query/optimizer/
windowed_sort.rs

1// Copyright 2023 Greptime Team
2//
3// Licensed under the Apache License, Version 2.0 (the "License");
4// you may not use this file except in compliance with the License.
5// You may obtain a copy of the License at
6//
7//     http://www.apache.org/licenses/LICENSE-2.0
8//
9// Unless required by applicable law or agreed to in writing, software
10// distributed under the License is distributed on an "AS IS" BASIS,
11// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12// See the License for the specific language governing permissions and
13// limitations under the License.
14
15use std::collections::HashSet;
16use std::sync::Arc;
17
18use datafusion::physical_optimizer::PhysicalOptimizerRule;
19use datafusion::physical_plan::coalesce_batches::CoalesceBatchesExec;
20use datafusion::physical_plan::coalesce_partitions::CoalescePartitionsExec;
21use datafusion::physical_plan::filter::FilterExec;
22use datafusion::physical_plan::projection::ProjectionExec;
23use datafusion::physical_plan::repartition::RepartitionExec;
24use datafusion::physical_plan::sorts::sort::SortExec;
25use datafusion::physical_plan::sorts::sort_preserving_merge::SortPreservingMergeExec;
26use datafusion::physical_plan::ExecutionPlan;
27use datafusion_common::tree_node::{Transformed, TreeNode};
28use datafusion_common::Result as DataFusionResult;
29use datafusion_physical_expr::expressions::Column as PhysicalColumn;
30use datafusion_physical_expr::LexOrdering;
31use store_api::region_engine::PartitionRange;
32use table::table::scan::RegionScanExec;
33
34use crate::part_sort::PartSortExec;
35use crate::window_sort::WindowedSortExec;
36
37/// Optimize rule for windowed sort.
38///
39/// This is expected to run after [`ScanHint`] and [`ParallelizeScan`].
40/// It would change the original sort to a custom plan. To make sure
41/// other rules are applied correctly, this rule can be run as later as
42/// possible.
43///
44/// [`ScanHint`]: crate::optimizer::scan_hint::ScanHintRule
45/// [`ParallelizeScan`]: crate::optimizer::parallelize_scan::ParallelizeScan
46#[derive(Debug)]
47pub struct WindowedSortPhysicalRule;
48
49impl PhysicalOptimizerRule for WindowedSortPhysicalRule {
50    fn optimize(
51        &self,
52        plan: Arc<dyn ExecutionPlan>,
53        config: &datafusion::config::ConfigOptions,
54    ) -> DataFusionResult<Arc<dyn ExecutionPlan>> {
55        Self::do_optimize(plan, config)
56    }
57
58    fn name(&self) -> &str {
59        "WindowedSortRule"
60    }
61
62    fn schema_check(&self) -> bool {
63        false
64    }
65}
66
67impl WindowedSortPhysicalRule {
68    fn do_optimize(
69        plan: Arc<dyn ExecutionPlan>,
70        _config: &datafusion::config::ConfigOptions,
71    ) -> DataFusionResult<Arc<dyn ExecutionPlan>> {
72        let result = plan
73            .transform_down(|plan| {
74                if let Some(sort_exec) = plan.as_any().downcast_ref::<SortExec>() {
75                    // TODO: support multiple expr in windowed sort
76                    if sort_exec.expr().len() != 1 {
77                        return Ok(Transformed::no(plan));
78                    }
79
80                    let preserve_partitioning = sort_exec.preserve_partitioning();
81
82                    let sort_input = remove_repartition(sort_exec.input().clone())?.data;
83                    let sort_input =
84                        remove_coalesce_batches_exec(sort_input, sort_exec.fetch())?.data;
85
86                    // Gets scanner info from the input without repartition before filter.
87                    let Some(scanner_info) = fetch_partition_range(sort_input.clone())? else {
88                        return Ok(Transformed::no(plan));
89                    };
90                    let input_schema = sort_input.schema();
91
92                    if let Some(first_sort_expr) = sort_exec.expr().first()
93                        && let Some(column_expr) = first_sort_expr
94                            .expr
95                            .as_any()
96                            .downcast_ref::<PhysicalColumn>()
97                        && scanner_info
98                            .time_index
99                            .contains(input_schema.field(column_expr.index()).name())
100                    {
101                    } else {
102                        return Ok(Transformed::no(plan));
103                    }
104                    let first_sort_expr = sort_exec.expr().first().unwrap().clone();
105
106                    // PartSortExec is unnecessary if:
107                    // - there is no tag column, and
108                    // - the sort is ascending on the time index column
109                    let new_input = if scanner_info.tag_columns.is_empty()
110                        && !first_sort_expr.options.descending
111                    {
112                        sort_input
113                    } else {
114                        Arc::new(PartSortExec::new(
115                            first_sort_expr.clone(),
116                            sort_exec.fetch(),
117                            scanner_info.partition_ranges.clone(),
118                            sort_input,
119                        ))
120                    };
121
122                    let windowed_sort_exec = WindowedSortExec::try_new(
123                        first_sort_expr,
124                        sort_exec.fetch(),
125                        scanner_info.partition_ranges,
126                        new_input,
127                    )?;
128
129                    if !preserve_partitioning {
130                        let order_preserving_merge = SortPreservingMergeExec::new(
131                            LexOrdering::new(sort_exec.expr().to_vec()),
132                            Arc::new(windowed_sort_exec),
133                        );
134                        return Ok(Transformed {
135                            data: Arc::new(order_preserving_merge),
136                            transformed: true,
137                            tnr: datafusion_common::tree_node::TreeNodeRecursion::Stop,
138                        });
139                    } else {
140                        return Ok(Transformed {
141                            data: Arc::new(windowed_sort_exec),
142                            transformed: true,
143                            tnr: datafusion_common::tree_node::TreeNodeRecursion::Stop,
144                        });
145                    }
146                }
147
148                Ok(Transformed::no(plan))
149            })?
150            .data;
151
152        Ok(result)
153    }
154}
155
156#[derive(Debug)]
157struct ScannerInfo {
158    partition_ranges: Vec<Vec<PartitionRange>>,
159    time_index: HashSet<String>,
160    tag_columns: Vec<String>,
161}
162
163fn fetch_partition_range(input: Arc<dyn ExecutionPlan>) -> DataFusionResult<Option<ScannerInfo>> {
164    let mut partition_ranges = None;
165    let mut time_index = HashSet::new();
166    let mut alias_map = Vec::new();
167    let mut tag_columns = None;
168    let mut is_batch_coalesced = false;
169
170    input.transform_up(|plan| {
171        // Unappliable case, reset the state.
172        if plan.as_any().is::<RepartitionExec>()
173            || plan.as_any().is::<CoalescePartitionsExec>()
174            || plan.as_any().is::<SortExec>()
175            || plan.as_any().is::<WindowedSortExec>()
176        {
177            partition_ranges = None;
178        }
179
180        if plan.as_any().is::<CoalesceBatchesExec>() {
181            is_batch_coalesced = true;
182        }
183
184        // TODO(discord9): do this in logical plan instead as it's lessy bugy there
185        // Collects alias of the time index column.
186        if let Some(projection) = plan.as_any().downcast_ref::<ProjectionExec>() {
187            for (expr, output_name) in projection.expr() {
188                if let Some(column_expr) = expr.as_any().downcast_ref::<PhysicalColumn>() {
189                    alias_map.push((column_expr.name().to_string(), output_name.clone()));
190                }
191            }
192            // resolve alias properly
193            time_index = resolve_alias(&alias_map, &time_index);
194        }
195
196        if let Some(region_scan_exec) = plan.as_any().downcast_ref::<RegionScanExec>() {
197            partition_ranges = Some(region_scan_exec.get_uncollapsed_partition_ranges());
198            // Reset time index column.
199            time_index = HashSet::from([region_scan_exec.time_index()]);
200            tag_columns = Some(region_scan_exec.tag_columns());
201
202            // set distinguish_partition_ranges to true, this is an incorrect workaround
203            if !is_batch_coalesced {
204                region_scan_exec.with_distinguish_partition_range(true);
205            }
206        }
207
208        Ok(Transformed::no(plan))
209    })?;
210
211    let result = try {
212        ScannerInfo {
213            partition_ranges: partition_ranges?,
214            time_index,
215            tag_columns: tag_columns?,
216        }
217    };
218
219    Ok(result)
220}
221
222/// Removes the repartition plan between the filter and region scan.
223fn remove_repartition(
224    plan: Arc<dyn ExecutionPlan>,
225) -> DataFusionResult<Transformed<Arc<dyn ExecutionPlan>>> {
226    plan.transform_down(|plan| {
227        if plan.as_any().is::<FilterExec>() {
228            // Checks child.
229            let maybe_repartition = plan.children()[0];
230            if maybe_repartition.as_any().is::<RepartitionExec>() {
231                let maybe_scan = maybe_repartition.children()[0];
232                if maybe_scan.as_any().is::<RegionScanExec>() {
233                    let new_filter = plan.clone().with_new_children(vec![maybe_scan.clone()])?;
234                    return Ok(Transformed::yes(new_filter));
235                }
236            }
237        }
238
239        Ok(Transformed::no(plan))
240    })
241}
242
243/// Remove `CoalesceBatchesExec` if the limit is less than the batch size.
244///
245/// so that if limit is too small we can avoid need to scan for more rows than necessary
246fn remove_coalesce_batches_exec(
247    plan: Arc<dyn ExecutionPlan>,
248    fetch: Option<usize>,
249) -> DataFusionResult<Transformed<Arc<dyn ExecutionPlan>>> {
250    let Some(fetch) = fetch else {
251        return Ok(Transformed::no(plan));
252    };
253
254    // Avoid removing multiple coalesce batches
255    let mut is_done = false;
256
257    plan.transform_down(|plan| {
258        if let Some(coalesce_batches_exec) = plan.as_any().downcast_ref::<CoalesceBatchesExec>() {
259            let target_batch_size = coalesce_batches_exec.target_batch_size();
260            if fetch < target_batch_size && !is_done {
261                is_done = true;
262                return Ok(Transformed::yes(coalesce_batches_exec.input().clone()));
263            }
264        }
265
266        Ok(Transformed::no(plan))
267    })
268}
269
270/// Resolves alias of the time index column.
271///
272/// i.e if a is time index, alias= {a:b, b:c}, then result should be {a, b}(not {a, c}) because projection is not transitive
273/// if alias={b:a} and a is time index, then return empty
274fn resolve_alias(alias_map: &[(String, String)], time_index: &HashSet<String>) -> HashSet<String> {
275    // available old name for time index
276    let mut avail_old_name = time_index.clone();
277    let mut new_time_index = HashSet::new();
278    for (old, new) in alias_map {
279        if time_index.contains(old) {
280            new_time_index.insert(new.clone());
281        } else if time_index.contains(new) && old != new {
282            // other alias to time index, remove the old name
283            avail_old_name.remove(new);
284            continue;
285        }
286    }
287    // add the remaining time index that is not in alias map
288    new_time_index.extend(avail_old_name);
289    new_time_index
290}
291
292#[cfg(test)]
293mod test {
294    use itertools::Itertools;
295
296    use super::*;
297
298    #[test]
299    fn test_alias() {
300        let testcases = [
301            // notice the old name is still in the result
302            (
303                vec![("a", "b"), ("b", "c")],
304                HashSet::from(["a"]),
305                HashSet::from(["a", "b"]),
306            ),
307            // alias swap
308            (
309                vec![("b", "a"), ("a", "b")],
310                HashSet::from(["a"]),
311                HashSet::from(["b"]),
312            ),
313            (
314                vec![("b", "a"), ("b", "c")],
315                HashSet::from(["a"]),
316                HashSet::from([]),
317            ),
318            // not in alias map
319            (
320                vec![("c", "d"), ("d", "c")],
321                HashSet::from(["a"]),
322                HashSet::from(["a"]),
323            ),
324            // no alias
325            (vec![], HashSet::from(["a"]), HashSet::from(["a"])),
326            // empty time index
327            (vec![], HashSet::from([]), HashSet::from([])),
328        ];
329        for (alias_map, time_index, expected) in testcases {
330            let alias_map = alias_map
331                .into_iter()
332                .map(|(k, v)| (k.to_string(), v.to_string()))
333                .collect_vec();
334            let time_index = time_index.into_iter().map(|i| i.to_string()).collect();
335            let expected: HashSet<String> = expected.into_iter().map(|i| i.to_string()).collect();
336
337            assert_eq!(
338                expected,
339                resolve_alias(&alias_map, &time_index),
340                "alias_map={:?}, time_index={:?}",
341                alias_map,
342                time_index
343            );
344        }
345    }
346}
query/optimizer/windowed_sort.rs

query/optimizer/
windowed_sort.rs
