Coverage Report

Created: 2026-05-09 10:24

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
#include "util/json/path_in_data.h"
63
64
namespace doris {
65
#include "common/compile_check_avoid_begin.h"
66
67
using ReadSource = TabletReadSource;
68
69
OlapScanner::OlapScanner(ScanLocalStateBase* parent, OlapScanner::Params&& params)
70
1.17M
        : Scanner(params.state, parent, params.limit, params.profile),
71
1.17M
          _key_ranges(std::move(params.key_ranges)),
72
1.17M
          _tablet_reader_params({.tablet = std::move(params.tablet),
73
1.17M
                                 .tablet_schema {},
74
1.17M
                                 .aggregation = params.aggregation,
75
1.17M
                                 .version = {0, params.version},
76
1.17M
                                 .start_key {},
77
1.17M
                                 .end_key {},
78
1.17M
                                 .predicates {},
79
1.17M
                                 .function_filters {},
80
1.17M
                                 .delete_predicates {},
81
1.17M
                                 .target_cast_type_for_variants {},
82
1.17M
                                 .all_access_paths {},
83
1.17M
                                 .predicate_access_paths {},
84
1.17M
                                 .rs_splits {},
85
1.17M
                                 .return_columns {},
86
1.17M
                                 .output_columns {},
87
1.17M
                                 .remaining_conjunct_roots {},
88
1.17M
                                 .common_expr_ctxs_push_down {},
89
1.17M
                                 .topn_filter_source_node_ids {},
90
1.17M
                                 .key_group_cluster_key_idxes {},
91
1.17M
                                 .virtual_column_exprs {},
92
1.17M
                                 .vir_cid_to_idx_in_block {},
93
1.17M
                                 .vir_col_idx_to_type {},
94
1.17M
                                 .score_runtime {},
95
1.17M
                                 .collection_statistics {},
96
1.17M
                                 .ann_topn_runtime {},
97
1.17M
                                 .condition_cache_digest = parent->get_condition_cache_digest()}) {
98
1.17M
    _tablet_reader_params.set_read_source(std::move(params.read_source),
99
1.17M
                                          _state->skip_delete_bitmap());
100
1.17M
    _has_prepared = false;
101
1.17M
    _vector_search_params = params.state->get_vector_search_params();
102
1.17M
}
103
104
static std::string read_columns_to_string(TabletSchemaSPtr tablet_schema,
105
2.78k
                                          const std::vector<uint32_t>& read_columns) {
106
    // avoid too long for one line,
107
    // it is hard to display in `show profile` stmt if one line is too long.
108
2.78k
    const int col_per_line = 10;
109
2.78k
    int i = 0;
110
2.78k
    std::string read_columns_string;
111
2.78k
    read_columns_string += "[";
112
13.5k
    for (auto it = read_columns.cbegin(); it != read_columns.cend(); it++) {
113
10.7k
        if (it != read_columns.cbegin()) {
114
8.01k
            read_columns_string += ", ";
115
8.01k
        }
116
10.7k
        read_columns_string += tablet_schema->columns().at(*it)->name();
117
10.7k
        if (i >= col_per_line) {
118
13
            read_columns_string += "\n";
119
13
            i = 0;
120
10.7k
        } else {
121
10.7k
            ++i;
122
10.7k
        }
123
10.7k
    }
124
2.78k
    read_columns_string += "]";
125
2.78k
    return read_columns_string;
126
2.78k
}
127
128
1.17M
Status OlapScanner::_prepare_impl() {
129
1.17M
    auto* local_state = static_cast<OlapScanLocalState*>(_local_state);
130
1.17M
    auto& tablet = _tablet_reader_params.tablet;
131
1.17M
    auto& tablet_schema = _tablet_reader_params.tablet_schema;
132
1.17M
    DBUG_EXECUTE_IF("CloudTablet.capture_rs_readers.return.e-230", {
133
1.17M
        LOG_WARNING("CloudTablet.capture_rs_readers.return e-230 init")
134
1.17M
                .tag("tablet_id", tablet->tablet_id());
135
1.17M
        return Status::Error<false>(-230, "injected error");
136
1.17M
    });
137
138
1.17M
    for (auto& ctx : local_state->_common_expr_ctxs_push_down) {
139
21.7k
        VExprContextSPtr context;
140
21.7k
        RETURN_IF_ERROR(ctx->clone(_state, context));
141
21.7k
        _common_expr_ctxs_push_down.emplace_back(context);
142
21.7k
        context->prepare_ann_range_search(_vector_search_params);
143
21.7k
    }
144
145
1.17M
    for (auto pair : local_state->_slot_id_to_virtual_column_expr) {
146
        // Scanner will be executed in a different thread, so we need to clone the context.
147
412
        VExprContextSPtr context;
148
412
        RETURN_IF_ERROR(pair.second->clone(_state, context));
149
412
        _slot_id_to_virtual_column_expr[pair.first] = context;
150
412
    }
151
152
1.17M
    _slot_id_to_index_in_block = local_state->_slot_id_to_index_in_block;
153
1.17M
    _slot_id_to_col_type = local_state->_slot_id_to_col_type;
154
1.17M
    _score_runtime = local_state->_score_runtime;
155
    // All scanners share the same ann_topn_runtime.
156
1.17M
    _ann_topn_runtime = local_state->_ann_topn_runtime;
157
158
    // set limit to reduce end of rowset and segment mem use
159
1.17M
    _tablet_reader = std::make_unique<BlockReader>();
160
    // batch size is passed down to segment iterator, use _state->batch_size()
161
    // instead of _parent->limit(), because if _parent->limit() is a very small
162
    // value (e.g. select a from t where a .. and b ... limit 1),
163
    // it will be very slow when reading data in segment iterator
164
1.17M
    _tablet_reader->set_batch_size(_state->batch_size());
165
    // Adaptive batch size: pass byte-budget settings to the storage reader.
166
    // The reader still uses batch_size() as the row ceiling.
167
1.17M
    _tablet_reader->set_preferred_block_size_bytes(_state->preferred_block_size_bytes());
168
1.17M
    {
169
1.17M
        TOlapScanNode& olap_scan_node = local_state->olap_scan_node();
170
171
        // Each scanner builds its own TabletSchema to avoid concurrent modification.
172
1.17M
        tablet_schema = std::make_shared<TabletSchema>();
173
1.17M
        tablet_schema->copy_from(*tablet->tablet_schema());
174
1.17M
        if (olap_scan_node.__isset.columns_desc && !olap_scan_node.columns_desc.empty() &&
175
1.17M
            olap_scan_node.columns_desc[0].col_unique_id >= 0) {
176
1.17M
            tablet_schema->clear_columns();
177
18.9M
            for (const auto& column_desc : olap_scan_node.columns_desc) {
178
18.9M
                tablet_schema->append_column(TabletColumn(column_desc));
179
18.9M
            }
180
1.17M
            if (olap_scan_node.__isset.schema_version) {
181
1.17M
                tablet_schema->set_schema_version(olap_scan_node.schema_version);
182
1.17M
            }
183
1.17M
        }
184
1.17M
        if (olap_scan_node.__isset.indexes_desc) {
185
1.17M
            tablet_schema->update_indexes_from_thrift(olap_scan_node.indexes_desc);
186
1.17M
        }
187
188
1.17M
        if (_tablet_reader_params.rs_splits.empty()) {
189
            // Non-pipeline mode, Tablet : Scanner = 1 : 1
190
            // acquire tablet rowset readers at the beginning of the scan node
191
            // to prevent this case: when there are lots of olap scanners to run for example 10000
192
            // the rowsets maybe compacted when the last olap scanner starts
193
0
            ReadSource read_source;
194
195
0
            if (config::is_cloud_mode()) {
196
                // FIXME(plat1ko): Avoid pointer cast
197
0
                ExecEnv::GetInstance()->storage_engine().to_cloud().tablet_hotspot().count(*tablet);
198
0
            }
199
200
0
            auto maybe_read_source = tablet->capture_read_source(
201
0
                    _tablet_reader_params.version,
202
0
                    {
203
0
                            .skip_missing_versions = _state->skip_missing_version(),
204
0
                            .enable_fetch_rowsets_from_peers =
205
0
                                    config::enable_fetch_rowsets_from_peer_replicas,
206
0
                            .enable_prefer_cached_rowset =
207
0
                                    config::is_cloud_mode() ? _state->enable_prefer_cached_rowset()
208
0
                                                            : false,
209
0
                            .query_freshness_tolerance_ms =
210
0
                                    config::is_cloud_mode() ? _state->query_freshness_tolerance_ms()
211
0
                                                            : -1,
212
0
                    });
213
0
            if (!maybe_read_source) {
214
0
                LOG(WARNING) << "fail to init reader. res=" << maybe_read_source.error();
215
0
                return maybe_read_source.error();
216
0
            }
217
218
0
            read_source = std::move(maybe_read_source.value());
219
220
0
            if (config::enable_mow_verbose_log && tablet->enable_unique_key_merge_on_write()) {
221
0
                LOG_INFO("finish capture_rs_readers for tablet={}, query_id={}",
222
0
                         tablet->tablet_id(), print_id(_state->query_id()));
223
0
            }
224
225
0
            if (!_state->skip_delete_predicate()) {
226
0
                read_source.fill_delete_predicates();
227
0
            }
228
0
            _tablet_reader_params.set_read_source(std::move(read_source));
229
0
        }
230
231
        // Initialize tablet_reader_params
232
1.17M
        RETURN_IF_ERROR(_init_tablet_reader_params(
233
1.17M
                local_state->_parent->cast<OlapScanOperatorX>()._slot_id_to_slot_desc, _key_ranges,
234
1.17M
                local_state->_slot_id_to_predicates, local_state->_push_down_functions));
235
1.17M
    }
236
237
    // add read columns in profile
238
1.17M
    if (_state->enable_profile()) {
239
2.78k
        _profile->add_info_string("ReadColumns",
240
2.78k
                                  read_columns_to_string(tablet_schema, _return_columns));
241
2.78k
    }
242
243
1.17M
    if (_tablet_reader_params.score_runtime) {
244
14
        SCOPED_TIMER(local_state->_statistics_collect_timer);
245
14
        _tablet_reader_params.collection_statistics = std::make_shared<CollectionStatistics>();
246
247
14
        io::IOContext io_ctx {
248
14
                .reader_type = ReaderType::READER_QUERY,
249
14
                .expiration_time = tablet->ttl_seconds(),
250
14
                .query_id = &_state->query_id(),
251
14
                .file_cache_stats = &_tablet_reader->mutable_stats()->file_cache_stats,
252
14
                .is_inverted_index = true,
253
14
        };
254
255
14
        RETURN_IF_ERROR(_tablet_reader_params.collection_statistics->collect(
256
14
                _state, _tablet_reader_params.rs_splits, _tablet_reader_params.tablet_schema,
257
14
                _tablet_reader_params.common_expr_ctxs_push_down, &io_ctx));
258
14
    }
259
260
1.17M
    _has_prepared = true;
261
1.17M
    return Status::OK();
262
1.17M
}
263
264
1.17M
Status OlapScanner::_open_impl(RuntimeState* state) {
265
1.17M
    RETURN_IF_ERROR(Scanner::_open_impl(state));
266
1.17M
    SCOPED_TIMER(_local_state->cast<OlapScanLocalState>()._reader_init_timer);
267
268
1.17M
    auto res = _tablet_reader->init(_tablet_reader_params);
269
1.17M
    if (!res.ok()) {
270
43
        res.append("failed to initialize storage reader. tablet=" +
271
43
                   std::to_string(_tablet_reader_params.tablet->tablet_id()) +
272
43
                   ", backend=" + BackendOptions::get_localhost());
273
43
        return res;
274
43
    }
275
276
    // Do not hold rs_splits any more to release memory.
277
1.17M
    _tablet_reader_params.rs_splits.clear();
278
279
1.17M
    return Status::OK();
280
1.17M
}
281
282
// it will be called under tablet read lock because capture rs readers need
283
Status OlapScanner::_init_tablet_reader_params(
284
        const phmap::flat_hash_map<int, SlotDescriptor*>& slot_id_to_slot_desc,
285
        const std::vector<OlapScanRange*>& key_ranges,
286
        const phmap::flat_hash_map<int, std::vector<std::shared_ptr<ColumnPredicate>>>&
287
                slot_to_predicates,
288
1.17M
        const std::vector<FunctionFilter>& function_filters) {
289
    // if the table with rowset [0-x] or [0-1] [2-y], and [0-1] is empty
290
1.17M
    const bool single_version = _tablet_reader_params.has_single_version();
291
292
1.17M
    auto* olap_local_state = static_cast<OlapScanLocalState*>(_local_state);
293
1.17M
    bool read_mor_as_dup = olap_local_state->olap_scan_node().__isset.read_mor_as_dup &&
294
1.17M
                           olap_local_state->olap_scan_node().read_mor_as_dup;
295
1.17M
    if (_state->skip_storage_engine_merge() || read_mor_as_dup) {
296
42
        _tablet_reader_params.direct_mode = true;
297
42
        _tablet_reader_params.aggregation = true;
298
1.17M
    } else {
299
1.17M
        auto push_down_agg_type = _local_state->get_push_down_agg_type();
300
1.17M
        _tablet_reader_params.direct_mode = _tablet_reader_params.aggregation || single_version ||
301
1.17M
                                            (push_down_agg_type != TPushAggOp::NONE &&
302
11.1k
                                             push_down_agg_type != TPushAggOp::COUNT_ON_INDEX);
303
1.17M
    }
304
305
1.17M
    RETURN_IF_ERROR(_init_variant_columns());
306
1.17M
    RETURN_IF_ERROR(_init_return_columns());
307
308
1.17M
    _tablet_reader_params.reader_type = ReaderType::READER_QUERY;
309
1.17M
    _tablet_reader_params.push_down_agg_type_opt = _local_state->get_push_down_agg_type();
310
311
    // TODO: If a new runtime filter arrives after `_conjuncts` move to `_common_expr_ctxs_push_down`,
312
1.17M
    if (_common_expr_ctxs_push_down.empty()) {
313
1.15M
        for (auto& conjunct : _conjuncts) {
314
3.72k
            _tablet_reader_params.remaining_conjunct_roots.emplace_back(conjunct->root());
315
3.72k
        }
316
1.15M
    } else {
317
21.6k
        for (auto& ctx : _common_expr_ctxs_push_down) {
318
21.6k
            _tablet_reader_params.remaining_conjunct_roots.emplace_back(ctx->root());
319
21.6k
        }
320
18.4k
    }
321
322
1.17M
    _tablet_reader_params.common_expr_ctxs_push_down = _common_expr_ctxs_push_down;
323
1.17M
    _tablet_reader_params.virtual_column_exprs = _virtual_column_exprs;
324
1.17M
    _tablet_reader_params.vir_cid_to_idx_in_block = _vir_cid_to_idx_in_block;
325
1.17M
    _tablet_reader_params.vir_col_idx_to_type = _vir_col_idx_to_type;
326
1.17M
    _tablet_reader_params.score_runtime = _score_runtime;
327
1.17M
    _tablet_reader_params.output_columns = ((OlapScanLocalState*)_local_state)->_output_column_ids;
328
1.17M
    _tablet_reader_params.ann_topn_runtime = _ann_topn_runtime;
329
1.17M
    for (const auto& ele : ((OlapScanLocalState*)_local_state)->_cast_types_for_variants) {
330
1.67k
        _tablet_reader_params.target_cast_type_for_variants[ele.first] = ele.second;
331
1.67k
    };
332
1.17M
    auto& tablet_schema = _tablet_reader_params.tablet_schema;
333
9.77M
    for (auto& predicates : slot_to_predicates) {
334
9.77M
        const int sid = predicates.first;
335
9.77M
        DCHECK(slot_id_to_slot_desc.contains(sid));
336
9.77M
        int32_t index =
337
9.77M
                tablet_schema->field_index(slot_id_to_slot_desc.find(sid)->second->col_name());
338
9.77M
        if (index < 0) {
339
0
            throw Exception(
340
0
                    Status::InternalError("Column {} not found in tablet schema",
341
0
                                          slot_id_to_slot_desc.find(sid)->second->col_name()));
342
0
        }
343
9.77M
        for (auto& predicate : predicates.second) {
344
958k
            _tablet_reader_params.predicates.push_back(predicate->clone(index));
345
958k
        }
346
9.77M
    }
347
348
1.17M
    std::copy(function_filters.cbegin(), function_filters.cend(),
349
1.17M
              std::inserter(_tablet_reader_params.function_filters,
350
1.17M
                            _tablet_reader_params.function_filters.begin()));
351
352
    // Merge the columns in delete predicate that not in latest schema in to current tablet schema
353
1.17M
    for (auto& del_pred : _tablet_reader_params.delete_predicates) {
354
7.30k
        tablet_schema->merge_dropped_columns(*del_pred->tablet_schema());
355
7.30k
    }
356
357
    // Push key ranges to the tablet reader.
358
    // Skip the "full scan" placeholder (has_lower_bound == false) — when no key
359
    // predicates exist, start_key/end_key remain empty and the reader does a full scan.
360
1.84M
    for (auto* key_range : key_ranges) {
361
1.84M
        if (!key_range->has_lower_bound) {
362
151k
            continue;
363
151k
        }
364
365
1.68M
        _tablet_reader_params.start_key_include = key_range->begin_include;
366
1.68M
        _tablet_reader_params.end_key_include = key_range->end_include;
367
368
1.68M
        _tablet_reader_params.start_key.push_back(key_range->begin_scan_range);
369
1.68M
        _tablet_reader_params.end_key.push_back(key_range->end_scan_range);
370
1.68M
    }
371
372
1.17M
    _tablet_reader_params.profile = _local_state->custom_profile();
373
1.17M
    _tablet_reader_params.runtime_state = _state;
374
375
1.17M
    _tablet_reader_params.origin_return_columns = &_return_columns;
376
1.17M
    _tablet_reader_params.tablet_columns_convert_to_null_set = &_tablet_columns_convert_to_null_set;
377
378
1.17M
    if (_tablet_reader_params.direct_mode) {
379
1.16M
        _tablet_reader_params.return_columns = _return_columns;
380
1.16M
    } else {
381
        // we need to fetch all key columns to do the right aggregation on storage engine side.
382
43.2k
        for (size_t i = 0; i < tablet_schema->num_key_columns(); ++i) {
383
30.1k
            _tablet_reader_params.return_columns.push_back(i);
384
30.1k
        }
385
53.8k
        for (auto index : _return_columns) {
386
53.8k
            if (tablet_schema->column(index).is_key()) {
387
21.0k
                continue;
388
21.0k
            }
389
32.7k
            _tablet_reader_params.return_columns.push_back(index);
390
32.7k
        }
391
        // expand the sequence column
392
13.1k
        if (tablet_schema->has_sequence_col() || tablet_schema->has_seq_map()) {
393
39
            bool has_replace_col = false;
394
88
            for (auto col : _return_columns) {
395
88
                if (tablet_schema->column(col).aggregation() ==
396
88
                    FieldAggregationMethod::OLAP_FIELD_AGGREGATION_REPLACE) {
397
38
                    has_replace_col = true;
398
38
                    break;
399
38
                }
400
88
            }
401
39
            if (auto sequence_col_idx = tablet_schema->sequence_col_idx();
402
39
                has_replace_col && tablet_schema->has_sequence_col() &&
403
39
                std::find(_return_columns.begin(), _return_columns.end(), sequence_col_idx) ==
404
26
                        _return_columns.end()) {
405
15
                _tablet_reader_params.return_columns.push_back(sequence_col_idx);
406
15
            }
407
39
            if (has_replace_col) {
408
38
                const auto& val_to_seq = tablet_schema->value_col_idx_to_seq_col_idx();
409
38
                std::set<uint32_t> return_seq_columns;
410
411
236
                for (auto col : _tablet_reader_params.return_columns) {
412
                    // we need to add the necessary sequence column in _return_columns, and
413
                    // Avoid adding the same seq column twice
414
236
                    const auto val_iter = val_to_seq.find(col);
415
236
                    if (val_iter != val_to_seq.end()) {
416
42
                        auto seq = val_iter->second;
417
42
                        if (std::find(_tablet_reader_params.return_columns.begin(),
418
42
                                      _tablet_reader_params.return_columns.end(),
419
42
                                      seq) == _tablet_reader_params.return_columns.end()) {
420
4
                            return_seq_columns.insert(seq);
421
4
                        }
422
42
                    }
423
236
                }
424
38
                _tablet_reader_params.return_columns.insert(
425
38
                        std::end(_tablet_reader_params.return_columns),
426
38
                        std::begin(return_seq_columns), std::end(return_seq_columns));
427
38
            }
428
39
        }
429
13.1k
    }
430
431
1.17M
    _tablet_reader_params.use_page_cache = _state->enable_page_cache();
432
433
1.17M
    DBUG_EXECUTE_IF("NewOlapScanner::_init_tablet_reader_params.block", DBUG_BLOCK);
434
435
1.17M
    if (!_state->skip_storage_engine_merge()) {
436
1.17M
        auto* olap_scan_local_state = (OlapScanLocalState*)_local_state;
437
1.17M
        TOlapScanNode& olap_scan_node = olap_scan_local_state->olap_scan_node();
438
439
        // Set MOR value predicate pushdown flag
440
1.17M
        if (olap_scan_node.__isset.enable_mor_value_predicate_pushdown &&
441
1.17M
            olap_scan_node.enable_mor_value_predicate_pushdown) {
442
25
            _tablet_reader_params.enable_mor_value_predicate_pushdown = true;
443
25
        }
444
445
1.17M
        const bool has_key_topn =
446
1.17M
                olap_scan_node.__isset.sort_info && !olap_scan_node.sort_info.is_asc_order.empty();
447
1.17M
        if (has_key_topn) {
448
2.14k
            _limit = _local_state->limit_per_scanner();
449
2.14k
        }
450
451
1.17M
        const bool no_runtime_filters = _total_rf_num == 0;
452
1.17M
        const bool segment_limit_enabled = _state->enable_segment_limit_pushdown();
453
1.17M
        const bool storage_no_merge = olap_scan_local_state->_storage_no_merge();
454
455
1.17M
        if (_limit > 0 && no_runtime_filters && segment_limit_enabled && storage_no_merge) {
456
3.91k
            for (const auto& conjunct : _conjuncts) {
457
0
                DORIS_CHECK(!olap_scan_local_state->_check_expr_storage_filter(
458
0
                        conjunct->root(), OlapScanLocalState::ExprStorageFilterCheckMode::
459
0
                                                  HAS_SEGMENT_EVALUABLE_EXPR));
460
0
            }
461
3.91k
        }
462
463
        // Segment LIMIT has only two legal states: completely disabled, or enabled after every
464
        // row-filtering conjunct has become a storage predicate or SegmentIterator common expr.
465
1.17M
        const bool can_push_down_segment_limit = _limit > 0 && no_runtime_filters &&
466
1.17M
                                                 _conjuncts.empty() && segment_limit_enabled &&
467
1.17M
                                                 storage_no_merge;
468
1.17M
        if (can_push_down_segment_limit) {
469
3.91k
            if (has_key_topn) {
470
2.04k
                _tablet_reader_params.read_orderby_key = true;
471
2.04k
                if (!olap_scan_node.sort_info.is_asc_order[0]) {
472
179
                    _tablet_reader_params.read_orderby_key_reverse = true;
473
179
                }
474
2.04k
                _tablet_reader_params.read_orderby_key_num_prefix_columns =
475
2.04k
                        olap_scan_node.sort_info.is_asc_order.size();
476
2.04k
                _tablet_reader_params.read_orderby_key_limit = _limit;
477
2.04k
            } else {
478
1.87k
                _tablet_reader_params.general_read_limit = _limit;
479
1.87k
            }
480
3.91k
        }
481
482
1.17M
        if (_tablet_reader_params.read_orderby_key_limit > 0 ||
483
1.17M
            _tablet_reader_params.general_read_limit > 0) {
484
3.91k
            DORIS_CHECK(can_push_down_segment_limit);
485
3.91k
            DORIS_CHECK(_conjuncts.empty());
486
3.91k
        }
487
488
        // A key TopN scan cannot share the plain LIMIT early-stop counter. If
489
        // storage TopN is pushed down, each scanner must produce its full local
490
        // candidates. If it is not pushed down for any reason, the upper TopN
491
        // still needs all rows from the scan.
492
1.17M
        if (has_key_topn) {
493
2.14k
            _shared_scan_limit = nullptr;
494
2.14k
            if (_tablet_reader_params.read_orderby_key_limit == 0) {
495
97
                _limit = -1;
496
97
            }
497
2.14k
        }
498
        // Note: _shared_scan_limit is intentionally not pushed into the
499
        // storage layer. SegmentIterator's _process_eof() is irreversible,
500
        // so a concurrently-decremented atomic could reach 0 while a segment
501
        // still has data needed by other scanners.
502
503
        // set push down topn filter
504
1.17M
        _tablet_reader_params.topn_filter_source_node_ids =
505
1.17M
                olap_scan_local_state->get_topn_filter_source_node_ids(_state, true);
506
1.17M
        if (!_tablet_reader_params.topn_filter_source_node_ids.empty()) {
507
4.94k
            _tablet_reader_params.topn_filter_target_node_id =
508
4.94k
                    olap_scan_local_state->parent()->node_id();
509
4.94k
        }
510
1.17M
    }
511
512
    // If this is a Two-Phase read query, and we need to delay the release of Rowset
513
    // by rowset->update_delayed_expired_timestamp().This could expand the lifespan of Rowset
514
1.17M
    if (tablet_schema->field_index(BeConsts::ROWID_COL) >= 0) {
515
84
        constexpr static int delayed_s = 60;
516
170
        for (auto rs_reader : _tablet_reader_params.rs_splits) {
517
170
            uint64_t delayed_expired_timestamp =
518
170
                    UnixSeconds() + _tablet_reader_params.runtime_state->execution_timeout() +
519
170
                    delayed_s;
520
170
            rs_reader.rs_reader->rowset()->update_delayed_expired_timestamp(
521
170
                    delayed_expired_timestamp);
522
170
            ExecEnv::GetInstance()->storage_engine().add_quering_rowset(
523
170
                    rs_reader.rs_reader->rowset());
524
170
        }
525
84
    }
526
527
1.17M
    if (tablet_schema->has_global_row_id()) {
528
5.82k
        auto& id_file_map = _state->get_id_file_map();
529
11.3k
        for (auto rs_reader : _tablet_reader_params.rs_splits) {
530
11.3k
            id_file_map->add_temp_rowset(rs_reader.rs_reader->rowset());
531
11.3k
        }
532
5.82k
    }
533
534
1.17M
    return Status::OK();
535
1.17M
}
536
537
1.17M
Status OlapScanner::_init_variant_columns() {
538
1.17M
    auto& tablet_schema = _tablet_reader_params.tablet_schema;
539
1.17M
    if (tablet_schema->num_variant_columns() == 0) {
540
1.16M
        return Status::OK();
541
1.16M
    }
542
    // Parent column has path info to distinction from each other
543
15.0k
    for (auto* slot : _output_tuple_desc->slots()) {
544
15.0k
        if (slot->type()->get_primitive_type() == PrimitiveType::TYPE_VARIANT) {
545
            // Such columns are not exist in frontend schema info, so we need to
546
            // add them into tablet_schema for later column indexing.
547
7.80k
            const auto& dt_variant =
548
7.80k
                    assert_cast<const DataTypeVariant&>(*remove_nullable(slot->type()));
549
7.80k
            TabletColumn subcol = TabletColumn::create_materialized_variant_column(
550
7.80k
                    tablet_schema->column_by_uid(slot->col_unique_id()).name_lower_case(),
551
7.80k
                    slot->column_paths(), slot->col_unique_id(),
552
7.80k
                    dt_variant.variant_max_subcolumns_count(), dt_variant.enable_doc_mode());
553
7.80k
            if (tablet_schema->field_index(*subcol.path_info_ptr()) < 0) {
554
5.87k
                tablet_schema->append_column(subcol, TabletSchema::ColumnType::VARIANT);
555
5.87k
            }
556
7.80k
        }
557
15.0k
    }
558
5.05k
    variant_util::inherit_column_attributes(tablet_schema);
559
5.05k
    return Status::OK();
560
1.17M
}
561
562
1.17M
Status OlapScanner::_init_return_columns() {
563
12.8M
    for (auto* slot : _output_tuple_desc->slots()) {
564
        // variant column using path to index a column
565
12.8M
        int32_t index = 0;
566
12.8M
        auto& tablet_schema = _tablet_reader_params.tablet_schema;
567
12.8M
        if (slot->type()->get_primitive_type() == PrimitiveType::TYPE_VARIANT) {
568
7.80k
            index = tablet_schema->field_index(PathInData(
569
7.80k
                    tablet_schema->column_by_uid(slot->col_unique_id()).name_lower_case(),
570
7.80k
                    slot->column_paths()));
571
12.8M
        } else {
572
12.8M
            index = slot->col_unique_id() >= 0 ? tablet_schema->field_index(slot->col_unique_id())
573
18.4E
                                               : tablet_schema->field_index(slot->col_name());
574
12.8M
        }
575
576
12.8M
        if (index < 0) {
577
0
            return Status::InternalError(
578
0
                    "field name is invalid. field={}, field_name_to_index={}, col_unique_id={}",
579
0
                    slot->col_name(), tablet_schema->get_all_field_names(), slot->col_unique_id());
580
0
        }
581
582
12.8M
        if (slot->get_virtual_column_expr()) {
583
412
            ColumnId virtual_column_cid = index;
584
412
            _virtual_column_exprs[virtual_column_cid] = _slot_id_to_virtual_column_expr[slot->id()];
585
412
            size_t idx_in_block = _slot_id_to_index_in_block[slot->id()];
586
412
            _vir_cid_to_idx_in_block[virtual_column_cid] = idx_in_block;
587
412
            _vir_col_idx_to_type[idx_in_block] = _slot_id_to_col_type[slot->id()];
588
589
412
            VLOG_DEBUG << fmt::format(
590
1
                    "Virtual column, slot id: {}, cid {}, column index: {}, type: {}", slot->id(),
591
1
                    virtual_column_cid, _vir_cid_to_idx_in_block[virtual_column_cid],
592
1
                    _vir_col_idx_to_type[idx_in_block]->get_name());
593
412
        }
594
595
12.8M
        const auto& column = tablet_schema->column(index);
596
12.8M
        int32_t unique_id =
597
12.8M
                column.unique_id() >= 0 ? column.unique_id() : column.parent_unique_id();
598
12.8M
        if (!slot->all_access_paths().empty()) {
599
73.5k
            _tablet_reader_params.all_access_paths.insert({unique_id, slot->all_access_paths()});
600
73.5k
        }
601
602
12.8M
        if (!slot->predicate_access_paths().empty()) {
603
5.71k
            _tablet_reader_params.predicate_access_paths.insert(
604
5.71k
                    {unique_id, slot->predicate_access_paths()});
605
5.71k
        }
606
607
12.8M
        if ((slot->type()->get_primitive_type() == PrimitiveType::TYPE_STRUCT ||
608
12.8M
             slot->type()->get_primitive_type() == PrimitiveType::TYPE_MAP ||
609
12.8M
             slot->type()->get_primitive_type() == PrimitiveType::TYPE_ARRAY) &&
610
12.8M
            !slot->all_access_paths().empty()) {
611
67.5k
            tablet_schema->add_pruned_columns_data_type(column.unique_id(), slot->type());
612
67.5k
        }
613
614
12.8M
        _return_columns.push_back(index);
615
12.8M
        if (slot->is_nullable() && !tablet_schema->column(index).is_nullable()) {
616
0
            _tablet_columns_convert_to_null_set.emplace(index);
617
12.8M
        } else if (!slot->is_nullable() && tablet_schema->column(index).is_nullable()) {
618
0
            return Status::Error<ErrorCode::INVALID_SCHEMA>(
619
0
                    "slot(id: {}, name: {})'s nullable does not match "
620
0
                    "column(tablet id: {}, index: {}, name: {}) ",
621
0
                    slot->id(), slot->col_name(), tablet_schema->table_id(), index,
622
0
                    tablet_schema->column(index).name());
623
0
        }
624
12.8M
    }
625
626
1.17M
    if (_return_columns.empty()) {
627
0
        return Status::InternalError("failed to build storage scanner, no materialized slot!");
628
0
    }
629
630
1.17M
    return Status::OK();
631
1.17M
}
632
633
1.22M
doris::TabletStorageType OlapScanner::get_storage_type() {
634
1.22M
    if (config::is_cloud_mode()) {
635
        // we don't have cold storage in cloud mode, all storage is treated as local
636
882k
        return doris::TabletStorageType::STORAGE_TYPE_LOCAL;
637
882k
    }
638
341k
    int local_reader = 0;
639
550k
    for (const auto& reader : _tablet_reader_params.rs_splits) {
640
550k
        local_reader += reader.rs_reader->rowset()->is_local();
641
550k
    }
642
341k
    int total_reader = _tablet_reader_params.rs_splits.size();
643
644
341k
    if (local_reader == total_reader) {
645
341k
        return doris::TabletStorageType::STORAGE_TYPE_LOCAL;
646
341k
    } else if (local_reader == 0) {
647
0
        return doris::TabletStorageType::STORAGE_TYPE_REMOTE;
648
0
    }
649
0
    return doris::TabletStorageType::STORAGE_TYPE_REMOTE_AND_LOCAL;
650
341k
}
651
652
1.43M
Status OlapScanner::_get_block_impl(RuntimeState* state, Block* block, bool* eof) {
653
    // Read one block from block reader
654
    // ATTN: Here we need to let the _get_block_impl method guarantee the semantics of the interface,
655
    // that is, eof can be set to true only when the returned block is empty.
656
1.43M
    RETURN_IF_ERROR(_tablet_reader->next_block_with_aggregation(block, eof));
657
1.43M
    if (block->rows() > 0) {
658
266k
        _tablet_reader_params.tablet->read_block_count.fetch_add(1, std::memory_order_relaxed);
659
266k
        *eof = false;
660
266k
    }
661
1.43M
    return Status::OK();
662
1.43M
}
663
664
1.18M
Status OlapScanner::close(RuntimeState* state) {
665
1.18M
    if (!_try_close()) {
666
150
        return Status::OK();
667
150
    }
668
1.18M
    RETURN_IF_ERROR(Scanner::close(state));
669
1.18M
    return Status::OK();
670
1.18M
}
671
672
1.22M
void OlapScanner::update_realtime_counters() {
673
1.22M
    if (!_has_prepared) {
674
        // Counter update need prepare successfully, or it maybe core. For example, olap scanner
675
        // will open tablet reader during prepare, if not prepare successfully, tablet reader == nullptr.
676
0
        return;
677
0
    }
678
1.22M
    OlapScanLocalState* local_state = static_cast<OlapScanLocalState*>(_local_state);
679
1.22M
    const OlapReaderStatistics& stats = _tablet_reader->stats();
680
1.22M
    COUNTER_UPDATE(local_state->_read_compressed_counter, stats.compressed_bytes_read);
681
1.22M
    COUNTER_UPDATE(local_state->_read_uncompressed_counter, stats.uncompressed_bytes_read);
682
1.22M
    COUNTER_UPDATE(local_state->_scan_bytes, stats.uncompressed_bytes_read);
683
1.22M
    COUNTER_UPDATE(local_state->_scan_rows, stats.raw_rows_read);
684
685
    // Make sure the scan bytes and scan rows counter in audit log is the same as the counter in
686
    // doris metrics.
687
    // ScanBytes is the uncompressed bytes read from local + remote
688
    // bytes_read_from_local is the compressed bytes read from local
689
    // bytes_read_from_remote is the compressed bytes read from remote
690
    // scan bytes > bytes_read_from_local + bytes_read_from_remote
691
1.22M
    _state->get_query_ctx()->resource_ctx()->io_context()->update_scan_rows(stats.raw_rows_read);
692
1.22M
    _state->get_query_ctx()->resource_ctx()->io_context()->update_scan_bytes(
693
1.22M
            stats.uncompressed_bytes_read);
694
695
    // In case of no cache, we still need to update the IO stats. uncompressed bytes read == local + remote
696
1.22M
    if (stats.file_cache_stats.bytes_read_from_local == 0 &&
697
1.22M
        stats.file_cache_stats.bytes_read_from_remote == 0) {
698
968k
        _state->get_query_ctx()->resource_ctx()->io_context()->update_scan_bytes_from_local_storage(
699
968k
                stats.compressed_bytes_read);
700
968k
        DorisMetrics::instance()->query_scan_bytes_from_local->increment(
701
968k
                stats.compressed_bytes_read);
702
968k
    } else {
703
253k
        _state->get_query_ctx()->resource_ctx()->io_context()->update_scan_bytes_from_local_storage(
704
253k
                stats.file_cache_stats.bytes_read_from_local - _bytes_read_from_local);
705
253k
        _state->get_query_ctx()
706
253k
                ->resource_ctx()
707
253k
                ->io_context()
708
253k
                ->update_scan_bytes_from_remote_storage(
709
253k
                        stats.file_cache_stats.bytes_read_from_remote - _bytes_read_from_remote);
710
711
253k
        DorisMetrics::instance()->query_scan_bytes_from_local->increment(
712
253k
                stats.file_cache_stats.bytes_read_from_local - _bytes_read_from_local);
713
253k
        DorisMetrics::instance()->query_scan_bytes_from_remote->increment(
714
253k
                stats.file_cache_stats.bytes_read_from_remote - _bytes_read_from_remote);
715
253k
    }
716
717
1.22M
    _tablet_reader->mutable_stats()->compressed_bytes_read = 0;
718
1.22M
    _tablet_reader->mutable_stats()->uncompressed_bytes_read = 0;
719
1.22M
    _tablet_reader->mutable_stats()->raw_rows_read = 0;
720
721
1.22M
    _bytes_read_from_local = _tablet_reader->stats().file_cache_stats.bytes_read_from_local;
722
1.22M
    _bytes_read_from_remote = _tablet_reader->stats().file_cache_stats.bytes_read_from_remote;
723
1.22M
}
724
725
1.17M
void OlapScanner::_collect_profile_before_close() {
726
    //  Please don't directly enable the profile here, we need to set QueryStatistics using the counter inside.
727
1.17M
    if (_has_updated_counter) {
728
0
        return;
729
0
    }
730
1.17M
    _has_updated_counter = true;
731
1.17M
    _tablet_reader->update_profile(_profile);
732
733
1.17M
    Scanner::_collect_profile_before_close();
734
735
    // Update counters for OlapScanner
736
    // Update counters from tablet reader's stats
737
1.17M
    auto& stats = _tablet_reader->stats();
738
1.17M
    auto* local_state = (OlapScanLocalState*)_local_state;
739
1.17M
    COUNTER_UPDATE(local_state->_io_timer, stats.io_ns);
740
1.17M
    COUNTER_UPDATE(local_state->_read_compressed_counter, stats.compressed_bytes_read);
741
1.17M
    COUNTER_UPDATE(local_state->_scan_bytes, stats.uncompressed_bytes_read);
742
1.17M
    COUNTER_UPDATE(local_state->_decompressor_timer, stats.decompress_ns);
743
1.17M
    COUNTER_UPDATE(local_state->_read_uncompressed_counter, stats.uncompressed_bytes_read);
744
1.17M
    COUNTER_UPDATE(local_state->_block_load_timer, stats.block_load_ns);
745
1.17M
    COUNTER_UPDATE(local_state->_block_load_counter, stats.blocks_load);
746
1.17M
    COUNTER_UPDATE(local_state->_block_fetch_timer, stats.block_fetch_ns);
747
1.17M
    COUNTER_UPDATE(local_state->_delete_bitmap_get_agg_timer, stats.delete_bitmap_get_agg_ns);
748
1.17M
    COUNTER_UPDATE(local_state->_scan_rows, stats.raw_rows_read);
749
1.17M
    COUNTER_UPDATE(local_state->_vec_cond_timer, stats.vec_cond_ns);
750
1.17M
    COUNTER_UPDATE(local_state->_short_cond_timer, stats.short_cond_ns);
751
1.17M
    COUNTER_UPDATE(local_state->_expr_filter_timer, stats.expr_filter_ns);
752
1.17M
    COUNTER_UPDATE(local_state->_block_init_timer, stats.block_init_ns);
753
1.17M
    COUNTER_UPDATE(local_state->_block_init_seek_timer, stats.block_init_seek_ns);
754
1.17M
    COUNTER_UPDATE(local_state->_block_init_seek_counter, stats.block_init_seek_num);
755
1.17M
    COUNTER_UPDATE(local_state->_segment_generate_row_range_by_keys_timer,
756
1.17M
                   stats.generate_row_ranges_by_keys_ns);
757
1.17M
    COUNTER_UPDATE(local_state->_segment_generate_row_range_by_column_conditions_timer,
758
1.17M
                   stats.generate_row_ranges_by_column_conditions_ns);
759
1.17M
    COUNTER_UPDATE(local_state->_segment_generate_row_range_by_bf_timer,
760
1.17M
                   stats.generate_row_ranges_by_bf_ns);
761
1.17M
    COUNTER_UPDATE(local_state->_collect_iterator_merge_next_timer,
762
1.17M
                   stats.collect_iterator_merge_next_timer);
763
1.17M
    COUNTER_UPDATE(local_state->_segment_generate_row_range_by_zonemap_timer,
764
1.17M
                   stats.generate_row_ranges_by_zonemap_ns);
765
1.17M
    COUNTER_UPDATE(local_state->_segment_generate_row_range_by_dict_timer,
766
1.17M
                   stats.generate_row_ranges_by_dict_ns);
767
1.17M
    COUNTER_UPDATE(local_state->_predicate_column_read_timer, stats.predicate_column_read_ns);
768
1.17M
    COUNTER_UPDATE(local_state->_non_predicate_column_read_timer, stats.non_predicate_read_ns);
769
1.17M
    COUNTER_UPDATE(local_state->_predicate_column_read_seek_timer,
770
1.17M
                   stats.predicate_column_read_seek_ns);
771
1.17M
    COUNTER_UPDATE(local_state->_predicate_column_read_seek_counter,
772
1.17M
                   stats.predicate_column_read_seek_num);
773
1.17M
    COUNTER_UPDATE(local_state->_lazy_read_timer, stats.lazy_read_ns);
774
1.17M
    COUNTER_UPDATE(local_state->_lazy_read_seek_timer, stats.block_lazy_read_seek_ns);
775
1.17M
    COUNTER_UPDATE(local_state->_lazy_read_seek_counter, stats.block_lazy_read_seek_num);
776
1.17M
    COUNTER_UPDATE(local_state->_output_col_timer, stats.output_col_ns);
777
1.17M
    COUNTER_UPDATE(local_state->_rows_vec_cond_filtered_counter, stats.rows_vec_cond_filtered);
778
1.17M
    COUNTER_UPDATE(local_state->_rows_short_circuit_cond_filtered_counter,
779
1.17M
                   stats.rows_short_circuit_cond_filtered);
780
1.17M
    COUNTER_UPDATE(local_state->_rows_expr_cond_filtered_counter, stats.rows_expr_cond_filtered);
781
1.17M
    COUNTER_UPDATE(local_state->_rows_vec_cond_input_counter, stats.vec_cond_input_rows);
782
1.17M
    COUNTER_UPDATE(local_state->_rows_short_circuit_cond_input_counter,
783
1.17M
                   stats.short_circuit_cond_input_rows);
784
1.17M
    COUNTER_UPDATE(local_state->_rows_expr_cond_input_counter, stats.expr_cond_input_rows);
785
1.17M
    COUNTER_UPDATE(local_state->_stats_filtered_counter, stats.rows_stats_filtered);
786
1.17M
    COUNTER_UPDATE(local_state->_stats_rp_filtered_counter, stats.rows_stats_rp_filtered);
787
1.17M
    COUNTER_UPDATE(local_state->_dict_filtered_counter, stats.segment_dict_filtered);
788
1.17M
    COUNTER_UPDATE(local_state->_bf_filtered_counter, stats.rows_bf_filtered);
789
1.17M
    COUNTER_UPDATE(local_state->_del_filtered_counter, stats.rows_del_filtered);
790
1.17M
    COUNTER_UPDATE(local_state->_del_filtered_counter, stats.rows_del_by_bitmap);
791
1.17M
    COUNTER_UPDATE(local_state->_del_filtered_counter, stats.rows_vec_del_cond_filtered);
792
1.17M
    COUNTER_UPDATE(local_state->_conditions_filtered_counter, stats.rows_conditions_filtered);
793
1.17M
    COUNTER_UPDATE(local_state->_key_range_filtered_counter, stats.rows_key_range_filtered);
794
1.17M
    COUNTER_UPDATE(local_state->_total_pages_num_counter, stats.total_pages_num);
795
1.17M
    COUNTER_UPDATE(local_state->_cached_pages_num_counter, stats.cached_pages_num);
796
1.17M
    COUNTER_UPDATE(local_state->_inverted_index_filter_counter, stats.rows_inverted_index_filtered);
797
1.17M
    COUNTER_UPDATE(local_state->_inverted_index_filter_timer, stats.inverted_index_filter_timer);
798
1.17M
    COUNTER_UPDATE(local_state->_inverted_index_query_cache_hit_counter,
799
1.17M
                   stats.inverted_index_query_cache_hit);
800
1.17M
    COUNTER_UPDATE(local_state->_inverted_index_query_cache_miss_counter,
801
1.17M
                   stats.inverted_index_query_cache_miss);
802
1.17M
    COUNTER_UPDATE(local_state->_inverted_index_query_timer, stats.inverted_index_query_timer);
803
1.17M
    COUNTER_UPDATE(local_state->_inverted_index_query_null_bitmap_timer,
804
1.17M
                   stats.inverted_index_query_null_bitmap_timer);
805
1.17M
    COUNTER_UPDATE(local_state->_inverted_index_query_bitmap_copy_timer,
806
1.17M
                   stats.inverted_index_query_bitmap_copy_timer);
807
1.17M
    COUNTER_UPDATE(local_state->_inverted_index_searcher_open_timer,
808
1.17M
                   stats.inverted_index_searcher_open_timer);
809
1.17M
    COUNTER_UPDATE(local_state->_inverted_index_searcher_search_timer,
810
1.17M
                   stats.inverted_index_searcher_search_timer);
811
1.17M
    COUNTER_UPDATE(local_state->_inverted_index_searcher_search_init_timer,
812
1.17M
                   stats.inverted_index_searcher_search_init_timer);
813
1.17M
    COUNTER_UPDATE(local_state->_inverted_index_searcher_search_exec_timer,
814
1.17M
                   stats.inverted_index_searcher_search_exec_timer);
815
1.17M
    COUNTER_UPDATE(local_state->_inverted_index_searcher_cache_hit_counter,
816
1.17M
                   stats.inverted_index_searcher_cache_hit);
817
1.17M
    COUNTER_UPDATE(local_state->_inverted_index_searcher_cache_miss_counter,
818
1.17M
                   stats.inverted_index_searcher_cache_miss);
819
1.17M
    COUNTER_UPDATE(local_state->_inverted_index_downgrade_count_counter,
820
1.17M
                   stats.inverted_index_downgrade_count);
821
1.17M
    COUNTER_UPDATE(local_state->_inverted_index_analyzer_timer,
822
1.17M
                   stats.inverted_index_analyzer_timer);
823
1.17M
    COUNTER_UPDATE(local_state->_inverted_index_lookup_timer, stats.inverted_index_lookup_timer);
824
1.17M
    COUNTER_UPDATE(local_state->_variant_scan_sparse_column_timer,
825
1.17M
                   stats.variant_scan_sparse_column_timer_ns);
826
1.17M
    COUNTER_UPDATE(local_state->_variant_scan_sparse_column_bytes,
827
1.17M
                   stats.variant_scan_sparse_column_bytes);
828
1.17M
    COUNTER_UPDATE(local_state->_variant_fill_path_from_sparse_column_timer,
829
1.17M
                   stats.variant_fill_path_from_sparse_column_timer_ns);
830
1.17M
    COUNTER_UPDATE(local_state->_variant_subtree_default_iter_count,
831
1.17M
                   stats.variant_subtree_default_iter_count);
832
1.17M
    COUNTER_UPDATE(local_state->_variant_subtree_leaf_iter_count,
833
1.17M
                   stats.variant_subtree_leaf_iter_count);
834
1.17M
    COUNTER_UPDATE(local_state->_variant_subtree_hierarchical_iter_count,
835
1.17M
                   stats.variant_subtree_hierarchical_iter_count);
836
1.17M
    COUNTER_UPDATE(local_state->_variant_subtree_sparse_iter_count,
837
1.17M
                   stats.variant_subtree_sparse_iter_count);
838
1.17M
    COUNTER_UPDATE(local_state->_variant_doc_value_column_iter_count,
839
1.17M
                   stats.variant_doc_value_column_iter_count);
840
841
1.17M
    if (stats.adaptive_batch_size_predict_max_rows > 0) {
842
929k
        local_state->_adaptive_batch_predict_min_rows_counter->set(
843
929k
                stats.adaptive_batch_size_predict_min_rows);
844
929k
        local_state->_adaptive_batch_predict_max_rows_counter->set(
845
929k
                stats.adaptive_batch_size_predict_max_rows);
846
929k
    }
847
848
1.17M
    InvertedIndexProfileReporter inverted_index_profile;
849
1.17M
    inverted_index_profile.update(local_state->_index_filter_profile.get(),
850
1.17M
                                  &stats.inverted_index_stats);
851
852
    // only cloud deploy mode will use file cache.
853
1.17M
    if (config::is_cloud_mode() && config::enable_file_cache) {
854
836k
        io::FileCacheProfileReporter cache_profile(local_state->_segment_profile.get());
855
836k
        cache_profile.update(&stats.file_cache_stats);
856
836k
        _state->get_query_ctx()->resource_ctx()->io_context()->update_bytes_write_into_cache(
857
836k
                stats.file_cache_stats.bytes_write_into_cache);
858
836k
    }
859
1.17M
    COUNTER_UPDATE(local_state->_output_index_result_column_timer,
860
1.17M
                   stats.output_index_result_column_timer);
861
1.17M
    COUNTER_UPDATE(local_state->_filtered_segment_counter, stats.filtered_segment_number);
862
1.17M
    COUNTER_UPDATE(local_state->_total_segment_counter, stats.total_segment_number);
863
1.17M
    COUNTER_UPDATE(local_state->_condition_cache_hit_counter, stats.condition_cache_hit_seg_nums);
864
1.17M
    COUNTER_UPDATE(local_state->_condition_cache_filtered_rows_counter,
865
1.17M
                   stats.condition_cache_filtered_rows);
866
867
1.17M
    COUNTER_UPDATE(local_state->_tablet_reader_init_timer, stats.tablet_reader_init_timer_ns);
868
1.17M
    COUNTER_UPDATE(local_state->_tablet_reader_capture_rs_readers_timer,
869
1.17M
                   stats.tablet_reader_capture_rs_readers_timer_ns);
870
1.17M
    COUNTER_UPDATE(local_state->_tablet_reader_init_return_columns_timer,
871
1.17M
                   stats.tablet_reader_init_return_columns_timer_ns);
872
1.17M
    COUNTER_UPDATE(local_state->_tablet_reader_init_keys_param_timer,
873
1.17M
                   stats.tablet_reader_init_keys_param_timer_ns);
874
1.17M
    COUNTER_UPDATE(local_state->_tablet_reader_init_orderby_keys_param_timer,
875
1.17M
                   stats.tablet_reader_init_orderby_keys_param_timer_ns);
876
1.17M
    COUNTER_UPDATE(local_state->_tablet_reader_init_conditions_param_timer,
877
1.17M
                   stats.tablet_reader_init_conditions_param_timer_ns);
878
1.17M
    COUNTER_UPDATE(local_state->_tablet_reader_init_delete_condition_param_timer,
879
1.17M
                   stats.tablet_reader_init_delete_condition_param_timer_ns);
880
1.17M
    COUNTER_UPDATE(local_state->_block_reader_vcollect_iter_init_timer,
881
1.17M
                   stats.block_reader_vcollect_iter_init_timer_ns);
882
1.17M
    COUNTER_UPDATE(local_state->_block_reader_rs_readers_init_timer,
883
1.17M
                   stats.block_reader_rs_readers_init_timer_ns);
884
1.17M
    COUNTER_UPDATE(local_state->_block_reader_build_heap_init_timer,
885
1.17M
                   stats.block_reader_build_heap_init_timer_ns);
886
887
1.17M
    COUNTER_UPDATE(local_state->_rowset_reader_get_segment_iterators_timer,
888
1.17M
                   stats.rowset_reader_get_segment_iterators_timer_ns);
889
1.17M
    COUNTER_UPDATE(local_state->_rowset_reader_create_iterators_timer,
890
1.17M
                   stats.rowset_reader_create_iterators_timer_ns);
891
1.17M
    COUNTER_UPDATE(local_state->_rowset_reader_init_iterators_timer,
892
1.17M
                   stats.rowset_reader_init_iterators_timer_ns);
893
1.17M
    COUNTER_UPDATE(local_state->_rowset_reader_load_segments_timer,
894
1.17M
                   stats.rowset_reader_load_segments_timer_ns);
895
896
1.17M
    COUNTER_UPDATE(local_state->_segment_iterator_init_timer, stats.segment_iterator_init_timer_ns);
897
1.17M
    COUNTER_UPDATE(local_state->_segment_iterator_init_return_column_iterators_timer,
898
1.17M
                   stats.segment_iterator_init_return_column_iterators_timer_ns);
899
1.17M
    COUNTER_UPDATE(local_state->_segment_iterator_init_index_iterators_timer,
900
1.17M
                   stats.segment_iterator_init_index_iterators_timer_ns);
901
1.17M
    COUNTER_UPDATE(local_state->_segment_iterator_init_segment_prefetchers_timer,
902
1.17M
                   stats.segment_iterator_init_segment_prefetchers_timer_ns);
903
904
1.17M
    COUNTER_UPDATE(local_state->_segment_create_column_readers_timer,
905
1.17M
                   stats.segment_create_column_readers_timer_ns);
906
1.17M
    COUNTER_UPDATE(local_state->_segment_load_index_timer, stats.segment_load_index_timer_ns);
907
908
    // Update metrics
909
1.17M
    DorisMetrics::instance()->query_scan_bytes->increment(
910
1.17M
            local_state->_read_uncompressed_counter->value());
911
1.17M
    DorisMetrics::instance()->query_scan_rows->increment(local_state->_scan_rows->value());
912
1.17M
    auto& tablet = _tablet_reader_params.tablet;
913
1.17M
    tablet->query_scan_bytes->increment(local_state->_read_uncompressed_counter->value());
914
1.17M
    tablet->query_scan_rows->increment(local_state->_scan_rows->value());
915
1.17M
    tablet->query_scan_count->increment(1);
916
917
1.17M
    COUNTER_UPDATE(local_state->_ann_range_search_filter_counter,
918
1.17M
                   stats.rows_ann_index_range_filtered);
919
1.17M
    COUNTER_UPDATE(local_state->_ann_topn_filter_counter, stats.rows_ann_index_topn_filtered);
920
1.17M
    COUNTER_UPDATE(local_state->_ann_index_load_costs, stats.ann_index_load_ns);
921
1.17M
    COUNTER_UPDATE(local_state->_ann_ivf_on_disk_load_costs, stats.ann_ivf_on_disk_load_ns);
922
1.17M
    COUNTER_UPDATE(local_state->_ann_ivf_on_disk_cache_hit_cnt,
923
1.17M
                   stats.ann_ivf_on_disk_cache_hit_cnt);
924
1.17M
    COUNTER_UPDATE(local_state->_ann_ivf_on_disk_cache_miss_cnt,
925
1.17M
                   stats.ann_ivf_on_disk_cache_miss_cnt);
926
1.17M
    COUNTER_UPDATE(local_state->_ann_range_search_costs, stats.ann_index_range_search_ns);
927
1.17M
    COUNTER_UPDATE(local_state->_ann_range_search_cnt, stats.ann_index_range_search_cnt);
928
1.17M
    COUNTER_UPDATE(local_state->_ann_range_engine_search_costs, stats.ann_range_engine_search_ns);
929
    // Engine prepare before search
930
1.17M
    COUNTER_UPDATE(local_state->_ann_range_pre_process_costs, stats.ann_range_pre_process_ns);
931
    // Post process parent: Doris result process + engine convert
932
1.17M
    COUNTER_UPDATE(local_state->_ann_range_post_process_costs,
933
1.17M
                   stats.ann_range_result_convert_ns + stats.ann_range_engine_convert_ns);
934
    // Engine convert (child under post-process)
935
1.17M
    COUNTER_UPDATE(local_state->_ann_range_engine_convert_costs, stats.ann_range_engine_convert_ns);
936
    // Doris-side result convert (child under post-process)
937
1.17M
    COUNTER_UPDATE(local_state->_ann_range_result_convert_costs, stats.ann_range_result_convert_ns);
938
939
1.17M
    COUNTER_UPDATE(local_state->_ann_topn_search_costs, stats.ann_topn_search_ns);
940
1.17M
    COUNTER_UPDATE(local_state->_ann_topn_search_cnt, stats.ann_index_topn_search_cnt);
941
942
    // Detailed ANN timers
943
    // ANN TopN timers with hierarchy
944
    // Engine search time (FAISS)
945
1.17M
    COUNTER_UPDATE(local_state->_ann_topn_engine_search_costs,
946
1.17M
                   stats.ann_index_topn_engine_search_ns);
947
    // Engine prepare time (allocations/buffer setup before search)
948
1.17M
    COUNTER_UPDATE(local_state->_ann_topn_pre_process_costs,
949
1.17M
                   stats.ann_index_topn_engine_prepare_ns);
950
    // Post process parent includes Doris result processing + engine convert
951
1.17M
    COUNTER_UPDATE(local_state->_ann_topn_post_process_costs,
952
1.17M
                   stats.ann_index_topn_result_process_ns + stats.ann_index_topn_engine_convert_ns);
953
    // Engine-side conversion time inside FAISS wrappers (child under post-process)
954
1.17M
    COUNTER_UPDATE(local_state->_ann_topn_engine_convert_costs,
955
1.17M
                   stats.ann_index_topn_engine_convert_ns);
956
957
    // Doris-side result convert costs (show separately as another child counter); use pure process time
958
1.17M
    COUNTER_UPDATE(local_state->_ann_topn_result_convert_costs,
959
1.17M
                   stats.ann_index_topn_result_process_ns);
960
961
1.17M
    COUNTER_UPDATE(local_state->_ann_fallback_brute_force_cnt, stats.ann_fall_back_brute_force_cnt);
962
963
    // Overhead counter removed; precise instrumentation is reported via engine_prepare above.
964
1.17M
}
965
966
#include "common/compile_check_avoid_end.h"
967
} // namespace doris