Coverage Report

Created: 2026-06-05 15:29

next uncovered line (L), next uncovered region (R), next uncovered branch (B)
be/src/exec/scan/parallel_scanner_builder.cpp
Line
Count
Source
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
#include "exec/scan/parallel_scanner_builder.h"
19
20
#include <cstddef>
21
22
#include "cloud/cloud_storage_engine.h"
23
#include "cloud/cloud_tablet_hotspot.h"
24
#include "cloud/config.h"
25
#include "common/status.h"
26
#include "exec/operator/olap_scan_operator.h"
27
#include "exec/scan/olap_scanner.h"
28
#include "storage/rowset/beta_rowset.h"
29
#include "storage/segment/segment_loader.h"
30
#include "storage/tablet/base_tablet.h"
31
32
namespace doris {
33
34
138k
Status ParallelScannerBuilder::build_scanners(std::list<ScannerSPtr>& scanners) {
35
138k
    RETURN_IF_ERROR(_load());
36
138k
    if (_scan_parallelism_by_per_segment) {
37
17
        return _build_scanners_by_per_segment(scanners);
38
140k
    } else if (_is_dup_mow_key) {
39
        // Default strategy for DUP/MOW tables: split by rowids within segments
40
140k
        return _build_scanners_by_rowid(scanners);
41
18.4E
    } else {
42
        // TODO: support to split by key range
43
18.4E
        return Status::NotSupported("split by key range not supported yet.");
44
18.4E
    }
45
138k
}
46
47
141k
Status ParallelScannerBuilder::_build_scanners_by_rowid(std::list<ScannerSPtr>& scanners) {
48
141k
    DCHECK_GE(_rows_per_scanner, _min_rows_per_scanner);
49
50
627k
    for (auto&& [tablet, version] : _tablets) {
51
627k
        DCHECK(_all_read_sources.contains(tablet->tablet_id()));
52
627k
        auto& entire_read_source = _all_read_sources[tablet->tablet_id()];
53
54
627k
        if (config::is_cloud_mode()) {
55
            // FIXME(plat1ko): Avoid pointer cast
56
445k
            ExecEnv::GetInstance()->storage_engine().to_cloud().tablet_hotspot().count(*tablet);
57
445k
        }
58
59
        // `rs_splits` in `entire read source` will be devided into several partitial read sources
60
        // to build several parallel scanners, based on segment rows number. All the partitial read sources
61
        // share the same delete predicates from their corresponding entire read source.
62
627k
        TabletReadSource partitial_read_source;
63
627k
        int64_t rows_collected = 0;
64
2.58M
        for (auto& rs_split : entire_read_source.rs_splits) {
65
2.58M
            auto reader = rs_split.rs_reader;
66
2.58M
            auto rowset = reader->rowset();
67
2.58M
            const auto rowset_id = rowset->rowset_id();
68
69
2.58M
            const auto& segments_rows = _all_segments_rows[rowset_id];
70
71
2.58M
            if (rowset->num_rows() == 0) {
72
1.60M
                continue;
73
1.60M
            }
74
75
979k
            int64_t segment_start = 0;
76
979k
            auto split = RowSetSplits(reader->clone());
77
78
1.95M
            for (size_t i = 0; i != segments_rows.size(); ++i) {
79
978k
                const size_t rows_of_segment = segments_rows[i];
80
978k
                RowRanges row_ranges;
81
978k
                int64_t offset_in_segment = 0;
82
83
                // try to split large segments into RowRanges
84
2.01M
                while (offset_in_segment < rows_of_segment) {
85
1.03M
                    const int64_t remaining_rows = rows_of_segment - offset_in_segment;
86
1.03M
                    auto rows_need = _rows_per_scanner - rows_collected;
87
88
                    // 0.9: try to avoid splitting the segments into excessively small parts.
89
1.03M
                    if (rows_need >= remaining_rows * 9 / 10) {
90
978k
                        rows_need = remaining_rows;
91
978k
                    }
92
1.03M
                    DCHECK_LE(rows_need, remaining_rows);
93
94
                    // RowRange stands for range: [From, To), From is inclusive, To is exclusive.
95
1.03M
                    row_ranges.add({offset_in_segment,
96
1.03M
                                    offset_in_segment + static_cast<int64_t>(rows_need)});
97
1.03M
                    rows_collected += rows_need;
98
1.03M
                    offset_in_segment += rows_need;
99
100
                    // If collected enough rows, build a new scanner
101
1.03M
                    if (rows_collected >= _rows_per_scanner) {
102
54.5k
                        split.segment_offsets.first = segment_start,
103
54.5k
                        split.segment_offsets.second = i + 1;
104
54.5k
                        split.segment_row_ranges.emplace_back(std::move(row_ranges));
105
106
54.5k
                        DCHECK_EQ(split.segment_offsets.second - split.segment_offsets.first,
107
54.5k
                                  split.segment_row_ranges.size());
108
109
54.5k
                        partitial_read_source.rs_splits.emplace_back(std::move(split));
110
111
54.5k
                        scanners.emplace_back(_build_scanner(
112
54.5k
                                tablet, version, _key_ranges,
113
54.5k
                                {.rs_splits = std::move(partitial_read_source.rs_splits),
114
54.5k
                                 .delete_predicates = entire_read_source.delete_predicates,
115
54.5k
                                 .delete_bitmap = entire_read_source.delete_bitmap}));
116
117
54.5k
                        partitial_read_source = {};
118
54.5k
                        split = RowSetSplits(reader->clone());
119
54.5k
                        row_ranges = RowRanges();
120
121
54.5k
                        segment_start = offset_in_segment < rows_of_segment ? i : i + 1;
122
54.5k
                        rows_collected = 0;
123
54.5k
                    }
124
1.03M
                }
125
126
                // The non-empty `row_ranges` means there are some rows left in this segment not added into `split`.
127
978k
                if (!row_ranges.is_empty()) {
128
976k
                    DCHECK_GT(rows_collected, 0);
129
976k
                    DCHECK_EQ(row_ranges.to(), rows_of_segment);
130
976k
                    split.segment_row_ranges.emplace_back(std::move(row_ranges));
131
976k
                }
132
978k
            }
133
134
979k
            DCHECK_LE(rows_collected, _rows_per_scanner);
135
979k
            if (rows_collected > 0) {
136
976k
                split.segment_offsets.first = segment_start;
137
976k
                split.segment_offsets.second = segments_rows.size();
138
976k
                DCHECK_GT(split.segment_offsets.second, split.segment_offsets.first);
139
976k
                DCHECK_EQ(split.segment_row_ranges.size(),
140
976k
                          split.segment_offsets.second - split.segment_offsets.first);
141
976k
                partitial_read_source.rs_splits.emplace_back(std::move(split));
142
976k
            }
143
979k
        } // end `for (auto& rowset : rowsets)`
144
145
627k
        DCHECK_LE(rows_collected, _rows_per_scanner);
146
627k
        if (rows_collected > 0) {
147
448k
            DCHECK_GT(partitial_read_source.rs_splits.size(), 0);
148
448k
#ifndef NDEBUG
149
896k
            for (auto& split : partitial_read_source.rs_splits) {
150
896k
                DCHECK(split.rs_reader != nullptr);
151
896k
                DCHECK_LT(split.segment_offsets.first, split.segment_offsets.second);
152
896k
                DCHECK_EQ(split.segment_row_ranges.size(),
153
896k
                          split.segment_offsets.second - split.segment_offsets.first);
154
896k
            }
155
448k
#endif
156
448k
            scanners.emplace_back(
157
448k
                    _build_scanner(tablet, version, _key_ranges,
158
448k
                                   {.rs_splits = std::move(partitial_read_source.rs_splits),
159
448k
                                    .delete_predicates = entire_read_source.delete_predicates,
160
448k
                                    .delete_bitmap = entire_read_source.delete_bitmap}));
161
448k
        }
162
627k
    }
163
164
141k
    return Status::OK();
165
141k
}
166
167
// Build scanners so that each segment is exclusively scanned by a single scanner.
168
// This guarantees the number of scanners equals the number of segments across all rowsets
169
// for the involved tablets. It preserves delete predicates and key ranges, and clones
170
// RowsetReader per scanner to avoid sharing between scanners.
171
17
Status ParallelScannerBuilder::_build_scanners_by_per_segment(std::list<ScannerSPtr>& scanners) {
172
17
    DCHECK_GE(_rows_per_scanner, _min_rows_per_scanner);
173
174
107
    for (auto&& [tablet, version] : _tablets) {
175
107
        DCHECK(_all_read_sources.contains(tablet->tablet_id()));
176
107
        auto& entire_read_source = _all_read_sources[tablet->tablet_id()];
177
178
107
        if (config::is_cloud_mode()) {
179
            // FIXME(plat1ko): Avoid pointer cast
180
107
            ExecEnv::GetInstance()->storage_engine().to_cloud().tablet_hotspot().count(*tablet);
181
107
        }
182
183
        // For each RowSet split in the read source, split by segment id and build
184
        // one scanner per segment. Keep delete predicates shared.
185
217
        for (auto& rs_split : entire_read_source.rs_splits) {
186
217
            auto reader = rs_split.rs_reader;
187
217
            auto rowset = reader->rowset();
188
217
            const auto rowset_id = rowset->rowset_id();
189
217
            const auto& segments_rows = _all_segments_rows[rowset_id];
190
217
            if (segments_rows.empty() || rowset->num_rows() == 0) {
191
169
                continue;
192
169
            }
193
194
            // Build scanners for [i, i+1) segment range, without row-range slicing.
195
96
            for (int64_t i = 0; i < rowset->num_segments(); ++i) {
196
48
                RowSetSplits split(reader->clone());
197
48
                split.segment_offsets.first = i;
198
48
                split.segment_offsets.second = i + 1;
199
                // No row-ranges slicing; scan whole segment i.
200
48
                DCHECK_GE(split.segment_offsets.second, split.segment_offsets.first + 1);
201
202
48
                TabletReadSource partitial_read_source;
203
48
                partitial_read_source.rs_splits.emplace_back(std::move(split));
204
205
48
                scanners.emplace_back(
206
48
                        _build_scanner(tablet, version, _key_ranges,
207
48
                                       {.rs_splits = std::move(partitial_read_source.rs_splits),
208
48
                                        .delete_predicates = entire_read_source.delete_predicates,
209
48
                                        .delete_bitmap = entire_read_source.delete_bitmap}));
210
48
            }
211
48
        }
212
107
    }
213
214
17
    return Status::OK();
215
17
}
216
217
/**
218
 * Load rowsets of each tablet with specified version, segments of each rowset.
219
 */
220
137k
Status ParallelScannerBuilder::_load() {
221
137k
    _total_rows = 0;
222
137k
    size_t idx = 0;
223
137k
    bool enable_segment_cache = _state->query_options().__isset.enable_segment_cache
224
137k
                                        ? _state->query_options().enable_segment_cache
225
137k
                                        : true;
226
623k
    for (auto&& [tablet, version] : _tablets) {
227
623k
        const auto tablet_id = tablet->tablet_id();
228
623k
        _all_read_sources[tablet_id] = _read_sources[idx];
229
623k
        const auto& read_source = _all_read_sources[tablet_id];
230
2.57M
        for (auto& rs_split : read_source.rs_splits) {
231
2.57M
            auto rowset = rs_split.rs_reader->rowset();
232
2.57M
            RETURN_IF_ERROR(rowset->load());
233
2.57M
            const auto rowset_id = rowset->rowset_id();
234
235
2.57M
            auto beta_rowset = std::dynamic_pointer_cast<BetaRowset>(rowset);
236
2.57M
            std::vector<uint32_t> segment_rows;
237
2.57M
            RETURN_IF_ERROR(beta_rowset->get_segment_num_rows(&segment_rows, enable_segment_cache,
238
2.57M
                                                              &_builder_stats));
239
2.57M
            auto segment_count = rowset->num_segments();
240
3.54M
            for (int64_t i = 0; i != segment_count; i++) {
241
974k
                _all_segments_rows[rowset_id].emplace_back(segment_rows[i]);
242
974k
            }
243
2.57M
            _total_rows += rowset->num_rows();
244
2.57M
        }
245
623k
        idx++;
246
623k
    }
247
248
137k
    _rows_per_scanner = _total_rows / _max_scanners_count;
249
137k
    _rows_per_scanner = std::max<size_t>(_rows_per_scanner, _min_rows_per_scanner);
250
251
137k
    return Status::OK();
252
137k
}
253
254
std::shared_ptr<OlapScanner> ParallelScannerBuilder::_build_scanner(
255
        BaseTabletSPtr tablet, int64_t version, const std::vector<OlapScanRange*>& key_ranges,
256
502k
        TabletReadSource&& read_source) {
257
502k
    OlapScanner::Params params {
258
502k
            _state,  _scanner_profile.get(), key_ranges,   std::move(tablet),
259
502k
            version, std::move(read_source), _limit,       _is_preaggregation,
260
502k
            false,   TBinlogScanType::NONE,  std::nullopt, std::nullopt,
261
502k
    };
262
502k
    return OlapScanner::create_shared(_parent, std::move(params));
263
502k
}
264
265
} // namespace doris