query/optimizer/
windowed_sort.rs

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
// Copyright 2023 Greptime Team
//
// Licensed 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 std::collections::HashSet;
use std::sync::Arc;

use datafusion::physical_optimizer::PhysicalOptimizerRule;
use datafusion::physical_plan::coalesce_batches::CoalesceBatchesExec;
use datafusion::physical_plan::coalesce_partitions::CoalescePartitionsExec;
use datafusion::physical_plan::filter::FilterExec;
use datafusion::physical_plan::projection::ProjectionExec;
use datafusion::physical_plan::repartition::RepartitionExec;
use datafusion::physical_plan::sorts::sort::SortExec;
use datafusion::physical_plan::sorts::sort_preserving_merge::SortPreservingMergeExec;
use datafusion::physical_plan::ExecutionPlan;
use datafusion_common::tree_node::{Transformed, TreeNode};
use datafusion_common::Result as DataFusionResult;
use datafusion_physical_expr::expressions::Column as PhysicalColumn;
use datafusion_physical_expr::LexOrdering;
use store_api::region_engine::PartitionRange;
use table::table::scan::RegionScanExec;

use crate::part_sort::PartSortExec;
use crate::window_sort::WindowedSortExec;

/// Optimize rule for windowed sort.
///
/// This is expected to run after [`ScanHint`] and [`ParallelizeScan`].
/// It would change the original sort to a custom plan. To make sure
/// other rules are applied correctly, this rule can be run as later as
/// possible.
///
/// [`ScanHint`]: crate::optimizer::scan_hint::ScanHintRule
/// [`ParallelizeScan`]: crate::optimizer::parallelize_scan::ParallelizeScan
#[derive(Debug)]
pub struct WindowedSortPhysicalRule;

impl PhysicalOptimizerRule for WindowedSortPhysicalRule {
    fn optimize(
        &self,
        plan: Arc<dyn ExecutionPlan>,
        config: &datafusion::config::ConfigOptions,
    ) -> DataFusionResult<Arc<dyn ExecutionPlan>> {
        Self::do_optimize(plan, config)
    }

    fn name(&self) -> &str {
        "WindowedSortRule"
    }

    fn schema_check(&self) -> bool {
        false
    }
}

impl WindowedSortPhysicalRule {
    fn do_optimize(
        plan: Arc<dyn ExecutionPlan>,
        _config: &datafusion::config::ConfigOptions,
    ) -> DataFusionResult<Arc<dyn ExecutionPlan>> {
        let result = plan
            .transform_down(|plan| {
                if let Some(sort_exec) = plan.as_any().downcast_ref::<SortExec>() {
                    // TODO: support multiple expr in windowed sort
                    if sort_exec.expr().len() != 1 {
                        return Ok(Transformed::no(plan));
                    }

                    let preserve_partitioning = sort_exec.preserve_partitioning();

                    let sort_input = remove_repartition(sort_exec.input().clone())?.data;
                    // Gets scanner info from the input without repartition before filter.
                    let Some(scanner_info) = fetch_partition_range(sort_input.clone())? else {
                        return Ok(Transformed::no(plan));
                    };

                    if let Some(first_sort_expr) = sort_exec.expr().first()
                        && let Some(column_expr) = first_sort_expr
                            .expr
                            .as_any()
                            .downcast_ref::<PhysicalColumn>()
                        && scanner_info.time_index.contains(column_expr.name())
                    {
                    } else {
                        return Ok(Transformed::no(plan));
                    }
                    let first_sort_expr = sort_exec.expr().first().unwrap().clone();

                    // PartSortExec is unnecessary if:
                    // - there is no tag column, and
                    // - the sort is ascending on the time index column
                    let new_input = if scanner_info.tag_columns.is_empty()
                        && !first_sort_expr.options.descending
                    {
                        sort_input
                    } else {
                        Arc::new(PartSortExec::new(
                            first_sort_expr.clone(),
                            sort_exec.fetch(),
                            scanner_info.partition_ranges.clone(),
                            sort_input,
                        ))
                    };

                    let windowed_sort_exec = WindowedSortExec::try_new(
                        first_sort_expr,
                        sort_exec.fetch(),
                        scanner_info.partition_ranges,
                        new_input,
                    )?;

                    if !preserve_partitioning {
                        let order_preserving_merge = SortPreservingMergeExec::new(
                            LexOrdering::new(sort_exec.expr().to_vec()),
                            Arc::new(windowed_sort_exec),
                        );
                        return Ok(Transformed {
                            data: Arc::new(order_preserving_merge),
                            transformed: true,
                            tnr: datafusion_common::tree_node::TreeNodeRecursion::Stop,
                        });
                    } else {
                        return Ok(Transformed {
                            data: Arc::new(windowed_sort_exec),
                            transformed: true,
                            tnr: datafusion_common::tree_node::TreeNodeRecursion::Stop,
                        });
                    }
                }

                Ok(Transformed::no(plan))
            })?
            .data;

        Ok(result)
    }
}

#[derive(Debug)]
struct ScannerInfo {
    partition_ranges: Vec<Vec<PartitionRange>>,
    time_index: HashSet<String>,
    tag_columns: Vec<String>,
}

fn fetch_partition_range(input: Arc<dyn ExecutionPlan>) -> DataFusionResult<Option<ScannerInfo>> {
    let mut partition_ranges = None;
    let mut time_index = HashSet::new();
    let mut tag_columns = None;
    let mut is_batch_coalesced = false;

    input.transform_up(|plan| {
        // Unappliable case, reset the state.
        if plan.as_any().is::<RepartitionExec>()
            || plan.as_any().is::<CoalescePartitionsExec>()
            || plan.as_any().is::<SortExec>()
            || plan.as_any().is::<WindowedSortExec>()
        {
            partition_ranges = None;
        }

        if plan.as_any().is::<CoalesceBatchesExec>() {
            is_batch_coalesced = true;
        }

        // Collects alias of the time index column.
        if let Some(projection) = plan.as_any().downcast_ref::<ProjectionExec>() {
            for (expr, output_name) in projection.expr() {
                if let Some(column_expr) = expr.as_any().downcast_ref::<PhysicalColumn>() {
                    if time_index.contains(column_expr.name()) {
                        time_index.insert(output_name.clone());
                    }
                }
            }
        }

        if let Some(region_scan_exec) = plan.as_any().downcast_ref::<RegionScanExec>() {
            partition_ranges = Some(region_scan_exec.get_uncollapsed_partition_ranges());
            // Reset time index column.
            time_index = HashSet::from([region_scan_exec.time_index()]);
            tag_columns = Some(region_scan_exec.tag_columns());

            // set distinguish_partition_ranges to true, this is an incorrect workaround
            if !is_batch_coalesced {
                region_scan_exec.with_distinguish_partition_range(true);
            }
        }

        Ok(Transformed::no(plan))
    })?;

    let result = try {
        ScannerInfo {
            partition_ranges: partition_ranges?,
            time_index,
            tag_columns: tag_columns?,
        }
    };

    Ok(result)
}

/// Removes the repartition plan between the filter and region scan.
fn remove_repartition(
    plan: Arc<dyn ExecutionPlan>,
) -> DataFusionResult<Transformed<Arc<dyn ExecutionPlan>>> {
    plan.transform_down(|plan| {
        if plan.as_any().is::<FilterExec>() {
            // Checks child.
            let maybe_repartition = plan.children()[0];
            if maybe_repartition.as_any().is::<RepartitionExec>() {
                let maybe_scan = maybe_repartition.children()[0];
                if maybe_scan.as_any().is::<RegionScanExec>() {
                    let new_filter = plan.clone().with_new_children(vec![maybe_scan.clone()])?;
                    return Ok(Transformed::yes(new_filter));
                }
            }
        }

        Ok(Transformed::no(plan))
    })
}