Coverage Report

Created: 2026-05-28 04:38

next uncovered line (L), next uncovered region (R), next uncovered branch (B)
be/src/exec/scan/olap_scanner.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/olap_scanner.h"
19
20
#include <gen_cpp/Descriptors_types.h>
21
#include <gen_cpp/PlanNodes_types.h>
22
#include <gen_cpp/Types_types.h>
23
#include <glog/logging.h>
24
#include <stdlib.h>
25
#include <thrift/protocol/TDebugProtocol.h>
26
27
#include <algorithm>
28
#include <atomic>
29
#include <iterator>
30
#include <ostream>
31
#include <set>
32
33
#include "cloud/cloud_storage_engine.h"
34
#include "cloud/cloud_tablet_hotspot.h"
35
#include "cloud/config.h"
36
#include "common/config.h"
37
#include "common/consts.h"
38
#include "common/logging.h"
39
#include "common/metrics/doris_metrics.h"
40
#include "core/block/block.h"
41
#include "exec/common/variant_util.h"
42
#include "exec/operator/olap_scan_operator.h"
43
#include "exec/scan/scan_node.h"
44
#include "exprs/function_filter.h"
45
#include "exprs/vexpr.h"
46
#include "exprs/vexpr_context.h"
47
#include "io/cache/block_file_cache_profile.h"
48
#include "io/io_common.h"
49
#include "runtime/descriptors.h"
50
#include "runtime/exec_env.h"
51
#include "runtime/runtime_profile.h"
52
#include "runtime/runtime_state.h"
53
#include "service/backend_options.h"
54
#include "storage/id_manager.h"
55
#include "storage/index/inverted/inverted_index_profile.h"
56
#include "storage/iterator/block_reader.h"
57
#include "storage/olap_common.h"
58
#include "storage/olap_tuple.h"
59
#include "storage/olap_utils.h"
60
#include "storage/storage_engine.h"
61
#include "storage/tablet/tablet_schema.h"
62
#ifndef NDEBUG
63
#include "util/debug_points.h"
64
#endif
65
#include "util/json/path_in_data.h"
66
67
namespace doris {
68
#include "common/compile_check_avoid_begin.h"
69
70
using ReadSource = TabletReadSource;
71
72
OlapScanner::OlapScanner(ScanLocalStateBase* parent, OlapScanner::Params&& params)
73
858k
        : Scanner(params.state, parent, params.limit, params.profile),
74
858k
          _key_ranges(std::move(params.key_ranges)),
75
858k
          _tablet_reader_params({.tablet = std::move(params.tablet),
76
858k
                                 .tablet_schema {},
77
858k
                                 .reader_type = params.read_row_binlog ? ReaderType::READER_BINLOG
78
858k
                                                                       : ReaderType::READER_QUERY,
79
858k
                                 .aggregation = params.aggregation,
80
858k
                                 .version = {0, params.version},
81
858k
                                 .start_key {},
82
858k
                                 .end_key {},
83
858k
                                 .predicates {},
84
858k
                                 .function_filters {},
85
858k
                                 .delete_predicates {},
86
858k
                                 .target_cast_type_for_variants {},
87
858k
                                 .all_access_paths {},
88
858k
                                 .predicate_access_paths {},
89
858k
                                 .rs_splits {},
90
858k
                                 .return_columns {},
91
858k
                                 .output_columns {},
92
858k
                                 .common_expr_ctxs_push_down {},
93
858k
                                 .topn_filter_source_node_ids {},
94
858k
                                 .key_group_cluster_key_idxes {},
95
858k
                                 .virtual_column_exprs {},
96
858k
                                 .vir_cid_to_idx_in_block {},
97
858k
                                 .vir_col_idx_to_type {},
98
858k
                                 .score_runtime {},
99
858k
                                 .collection_statistics {},
100
858k
                                 .ann_topn_runtime {},
101
858k
                                 .condition_cache_digest = parent->get_condition_cache_digest()}) {
102
858k
    _tablet_reader_params.set_read_source(std::move(params.read_source),
103
858k
                                          _state->skip_delete_bitmap());
104
858k
    _has_prepared = false;
105
858k
    _vector_search_params = params.state->get_vector_search_params();
106
858k
}
107
108
static std::string read_columns_to_string(TabletSchemaSPtr tablet_schema,
109
2.90k
                                          const std::vector<uint32_t>& read_columns) {
110
    // avoid too long for one line,
111
    // it is hard to display in `show profile` stmt if one line is too long.
112
2.90k
    const int col_per_line = 10;
113
2.90k
    int i = 0;
114
2.90k
    std::string read_columns_string;
115
2.90k
    read_columns_string += "[";
116
15.5k
    for (auto it = read_columns.cbegin(); it != read_columns.cend(); it++) {
117
12.6k
        if (it != read_columns.cbegin()) {
118
9.77k
            read_columns_string += ", ";
119
9.77k
        }
120
12.6k
        read_columns_string += tablet_schema->columns().at(*it)->name();
121
12.6k
        if (i >= col_per_line) {
122
13
            read_columns_string += "\n";
123
13
            i = 0;
124
12.6k
        } else {
125
12.6k
            ++i;
126
12.6k
        }
127
12.6k
    }
128
2.90k
    read_columns_string += "]";
129
2.90k
    return read_columns_string;
130
2.90k
}
131
132
855k
Status OlapScanner::_prepare_impl() {
133
855k
    auto* local_state = static_cast<OlapScanLocalState*>(_local_state);
134
855k
    auto& tablet = _tablet_reader_params.tablet;
135
855k
    auto& tablet_schema = _tablet_reader_params.tablet_schema;
136
855k
    DBUG_EXECUTE_IF("CloudTablet.capture_rs_readers.return.e-230", {
137
855k
        LOG_WARNING("CloudTablet.capture_rs_readers.return e-230 init")
138
855k
                .tag("tablet_id", tablet->tablet_id());
139
855k
        return Status::Error<false>(-230, "injected error");
140
855k
    });
141
142
855k
    for (auto& ctx : local_state->_common_expr_ctxs_push_down) {
143
20.7k
        VExprContextSPtr context;
144
20.7k
        RETURN_IF_ERROR(ctx->clone(_state, context));
145
20.7k
        _common_expr_ctxs_push_down.emplace_back(context);
146
20.7k
        context->prepare_ann_range_search(_vector_search_params);
147
20.7k
    }
148
149
855k
    for (auto pair : local_state->_slot_id_to_virtual_column_expr) {
150
        // Scanner will be executed in a different thread, so we need to clone the context.
151
411
        VExprContextSPtr context;
152
411
        RETURN_IF_ERROR(pair.second->clone(_state, context));
153
411
        _slot_id_to_virtual_column_expr[pair.first] = context;
154
411
    }
155
156
855k
    _slot_id_to_index_in_block = local_state->_slot_id_to_index_in_block;
157
855k
    _slot_id_to_col_type = local_state->_slot_id_to_col_type;
158
855k
    _score_runtime = local_state->_score_runtime;
159
    // All scanners share the same ann_topn_runtime.
160
855k
    _ann_topn_runtime = local_state->_ann_topn_runtime;
161
162
    // set limit to reduce end of rowset and segment mem use
163
855k
    _tablet_reader = std::make_unique<BlockReader>();
164
    // batch size is passed down to segment iterator, use _state->batch_size()
165
    // instead of _parent->limit(), because if _parent->limit() is a very small
166
    // value (e.g. select a from t where a .. and b ... limit 1),
167
    // it will be very slow when reading data in segment iterator
168
855k
    _tablet_reader->set_batch_size(_state->batch_size());
169
    // Adaptive batch size: pass byte-budget settings to the storage reader.
170
    // The reader still uses batch_size() as the row ceiling.
171
855k
    _tablet_reader->set_preferred_block_size_bytes(_state->preferred_block_size_bytes());
172
855k
    {
173
855k
        TOlapScanNode& olap_scan_node = local_state->olap_scan_node();
174
855k
        TabletSchemaSPtr source_tablet_schema =
175
855k
                _tablet_reader_params.reader_type == ReaderType::READER_BINLOG
176
855k
                        ? tablet->row_binlog_tablet_schema()
177
855k
                        : tablet->tablet_schema();
178
179
855k
        tablet_schema = std::make_shared<TabletSchema>();
180
855k
        tablet_schema->copy_from(*source_tablet_schema);
181
855k
        if (olap_scan_node.__isset.columns_desc && !olap_scan_node.columns_desc.empty() &&
182
855k
            olap_scan_node.columns_desc[0].col_unique_id >= 0) {
183
854k
            tablet_schema->clear_columns();
184
13.3M
            for (const auto& column_desc : olap_scan_node.columns_desc) {
185
13.3M
                tablet_schema->append_column(TabletColumn(column_desc));
186
13.3M
            }
187
855k
            if (olap_scan_node.__isset.schema_version) {
188
855k
                tablet_schema->set_schema_version(olap_scan_node.schema_version);
189
855k
            }
190
854k
        }
191
855k
        if (olap_scan_node.__isset.indexes_desc) {
192
855k
            tablet_schema->update_indexes_from_thrift(olap_scan_node.indexes_desc);
193
855k
        }
194
195
855k
        if (_tablet_reader_params.rs_splits.empty()) {
196
            // Non-pipeline mode, Tablet : Scanner = 1 : 1
197
            // acquire tablet rowset readers at the beginning of the scan node
198
            // to prevent this case: when there are lots of olap scanners to run for example 10000
199
            // the rowsets maybe compacted when the last olap scanner starts
200
0
            ReadSource read_source;
201
202
0
            if (config::is_cloud_mode()) {
203
                // FIXME(plat1ko): Avoid pointer cast
204
0
                ExecEnv::GetInstance()->storage_engine().to_cloud().tablet_hotspot().count(*tablet);
205
0
            }
206
207
0
            auto maybe_read_source = tablet->capture_read_source(
208
0
                    _tablet_reader_params.version,
209
0
                    {
210
0
                            .skip_missing_versions = _state->skip_missing_version(),
211
0
                            .enable_fetch_rowsets_from_peers =
212
0
                                    config::enable_fetch_rowsets_from_peer_replicas,
213
0
                            .capture_row_binlog =
214
0
                                    _tablet_reader_params.reader_type == ReaderType::READER_BINLOG,
215
0
                            .enable_prefer_cached_rowset =
216
0
                                    config::is_cloud_mode() ? _state->enable_prefer_cached_rowset()
217
0
                                                            : false,
218
0
                            .query_freshness_tolerance_ms =
219
0
                                    config::is_cloud_mode() ? _state->query_freshness_tolerance_ms()
220
0
                                                            : -1,
221
0
                    });
222
0
            if (!maybe_read_source) {
223
0
                LOG(WARNING) << "fail to init reader. res=" << maybe_read_source.error();
224
0
                return maybe_read_source.error();
225
0
            }
226
0
            read_source = std::move(maybe_read_source.value());
227
228
0
            if (config::enable_mow_verbose_log && tablet->enable_unique_key_merge_on_write()) {
229
0
                LOG_INFO("finish capture_rs_readers for tablet={}, query_id={}",
230
0
                         tablet->tablet_id(), print_id(_state->query_id()));
231
0
            }
232
233
0
            if (!_state->skip_delete_predicate()) {
234
0
                read_source.fill_delete_predicates();
235
0
            }
236
0
            _tablet_reader_params.set_read_source(std::move(read_source));
237
0
        }
238
239
        // Initialize tablet_reader_params
240
855k
        RETURN_IF_ERROR(_init_tablet_reader_params(
241
855k
                local_state->_parent->cast<OlapScanOperatorX>()._slot_id_to_slot_desc, _key_ranges,
242
855k
                local_state->_slot_id_to_predicates, local_state->_push_down_functions));
243
855k
    }
244
245
    // add read columns in profile
246
855k
    if (_state->enable_profile()) {
247
2.89k
        _profile->add_info_string("ReadColumns",
248
2.89k
                                  read_columns_to_string(tablet_schema, _return_columns));
249
2.89k
    }
250
251
855k
    if (_tablet_reader_params.score_runtime) {
252
14
        SCOPED_TIMER(local_state->_statistics_collect_timer);
253
14
        _tablet_reader_params.collection_statistics = std::make_shared<CollectionStatistics>();
254
255
14
        io::IOContext io_ctx {
256
14
                .reader_type = _tablet_reader_params.reader_type,
257
14
                .expiration_time = tablet->ttl_seconds(),
258
14
                .query_id = &_state->query_id(),
259
14
                .file_cache_stats = &_tablet_reader->mutable_stats()->file_cache_stats,
260
14
                .is_inverted_index = true,
261
14
        };
262
263
14
        RETURN_IF_ERROR(_tablet_reader_params.collection_statistics->collect(
264
14
                _state, _tablet_reader_params.rs_splits, _tablet_reader_params.tablet_schema,
265
14
                _tablet_reader_params.common_expr_ctxs_push_down, &io_ctx));
266
14
    }
267
268
855k
    _has_prepared = true;
269
855k
    return Status::OK();
270
855k
}
271
272
853k
Status OlapScanner::_open_impl(RuntimeState* state) {
273
853k
    RETURN_IF_ERROR(Scanner::_open_impl(state));
274
853k
    SCOPED_TIMER(_local_state->cast<OlapScanLocalState>()._reader_init_timer);
275
276
853k
    auto res = _tablet_reader->init(_tablet_reader_params);
277
853k
    if (!res.ok()) {
278
64
        res.append("failed to initialize storage reader. tablet=" +
279
64
                   std::to_string(_tablet_reader_params.tablet->tablet_id()) +
280
64
                   ", backend=" + BackendOptions::get_localhost());
281
64
        return res;
282
64
    }
283
284
    // Do not hold rs_splits any more to release memory.
285
853k
    _tablet_reader_params.rs_splits.clear();
286
287
853k
    return Status::OK();
288
853k
}
289
290
// it will be called under tablet read lock because capture rs readers need
291
Status OlapScanner::_init_tablet_reader_params(
292
        const phmap::flat_hash_map<int, SlotDescriptor*>& slot_id_to_slot_desc,
293
        const std::vector<OlapScanRange*>& key_ranges,
294
        const phmap::flat_hash_map<int, std::vector<std::shared_ptr<ColumnPredicate>>>&
295
                slot_to_predicates,
296
854k
        const std::vector<FunctionFilter>& function_filters) {
297
    // if the table with rowset [0-x] or [0-1] [2-y], and [0-1] is empty
298
854k
    const bool single_version = _tablet_reader_params.has_single_version();
299
300
854k
    auto* olap_local_state = static_cast<OlapScanLocalState*>(_local_state);
301
854k
    bool read_mor_as_dup = olap_local_state->olap_scan_node().__isset.read_mor_as_dup &&
302
854k
                           olap_local_state->olap_scan_node().read_mor_as_dup;
303
854k
    if (_state->skip_storage_engine_merge() || read_mor_as_dup) {
304
49
        _tablet_reader_params.direct_mode = true;
305
49
        _tablet_reader_params.aggregation = true;
306
854k
    } else {
307
854k
        auto push_down_agg_type = _local_state->get_push_down_agg_type();
308
854k
        _tablet_reader_params.direct_mode = _tablet_reader_params.aggregation || single_version ||
309
854k
                                            (push_down_agg_type != TPushAggOp::NONE &&
310
10.7k
                                             push_down_agg_type != TPushAggOp::COUNT_ON_INDEX);
311
854k
    }
312
313
854k
    RETURN_IF_ERROR(_init_variant_columns());
314
854k
    RETURN_IF_ERROR(_init_return_columns());
315
316
854k
    _tablet_reader_params.push_down_agg_type_opt = _local_state->get_push_down_agg_type();
317
318
854k
    _tablet_reader_params.common_expr_ctxs_push_down = _common_expr_ctxs_push_down;
319
854k
    _tablet_reader_params.virtual_column_exprs = _virtual_column_exprs;
320
854k
    _tablet_reader_params.vir_cid_to_idx_in_block = _vir_cid_to_idx_in_block;
321
854k
    _tablet_reader_params.vir_col_idx_to_type = _vir_col_idx_to_type;
322
854k
    _tablet_reader_params.score_runtime = _score_runtime;
323
854k
    _tablet_reader_params.output_columns = ((OlapScanLocalState*)_local_state)->_output_column_ids;
324
854k
    _tablet_reader_params.ann_topn_runtime = _ann_topn_runtime;
325
854k
    for (const auto& ele : ((OlapScanLocalState*)_local_state)->_cast_types_for_variants) {
326
1.23k
        _tablet_reader_params.target_cast_type_for_variants[ele.first] = ele.second;
327
1.23k
    };
328
854k
    auto& tablet_schema = _tablet_reader_params.tablet_schema;
329
6.00M
    for (auto& predicates : slot_to_predicates) {
330
6.00M
        const int sid = predicates.first;
331
6.00M
        DCHECK(slot_id_to_slot_desc.contains(sid));
332
6.00M
        int32_t index =
333
6.00M
                tablet_schema->field_index(slot_id_to_slot_desc.find(sid)->second->col_name());
334
6.00M
        if (index < 0) {
335
0
            throw Exception(
336
0
                    Status::InternalError("Column {} not found in tablet schema",
337
0
                                          slot_id_to_slot_desc.find(sid)->second->col_name()));
338
0
        }
339
6.00M
        for (auto& predicate : predicates.second) {
340
659k
            _tablet_reader_params.predicates.push_back(predicate->clone(index));
341
659k
        }
342
6.00M
    }
343
344
854k
    std::copy(function_filters.cbegin(), function_filters.cend(),
345
854k
              std::inserter(_tablet_reader_params.function_filters,
346
854k
                            _tablet_reader_params.function_filters.begin()));
347
348
    // Merge the columns in delete predicate that not in latest schema in to current tablet schema
349
854k
    for (auto& del_pred : _tablet_reader_params.delete_predicates) {
350
6.92k
        tablet_schema->merge_dropped_columns(*del_pred->tablet_schema());
351
6.92k
    }
352
353
    // Push key ranges to the tablet reader.
354
    // Skip the "full scan" placeholder (has_lower_bound == false) — when no key
355
    // predicates exist, start_key/end_key remain empty and the reader does a full scan.
356
1.47M
    for (auto* key_range : key_ranges) {
357
1.47M
        if (!key_range->has_lower_bound) {
358
142k
            continue;
359
142k
        }
360
361
1.33M
        _tablet_reader_params.start_key_include = key_range->begin_include;
362
1.33M
        _tablet_reader_params.end_key_include = key_range->end_include;
363
364
1.33M
        _tablet_reader_params.start_key.push_back(key_range->begin_scan_range);
365
1.33M
        _tablet_reader_params.end_key.push_back(key_range->end_scan_range);
366
1.33M
    }
367
368
854k
    _tablet_reader_params.profile = _local_state->custom_profile();
369
854k
    _tablet_reader_params.runtime_state = _state;
370
371
854k
    _tablet_reader_params.origin_return_columns = &_return_columns;
372
854k
    _tablet_reader_params.tablet_columns_convert_to_null_set = &_tablet_columns_convert_to_null_set;
373
374
854k
    if (_tablet_reader_params.direct_mode) {
375
842k
        _tablet_reader_params.return_columns = _return_columns;
376
842k
    } else {
377
        // we need to fetch all key columns to do the right aggregation on storage engine side.
378
42.2k
        for (size_t i = 0; i < tablet_schema->num_key_columns(); ++i) {
379
30.1k
            _tablet_reader_params.return_columns.push_back(i);
380
30.1k
        }
381
54.1k
        for (auto index : _return_columns) {
382
54.1k
            if (tablet_schema->column(index).is_key()) {
383
21.2k
                continue;
384
21.2k
            }
385
32.9k
            _tablet_reader_params.return_columns.push_back(index);
386
32.9k
        }
387
        // expand the sequence column
388
12.0k
        if (tablet_schema->has_sequence_col() || tablet_schema->has_seq_map()) {
389
40
            bool has_replace_col = false;
390
91
            for (auto col : _return_columns) {
391
91
                if (tablet_schema->column(col).aggregation() ==
392
91
                    FieldAggregationMethod::OLAP_FIELD_AGGREGATION_REPLACE) {
393
40
                    has_replace_col = true;
394
40
                    break;
395
40
                }
396
91
            }
397
40
            if (auto sequence_col_idx = tablet_schema->sequence_col_idx();
398
40
                has_replace_col && tablet_schema->has_sequence_col() &&
399
40
                std::find(_return_columns.begin(), _return_columns.end(), sequence_col_idx) ==
400
28
                        _return_columns.end()) {
401
16
                _tablet_reader_params.return_columns.push_back(sequence_col_idx);
402
16
            }
403
40
            if (has_replace_col) {
404
40
                const auto& val_to_seq = tablet_schema->value_col_idx_to_seq_col_idx();
405
40
                std::set<uint32_t> return_seq_columns;
406
407
244
                for (auto col : _tablet_reader_params.return_columns) {
408
                    // we need to add the necessary sequence column in _return_columns, and
409
                    // Avoid adding the same seq column twice
410
244
                    const auto val_iter = val_to_seq.find(col);
411
244
                    if (val_iter != val_to_seq.end()) {
412
42
                        auto seq = val_iter->second;
413
42
                        if (std::find(_tablet_reader_params.return_columns.begin(),
414
42
                                      _tablet_reader_params.return_columns.end(),
415
42
                                      seq) == _tablet_reader_params.return_columns.end()) {
416
4
                            return_seq_columns.insert(seq);
417
4
                        }
418
42
                    }
419
244
                }
420
40
                _tablet_reader_params.return_columns.insert(
421
40
                        std::end(_tablet_reader_params.return_columns),
422
40
                        std::begin(return_seq_columns), std::end(return_seq_columns));
423
40
            }
424
40
        }
425
12.0k
    }
426
427
854k
    _tablet_reader_params.use_page_cache = _state->enable_page_cache();
428
429
854k
    DBUG_EXECUTE_IF("NewOlapScanner::_init_tablet_reader_params.block", DBUG_BLOCK);
430
431
854k
    if (!_state->skip_storage_engine_merge()) {
432
852k
        auto* olap_scan_local_state = (OlapScanLocalState*)_local_state;
433
852k
        TOlapScanNode& olap_scan_node = olap_scan_local_state->olap_scan_node();
434
435
        // Set MOR value predicate pushdown flag
436
852k
        if (olap_scan_node.__isset.enable_mor_value_predicate_pushdown &&
437
852k
            olap_scan_node.enable_mor_value_predicate_pushdown) {
438
25
            _tablet_reader_params.enable_mor_value_predicate_pushdown = true;
439
25
        }
440
441
852k
        const bool has_key_topn =
442
852k
                olap_scan_node.__isset.sort_info && !olap_scan_node.sort_info.is_asc_order.empty();
443
852k
        if (has_key_topn) {
444
2.79k
            _limit = _local_state->limit_per_scanner();
445
2.79k
        }
446
447
852k
        const bool no_runtime_filters = _total_rf_num == 0;
448
852k
        const bool segment_limit_enabled = _state->enable_segment_limit_pushdown();
449
852k
        const bool storage_no_merge = olap_scan_local_state->_storage_no_merge();
450
451
852k
        if (_limit > 0 && no_runtime_filters && segment_limit_enabled && storage_no_merge) {
452
4.48k
            for (const auto& conjunct : _conjuncts) {
453
0
                DORIS_CHECK(!olap_scan_local_state->_check_expr_storage_filter(
454
0
                        conjunct->root(), OlapScanLocalState::ExprStorageFilterCheckMode::
455
0
                                                  HAS_SEGMENT_EVALUABLE_EXPR));
456
0
            }
457
4.48k
        }
458
459
        // Segment LIMIT has only two legal states: completely disabled, or enabled after every
460
        // row-filtering conjunct has become a storage predicate or SegmentIterator common expr.
461
852k
        const bool can_push_down_segment_limit = _limit > 0 && no_runtime_filters &&
462
852k
                                                 _conjuncts.empty() && segment_limit_enabled &&
463
852k
                                                 storage_no_merge;
464
852k
        if (can_push_down_segment_limit) {
465
4.49k
            if (has_key_topn) {
466
2.70k
                _tablet_reader_params.read_orderby_key = true;
467
2.70k
                if (!olap_scan_node.sort_info.is_asc_order[0]) {
468
659
                    _tablet_reader_params.read_orderby_key_reverse = true;
469
659
                }
470
2.70k
                _tablet_reader_params.read_orderby_key_num_prefix_columns =
471
2.70k
                        olap_scan_node.sort_info.is_asc_order.size();
472
2.70k
                _tablet_reader_params.read_orderby_key_limit = _limit;
473
2.70k
            } else {
474
1.79k
                _tablet_reader_params.general_read_limit = _limit;
475
1.79k
            }
476
4.49k
        }
477
478
852k
        if (_tablet_reader_params.read_orderby_key_limit > 0 ||
479
852k
            _tablet_reader_params.general_read_limit > 0) {
480
4.49k
            DORIS_CHECK(can_push_down_segment_limit);
481
4.49k
            DORIS_CHECK(_conjuncts.empty());
482
4.49k
        }
483
484
        // A key TopN scan cannot share the plain LIMIT early-stop counter. If
485
        // storage TopN is pushed down, each scanner must produce its full local
486
        // candidates. If it is not pushed down for any reason, the upper TopN
487
        // still needs all rows from the scan.
488
852k
        if (has_key_topn) {
489
2.80k
            _shared_scan_limit = nullptr;
490
2.80k
            if (_tablet_reader_params.read_orderby_key_limit == 0) {
491
106
                _limit = -1;
492
106
            }
493
2.80k
        }
494
        // Note: _shared_scan_limit is intentionally not pushed into the
495
        // storage layer. SegmentIterator's _process_eof() is irreversible,
496
        // so a concurrently-decremented atomic could reach 0 while a segment
497
        // still has data needed by other scanners.
498
499
        // set push down topn filter
500
852k
        _tablet_reader_params.topn_filter_source_node_ids =
501
852k
                olap_scan_local_state->get_topn_filter_source_node_ids(_state, true);
502
852k
        if (!_tablet_reader_params.topn_filter_source_node_ids.empty()) {
503
5.62k
            _tablet_reader_params.topn_filter_target_node_id =
504
5.62k
                    olap_scan_local_state->parent()->node_id();
505
5.62k
        }
506
852k
    }
507
508
    // If this is a Two-Phase read query, and we need to delay the release of Rowset
509
    // by rowset->update_delayed_expired_timestamp().This could expand the lifespan of Rowset
510
854k
    if (tablet_schema->field_index(BeConsts::ROWID_COL) >= 0) {
511
0
        constexpr static int delayed_s = 60;
512
0
        for (auto rs_reader : _tablet_reader_params.rs_splits) {
513
0
            uint64_t delayed_expired_timestamp =
514
0
                    UnixSeconds() + _tablet_reader_params.runtime_state->execution_timeout() +
515
0
                    delayed_s;
516
0
            rs_reader.rs_reader->rowset()->update_delayed_expired_timestamp(
517
0
                    delayed_expired_timestamp);
518
0
            ExecEnv::GetInstance()->storage_engine().add_quering_rowset(
519
0
                    rs_reader.rs_reader->rowset());
520
0
        }
521
0
    }
522
523
854k
    if (tablet_schema->has_global_row_id()) {
524
7.44k
        auto& id_file_map = _state->get_id_file_map();
525
14.1k
        for (auto rs_reader : _tablet_reader_params.rs_splits) {
526
14.1k
            id_file_map->add_temp_rowset(rs_reader.rs_reader->rowset());
527
14.1k
        }
528
7.44k
    }
529
530
854k
    return Status::OK();
531
854k
}
532
533
853k
Status OlapScanner::_init_variant_columns() {
534
853k
    auto& tablet_schema = _tablet_reader_params.tablet_schema;
535
853k
    if (tablet_schema->num_variant_columns() == 0) {
536
847k
        return Status::OK();
537
847k
    }
538
    // Parent column has path info to distinction from each other
539
14.5k
    for (auto* slot : _output_tuple_desc->slots()) {
540
14.5k
        if (slot->type()->get_primitive_type() == PrimitiveType::TYPE_VARIANT) {
541
            // Such columns are not exist in frontend schema info, so we need to
542
            // add them into tablet_schema for later column indexing.
543
6.95k
            const auto& dt_variant =
544
6.95k
                    assert_cast<const DataTypeVariant&>(*remove_nullable(slot->type()));
545
6.95k
            TabletColumn subcol = TabletColumn::create_materialized_variant_column(
546
6.95k
                    tablet_schema->column_by_uid(slot->col_unique_id()).name_lower_case(),
547
6.95k
                    slot->column_paths(), slot->col_unique_id(),
548
6.95k
                    dt_variant.variant_max_subcolumns_count(), dt_variant.enable_doc_mode());
549
6.95k
            if (tablet_schema->field_index(*subcol.path_info_ptr()) < 0) {
550
5.44k
                tablet_schema->append_column(subcol, TabletSchema::ColumnType::VARIANT);
551
5.44k
            }
552
6.95k
        }
553
14.5k
    }
554
5.65k
    variant_util::inherit_column_attributes(tablet_schema);
555
5.65k
    return Status::OK();
556
853k
}
557
558
853k
Status OlapScanner::_init_return_columns() {
559
7.85M
    for (auto* slot : _output_tuple_desc->slots()) {
560
        // variant column using path to index a column
561
7.85M
        int32_t index = 0;
562
7.85M
        auto& tablet_schema = _tablet_reader_params.tablet_schema;
563
7.85M
        if (slot->type()->get_primitive_type() == PrimitiveType::TYPE_VARIANT) {
564
6.95k
            index = tablet_schema->field_index(PathInData(
565
6.95k
                    tablet_schema->column_by_uid(slot->col_unique_id()).name_lower_case(),
566
6.95k
                    slot->column_paths()));
567
7.85M
        } else {
568
7.85M
            index = slot->col_unique_id() >= 0 ? tablet_schema->field_index(slot->col_unique_id())
569
18.4E
                                               : tablet_schema->field_index(slot->col_name());
570
7.85M
        }
571
572
7.85M
        if (index < 0) {
573
0
            return Status::InternalError(
574
0
                    "field name is invalid. field={}, field_name_to_index={}, col_unique_id={}",
575
0
                    slot->col_name(), tablet_schema->get_all_field_names(), slot->col_unique_id());
576
0
        }
577
578
7.85M
        if (slot->get_virtual_column_expr()) {
579
409
            ColumnId virtual_column_cid = index;
580
409
            _virtual_column_exprs[virtual_column_cid] = _slot_id_to_virtual_column_expr[slot->id()];
581
409
            size_t idx_in_block = _slot_id_to_index_in_block[slot->id()];
582
409
            _vir_cid_to_idx_in_block[virtual_column_cid] = idx_in_block;
583
409
            _vir_col_idx_to_type[idx_in_block] = _slot_id_to_col_type[slot->id()];
584
585
409
            VLOG_DEBUG << fmt::format(
586
5
                    "Virtual column, slot id: {}, cid {}, column index: {}, type: {}", slot->id(),
587
5
                    virtual_column_cid, _vir_cid_to_idx_in_block[virtual_column_cid],
588
5
                    _vir_col_idx_to_type[idx_in_block]->get_name());
589
409
        }
590
591
7.85M
        const auto& column = tablet_schema->column(index);
592
7.85M
        int32_t unique_id =
593
7.85M
                column.unique_id() >= 0 ? column.unique_id() : column.parent_unique_id();
594
7.85M
        if (!slot->all_access_paths().empty()) {
595
72.7k
            _tablet_reader_params.all_access_paths.insert({unique_id, slot->all_access_paths()});
596
72.7k
        }
597
598
7.85M
        if (!slot->predicate_access_paths().empty()) {
599
8.10k
            _tablet_reader_params.predicate_access_paths.insert(
600
8.10k
                    {unique_id, slot->predicate_access_paths()});
601
8.10k
        }
602
603
7.85M
        if ((slot->type()->get_primitive_type() == PrimitiveType::TYPE_STRUCT ||
604
7.86M
             slot->type()->get_primitive_type() == PrimitiveType::TYPE_MAP ||
605
7.86M
             slot->type()->get_primitive_type() == PrimitiveType::TYPE_ARRAY) &&
606
7.85M
            !slot->all_access_paths().empty()) {
607
61.0k
            tablet_schema->add_pruned_columns_data_type(column.unique_id(), slot->type());
608
61.0k
        }
609
610
7.85M
        _return_columns.push_back(index);
611
7.85M
        if (slot->is_nullable() && !tablet_schema->column(index).is_nullable()) {
612
0
            _tablet_columns_convert_to_null_set.emplace(index);
613
7.85M
        } else if (!slot->is_nullable() && tablet_schema->column(index).is_nullable()) {
614
0
            return Status::Error<ErrorCode::INVALID_SCHEMA>(
615
0
                    "slot(id: {}, name: {})'s nullable does not match "
616
0
                    "column(tablet id: {}, index: {}, name: {}) ",
617
0
                    slot->id(), slot->col_name(), tablet_schema->table_id(), index,
618
0
                    tablet_schema->column(index).name());
619
0
        }
620
7.85M
    }
621
622
853k
    if (_return_columns.empty()) {
623
0
        return Status::InternalError("failed to build storage scanner, no materialized slot!");
624
0
    }
625
626
853k
    return Status::OK();
627
853k
}
628
629
899k
doris::TabletStorageType OlapScanner::get_storage_type() {
630
899k
    if (config::is_cloud_mode()) {
631
        // we don't have cold storage in cloud mode, all storage is treated as local
632
896k
        return doris::TabletStorageType::STORAGE_TYPE_LOCAL;
633
896k
    }
634
3.29k
    int local_reader = 0;
635
9.32k
    for (const auto& reader : _tablet_reader_params.rs_splits) {
636
9.32k
        local_reader += reader.rs_reader->rowset()->is_local();
637
9.32k
    }
638
3.29k
    int total_reader = _tablet_reader_params.rs_splits.size();
639
640
3.29k
    if (local_reader == total_reader) {
641
3.29k
        return doris::TabletStorageType::STORAGE_TYPE_LOCAL;
642
3.29k
    } else if (local_reader == 0) {
643
0
        return doris::TabletStorageType::STORAGE_TYPE_REMOTE;
644
0
    }
645
0
    return doris::TabletStorageType::STORAGE_TYPE_REMOTE_AND_LOCAL;
646
3.29k
}
647
648
1.10M
Status OlapScanner::_get_block_impl(RuntimeState* state, Block* block, bool* eof) {
649
    // Read one block from block reader
650
    // ATTN: Here we need to let the _get_block_impl method guarantee the semantics of the interface,
651
    // that is, eof can be set to true only when the returned block is empty.
652
1.10M
    RETURN_IF_ERROR(_tablet_reader->next_block_with_aggregation(block, eof));
653
1.10M
    if (block->rows() > 0) {
654
252k
        _tablet_reader_params.tablet->read_block_count.fetch_add(1, std::memory_order_relaxed);
655
252k
        *eof = false;
656
252k
    }
657
1.10M
#ifndef NDEBUG
658
1.10M
    RETURN_IF_ERROR(_check_ann_cache_hit_debug_points(_tablet_reader->stats()));
659
1.10M
#endif
660
1.10M
    return Status::OK();
661
1.10M
}
662
663
861k
Status OlapScanner::close(RuntimeState* state) {
664
861k
    if (!_try_close()) {
665
150
        return Status::OK();
666
150
    }
667
861k
    RETURN_IF_ERROR(Scanner::close(state));
668
861k
    return Status::OK();
669
861k
}
670
671
897k
void OlapScanner::update_realtime_counters() {
672
897k
    if (!_has_prepared) {
673
        // Counter update need prepare successfully, or it maybe core. For example, olap scanner
674
        // will open tablet reader during prepare, if not prepare successfully, tablet reader == nullptr.
675
0
        return;
676
0
    }
677
897k
    OlapScanLocalState* local_state = static_cast<OlapScanLocalState*>(_local_state);
678
897k
    const OlapReaderStatistics& stats = _tablet_reader->stats();
679
897k
    COUNTER_UPDATE(local_state->_read_compressed_counter, stats.compressed_bytes_read);
680
897k
    COUNTER_UPDATE(local_state->_read_uncompressed_counter, stats.uncompressed_bytes_read);
681
897k
    COUNTER_UPDATE(local_state->_scan_bytes, stats.uncompressed_bytes_read);
682
897k
    COUNTER_UPDATE(local_state->_scan_rows, stats.raw_rows_read);
683
684
    // Make sure the scan bytes and scan rows counter in audit log is the same as the counter in
685
    // doris metrics.
686
    // ScanBytes is the uncompressed bytes read from local + remote
687
    // bytes_read_from_local is the compressed bytes read from local
688
    // bytes_read_from_remote is the compressed bytes read from remote
689
    // scan bytes > bytes_read_from_local + bytes_read_from_remote
690
897k
    _state->get_query_ctx()->resource_ctx()->io_context()->update_scan_rows(stats.raw_rows_read);
691
897k
    _state->get_query_ctx()->resource_ctx()->io_context()->update_scan_bytes(
692
897k
            stats.uncompressed_bytes_read);
693
694
    // In case of no cache, we still need to update the IO stats. uncompressed bytes read == local + remote
695
897k
    if (stats.file_cache_stats.bytes_read_from_local == 0 &&
696
897k
        stats.file_cache_stats.bytes_read_from_remote == 0) {
697
787k
        _state->get_query_ctx()->resource_ctx()->io_context()->update_scan_bytes_from_local_storage(
698
787k
                stats.compressed_bytes_read);
699
787k
        DorisMetrics::instance()->query_scan_bytes_from_local->increment(
700
787k
                stats.compressed_bytes_read);
701
787k
    } else {
702
109k
        _state->get_query_ctx()->resource_ctx()->io_context()->update_scan_bytes_from_local_storage(
703
109k
                stats.file_cache_stats.bytes_read_from_local - _bytes_read_from_local);
704
109k
        _state->get_query_ctx()
705
109k
                ->resource_ctx()
706
109k
                ->io_context()
707
109k
                ->update_scan_bytes_from_remote_storage(
708
109k
                        stats.file_cache_stats.bytes_read_from_remote - _bytes_read_from_remote);
709
710
109k
        DorisMetrics::instance()->query_scan_bytes_from_local->increment(
711
109k
                stats.file_cache_stats.bytes_read_from_local - _bytes_read_from_local);
712
109k
        DorisMetrics::instance()->query_scan_bytes_from_remote->increment(
713
109k
                stats.file_cache_stats.bytes_read_from_remote - _bytes_read_from_remote);
714
109k
    }
715
716
897k
    _tablet_reader->mutable_stats()->compressed_bytes_read = 0;
717
897k
    _tablet_reader->mutable_stats()->uncompressed_bytes_read = 0;
718
897k
    _tablet_reader->mutable_stats()->raw_rows_read = 0;
719
720
897k
    _bytes_read_from_local = _tablet_reader->stats().file_cache_stats.bytes_read_from_local;
721
897k
    _bytes_read_from_remote = _tablet_reader->stats().file_cache_stats.bytes_read_from_remote;
722
897k
}
723
724
852k
void OlapScanner::_collect_profile_before_close() {
725
    //  Please don't directly enable the profile here, we need to set QueryStatistics using the counter inside.
726
852k
    if (_has_updated_counter) {
727
0
        return;
728
0
    }
729
852k
    _has_updated_counter = true;
730
852k
    _tablet_reader->update_profile(_profile);
731
732
852k
    Scanner::_collect_profile_before_close();
733
734
    // Update counters for OlapScanner
735
    // Update counters from tablet reader's stats
736
852k
    auto& stats = _tablet_reader->stats();
737
852k
    auto* local_state = (OlapScanLocalState*)_local_state;
738
852k
    COUNTER_UPDATE(local_state->_io_timer, stats.io_ns);
739
852k
    COUNTER_UPDATE(local_state->_read_compressed_counter, stats.compressed_bytes_read);
740
852k
    COUNTER_UPDATE(local_state->_scan_bytes, stats.uncompressed_bytes_read);
741
852k
    COUNTER_UPDATE(local_state->_decompressor_timer, stats.decompress_ns);
742
852k
    COUNTER_UPDATE(local_state->_read_uncompressed_counter, stats.uncompressed_bytes_read);
743
852k
    COUNTER_UPDATE(local_state->_block_load_timer, stats.block_load_ns);
744
852k
    COUNTER_UPDATE(local_state->_block_load_counter, stats.blocks_load);
745
852k
    COUNTER_UPDATE(local_state->_block_fetch_timer, stats.block_fetch_ns);
746
852k
    COUNTER_UPDATE(local_state->_delete_bitmap_get_agg_timer, stats.delete_bitmap_get_agg_ns);
747
852k
    COUNTER_UPDATE(local_state->_scan_rows, stats.raw_rows_read);
748
852k
    COUNTER_UPDATE(local_state->_vec_cond_timer, stats.vec_cond_ns);
749
852k
    COUNTER_UPDATE(local_state->_short_cond_timer, stats.short_cond_ns);
750
852k
    COUNTER_UPDATE(local_state->_expr_filter_timer, stats.expr_filter_ns);
751
852k
    COUNTER_UPDATE(local_state->_block_init_timer, stats.block_init_ns);
752
852k
    COUNTER_UPDATE(local_state->_block_init_seek_timer, stats.block_init_seek_ns);
753
852k
    COUNTER_UPDATE(local_state->_block_init_seek_counter, stats.block_init_seek_num);
754
852k
    COUNTER_UPDATE(local_state->_segment_generate_row_range_by_keys_timer,
755
852k
                   stats.generate_row_ranges_by_keys_ns);
756
852k
    COUNTER_UPDATE(local_state->_segment_generate_row_range_by_column_conditions_timer,
757
852k
                   stats.generate_row_ranges_by_column_conditions_ns);
758
852k
    COUNTER_UPDATE(local_state->_segment_generate_row_range_by_bf_timer,
759
852k
                   stats.generate_row_ranges_by_bf_ns);
760
852k
    COUNTER_UPDATE(local_state->_collect_iterator_merge_next_timer,
761
852k
                   stats.collect_iterator_merge_next_timer);
762
852k
    COUNTER_UPDATE(local_state->_segment_generate_row_range_by_zonemap_timer,
763
852k
                   stats.generate_row_ranges_by_zonemap_ns);
764
852k
    COUNTER_UPDATE(local_state->_segment_generate_row_range_by_dict_timer,
765
852k
                   stats.generate_row_ranges_by_dict_ns);
766
852k
    COUNTER_UPDATE(local_state->_predicate_column_read_timer, stats.predicate_column_read_ns);
767
852k
    COUNTER_UPDATE(local_state->_non_predicate_column_read_timer, stats.non_predicate_read_ns);
768
852k
    COUNTER_UPDATE(local_state->_predicate_column_read_seek_timer,
769
852k
                   stats.predicate_column_read_seek_ns);
770
852k
    COUNTER_UPDATE(local_state->_predicate_column_read_seek_counter,
771
852k
                   stats.predicate_column_read_seek_num);
772
852k
    COUNTER_UPDATE(local_state->_lazy_read_timer, stats.lazy_read_ns);
773
852k
    COUNTER_UPDATE(local_state->_lazy_read_seek_timer, stats.block_lazy_read_seek_ns);
774
852k
    COUNTER_UPDATE(local_state->_lazy_read_seek_counter, stats.block_lazy_read_seek_num);
775
852k
    COUNTER_UPDATE(local_state->_output_col_timer, stats.output_col_ns);
776
852k
    COUNTER_UPDATE(local_state->_rows_vec_cond_filtered_counter, stats.rows_vec_cond_filtered);
777
852k
    COUNTER_UPDATE(local_state->_rows_short_circuit_cond_filtered_counter,
778
852k
                   stats.rows_short_circuit_cond_filtered);
779
852k
    COUNTER_UPDATE(local_state->_rows_expr_cond_filtered_counter, stats.rows_expr_cond_filtered);
780
852k
    COUNTER_UPDATE(local_state->_rows_vec_cond_input_counter, stats.vec_cond_input_rows);
781
852k
    COUNTER_UPDATE(local_state->_rows_short_circuit_cond_input_counter,
782
852k
                   stats.short_circuit_cond_input_rows);
783
852k
    COUNTER_UPDATE(local_state->_rows_expr_cond_input_counter, stats.expr_cond_input_rows);
784
852k
    COUNTER_UPDATE(local_state->_stats_filtered_counter, stats.rows_stats_filtered);
785
852k
    COUNTER_UPDATE(local_state->_stats_rp_filtered_counter, stats.rows_stats_rp_filtered);
786
852k
    COUNTER_UPDATE(local_state->_dict_filtered_counter, stats.segment_dict_filtered);
787
852k
    COUNTER_UPDATE(local_state->_bf_filtered_counter, stats.rows_bf_filtered);
788
852k
    COUNTER_UPDATE(local_state->_del_filtered_counter, stats.rows_del_filtered);
789
852k
    COUNTER_UPDATE(local_state->_del_filtered_counter, stats.rows_del_by_bitmap);
790
852k
    COUNTER_UPDATE(local_state->_del_filtered_counter, stats.rows_vec_del_cond_filtered);
791
852k
    COUNTER_UPDATE(local_state->_conditions_filtered_counter, stats.rows_conditions_filtered);
792
852k
    COUNTER_UPDATE(local_state->_key_range_filtered_counter, stats.rows_key_range_filtered);
793
852k
    COUNTER_UPDATE(local_state->_total_pages_num_counter, stats.total_pages_num);
794
852k
    COUNTER_UPDATE(local_state->_cached_pages_num_counter, stats.cached_pages_num);
795
852k
    COUNTER_UPDATE(local_state->_inverted_index_filter_counter, stats.rows_inverted_index_filtered);
796
852k
    COUNTER_UPDATE(local_state->_inverted_index_filter_timer, stats.inverted_index_filter_timer);
797
852k
    COUNTER_UPDATE(local_state->_inverted_index_query_cache_hit_counter,
798
852k
                   stats.inverted_index_query_cache_hit);
799
852k
    COUNTER_UPDATE(local_state->_inverted_index_query_cache_miss_counter,
800
852k
                   stats.inverted_index_query_cache_miss);
801
852k
    COUNTER_UPDATE(local_state->_inverted_index_query_timer, stats.inverted_index_query_timer);
802
852k
    COUNTER_UPDATE(local_state->_inverted_index_query_null_bitmap_timer,
803
852k
                   stats.inverted_index_query_null_bitmap_timer);
804
852k
    COUNTER_UPDATE(local_state->_inverted_index_query_bitmap_copy_timer,
805
852k
                   stats.inverted_index_query_bitmap_copy_timer);
806
852k
    COUNTER_UPDATE(local_state->_inverted_index_searcher_open_timer,
807
852k
                   stats.inverted_index_searcher_open_timer);
808
852k
    COUNTER_UPDATE(local_state->_inverted_index_searcher_search_timer,
809
852k
                   stats.inverted_index_searcher_search_timer);
810
852k
    COUNTER_UPDATE(local_state->_inverted_index_searcher_search_init_timer,
811
852k
                   stats.inverted_index_searcher_search_init_timer);
812
852k
    COUNTER_UPDATE(local_state->_inverted_index_searcher_search_exec_timer,
813
852k
                   stats.inverted_index_searcher_search_exec_timer);
814
852k
    COUNTER_UPDATE(local_state->_inverted_index_searcher_cache_hit_counter,
815
852k
                   stats.inverted_index_searcher_cache_hit);
816
852k
    COUNTER_UPDATE(local_state->_inverted_index_searcher_cache_miss_counter,
817
852k
                   stats.inverted_index_searcher_cache_miss);
818
852k
    COUNTER_UPDATE(local_state->_inverted_index_downgrade_count_counter,
819
852k
                   stats.inverted_index_downgrade_count);
820
852k
    COUNTER_UPDATE(local_state->_inverted_index_analyzer_timer,
821
852k
                   stats.inverted_index_analyzer_timer);
822
852k
    COUNTER_UPDATE(local_state->_inverted_index_lookup_timer, stats.inverted_index_lookup_timer);
823
852k
    COUNTER_UPDATE(local_state->_variant_scan_sparse_column_timer,
824
852k
                   stats.variant_scan_sparse_column_timer_ns);
825
852k
    COUNTER_UPDATE(local_state->_variant_scan_sparse_column_bytes,
826
852k
                   stats.variant_scan_sparse_column_bytes);
827
852k
    COUNTER_UPDATE(local_state->_variant_fill_path_from_sparse_column_timer,
828
852k
                   stats.variant_fill_path_from_sparse_column_timer_ns);
829
852k
    COUNTER_UPDATE(local_state->_variant_subtree_default_iter_count,
830
852k
                   stats.variant_subtree_default_iter_count);
831
852k
    COUNTER_UPDATE(local_state->_variant_subtree_leaf_iter_count,
832
852k
                   stats.variant_subtree_leaf_iter_count);
833
852k
    COUNTER_UPDATE(local_state->_variant_subtree_hierarchical_iter_count,
834
852k
                   stats.variant_subtree_hierarchical_iter_count);
835
852k
    COUNTER_UPDATE(local_state->_variant_subtree_sparse_iter_count,
836
852k
                   stats.variant_subtree_sparse_iter_count);
837
852k
    COUNTER_UPDATE(local_state->_variant_doc_value_column_iter_count,
838
852k
                   stats.variant_doc_value_column_iter_count);
839
840
852k
    if (stats.adaptive_batch_size_predict_max_rows > 0) {
841
609k
        local_state->_adaptive_batch_predict_min_rows_counter->set(
842
609k
                stats.adaptive_batch_size_predict_min_rows);
843
609k
        local_state->_adaptive_batch_predict_max_rows_counter->set(
844
609k
                stats.adaptive_batch_size_predict_max_rows);
845
609k
    }
846
847
852k
    InvertedIndexProfileReporter inverted_index_profile;
848
852k
    inverted_index_profile.update(local_state->_index_filter_profile.get(),
849
852k
                                  &stats.inverted_index_stats);
850
851
    // only cloud deploy mode will use file cache.
852
852k
    if (config::is_cloud_mode() && config::enable_file_cache) {
853
851k
        io::FileCacheProfileReporter cache_profile(local_state->_segment_profile.get());
854
851k
        cache_profile.update(&stats.file_cache_stats);
855
851k
        _state->get_query_ctx()->resource_ctx()->io_context()->update_bytes_write_into_cache(
856
851k
                stats.file_cache_stats.bytes_write_into_cache);
857
851k
    }
858
852k
    COUNTER_UPDATE(local_state->_output_index_result_column_timer,
859
852k
                   stats.output_index_result_column_timer);
860
852k
    COUNTER_UPDATE(local_state->_filtered_segment_counter, stats.filtered_segment_number);
861
852k
    COUNTER_UPDATE(local_state->_total_segment_counter, stats.total_segment_number);
862
852k
    COUNTER_UPDATE(local_state->_condition_cache_hit_counter, stats.condition_cache_hit_seg_nums);
863
852k
    COUNTER_UPDATE(local_state->_condition_cache_filtered_rows_counter,
864
852k
                   stats.condition_cache_filtered_rows);
865
866
852k
    COUNTER_UPDATE(local_state->_tablet_reader_init_timer, stats.tablet_reader_init_timer_ns);
867
852k
    COUNTER_UPDATE(local_state->_tablet_reader_capture_rs_readers_timer,
868
852k
                   stats.tablet_reader_capture_rs_readers_timer_ns);
869
852k
    COUNTER_UPDATE(local_state->_tablet_reader_init_return_columns_timer,
870
852k
                   stats.tablet_reader_init_return_columns_timer_ns);
871
852k
    COUNTER_UPDATE(local_state->_tablet_reader_init_keys_param_timer,
872
852k
                   stats.tablet_reader_init_keys_param_timer_ns);
873
852k
    COUNTER_UPDATE(local_state->_tablet_reader_init_orderby_keys_param_timer,
874
852k
                   stats.tablet_reader_init_orderby_keys_param_timer_ns);
875
852k
    COUNTER_UPDATE(local_state->_tablet_reader_init_conditions_param_timer,
876
852k
                   stats.tablet_reader_init_conditions_param_timer_ns);
877
852k
    COUNTER_UPDATE(local_state->_tablet_reader_init_delete_condition_param_timer,
878
852k
                   stats.tablet_reader_init_delete_condition_param_timer_ns);
879
852k
    COUNTER_UPDATE(local_state->_block_reader_vcollect_iter_init_timer,
880
852k
                   stats.block_reader_vcollect_iter_init_timer_ns);
881
852k
    COUNTER_UPDATE(local_state->_block_reader_rs_readers_init_timer,
882
852k
                   stats.block_reader_rs_readers_init_timer_ns);
883
852k
    COUNTER_UPDATE(local_state->_block_reader_build_heap_init_timer,
884
852k
                   stats.block_reader_build_heap_init_timer_ns);
885
886
852k
    COUNTER_UPDATE(local_state->_rowset_reader_get_segment_iterators_timer,
887
852k
                   stats.rowset_reader_get_segment_iterators_timer_ns);
888
852k
    COUNTER_UPDATE(local_state->_rowset_reader_create_iterators_timer,
889
852k
                   stats.rowset_reader_create_iterators_timer_ns);
890
852k
    COUNTER_UPDATE(local_state->_rowset_reader_init_iterators_timer,
891
852k
                   stats.rowset_reader_init_iterators_timer_ns);
892
852k
    COUNTER_UPDATE(local_state->_rowset_reader_load_segments_timer,
893
852k
                   stats.rowset_reader_load_segments_timer_ns);
894
895
852k
    COUNTER_UPDATE(local_state->_segment_iterator_init_timer, stats.segment_iterator_init_timer_ns);
896
852k
    COUNTER_UPDATE(local_state->_segment_iterator_init_return_column_iterators_timer,
897
852k
                   stats.segment_iterator_init_return_column_iterators_timer_ns);
898
852k
    COUNTER_UPDATE(local_state->_segment_iterator_init_index_iterators_timer,
899
852k
                   stats.segment_iterator_init_index_iterators_timer_ns);
900
852k
    COUNTER_UPDATE(local_state->_segment_iterator_init_segment_prefetchers_timer,
901
852k
                   stats.segment_iterator_init_segment_prefetchers_timer_ns);
902
903
852k
    COUNTER_UPDATE(local_state->_segment_create_column_readers_timer,
904
852k
                   stats.segment_create_column_readers_timer_ns);
905
852k
    COUNTER_UPDATE(local_state->_segment_load_index_timer, stats.segment_load_index_timer_ns);
906
907
    // Update metrics
908
852k
    DorisMetrics::instance()->query_scan_bytes->increment(
909
852k
            local_state->_read_uncompressed_counter->value());
910
852k
    DorisMetrics::instance()->query_scan_rows->increment(local_state->_scan_rows->value());
911
852k
    auto& tablet = _tablet_reader_params.tablet;
912
852k
    tablet->query_scan_bytes->increment(local_state->_read_uncompressed_counter->value());
913
852k
    tablet->query_scan_rows->increment(local_state->_scan_rows->value());
914
852k
    tablet->query_scan_count->increment(1);
915
916
852k
    COUNTER_UPDATE(local_state->_ann_range_search_filter_counter,
917
852k
                   stats.rows_ann_index_range_filtered);
918
852k
    COUNTER_UPDATE(local_state->_ann_topn_filter_counter, stats.rows_ann_index_topn_filtered);
919
852k
    COUNTER_UPDATE(local_state->_ann_index_load_costs, stats.ann_index_load_ns);
920
852k
    COUNTER_UPDATE(local_state->_ann_ivf_on_disk_load_costs, stats.ann_ivf_on_disk_load_ns);
921
852k
    COUNTER_UPDATE(local_state->_ann_ivf_on_disk_cache_hit_cnt,
922
852k
                   stats.ann_ivf_on_disk_cache_hit_cnt);
923
852k
    COUNTER_UPDATE(local_state->_ann_ivf_on_disk_cache_miss_cnt,
924
852k
                   stats.ann_ivf_on_disk_cache_miss_cnt);
925
852k
    COUNTER_UPDATE(local_state->_ann_range_search_costs, stats.ann_index_range_search_ns);
926
852k
    COUNTER_UPDATE(local_state->_ann_range_search_cnt, stats.ann_index_range_search_cnt);
927
852k
    COUNTER_UPDATE(local_state->_ann_range_engine_search_costs, stats.ann_range_engine_search_ns);
928
    // Engine prepare before search
929
852k
    COUNTER_UPDATE(local_state->_ann_range_pre_process_costs, stats.ann_range_pre_process_ns);
930
    // Post process parent: Doris result process + engine convert
931
852k
    COUNTER_UPDATE(local_state->_ann_range_post_process_costs,
932
852k
                   stats.ann_range_result_convert_ns + stats.ann_range_engine_convert_ns);
933
    // Engine convert (child under post-process)
934
852k
    COUNTER_UPDATE(local_state->_ann_range_engine_convert_costs, stats.ann_range_engine_convert_ns);
935
    // Doris-side result convert (child under post-process)
936
852k
    COUNTER_UPDATE(local_state->_ann_range_result_convert_costs, stats.ann_range_result_convert_ns);
937
938
852k
    COUNTER_UPDATE(local_state->_ann_topn_search_costs, stats.ann_topn_search_ns);
939
852k
    COUNTER_UPDATE(local_state->_ann_topn_search_cnt, stats.ann_index_topn_search_cnt);
940
852k
    COUNTER_UPDATE(local_state->_ann_cache_hit_cnt, stats.ann_index_cache_hits);
941
852k
    COUNTER_UPDATE(local_state->_ann_range_cache_hit_cnt, stats.ann_index_range_cache_hits);
942
943
    // Detailed ANN timers
944
    // ANN TopN timers with hierarchy
945
    // Engine search time (FAISS)
946
852k
    COUNTER_UPDATE(local_state->_ann_topn_engine_search_costs,
947
852k
                   stats.ann_index_topn_engine_search_ns);
948
    // Engine prepare time (allocations/buffer setup before search)
949
852k
    COUNTER_UPDATE(local_state->_ann_topn_pre_process_costs,
950
852k
                   stats.ann_index_topn_engine_prepare_ns);
951
    // Post process parent includes Doris result processing + engine convert
952
852k
    COUNTER_UPDATE(local_state->_ann_topn_post_process_costs,
953
852k
                   stats.ann_index_topn_result_process_ns + stats.ann_index_topn_engine_convert_ns);
954
    // Engine-side conversion time inside FAISS wrappers (child under post-process)
955
852k
    COUNTER_UPDATE(local_state->_ann_topn_engine_convert_costs,
956
852k
                   stats.ann_index_topn_engine_convert_ns);
957
958
    // Doris-side result convert costs (show separately as another child counter); use pure process time
959
852k
    COUNTER_UPDATE(local_state->_ann_topn_result_convert_costs,
960
852k
                   stats.ann_index_topn_result_process_ns);
961
962
852k
    COUNTER_UPDATE(local_state->_ann_fallback_brute_force_cnt, stats.ann_fall_back_brute_force_cnt);
963
964
    // Overhead counter removed; precise instrumentation is reported via engine_prepare above.
965
852k
}
966
967
#ifndef NDEBUG
968
1.10M
Status OlapScanner::_check_ann_cache_hit_debug_points(const OlapReaderStatistics& stats) {
969
1.10M
    DBUG_EXECUTE_IF("olap_scanner.ann_topn_cache_hits", {
970
1.10M
        auto expected_hits = dp->param<int32_t>("expected_hits", -1);
971
1.10M
        auto min_hits = dp->param<int32_t>("min_hits", -1);
972
1.10M
        if (expected_hits >= 0 && stats.ann_index_cache_hits != expected_hits) {
973
1.10M
            return Status::Error<ErrorCode::INTERNAL_ERROR>(
974
1.10M
                    "ann_index_cache_hits: {} not equal to expected: {}",
975
1.10M
                    stats.ann_index_cache_hits, expected_hits);
976
1.10M
        }
977
1.10M
        if (min_hits >= 0 && stats.ann_index_cache_hits < min_hits) {
978
1.10M
            return Status::Error<ErrorCode::INTERNAL_ERROR>(
979
1.10M
                    "ann_index_cache_hits: {} less than expected min: {}",
980
1.10M
                    stats.ann_index_cache_hits, min_hits);
981
1.10M
        }
982
1.10M
    })
983
1.10M
    DBUG_EXECUTE_IF("olap_scanner.ann_range_cache_hits", {
984
1.10M
        auto expected_hits = dp->param<int32_t>("expected_hits", -1);
985
1.10M
        auto min_hits = dp->param<int32_t>("min_hits", -1);
986
1.10M
        if (expected_hits >= 0 && stats.ann_index_range_cache_hits != expected_hits) {
987
1.10M
            return Status::Error<ErrorCode::INTERNAL_ERROR>(
988
1.10M
                    "ann_index_range_cache_hits: {} not equal to expected: {}",
989
1.10M
                    stats.ann_index_range_cache_hits, expected_hits);
990
1.10M
        }
991
1.10M
        if (min_hits >= 0 && stats.ann_index_range_cache_hits < min_hits) {
992
1.10M
            return Status::Error<ErrorCode::INTERNAL_ERROR>(
993
1.10M
                    "ann_index_range_cache_hits: {} less than expected min: {}",
994
1.10M
                    stats.ann_index_range_cache_hits, min_hits);
995
1.10M
        }
996
1.10M
    })
997
1.10M
    return Status::OK();
998
1.10M
}
999
#endif
1000
1001
#include "common/compile_check_avoid_end.h"
1002
} // namespace doris