: Arc<WindowFrame>,
54    ) -> Self {
55        Self {
56            expr,
57            partition_by: partition_by.to_vec(),
58            order_by: order_by.clone(),
59            window_frame,
60        }
61    }
62
63    /// Get StandardWindowFunction expr of StandardWindowExpr
64    pub fn get_standard_func_expr(&self) -> &Arc<dyn StandardWindowFunctionExpr> {
65        &self.expr
66    }
67
68    /// Adds any equivalent orderings generated by `self.expr` to `builder`.
69    ///
70    /// If `self.expr` doesn't have an ordering, ordering equivalence properties
71    /// are not updated. Otherwise, ordering equivalence properties are updated
72    /// by the ordering of `self.expr`.
73    pub fn add_equal_orderings(&self, eq_properties: &mut EquivalenceProperties) {
74        let schema = eq_properties.schema();
75        if let Some(fn_res_ordering) = self.expr.get_result_ordering(schema) {
76            add_new_ordering_expr_with_partition_by(
77                eq_properties,
78                fn_res_ordering,
79                &self.partition_by,
80            );
81        }
82    }
83}
84
85impl WindowExpr for StandardWindowExpr {
86    /// Return a reference to Any that can be used for downcasting
87    fn as_any(&self) -> &dyn Any {
88        self
89    }
90
91    fn name(&self) -> &str {
92        self.expr.name()
93    }
94
95    fn field(&self) -> Result<Field> {
96        self.expr.field()
97    }
98
99    fn expressions(&self) -> Vec<Arc<dyn PhysicalExpr>> {
100        self.expr.expressions()
101    }
102
103    fn partition_by(&self) -> &[Arc<dyn PhysicalExpr>] {
104        &self.partition_by
105    }
106
107    fn order_by(&self) -> &LexOrdering {
108        self.order_by.as_ref()
109    }
110
111    fn evaluate(&self, batch: &RecordBatch) -> Result<ArrayRef> {
112        let mut evaluator = self.expr.create_evaluator()?;
113        let num_rows = batch.num_rows();
114        if evaluator.uses_window_frame() {
115            let sort_options: Vec<SortOptions> =
116                self.order_by.iter().map(|o| o.options).collect();
117            let mut row_wise_results = vec![];
118
119            let mut values = self.evaluate_args(batch)?;
120            let order_bys = get_orderby_values(self.order_by_columns(batch)?);
121            let n_args = values.len();
122            values.extend(order_bys);
123            let order_bys_ref = &values[n_args..];
124
125            let mut window_frame_ctx =
126                WindowFrameContext::new(Arc::clone(&self.window_frame), sort_options);
127            let mut last_range = Range { start: 0, end: 0 };
128            // We iterate on each row to calculate window frame range and and window function result
129            for idx in 0..num_rows {
130                let range = window_frame_ctx.calculate_range(
131                    order_bys_ref,
132                    &last_range,
133                    num_rows,
134                    idx,
135                )?;
136                let value = evaluator.evaluate(&values, &range)?;
137                row_wise_results.push(value);
138                last_range = range;
139            }
140            ScalarValue::iter_to_array(row_wise_results)
141        } else if evaluator.include_rank() {
142            let columns = self.order_by_columns(batch)?;
143            let sort_partition_points = evaluate_partition_ranges(num_rows, &columns)?;
144            evaluator.evaluate_all_with_rank(num_rows, &sort_partition_points)
145        } else {
146            let values = self.evaluate_args(batch)?;
147            evaluator.evaluate_all(&values, num_rows)
148        }
149    }
150
151    /// Evaluate the window function against the batch. This function facilitates
152    /// stateful, bounded-memory implementations.
153    fn evaluate_stateful(
154        &self,
155        partition_batches: &PartitionBatches,
156        window_agg_state: &mut PartitionWindowAggStates,
157    ) -> Result<()> {
158        let field = self.expr.field()?;
159        let out_type = field.data_type();
160        let sort_options = self.order_by.iter().map(|o| o.options).collect::<Vec<_>>();
161        for (partition_row, partition_batch_state) in partition_batches.iter() {
162            let window_state =
163                if let Some(window_state) = window_agg_state.get_mut(partition_row) {
164                    window_state
165                } else {
166                    let evaluator = self.expr.create_evaluator()?;
167                    window_agg_state
168                        .entry(partition_row.clone())
169                        .or_insert(WindowState {
170                            state: WindowAggState::new(out_type)?,
171                            window_fn: WindowFn::Builtin(evaluator),
172                        })
173                };
174            let evaluator = match &mut window_state.window_fn {
175                WindowFn::Builtin(evaluator) => evaluator,
176                _ => unreachable!(),
177            };
178            let state = &mut window_state.state;
179
180            let batch_ref = &partition_batch_state.record_batch;
181            let mut values = self.evaluate_args(batch_ref)?;
182            let order_bys = if evaluator.uses_window_frame() || evaluator.include_rank() {
183                get_orderby_values(self.order_by_columns(batch_ref)?)
184            } else {
185                vec![]
186            };
187            let n_args = values.len();
188            values.extend(order_bys);
189            let order_bys_ref = &values[n_args..];
190
191            // We iterate on each row to perform a running calculation.
192            let record_batch = &partition_batch_state.record_batch;
193            let num_rows = record_batch.num_rows();
194            let mut row_wise_results: Vec<ScalarValue> = vec![];
195            let is_causal = if evaluator.uses_window_frame() {
196                self.window_frame.is_causal()
197            } else {
198                evaluator.is_causal()
199            };
200            for idx in state.last_calculated_index..num_rows {
201                let frame_range = if evaluator.uses_window_frame() {
202                    state
203                        .window_frame_ctx
204                        .get_or_insert_with(|| {
205                            WindowFrameContext::new(
206                                Arc::clone(&self.window_frame),
207                                sort_options.clone(),
208                            )
209                        })
210                        .calculate_range(
211                            order_bys_ref,
212                            // Start search from the last range
213                            &state.window_frame_range,
214                            num_rows,
215                            idx,
216                        )
217                } else {
218                    evaluator.get_range(idx, num_rows)
219                }?;
220
221                // Exit if the range is non-causal and extends all the way:
222                if frame_range.end == num_rows
223                    && !is_causal
224                    && !partition_batch_state.is_end
225                {
226                    break;
227                }
228                // Update last range
229                state.window_frame_range = frame_range;
230                row_wise_results
231                    .push(evaluator.evaluate(&values, &state.window_frame_range)?);
232            }
233            let out_col = if row_wise_results.is_empty() {
234                new_empty_array(out_type)
235            } else {
236                ScalarValue::iter_to_array(row_wise_results.into_iter())?
237            };
238
239            state.update(&out_col, partition_batch_state)?;
240            if self.window_frame.start_bound.is_unbounded() {
241                evaluator.memoize(state)?;
242            }
243        }
244        Ok(())
245    }
246
247    fn get_window_frame(&self) -> &Arc<WindowFrame> {
248        &self.window_frame
249    }
250
251    fn get_reverse_expr(&self) -> Option<Arc<dyn WindowExpr>> {
252        self.expr.reverse_expr().map(|reverse_expr| {
253            Arc::new(StandardWindowExpr::new(
254                reverse_expr,
255                &self.partition_by.clone(),
256                reverse_order_bys(self.order_by.as_ref()).as_ref(),
257                Arc::new(self.window_frame.reverse()),
258            )) as _
259        })
260    }
261
262    fn uses_bounded_memory(&self) -> bool {
263        if let Ok(evaluator) = self.expr.create_evaluator() {
264            evaluator.supports_bounded_execution()
265                && (!evaluator.uses_window_frame()
266                    || !self.window_frame.end_bound.is_unbounded())
267        } else {
268            false
269        }
270    }
271}
272
273/// Adds a new ordering expression into existing ordering equivalence class(es) based on
274/// PARTITION BY information (if it exists).
275pub(crate) fn add_new_ordering_expr_with_partition_by(
276    eqp: &mut EquivalenceProperties,
277    expr: PhysicalSortExpr,
278    partition_by: &[Arc<dyn PhysicalExpr>],
279) {
280    if partition_by.is_empty() {
281        // In the absence of a PARTITION BY, ordering of `self.expr` is global:
282        eqp.add_new_orderings([LexOrdering::new(vec![expr])]);
283    } else {
284        // If we have a PARTITION BY, standard functions can not introduce
285        // a global ordering unless the existing ordering is compatible
286        // with PARTITION BY expressions. To elaborate, when PARTITION BY
287        // expressions and existing ordering expressions are equal (w.r.t.
288        // set equality), we can prefix the ordering of `self.expr` with
289        // the existing ordering.
290        let (mut ordering, _) = eqp.find_longest_permutation(partition_by);
291        if ordering.len() == partition_by.len() {
292            ordering.push(expr);
293            eqp.add_new_orderings([ordering]);
294        }
295    }
296}
datafusion_physical_expr/window/standard.rs

datafusion_physical_expr/window/
standard.rs
