Coverage Report

Created: 2026-05-16 12:16

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
1.20M
        : Scanner(params.state, parent, params.limit, params.profile),
74
1.20M
          _key_ranges(std::move(params.key_ranges)),
75
1.20M
          _tablet_reader_params({.tablet = std::move(params.tablet),
76
1.20M
                                 .tablet_schema {},
77
1.20M
                                 .reader_type = params.read_row_binlog ? ReaderType::READER_BINLOG
78
1.20M
                                                                       : ReaderType::READER_QUERY,
79
1.20M
                                 .aggregation = params.aggregation,
80
1.20M
                                 .version = {0, params.version},
81
1.20M
                                 .start_key {},
82
1.20M
                                 .end_key {},
83
1.20M
                                 .predicates {},
84
1.20M
                                 .function_filters {},
85
1.20M
                                 .delete_predicates {},
86
1.20M
                                 .target_cast_type_for_variants {},
87
1.20M
                                 .all_access_paths {},
88
1.20M
                                 .predicate_access_paths {},
89
1.20M
                                 .rs_splits {},
90
1.20M
                                 .return_columns {},
91
1.20M
                                 .output_columns {},
92
1.20M
                                 .remaining_conjunct_roots {},
93
1.20M
                                 .common_expr_ctxs_push_down {},
94
1.20M
                                 .topn_filter_source_node_ids {},
95
1.20M
                                 .key_group_cluster_key_idxes {},
96
1.20M
                                 .virtual_column_exprs {},
97
1.20M
                                 .vir_cid_to_idx_in_block {},
98
1.20M
                                 .vir_col_idx_to_type {},
99
1.20M
                                 .score_runtime {},
100
1.20M
                                 .collection_statistics {},
101
1.20M
                                 .ann_topn_runtime {},
102
1.20M
                                 .condition_cache_digest = parent->get_condition_cache_digest()}) {
103
1.20M
    _tablet_reader_params.set_read_source(std::move(params.read_source),
104
1.20M
                                          _state->skip_delete_bitmap());
105
1.20M
    _has_prepared = false;
106
1.20M
    _vector_search_params = params.state->get_vector_search_params();
107
1.20M
}
108
109
static std::string read_columns_to_string(TabletSchemaSPtr tablet_schema,
110
2.87k
                                          const std::vector<uint32_t>& read_columns) {
111
    // avoid too long for one line,
112
    // it is hard to display in `show profile` stmt if one line is too long.
113
2.87k
    const int col_per_line = 10;
114
2.87k
    int i = 0;
115
2.87k
    std::string read_columns_string;
116
2.87k
    read_columns_string += "[";
117
15.9k
    for (auto it = read_columns.cbegin(); it != read_columns.cend(); it++) {
118
13.0k
        if (it != read_columns.cbegin()) {
119
10.1k
            read_columns_string += ", ";
120
10.1k
        }
121
13.0k
        read_columns_string += tablet_schema->columns().at(*it)->name();
122
13.0k
        if (i >= col_per_line) {
123
13
            read_columns_string += "\n";
124
13
            i = 0;
125
13.0k
        } else {
126
13.0k
            ++i;
127
13.0k
        }
128
13.0k
    }
129
2.87k
    read_columns_string += "]";
130
2.87k
    return read_columns_string;
131
2.87k
}
132
133
1.20M
Status OlapScanner::_prepare_impl() {
134
1.20M
    auto* local_state = static_cast<OlapScanLocalState*>(_local_state);
135
1.20M
    auto& tablet = _tablet_reader_params.tablet;
136
1.20M
    auto& tablet_schema = _tablet_reader_params.tablet_schema;
137
1.20M
    DBUG_EXECUTE_IF("CloudTablet.capture_rs_readers.return.e-230", {
138
1.20M
        LOG_WARNING("CloudTablet.capture_rs_readers.return e-230 init")
139
1.20M
                .tag("tablet_id", tablet->tablet_id());
140
1.20M
        return Status::Error<false>(-230, "injected error");
141
1.20M
    });
142
143
1.20M
    for (auto& ctx : local_state->_common_expr_ctxs_push_down) {
144
22.8k
        VExprContextSPtr context;
145
22.8k
        RETURN_IF_ERROR(ctx->clone(_state, context));
146
22.8k
        _common_expr_ctxs_push_down.emplace_back(context);
147
22.8k
        context->prepare_ann_range_search(_vector_search_params);
148
22.8k
    }
149
150
1.20M
    for (auto pair : local_state->_slot_id_to_virtual_column_expr) {
151
        // Scanner will be executed in a different thread, so we need to clone the context.
152
284
        VExprContextSPtr context;
153
284
        RETURN_IF_ERROR(pair.second->clone(_state, context));
154
284
        _slot_id_to_virtual_column_expr[pair.first] = context;
155
284
    }
156
157
1.20M
    _slot_id_to_index_in_block = local_state->_slot_id_to_index_in_block;
158
1.20M
    _slot_id_to_col_type = local_state->_slot_id_to_col_type;
159
1.20M
    _score_runtime = local_state->_score_runtime;
160
    // All scanners share the same ann_topn_runtime.
161
1.20M
    _ann_topn_runtime = local_state->_ann_topn_runtime;
162
163
    // set limit to reduce end of rowset and segment mem use
164
1.20M
    _tablet_reader = std::make_unique<BlockReader>();
165
    // batch size is passed down to segment iterator, use _state->batch_size()
166
    // instead of _parent->limit(), because if _parent->limit() is a very small
167
    // value (e.g. select a from t where a .. and b ... limit 1),
168
    // it will be very slow when reading data in segment iterator
169
1.20M
    _tablet_reader->set_batch_size(_state->batch_size());
170
    // Adaptive batch size: pass byte-budget settings to the storage reader.
171
    // The reader still uses batch_size() as the row ceiling.
172
1.20M
    _tablet_reader->set_preferred_block_size_bytes(_state->preferred_block_size_bytes());
173
1.20M
    {
174
1.20M
        TOlapScanNode& olap_scan_node = local_state->olap_scan_node();
175
1.20M
        TabletSchemaSPtr source_tablet_schema =
176
1.20M
                _tablet_reader_params.reader_type == ReaderType::READER_BINLOG
177
1.20M
                        ? tablet->row_binlog_tablet_schema()
178
1.20M
                        : tablet->tablet_schema();
179
180
1.20M
        tablet_schema = std::make_shared<TabletSchema>();
181
1.20M
        tablet_schema->copy_from(*source_tablet_schema);
182
1.20M
        if (olap_scan_node.__isset.columns_desc && !olap_scan_node.columns_desc.empty() &&
183
1.20M
            olap_scan_node.columns_desc[0].col_unique_id >= 0) {
184
1.20M
            tablet_schema->clear_columns();
185
19.2M
            for (const auto& column_desc : olap_scan_node.columns_desc) {
186
19.2M
                tablet_schema->append_column(TabletColumn(column_desc));
187
19.2M
            }
188
1.20M
            if (olap_scan_node.__isset.schema_version) {
189
1.20M
                tablet_schema->set_schema_version(olap_scan_node.schema_version);
190
1.20M
            }
191
1.20M
        }
192
1.20M
        if (olap_scan_node.__isset.indexes_desc) {
193
1.20M
            tablet_schema->update_indexes_from_thrift(olap_scan_node.indexes_desc);
194
1.20M
        }
195
196
1.20M
        if (_tablet_reader_params.rs_splits.empty()) {
197
            // Non-pipeline mode, Tablet : Scanner = 1 : 1
198
            // acquire tablet rowset readers at the beginning of the scan node
199
            // to prevent this case: when there are lots of olap scanners to run for example 10000
200
            // the rowsets maybe compacted when the last olap scanner starts
201
0
            ReadSource read_source;
202
203
0
            if (config::is_cloud_mode()) {
204
                // FIXME(plat1ko): Avoid pointer cast
205
0
                ExecEnv::GetInstance()->storage_engine().to_cloud().tablet_hotspot().count(*tablet);
206
0
            }
207
208
0
            auto maybe_read_source = tablet->capture_read_source(
209
0
                    _tablet_reader_params.version,
210
0
                    {
211
0
                            .skip_missing_versions = _state->skip_missing_version(),
212
0
                            .enable_fetch_rowsets_from_peers =
213
0
                                    config::enable_fetch_rowsets_from_peer_replicas,
214
0
                            .capture_row_binlog =
215
0
                                    _tablet_reader_params.reader_type == ReaderType::READER_BINLOG,
216
0
                            .enable_prefer_cached_rowset =
217
0
                                    config::is_cloud_mode() ? _state->enable_prefer_cached_rowset()
218
0
                                                            : false,
219
0
                            .query_freshness_tolerance_ms =
220
0
                                    config::is_cloud_mode() ? _state->query_freshness_tolerance_ms()
221
0
                                                            : -1,
222
0
                    });
223
0
            if (!maybe_read_source) {
224
0
                LOG(WARNING) << "fail to init reader. res=" << maybe_read_source.error();
225
0
                return maybe_read_source.error();
226
0
            }
227
0
            read_source = std::move(maybe_read_source.value());
228
229
0
            if (config::enable_mow_verbose_log && tablet->enable_unique_key_merge_on_write()) {
230
0
                LOG_INFO("finish capture_rs_readers for tablet={}, query_id={}",
231
0
                         tablet->tablet_id(), print_id(_state->query_id()));
232
0
            }
233
234
0
            if (!_state->skip_delete_predicate()) {
235
0
                read_source.fill_delete_predicates();
236
0
            }
237
0
            _tablet_reader_params.set_read_source(std::move(read_source));
238
0
        }
239
240
        // Initialize tablet_reader_params
241
1.20M
        RETURN_IF_ERROR(_init_tablet_reader_params(
242
1.20M
                local_state->_parent->cast<OlapScanOperatorX>()._slot_id_to_slot_desc, _key_ranges,
243
1.20M
                local_state->_slot_id_to_predicates, local_state->_push_down_functions));
244
1.20M
    }
245
246
    // add read columns in profile
247
1.20M
    if (_state->enable_profile()) {
248
2.87k
        _profile->add_info_string("ReadColumns",
249
2.87k
                                  read_columns_to_string(tablet_schema, _return_columns));
250
2.87k
    }
251
252
1.20M
    if (_tablet_reader_params.score_runtime) {
253
14
        SCOPED_TIMER(local_state->_statistics_collect_timer);
254
14
        _tablet_reader_params.collection_statistics = std::make_shared<CollectionStatistics>();
255
256
14
        io::IOContext io_ctx {
257
14
                .reader_type = _tablet_reader_params.reader_type,
258
14
                .expiration_time = tablet->ttl_seconds(),
259
14
                .query_id = &_state->query_id(),
260
14
                .file_cache_stats = &_tablet_reader->mutable_stats()->file_cache_stats,
261
14
                .is_inverted_index = true,
262
14
        };
263
264
14
        RETURN_IF_ERROR(_tablet_reader_params.collection_statistics->collect(
265
14
                _state, _tablet_reader_params.rs_splits, _tablet_reader_params.tablet_schema,
266
14
                _tablet_reader_params.common_expr_ctxs_push_down, &io_ctx));
267
14
    }
268
269
1.20M
    _has_prepared = true;
270
1.20M
    return Status::OK();
271
1.20M
}
272
273
1.20M
Status OlapScanner::_open_impl(RuntimeState* state) {
274
1.20M
    RETURN_IF_ERROR(Scanner::_open_impl(state));
275
1.20M
    SCOPED_TIMER(_local_state->cast<OlapScanLocalState>()._reader_init_timer);
276
277
1.20M
    auto res = _tablet_reader->init(_tablet_reader_params);
278
1.20M
    if (!res.ok()) {
279
76
        res.append("failed to initialize storage reader. tablet=" +
280
76
                   std::to_string(_tablet_reader_params.tablet->tablet_id()) +
281
76
                   ", backend=" + BackendOptions::get_localhost());
282
76
        return res;
283
76
    }
284
285
    // Do not hold rs_splits any more to release memory.
286
1.20M
    _tablet_reader_params.rs_splits.clear();
287
288
1.20M
    return Status::OK();
289
1.20M
}
290
291
// it will be called under tablet read lock because capture rs readers need
292
Status OlapScanner::_init_tablet_reader_params(
293
        const phmap::flat_hash_map<int, SlotDescriptor*>& slot_id_to_slot_desc,
294
        const std::vector<OlapScanRange*>& key_ranges,
295
        const phmap::flat_hash_map<int, std::vector<std::shared_ptr<ColumnPredicate>>>&
296
                slot_to_predicates,
297
1.20M
        const std::vector<FunctionFilter>& function_filters) {
298
    // if the table with rowset [0-x] or [0-1] [2-y], and [0-1] is empty
299
1.20M
    const bool single_version = _tablet_reader_params.has_single_version();
300
301
1.20M
    auto* olap_local_state = static_cast<OlapScanLocalState*>(_local_state);
302
1.20M
    bool read_mor_as_dup = olap_local_state->olap_scan_node().__isset.read_mor_as_dup &&
303
1.20M
                           olap_local_state->olap_scan_node().read_mor_as_dup;
304
1.20M
    if (_state->skip_storage_engine_merge() || read_mor_as_dup) {
305
43
        _tablet_reader_params.direct_mode = true;
306
43
        _tablet_reader_params.aggregation = true;
307
1.20M
    } else {
308
1.20M
        auto push_down_agg_type = _local_state->get_push_down_agg_type();
309
1.20M
        _tablet_reader_params.direct_mode = _tablet_reader_params.aggregation || single_version ||
310
1.20M
                                            (push_down_agg_type != TPushAggOp::NONE &&
311
10.6k
                                             push_down_agg_type != TPushAggOp::COUNT_ON_INDEX);
312
1.20M
    }
313
314
1.20M
    RETURN_IF_ERROR(_init_variant_columns());
315
1.20M
    RETURN_IF_ERROR(_init_return_columns());
316
317
1.20M
    _tablet_reader_params.push_down_agg_type_opt = _local_state->get_push_down_agg_type();
318
319
    // TODO: If a new runtime filter arrives after `_conjuncts` move to `_common_expr_ctxs_push_down`,
320
1.20M
    if (_common_expr_ctxs_push_down.empty()) {
321
1.18M
        for (auto& conjunct : _conjuncts) {
322
3.75k
            _tablet_reader_params.remaining_conjunct_roots.emplace_back(conjunct->root());
323
3.75k
        }
324
1.18M
    } else {
325
22.7k
        for (auto& ctx : _common_expr_ctxs_push_down) {
326
22.7k
            _tablet_reader_params.remaining_conjunct_roots.emplace_back(ctx->root());
327
22.7k
        }
328
19.6k
    }
329
330
1.20M
    _tablet_reader_params.common_expr_ctxs_push_down = _common_expr_ctxs_push_down;
331
1.20M
    _tablet_reader_params.virtual_column_exprs = _virtual_column_exprs;
332
1.20M
    _tablet_reader_params.vir_cid_to_idx_in_block = _vir_cid_to_idx_in_block;
333
1.20M
    _tablet_reader_params.vir_col_idx_to_type = _vir_col_idx_to_type;
334
1.20M
    _tablet_reader_params.score_runtime = _score_runtime;
335
1.20M
    _tablet_reader_params.output_columns = ((OlapScanLocalState*)_local_state)->_output_column_ids;
336
1.20M
    _tablet_reader_params.ann_topn_runtime = _ann_topn_runtime;
337
1.20M
    for (const auto& ele : ((OlapScanLocalState*)_local_state)->_cast_types_for_variants) {
338
1.70k
        _tablet_reader_params.target_cast_type_for_variants[ele.first] = ele.second;
339
1.70k
    };
340
1.20M
    auto& tablet_schema = _tablet_reader_params.tablet_schema;
341
9.70M
    for (auto& predicates : slot_to_predicates) {
342
9.70M
        const int sid = predicates.first;
343
9.70M
        DCHECK(slot_id_to_slot_desc.contains(sid));
344
9.70M
        int32_t index =
345
9.70M
                tablet_schema->field_index(slot_id_to_slot_desc.find(sid)->second->col_name());
346
9.70M
        if (index < 0) {
347
0
            throw Exception(
348
0
                    Status::InternalError("Column {} not found in tablet schema",
349
0
                                          slot_id_to_slot_desc.find(sid)->second->col_name()));
350
0
        }
351
9.70M
        for (auto& predicate : predicates.second) {
352
971k
            _tablet_reader_params.predicates.push_back(predicate->clone(index));
353
971k
        }
354
9.70M
    }
355
356
1.20M
    std::copy(function_filters.cbegin(), function_filters.cend(),
357
1.20M
              std::inserter(_tablet_reader_params.function_filters,
358
1.20M
                            _tablet_reader_params.function_filters.begin()));
359
360
    // Merge the columns in delete predicate that not in latest schema in to current tablet schema
361
1.20M
    for (auto& del_pred : _tablet_reader_params.delete_predicates) {
362
6.71k
        tablet_schema->merge_dropped_columns(*del_pred->tablet_schema());
363
6.71k
    }
364
365
    // Push key ranges to the tablet reader.
366
    // Skip the "full scan" placeholder (has_lower_bound == false) — when no key
367
    // predicates exist, start_key/end_key remain empty and the reader does a full scan.
368
1.77M
    for (auto* key_range : key_ranges) {
369
1.77M
        if (!key_range->has_lower_bound) {
370
167k
            continue;
371
167k
        }
372
373
1.60M
        _tablet_reader_params.start_key_include = key_range->begin_include;
374
1.60M
        _tablet_reader_params.end_key_include = key_range->end_include;
375
376
1.60M
        _tablet_reader_params.start_key.push_back(key_range->begin_scan_range);
377
1.60M
        _tablet_reader_params.end_key.push_back(key_range->end_scan_range);
378
1.60M
    }
379
380
1.20M
    _tablet_reader_params.profile = _local_state->custom_profile();
381
1.20M
    _tablet_reader_params.runtime_state = _state;
382
383
1.20M
    _tablet_reader_params.origin_return_columns = &_return_columns;
384
1.20M
    _tablet_reader_params.tablet_columns_convert_to_null_set = &_tablet_columns_convert_to_null_set;
385
386
1.20M
    if (_tablet_reader_params.direct_mode) {
387
1.19M
        _tablet_reader_params.return_columns = _return_columns;
388
1.19M
    } else {
389
        // we need to fetch all key columns to do the right aggregation on storage engine side.
390
38.4k
        for (size_t i = 0; i < tablet_schema->num_key_columns(); ++i) {
391
27.1k
            _tablet_reader_params.return_columns.push_back(i);
392
27.1k
        }
393
49.5k
        for (auto index : _return_columns) {
394
49.5k
            if (tablet_schema->column(index).is_key()) {
395
18.1k
                continue;
396
18.1k
            }
397
31.3k
            _tablet_reader_params.return_columns.push_back(index);
398
31.3k
        }
399
        // expand the sequence column
400
11.2k
        if (tablet_schema->has_sequence_col() || tablet_schema->has_seq_map()) {
401
40
            bool has_replace_col = false;
402
91
            for (auto col : _return_columns) {
403
91
                if (tablet_schema->column(col).aggregation() ==
404
91
                    FieldAggregationMethod::OLAP_FIELD_AGGREGATION_REPLACE) {
405
40
                    has_replace_col = true;
406
40
                    break;
407
40
                }
408
91
            }
409
40
            if (auto sequence_col_idx = tablet_schema->sequence_col_idx();
410
40
                has_replace_col && tablet_schema->has_sequence_col() &&
411
40
                std::find(_return_columns.begin(), _return_columns.end(), sequence_col_idx) ==
412
28
                        _return_columns.end()) {
413
16
                _tablet_reader_params.return_columns.push_back(sequence_col_idx);
414
16
            }
415
40
            if (has_replace_col) {
416
40
                const auto& val_to_seq = tablet_schema->value_col_idx_to_seq_col_idx();
417
40
                std::set<uint32_t> return_seq_columns;
418
419
244
                for (auto col : _tablet_reader_params.return_columns) {
420
                    // we need to add the necessary sequence column in _return_columns, and
421
                    // Avoid adding the same seq column twice
422
244
                    const auto val_iter = val_to_seq.find(col);
423
244
                    if (val_iter != val_to_seq.end()) {
424
42
                        auto seq = val_iter->second;
425
42
                        if (std::find(_tablet_reader_params.return_columns.begin(),
426
42
                                      _tablet_reader_params.return_columns.end(),
427
42
                                      seq) == _tablet_reader_params.return_columns.end()) {
428
4
                            return_seq_columns.insert(seq);
429
4
                        }
430
42
                    }
431
244
                }
432
40
                _tablet_reader_params.return_columns.insert(
433
40
                        std::end(_tablet_reader_params.return_columns),
434
40
                        std::begin(return_seq_columns), std::end(return_seq_columns));
435
40
            }
436
40
        }
437
11.2k
    }
438
439
1.20M
    _tablet_reader_params.use_page_cache = _state->enable_page_cache();
440
441
1.20M
    DBUG_EXECUTE_IF("NewOlapScanner::_init_tablet_reader_params.block", DBUG_BLOCK);
442
443
1.20M
    if (!_state->skip_storage_engine_merge()) {
444
1.20M
        auto* olap_scan_local_state = (OlapScanLocalState*)_local_state;
445
1.20M
        TOlapScanNode& olap_scan_node = olap_scan_local_state->olap_scan_node();
446
447
        // Set MOR value predicate pushdown flag
448
1.20M
        if (olap_scan_node.__isset.enable_mor_value_predicate_pushdown &&
449
1.20M
            olap_scan_node.enable_mor_value_predicate_pushdown) {
450
25
            _tablet_reader_params.enable_mor_value_predicate_pushdown = true;
451
25
        }
452
453
1.20M
        const bool has_key_topn =
454
1.20M
                olap_scan_node.__isset.sort_info && !olap_scan_node.sort_info.is_asc_order.empty();
455
1.20M
        if (has_key_topn) {
456
2.24k
            _limit = _local_state->limit_per_scanner();
457
2.24k
        }
458
459
1.20M
        const bool no_runtime_filters = _total_rf_num == 0;
460
1.20M
        const bool segment_limit_enabled = _state->enable_segment_limit_pushdown();
461
1.20M
        const bool storage_no_merge = olap_scan_local_state->_storage_no_merge();
462
463
1.20M
        if (_limit > 0 && no_runtime_filters && segment_limit_enabled && storage_no_merge) {
464
4.00k
            for (const auto& conjunct : _conjuncts) {
465
0
                DORIS_CHECK(!olap_scan_local_state->_check_expr_storage_filter(
466
0
                        conjunct->root(), OlapScanLocalState::ExprStorageFilterCheckMode::
467
0
                                                  HAS_SEGMENT_EVALUABLE_EXPR));
468
0
            }
469
4.00k
        }
470
471
        // Segment LIMIT has only two legal states: completely disabled, or enabled after every
472
        // row-filtering conjunct has become a storage predicate or SegmentIterator common expr.
473
1.20M
        const bool can_push_down_segment_limit = _limit > 0 && no_runtime_filters &&
474
1.20M
                                                 _conjuncts.empty() && segment_limit_enabled &&
475
1.20M
                                                 storage_no_merge;
476
1.20M
        if (can_push_down_segment_limit) {
477
3.99k
            if (has_key_topn) {
478
2.13k
                _tablet_reader_params.read_orderby_key = true;
479
2.13k
                if (!olap_scan_node.sort_info.is_asc_order[0]) {
480
460
                    _tablet_reader_params.read_orderby_key_reverse = true;
481
460
                }
482
2.13k
                _tablet_reader_params.read_orderby_key_num_prefix_columns =
483
2.13k
                        olap_scan_node.sort_info.is_asc_order.size();
484
2.13k
                _tablet_reader_params.read_orderby_key_limit = _limit;
485
2.13k
            } else {
486
1.86k
                _tablet_reader_params.general_read_limit = _limit;
487
1.86k
            }
488
3.99k
        }
489
490
1.20M
        if (_tablet_reader_params.read_orderby_key_limit > 0 ||
491
1.20M
            _tablet_reader_params.general_read_limit > 0) {
492
3.99k
            DORIS_CHECK(can_push_down_segment_limit);
493
3.99k
            DORIS_CHECK(_conjuncts.empty());
494
3.99k
        }
495
496
        // A key TopN scan cannot share the plain LIMIT early-stop counter. If
497
        // storage TopN is pushed down, each scanner must produce its full local
498
        // candidates. If it is not pushed down for any reason, the upper TopN
499
        // still needs all rows from the scan.
500
1.20M
        if (has_key_topn) {
501
2.24k
            _shared_scan_limit = nullptr;
502
2.24k
            if (_tablet_reader_params.read_orderby_key_limit == 0) {
503
107
                _limit = -1;
504
107
            }
505
2.24k
        }
506
        // Note: _shared_scan_limit is intentionally not pushed into the
507
        // storage layer. SegmentIterator's _process_eof() is irreversible,
508
        // so a concurrently-decremented atomic could reach 0 while a segment
509
        // still has data needed by other scanners.
510
511
        // set push down topn filter
512
1.20M
        _tablet_reader_params.topn_filter_source_node_ids =
513
1.20M
                olap_scan_local_state->get_topn_filter_source_node_ids(_state, true);
514
1.20M
        if (!_tablet_reader_params.topn_filter_source_node_ids.empty()) {
515
6.27k
            _tablet_reader_params.topn_filter_target_node_id =
516
6.27k
                    olap_scan_local_state->parent()->node_id();
517
6.27k
        }
518
1.20M
    }
519
520
    // If this is a Two-Phase read query, and we need to delay the release of Rowset
521
    // by rowset->update_delayed_expired_timestamp().This could expand the lifespan of Rowset
522
1.20M
    if (tablet_schema->field_index(BeConsts::ROWID_COL) >= 0) {
523
0
        constexpr static int delayed_s = 60;
524
0
        for (auto rs_reader : _tablet_reader_params.rs_splits) {
525
0
            uint64_t delayed_expired_timestamp =
526
0
                    UnixSeconds() + _tablet_reader_params.runtime_state->execution_timeout() +
527
0
                    delayed_s;
528
0
            rs_reader.rs_reader->rowset()->update_delayed_expired_timestamp(
529
0
                    delayed_expired_timestamp);
530
0
            ExecEnv::GetInstance()->storage_engine().add_quering_rowset(
531
0
                    rs_reader.rs_reader->rowset());
532
0
        }
533
0
    }
534
535
1.20M
    if (tablet_schema->has_global_row_id()) {
536
7.52k
        auto& id_file_map = _state->get_id_file_map();
537
13.5k
        for (auto rs_reader : _tablet_reader_params.rs_splits) {
538
13.5k
            id_file_map->add_temp_rowset(rs_reader.rs_reader->rowset());
539
13.5k
        }
540
7.52k
    }
541
542
1.20M
    return Status::OK();
543
1.20M
}
544
545
1.20M
Status OlapScanner::_init_variant_columns() {
546
1.20M
    auto& tablet_schema = _tablet_reader_params.tablet_schema;
547
1.20M
    if (tablet_schema->num_variant_columns() == 0) {
548
1.19M
        return Status::OK();
549
1.19M
    }
550
    // Parent column has path info to distinction from each other
551
14.3k
    for (auto* slot : _output_tuple_desc->slots()) {
552
14.3k
        if (slot->type()->get_primitive_type() == PrimitiveType::TYPE_VARIANT) {
553
            // Such columns are not exist in frontend schema info, so we need to
554
            // add them into tablet_schema for later column indexing.
555
7.11k
            const auto& dt_variant =
556
7.11k
                    assert_cast<const DataTypeVariant&>(*remove_nullable(slot->type()));
557
7.11k
            TabletColumn subcol = TabletColumn::create_materialized_variant_column(
558
7.11k
                    tablet_schema->column_by_uid(slot->col_unique_id()).name_lower_case(),
559
7.11k
                    slot->column_paths(), slot->col_unique_id(),
560
7.11k
                    dt_variant.variant_max_subcolumns_count(), dt_variant.enable_doc_mode());
561
7.11k
            if (tablet_schema->field_index(*subcol.path_info_ptr()) < 0) {
562
5.40k
                tablet_schema->append_column(subcol, TabletSchema::ColumnType::VARIANT);
563
5.40k
            }
564
7.11k
        }
565
14.3k
    }
566
5.38k
    variant_util::inherit_column_attributes(tablet_schema);
567
5.38k
    return Status::OK();
568
1.20M
}
569
570
1.20M
Status OlapScanner::_init_return_columns() {
571
12.7M
    for (auto* slot : _output_tuple_desc->slots()) {
572
        // variant column using path to index a column
573
12.7M
        int32_t index = 0;
574
12.7M
        auto& tablet_schema = _tablet_reader_params.tablet_schema;
575
12.7M
        if (slot->type()->get_primitive_type() == PrimitiveType::TYPE_VARIANT) {
576
7.11k
            index = tablet_schema->field_index(PathInData(
577
7.11k
                    tablet_schema->column_by_uid(slot->col_unique_id()).name_lower_case(),
578
7.11k
                    slot->column_paths()));
579
12.7M
        } else {
580
12.7M
            index = slot->col_unique_id() >= 0 ? tablet_schema->field_index(slot->col_unique_id())
581
18.4E
                                               : tablet_schema->field_index(slot->col_name());
582
12.7M
        }
583
584
12.7M
        if (index < 0) {
585
0
            return Status::InternalError(
586
0
                    "field name is invalid. field={}, field_name_to_index={}, col_unique_id={}",
587
0
                    slot->col_name(), tablet_schema->get_all_field_names(), slot->col_unique_id());
588
0
        }
589
590
12.7M
        if (slot->get_virtual_column_expr()) {
591
285
            ColumnId virtual_column_cid = index;
592
285
            _virtual_column_exprs[virtual_column_cid] = _slot_id_to_virtual_column_expr[slot->id()];
593
285
            size_t idx_in_block = _slot_id_to_index_in_block[slot->id()];
594
285
            _vir_cid_to_idx_in_block[virtual_column_cid] = idx_in_block;
595
285
            _vir_col_idx_to_type[idx_in_block] = _slot_id_to_col_type[slot->id()];
596
597
285
            VLOG_DEBUG << fmt::format(
598
1
                    "Virtual column, slot id: {}, cid {}, column index: {}, type: {}", slot->id(),
599
1
                    virtual_column_cid, _vir_cid_to_idx_in_block[virtual_column_cid],
600
1
                    _vir_col_idx_to_type[idx_in_block]->get_name());
601
285
        }
602
603
12.7M
        const auto& column = tablet_schema->column(index);
604
12.7M
        int32_t unique_id =
605
12.7M
                column.unique_id() >= 0 ? column.unique_id() : column.parent_unique_id();
606
12.7M
        if (!slot->all_access_paths().empty()) {
607
84.8k
            _tablet_reader_params.all_access_paths.insert({unique_id, slot->all_access_paths()});
608
84.8k
        }
609
610
12.7M
        if (!slot->predicate_access_paths().empty()) {
611
14.4k
            _tablet_reader_params.predicate_access_paths.insert(
612
14.4k
                    {unique_id, slot->predicate_access_paths()});
613
14.4k
        }
614
615
12.7M
        if ((slot->type()->get_primitive_type() == PrimitiveType::TYPE_STRUCT ||
616
12.7M
             slot->type()->get_primitive_type() == PrimitiveType::TYPE_MAP ||
617
12.7M
             slot->type()->get_primitive_type() == PrimitiveType::TYPE_ARRAY) &&
618
12.7M
            !slot->all_access_paths().empty()) {
619
70.7k
            tablet_schema->add_pruned_columns_data_type(column.unique_id(), slot->type());
620
70.7k
        }
621
622
12.7M
        _return_columns.push_back(index);
623
12.7M
        if (slot->is_nullable() && !tablet_schema->column(index).is_nullable()) {
624
0
            _tablet_columns_convert_to_null_set.emplace(index);
625
12.7M
        } else if (!slot->is_nullable() && tablet_schema->column(index).is_nullable()) {
626
0
            return Status::Error<ErrorCode::INVALID_SCHEMA>(
627
0
                    "slot(id: {}, name: {})'s nullable does not match "
628
0
                    "column(tablet id: {}, index: {}, name: {}) ",
629
0
                    slot->id(), slot->col_name(), tablet_schema->table_id(), index,
630
0
                    tablet_schema->column(index).name());
631
0
        }
632
12.7M
    }
633
634
1.20M
    if (_return_columns.empty()) {
635
0
        return Status::InternalError("failed to build storage scanner, no materialized slot!");
636
0
    }
637
638
1.20M
    return Status::OK();
639
1.20M
}
640
641
1.25M
doris::TabletStorageType OlapScanner::get_storage_type() {
642
1.25M
    if (config::is_cloud_mode()) {
643
        // we don't have cold storage in cloud mode, all storage is treated as local
644
909k
        return doris::TabletStorageType::STORAGE_TYPE_LOCAL;
645
909k
    }
646
341k
    int local_reader = 0;
647
521k
    for (const auto& reader : _tablet_reader_params.rs_splits) {
648
521k
        local_reader += reader.rs_reader->rowset()->is_local();
649
521k
    }
650
341k
    int total_reader = _tablet_reader_params.rs_splits.size();
651
652
341k
    if (local_reader == total_reader) {
653
341k
        return doris::TabletStorageType::STORAGE_TYPE_LOCAL;
654
341k
    } else if (local_reader == 0) {
655
0
        return doris::TabletStorageType::STORAGE_TYPE_REMOTE;
656
0
    }
657
0
    return doris::TabletStorageType::STORAGE_TYPE_REMOTE_AND_LOCAL;
658
341k
}
659
660
1.47M
Status OlapScanner::_get_block_impl(RuntimeState* state, Block* block, bool* eof) {
661
    // Read one block from block reader
662
    // ATTN: Here we need to let the _get_block_impl method guarantee the semantics of the interface,
663
    // that is, eof can be set to true only when the returned block is empty.
664
1.47M
    RETURN_IF_ERROR(_tablet_reader->next_block_with_aggregation(block, eof));
665
1.47M
    if (block->rows() > 0) {
666
267k
        _tablet_reader_params.tablet->read_block_count.fetch_add(1, std::memory_order_relaxed);
667
267k
        *eof = false;
668
267k
    }
669
1.47M
#ifndef NDEBUG
670
1.47M
    RETURN_IF_ERROR(_check_ann_cache_hit_debug_points(_tablet_reader->stats()));
671
1.47M
#endif
672
1.47M
    return Status::OK();
673
1.47M
}
674
675
1.20M
Status OlapScanner::close(RuntimeState* state) {
676
1.20M
    if (!_try_close()) {
677
150
        return Status::OK();
678
150
    }
679
1.20M
    RETURN_IF_ERROR(Scanner::close(state));
680
1.20M
    return Status::OK();
681
1.20M
}
682
683
1.24M
void OlapScanner::update_realtime_counters() {
684
1.24M
    if (!_has_prepared) {
685
        // Counter update need prepare successfully, or it maybe core. For example, olap scanner
686
        // will open tablet reader during prepare, if not prepare successfully, tablet reader == nullptr.
687
0
        return;
688
0
    }
689
1.24M
    OlapScanLocalState* local_state = static_cast<OlapScanLocalState*>(_local_state);
690
1.24M
    const OlapReaderStatistics& stats = _tablet_reader->stats();
691
1.24M
    COUNTER_UPDATE(local_state->_read_compressed_counter, stats.compressed_bytes_read);
692
1.24M
    COUNTER_UPDATE(local_state->_read_uncompressed_counter, stats.uncompressed_bytes_read);
693
1.24M
    COUNTER_UPDATE(local_state->_scan_bytes, stats.uncompressed_bytes_read);
694
1.24M
    COUNTER_UPDATE(local_state->_scan_rows, stats.raw_rows_read);
695
696
    // Make sure the scan bytes and scan rows counter in audit log is the same as the counter in
697
    // doris metrics.
698
    // ScanBytes is the uncompressed bytes read from local + remote
699
    // bytes_read_from_local is the compressed bytes read from local
700
    // bytes_read_from_remote is the compressed bytes read from remote
701
    // scan bytes > bytes_read_from_local + bytes_read_from_remote
702
1.24M
    _state->get_query_ctx()->resource_ctx()->io_context()->update_scan_rows(stats.raw_rows_read);
703
1.24M
    _state->get_query_ctx()->resource_ctx()->io_context()->update_scan_bytes(
704
1.24M
            stats.uncompressed_bytes_read);
705
706
    // In case of no cache, we still need to update the IO stats. uncompressed bytes read == local + remote
707
1.24M
    if (stats.file_cache_stats.bytes_read_from_local == 0 &&
708
1.24M
        stats.file_cache_stats.bytes_read_from_remote == 0) {
709
994k
        _state->get_query_ctx()->resource_ctx()->io_context()->update_scan_bytes_from_local_storage(
710
994k
                stats.compressed_bytes_read);
711
994k
        DorisMetrics::instance()->query_scan_bytes_from_local->increment(
712
994k
                stats.compressed_bytes_read);
713
994k
    } else {
714
254k
        _state->get_query_ctx()->resource_ctx()->io_context()->update_scan_bytes_from_local_storage(
715
254k
                stats.file_cache_stats.bytes_read_from_local - _bytes_read_from_local);
716
254k
        _state->get_query_ctx()
717
254k
                ->resource_ctx()
718
254k
                ->io_context()
719
254k
                ->update_scan_bytes_from_remote_storage(
720
254k
                        stats.file_cache_stats.bytes_read_from_remote - _bytes_read_from_remote);
721
722
254k
        DorisMetrics::instance()->query_scan_bytes_from_local->increment(
723
254k
                stats.file_cache_stats.bytes_read_from_local - _bytes_read_from_local);
724
254k
        DorisMetrics::instance()->query_scan_bytes_from_remote->increment(
725
254k
                stats.file_cache_stats.bytes_read_from_remote - _bytes_read_from_remote);
726
254k
    }
727
728
1.24M
    _tablet_reader->mutable_stats()->compressed_bytes_read = 0;
729
1.24M
    _tablet_reader->mutable_stats()->uncompressed_bytes_read = 0;
730
1.24M
    _tablet_reader->mutable_stats()->raw_rows_read = 0;
731
732
1.24M
    _bytes_read_from_local = _tablet_reader->stats().file_cache_stats.bytes_read_from_local;
733
1.24M
    _bytes_read_from_remote = _tablet_reader->stats().file_cache_stats.bytes_read_from_remote;
734
1.24M
}
735
736
1.20M
void OlapScanner::_collect_profile_before_close() {
737
    //  Please don't directly enable the profile here, we need to set QueryStatistics using the counter inside.
738
1.20M
    if (_has_updated_counter) {
739
0
        return;
740
0
    }
741
1.20M
    _has_updated_counter = true;
742
1.20M
    _tablet_reader->update_profile(_profile);
743
744
1.20M
    Scanner::_collect_profile_before_close();
745
746
    // Update counters for OlapScanner
747
    // Update counters from tablet reader's stats
748
1.20M
    auto& stats = _tablet_reader->stats();
749
1.20M
    auto* local_state = (OlapScanLocalState*)_local_state;
750
1.20M
    COUNTER_UPDATE(local_state->_io_timer, stats.io_ns);
751
1.20M
    COUNTER_UPDATE(local_state->_read_compressed_counter, stats.compressed_bytes_read);
752
1.20M
    COUNTER_UPDATE(local_state->_scan_bytes, stats.uncompressed_bytes_read);
753
1.20M
    COUNTER_UPDATE(local_state->_decompressor_timer, stats.decompress_ns);
754
1.20M
    COUNTER_UPDATE(local_state->_read_uncompressed_counter, stats.uncompressed_bytes_read);
755
1.20M
    COUNTER_UPDATE(local_state->_block_load_timer, stats.block_load_ns);
756
1.20M
    COUNTER_UPDATE(local_state->_block_load_counter, stats.blocks_load);
757
1.20M
    COUNTER_UPDATE(local_state->_block_fetch_timer, stats.block_fetch_ns);
758
1.20M
    COUNTER_UPDATE(local_state->_delete_bitmap_get_agg_timer, stats.delete_bitmap_get_agg_ns);
759
1.20M
    COUNTER_UPDATE(local_state->_scan_rows, stats.raw_rows_read);
760
1.20M
    COUNTER_UPDATE(local_state->_vec_cond_timer, stats.vec_cond_ns);
761
1.20M
    COUNTER_UPDATE(local_state->_short_cond_timer, stats.short_cond_ns);
762
1.20M
    COUNTER_UPDATE(local_state->_expr_filter_timer, stats.expr_filter_ns);
763
1.20M
    COUNTER_UPDATE(local_state->_block_init_timer, stats.block_init_ns);
764
1.20M
    COUNTER_UPDATE(local_state->_block_init_seek_timer, stats.block_init_seek_ns);
765
1.20M
    COUNTER_UPDATE(local_state->_block_init_seek_counter, stats.block_init_seek_num);
766
1.20M
    COUNTER_UPDATE(local_state->_segment_generate_row_range_by_keys_timer,
767
1.20M
                   stats.generate_row_ranges_by_keys_ns);
768
1.20M
    COUNTER_UPDATE(local_state->_segment_generate_row_range_by_column_conditions_timer,
769
1.20M
                   stats.generate_row_ranges_by_column_conditions_ns);
770
1.20M
    COUNTER_UPDATE(local_state->_segment_generate_row_range_by_bf_timer,
771
1.20M
                   stats.generate_row_ranges_by_bf_ns);
772
1.20M
    COUNTER_UPDATE(local_state->_collect_iterator_merge_next_timer,
773
1.20M
                   stats.collect_iterator_merge_next_timer);
774
1.20M
    COUNTER_UPDATE(local_state->_segment_generate_row_range_by_zonemap_timer,
775
1.20M
                   stats.generate_row_ranges_by_zonemap_ns);
776
1.20M
    COUNTER_UPDATE(local_state->_segment_generate_row_range_by_dict_timer,
777
1.20M
                   stats.generate_row_ranges_by_dict_ns);
778
1.20M
    COUNTER_UPDATE(local_state->_predicate_column_read_timer, stats.predicate_column_read_ns);
779
1.20M
    COUNTER_UPDATE(local_state->_non_predicate_column_read_timer, stats.non_predicate_read_ns);
780
1.20M
    COUNTER_UPDATE(local_state->_predicate_column_read_seek_timer,
781
1.20M
                   stats.predicate_column_read_seek_ns);
782
1.20M
    COUNTER_UPDATE(local_state->_predicate_column_read_seek_counter,
783
1.20M
                   stats.predicate_column_read_seek_num);
784
1.20M
    COUNTER_UPDATE(local_state->_lazy_read_timer, stats.lazy_read_ns);
785
1.20M
    COUNTER_UPDATE(local_state->_lazy_read_seek_timer, stats.block_lazy_read_seek_ns);
786
1.20M
    COUNTER_UPDATE(local_state->_lazy_read_seek_counter, stats.block_lazy_read_seek_num);
787
1.20M
    COUNTER_UPDATE(local_state->_output_col_timer, stats.output_col_ns);
788
1.20M
    COUNTER_UPDATE(local_state->_rows_vec_cond_filtered_counter, stats.rows_vec_cond_filtered);
789
1.20M
    COUNTER_UPDATE(local_state->_rows_short_circuit_cond_filtered_counter,
790
1.20M
                   stats.rows_short_circuit_cond_filtered);
791
1.20M
    COUNTER_UPDATE(local_state->_rows_expr_cond_filtered_counter, stats.rows_expr_cond_filtered);
792
1.20M
    COUNTER_UPDATE(local_state->_rows_vec_cond_input_counter, stats.vec_cond_input_rows);
793
1.20M
    COUNTER_UPDATE(local_state->_rows_short_circuit_cond_input_counter,
794
1.20M
                   stats.short_circuit_cond_input_rows);
795
1.20M
    COUNTER_UPDATE(local_state->_rows_expr_cond_input_counter, stats.expr_cond_input_rows);
796
1.20M
    COUNTER_UPDATE(local_state->_stats_filtered_counter, stats.rows_stats_filtered);
797
1.20M
    COUNTER_UPDATE(local_state->_stats_rp_filtered_counter, stats.rows_stats_rp_filtered);
798
1.20M
    COUNTER_UPDATE(local_state->_dict_filtered_counter, stats.segment_dict_filtered);
799
1.20M
    COUNTER_UPDATE(local_state->_bf_filtered_counter, stats.rows_bf_filtered);
800
1.20M
    COUNTER_UPDATE(local_state->_del_filtered_counter, stats.rows_del_filtered);
801
1.20M
    COUNTER_UPDATE(local_state->_del_filtered_counter, stats.rows_del_by_bitmap);
802
1.20M
    COUNTER_UPDATE(local_state->_del_filtered_counter, stats.rows_vec_del_cond_filtered);
803
1.20M
    COUNTER_UPDATE(local_state->_conditions_filtered_counter, stats.rows_conditions_filtered);
804
1.20M
    COUNTER_UPDATE(local_state->_key_range_filtered_counter, stats.rows_key_range_filtered);
805
1.20M
    COUNTER_UPDATE(local_state->_total_pages_num_counter, stats.total_pages_num);
806
1.20M
    COUNTER_UPDATE(local_state->_cached_pages_num_counter, stats.cached_pages_num);
807
1.20M
    COUNTER_UPDATE(local_state->_inverted_index_filter_counter, stats.rows_inverted_index_filtered);
808
1.20M
    COUNTER_UPDATE(local_state->_inverted_index_filter_timer, stats.inverted_index_filter_timer);
809
1.20M
    COUNTER_UPDATE(local_state->_inverted_index_query_cache_hit_counter,
810
1.20M
                   stats.inverted_index_query_cache_hit);
811
1.20M
    COUNTER_UPDATE(local_state->_inverted_index_query_cache_miss_counter,
812
1.20M
                   stats.inverted_index_query_cache_miss);
813
1.20M
    COUNTER_UPDATE(local_state->_inverted_index_query_timer, stats.inverted_index_query_timer);
814
1.20M
    COUNTER_UPDATE(local_state->_inverted_index_query_null_bitmap_timer,
815
1.20M
                   stats.inverted_index_query_null_bitmap_timer);
816
1.20M
    COUNTER_UPDATE(local_state->_inverted_index_query_bitmap_copy_timer,
817
1.20M
                   stats.inverted_index_query_bitmap_copy_timer);
818
1.20M
    COUNTER_UPDATE(local_state->_inverted_index_searcher_open_timer,
819
1.20M
                   stats.inverted_index_searcher_open_timer);
820
1.20M
    COUNTER_UPDATE(local_state->_inverted_index_searcher_search_timer,
821
1.20M
                   stats.inverted_index_searcher_search_timer);
822
1.20M
    COUNTER_UPDATE(local_state->_inverted_index_searcher_search_init_timer,
823
1.20M
                   stats.inverted_index_searcher_search_init_timer);
824
1.20M
    COUNTER_UPDATE(local_state->_inverted_index_searcher_search_exec_timer,
825
1.20M
                   stats.inverted_index_searcher_search_exec_timer);
826
1.20M
    COUNTER_UPDATE(local_state->_inverted_index_searcher_cache_hit_counter,
827
1.20M
                   stats.inverted_index_searcher_cache_hit);
828
1.20M
    COUNTER_UPDATE(local_state->_inverted_index_searcher_cache_miss_counter,
829
1.20M
                   stats.inverted_index_searcher_cache_miss);
830
1.20M
    COUNTER_UPDATE(local_state->_inverted_index_downgrade_count_counter,
831
1.20M
                   stats.inverted_index_downgrade_count);
832
1.20M
    COUNTER_UPDATE(local_state->_inverted_index_analyzer_timer,
833
1.20M
                   stats.inverted_index_analyzer_timer);
834
1.20M
    COUNTER_UPDATE(local_state->_inverted_index_lookup_timer, stats.inverted_index_lookup_timer);
835
1.20M
    COUNTER_UPDATE(local_state->_variant_scan_sparse_column_timer,
836
1.20M
                   stats.variant_scan_sparse_column_timer_ns);
837
1.20M
    COUNTER_UPDATE(local_state->_variant_scan_sparse_column_bytes,
838
1.20M
                   stats.variant_scan_sparse_column_bytes);
839
1.20M
    COUNTER_UPDATE(local_state->_variant_fill_path_from_sparse_column_timer,
840
1.20M
                   stats.variant_fill_path_from_sparse_column_timer_ns);
841
1.20M
    COUNTER_UPDATE(local_state->_variant_subtree_default_iter_count,
842
1.20M
                   stats.variant_subtree_default_iter_count);
843
1.20M
    COUNTER_UPDATE(local_state->_variant_subtree_leaf_iter_count,
844
1.20M
                   stats.variant_subtree_leaf_iter_count);
845
1.20M
    COUNTER_UPDATE(local_state->_variant_subtree_hierarchical_iter_count,
846
1.20M
                   stats.variant_subtree_hierarchical_iter_count);
847
1.20M
    COUNTER_UPDATE(local_state->_variant_subtree_sparse_iter_count,
848
1.20M
                   stats.variant_subtree_sparse_iter_count);
849
1.20M
    COUNTER_UPDATE(local_state->_variant_doc_value_column_iter_count,
850
1.20M
                   stats.variant_doc_value_column_iter_count);
851
852
1.20M
    if (stats.adaptive_batch_size_predict_max_rows > 0) {
853
924k
        local_state->_adaptive_batch_predict_min_rows_counter->set(
854
924k
                stats.adaptive_batch_size_predict_min_rows);
855
924k
        local_state->_adaptive_batch_predict_max_rows_counter->set(
856
924k
                stats.adaptive_batch_size_predict_max_rows);
857
924k
    }
858
859
1.20M
    InvertedIndexProfileReporter inverted_index_profile;
860
1.20M
    inverted_index_profile.update(local_state->_index_filter_profile.get(),
861
1.20M
                                  &stats.inverted_index_stats);
862
863
    // only cloud deploy mode will use file cache.
864
1.20M
    if (config::is_cloud_mode() && config::enable_file_cache) {
865
865k
        io::FileCacheProfileReporter cache_profile(local_state->_segment_profile.get());
866
865k
        cache_profile.update(&stats.file_cache_stats);
867
865k
        _state->get_query_ctx()->resource_ctx()->io_context()->update_bytes_write_into_cache(
868
865k
                stats.file_cache_stats.bytes_write_into_cache);
869
865k
    }
870
1.20M
    COUNTER_UPDATE(local_state->_output_index_result_column_timer,
871
1.20M
                   stats.output_index_result_column_timer);
872
1.20M
    COUNTER_UPDATE(local_state->_filtered_segment_counter, stats.filtered_segment_number);
873
1.20M
    COUNTER_UPDATE(local_state->_total_segment_counter, stats.total_segment_number);
874
1.20M
    COUNTER_UPDATE(local_state->_condition_cache_hit_counter, stats.condition_cache_hit_seg_nums);
875
1.20M
    COUNTER_UPDATE(local_state->_condition_cache_filtered_rows_counter,
876
1.20M
                   stats.condition_cache_filtered_rows);
877
878
1.20M
    COUNTER_UPDATE(local_state->_tablet_reader_init_timer, stats.tablet_reader_init_timer_ns);
879
1.20M
    COUNTER_UPDATE(local_state->_tablet_reader_capture_rs_readers_timer,
880
1.20M
                   stats.tablet_reader_capture_rs_readers_timer_ns);
881
1.20M
    COUNTER_UPDATE(local_state->_tablet_reader_init_return_columns_timer,
882
1.20M
                   stats.tablet_reader_init_return_columns_timer_ns);
883
1.20M
    COUNTER_UPDATE(local_state->_tablet_reader_init_keys_param_timer,
884
1.20M
                   stats.tablet_reader_init_keys_param_timer_ns);
885
1.20M
    COUNTER_UPDATE(local_state->_tablet_reader_init_orderby_keys_param_timer,
886
1.20M
                   stats.tablet_reader_init_orderby_keys_param_timer_ns);
887
1.20M
    COUNTER_UPDATE(local_state->_tablet_reader_init_conditions_param_timer,
888
1.20M
                   stats.tablet_reader_init_conditions_param_timer_ns);
889
1.20M
    COUNTER_UPDATE(local_state->_tablet_reader_init_delete_condition_param_timer,
890
1.20M
                   stats.tablet_reader_init_delete_condition_param_timer_ns);
891
1.20M
    COUNTER_UPDATE(local_state->_block_reader_vcollect_iter_init_timer,
892
1.20M
                   stats.block_reader_vcollect_iter_init_timer_ns);
893
1.20M
    COUNTER_UPDATE(local_state->_block_reader_rs_readers_init_timer,
894
1.20M
                   stats.block_reader_rs_readers_init_timer_ns);
895
1.20M
    COUNTER_UPDATE(local_state->_block_reader_build_heap_init_timer,
896
1.20M
                   stats.block_reader_build_heap_init_timer_ns);
897
898
1.20M
    COUNTER_UPDATE(local_state->_rowset_reader_get_segment_iterators_timer,
899
1.20M
                   stats.rowset_reader_get_segment_iterators_timer_ns);
900
1.20M
    COUNTER_UPDATE(local_state->_rowset_reader_create_iterators_timer,
901
1.20M
                   stats.rowset_reader_create_iterators_timer_ns);
902
1.20M
    COUNTER_UPDATE(local_state->_rowset_reader_init_iterators_timer,
903
1.20M
                   stats.rowset_reader_init_iterators_timer_ns);
904
1.20M
    COUNTER_UPDATE(local_state->_rowset_reader_load_segments_timer,
905
1.20M
                   stats.rowset_reader_load_segments_timer_ns);
906
907
1.20M
    COUNTER_UPDATE(local_state->_segment_iterator_init_timer, stats.segment_iterator_init_timer_ns);
908
1.20M
    COUNTER_UPDATE(local_state->_segment_iterator_init_return_column_iterators_timer,
909
1.20M
                   stats.segment_iterator_init_return_column_iterators_timer_ns);
910
1.20M
    COUNTER_UPDATE(local_state->_segment_iterator_init_index_iterators_timer,
911
1.20M
                   stats.segment_iterator_init_index_iterators_timer_ns);
912
1.20M
    COUNTER_UPDATE(local_state->_segment_iterator_init_segment_prefetchers_timer,
913
1.20M
                   stats.segment_iterator_init_segment_prefetchers_timer_ns);
914
915
1.20M
    COUNTER_UPDATE(local_state->_segment_create_column_readers_timer,
916
1.20M
                   stats.segment_create_column_readers_timer_ns);
917
1.20M
    COUNTER_UPDATE(local_state->_segment_load_index_timer, stats.segment_load_index_timer_ns);
918
919
    // Update metrics
920
1.20M
    DorisMetrics::instance()->query_scan_bytes->increment(
921
1.20M
            local_state->_read_uncompressed_counter->value());
922
1.20M
    DorisMetrics::instance()->query_scan_rows->increment(local_state->_scan_rows->value());
923
1.20M
    auto& tablet = _tablet_reader_params.tablet;
924
1.20M
    tablet->query_scan_bytes->increment(local_state->_read_uncompressed_counter->value());
925
1.20M
    tablet->query_scan_rows->increment(local_state->_scan_rows->value());
926
1.20M
    tablet->query_scan_count->increment(1);
927
928
1.20M
    COUNTER_UPDATE(local_state->_ann_range_search_filter_counter,
929
1.20M
                   stats.rows_ann_index_range_filtered);
930
1.20M
    COUNTER_UPDATE(local_state->_ann_topn_filter_counter, stats.rows_ann_index_topn_filtered);
931
1.20M
    COUNTER_UPDATE(local_state->_ann_index_load_costs, stats.ann_index_load_ns);
932
1.20M
    COUNTER_UPDATE(local_state->_ann_ivf_on_disk_load_costs, stats.ann_ivf_on_disk_load_ns);
933
1.20M
    COUNTER_UPDATE(local_state->_ann_ivf_on_disk_cache_hit_cnt,
934
1.20M
                   stats.ann_ivf_on_disk_cache_hit_cnt);
935
1.20M
    COUNTER_UPDATE(local_state->_ann_ivf_on_disk_cache_miss_cnt,
936
1.20M
                   stats.ann_ivf_on_disk_cache_miss_cnt);
937
1.20M
    COUNTER_UPDATE(local_state->_ann_range_search_costs, stats.ann_index_range_search_ns);
938
1.20M
    COUNTER_UPDATE(local_state->_ann_range_search_cnt, stats.ann_index_range_search_cnt);
939
1.20M
    COUNTER_UPDATE(local_state->_ann_range_engine_search_costs, stats.ann_range_engine_search_ns);
940
    // Engine prepare before search
941
1.20M
    COUNTER_UPDATE(local_state->_ann_range_pre_process_costs, stats.ann_range_pre_process_ns);
942
    // Post process parent: Doris result process + engine convert
943
1.20M
    COUNTER_UPDATE(local_state->_ann_range_post_process_costs,
944
1.20M
                   stats.ann_range_result_convert_ns + stats.ann_range_engine_convert_ns);
945
    // Engine convert (child under post-process)
946
1.20M
    COUNTER_UPDATE(local_state->_ann_range_engine_convert_costs, stats.ann_range_engine_convert_ns);
947
    // Doris-side result convert (child under post-process)
948
1.20M
    COUNTER_UPDATE(local_state->_ann_range_result_convert_costs, stats.ann_range_result_convert_ns);
949
950
1.20M
    COUNTER_UPDATE(local_state->_ann_topn_search_costs, stats.ann_topn_search_ns);
951
1.20M
    COUNTER_UPDATE(local_state->_ann_topn_search_cnt, stats.ann_index_topn_search_cnt);
952
1.20M
    COUNTER_UPDATE(local_state->_ann_cache_hit_cnt, stats.ann_index_cache_hits);
953
1.20M
    COUNTER_UPDATE(local_state->_ann_range_cache_hit_cnt, stats.ann_index_range_cache_hits);
954
955
    // Detailed ANN timers
956
    // ANN TopN timers with hierarchy
957
    // Engine search time (FAISS)
958
1.20M
    COUNTER_UPDATE(local_state->_ann_topn_engine_search_costs,
959
1.20M
                   stats.ann_index_topn_engine_search_ns);
960
    // Engine prepare time (allocations/buffer setup before search)
961
1.20M
    COUNTER_UPDATE(local_state->_ann_topn_pre_process_costs,
962
1.20M
                   stats.ann_index_topn_engine_prepare_ns);
963
    // Post process parent includes Doris result processing + engine convert
964
1.20M
    COUNTER_UPDATE(local_state->_ann_topn_post_process_costs,
965
1.20M
                   stats.ann_index_topn_result_process_ns + stats.ann_index_topn_engine_convert_ns);
966
    // Engine-side conversion time inside FAISS wrappers (child under post-process)
967
1.20M
    COUNTER_UPDATE(local_state->_ann_topn_engine_convert_costs,
968
1.20M
                   stats.ann_index_topn_engine_convert_ns);
969
970
    // Doris-side result convert costs (show separately as another child counter); use pure process time
971
1.20M
    COUNTER_UPDATE(local_state->_ann_topn_result_convert_costs,
972
1.20M
                   stats.ann_index_topn_result_process_ns);
973
974
1.20M
    COUNTER_UPDATE(local_state->_ann_fallback_brute_force_cnt, stats.ann_fall_back_brute_force_cnt);
975
976
    // Overhead counter removed; precise instrumentation is reported via engine_prepare above.
977
1.20M
}
978
979
#ifndef NDEBUG
980
1.46M
Status OlapScanner::_check_ann_cache_hit_debug_points(const OlapReaderStatistics& stats) {
981
1.46M
    DBUG_EXECUTE_IF("olap_scanner.ann_topn_cache_hits", {
982
1.46M
        auto expected_hits = dp->param<int32_t>("expected_hits", -1);
983
1.46M
        auto min_hits = dp->param<int32_t>("min_hits", -1);
984
1.46M
        if (expected_hits >= 0 && stats.ann_index_cache_hits != expected_hits) {
985
1.46M
            return Status::Error<ErrorCode::INTERNAL_ERROR>(
986
1.46M
                    "ann_index_cache_hits: {} not equal to expected: {}",
987
1.46M
                    stats.ann_index_cache_hits, expected_hits);
988
1.46M
        }
989
1.46M
        if (min_hits >= 0 && stats.ann_index_cache_hits < min_hits) {
990
1.46M
            return Status::Error<ErrorCode::INTERNAL_ERROR>(
991
1.46M
                    "ann_index_cache_hits: {} less than expected min: {}",
992
1.46M
                    stats.ann_index_cache_hits, min_hits);
993
1.46M
        }
994
1.46M
    })
995
1.46M
    DBUG_EXECUTE_IF("olap_scanner.ann_range_cache_hits", {
996
1.46M
        auto expected_hits = dp->param<int32_t>("expected_hits", -1);
997
1.46M
        auto min_hits = dp->param<int32_t>("min_hits", -1);
998
1.46M
        if (expected_hits >= 0 && stats.ann_index_range_cache_hits != expected_hits) {
999
1.46M
            return Status::Error<ErrorCode::INTERNAL_ERROR>(
1000
1.46M
                    "ann_index_range_cache_hits: {} not equal to expected: {}",
1001
1.46M
                    stats.ann_index_range_cache_hits, expected_hits);
1002
1.46M
        }
1003
1.46M
        if (min_hits >= 0 && stats.ann_index_range_cache_hits < min_hits) {
1004
1.46M
            return Status::Error<ErrorCode::INTERNAL_ERROR>(
1005
1.46M
                    "ann_index_range_cache_hits: {} less than expected min: {}",
1006
1.46M
                    stats.ann_index_range_cache_hits, min_hits);
1007
1.46M
        }
1008
1.46M
    })
1009
1.46M
    return Status::OK();
1010
1.46M
}
1011
#endif
1012
1013
#include "common/compile_check_avoid_end.h"
1014
} // namespace doris