Coverage Report

Created: 2026-05-14 07:56

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
1.18M
        : Scanner(params.state, parent, params.limit, params.profile),
74
1.18M
          _key_ranges(std::move(params.key_ranges)),
75
1.18M
          _tablet_reader_params({.tablet = std::move(params.tablet),
76
1.18M
                                 .tablet_schema {},
77
1.18M
                                 .aggregation = params.aggregation,
78
1.18M
                                 .version = {0, params.version},
79
1.18M
                                 .start_key {},
80
1.18M
                                 .end_key {},
81
1.18M
                                 .predicates {},
82
1.18M
                                 .function_filters {},
83
1.18M
                                 .delete_predicates {},
84
1.18M
                                 .target_cast_type_for_variants {},
85
1.18M
                                 .all_access_paths {},
86
1.18M
                                 .predicate_access_paths {},
87
1.18M
                                 .rs_splits {},
88
1.18M
                                 .return_columns {},
89
1.18M
                                 .output_columns {},
90
1.18M
                                 .remaining_conjunct_roots {},
91
1.18M
                                 .common_expr_ctxs_push_down {},
92
1.18M
                                 .topn_filter_source_node_ids {},
93
1.18M
                                 .key_group_cluster_key_idxes {},
94
1.18M
                                 .virtual_column_exprs {},
95
1.18M
                                 .vir_cid_to_idx_in_block {},
96
1.18M
                                 .vir_col_idx_to_type {},
97
1.18M
                                 .score_runtime {},
98
1.18M
                                 .collection_statistics {},
99
1.18M
                                 .ann_topn_runtime {},
100
1.18M
                                 .condition_cache_digest = parent->get_condition_cache_digest()}) {
101
1.18M
    _tablet_reader_params.set_read_source(std::move(params.read_source),
102
1.18M
                                          _state->skip_delete_bitmap());
103
1.18M
    _has_prepared = false;
104
1.18M
    _vector_search_params = params.state->get_vector_search_params();
105
1.18M
}
106
107
static std::string read_columns_to_string(TabletSchemaSPtr tablet_schema,
108
2.70k
                                          const std::vector<uint32_t>& read_columns) {
109
    // avoid too long for one line,
110
    // it is hard to display in `show profile` stmt if one line is too long.
111
2.70k
    const int col_per_line = 10;
112
2.70k
    int i = 0;
113
2.70k
    std::string read_columns_string;
114
2.70k
    read_columns_string += "[";
115
12.7k
    for (auto it = read_columns.cbegin(); it != read_columns.cend(); it++) {
116
10.0k
        if (it != read_columns.cbegin()) {
117
7.41k
            read_columns_string += ", ";
118
7.41k
        }
119
10.0k
        read_columns_string += tablet_schema->columns().at(*it)->name();
120
10.0k
        if (i >= col_per_line) {
121
13
            read_columns_string += "\n";
122
13
            i = 0;
123
10.0k
        } else {
124
10.0k
            ++i;
125
10.0k
        }
126
10.0k
    }
127
2.70k
    read_columns_string += "]";
128
2.70k
    return read_columns_string;
129
2.70k
}
130
131
1.18M
Status OlapScanner::_prepare_impl() {
132
1.18M
    auto* local_state = static_cast<OlapScanLocalState*>(_local_state);
133
1.18M
    auto& tablet = _tablet_reader_params.tablet;
134
1.18M
    auto& tablet_schema = _tablet_reader_params.tablet_schema;
135
1.18M
    DBUG_EXECUTE_IF("CloudTablet.capture_rs_readers.return.e-230", {
136
1.18M
        LOG_WARNING("CloudTablet.capture_rs_readers.return e-230 init")
137
1.18M
                .tag("tablet_id", tablet->tablet_id());
138
1.18M
        return Status::Error<false>(-230, "injected error");
139
1.18M
    });
140
141
1.18M
    for (auto& ctx : local_state->_common_expr_ctxs_push_down) {
142
21.9k
        VExprContextSPtr context;
143
21.9k
        RETURN_IF_ERROR(ctx->clone(_state, context));
144
21.9k
        _common_expr_ctxs_push_down.emplace_back(context);
145
21.9k
        context->prepare_ann_range_search(_vector_search_params);
146
21.9k
    }
147
148
1.18M
    for (auto pair : local_state->_slot_id_to_virtual_column_expr) {
149
        // Scanner will be executed in a different thread, so we need to clone the context.
150
405
        VExprContextSPtr context;
151
405
        RETURN_IF_ERROR(pair.second->clone(_state, context));
152
405
        _slot_id_to_virtual_column_expr[pair.first] = context;
153
405
    }
154
155
1.18M
    _slot_id_to_index_in_block = local_state->_slot_id_to_index_in_block;
156
1.18M
    _slot_id_to_col_type = local_state->_slot_id_to_col_type;
157
1.18M
    _score_runtime = local_state->_score_runtime;
158
    // All scanners share the same ann_topn_runtime.
159
1.18M
    _ann_topn_runtime = local_state->_ann_topn_runtime;
160
161
    // set limit to reduce end of rowset and segment mem use
162
1.18M
    _tablet_reader = std::make_unique<BlockReader>();
163
    // batch size is passed down to segment iterator, use _state->batch_size()
164
    // instead of _parent->limit(), because if _parent->limit() is a very small
165
    // value (e.g. select a from t where a .. and b ... limit 1),
166
    // it will be very slow when reading data in segment iterator
167
1.18M
    _tablet_reader->set_batch_size(_state->batch_size());
168
    // Adaptive batch size: pass byte-budget settings to the storage reader.
169
    // The reader still uses batch_size() as the row ceiling.
170
1.18M
    _tablet_reader->set_preferred_block_size_bytes(_state->preferred_block_size_bytes());
171
1.18M
    {
172
1.18M
        TOlapScanNode& olap_scan_node = local_state->olap_scan_node();
173
174
        // Each scanner builds its own TabletSchema to avoid concurrent modification.
175
1.18M
        tablet_schema = std::make_shared<TabletSchema>();
176
1.18M
        tablet_schema->copy_from(*tablet->tablet_schema());
177
1.18M
        if (olap_scan_node.__isset.columns_desc && !olap_scan_node.columns_desc.empty() &&
178
1.18M
            olap_scan_node.columns_desc[0].col_unique_id >= 0) {
179
1.18M
            tablet_schema->clear_columns();
180
18.9M
            for (const auto& column_desc : olap_scan_node.columns_desc) {
181
18.9M
                tablet_schema->append_column(TabletColumn(column_desc));
182
18.9M
            }
183
1.18M
            if (olap_scan_node.__isset.schema_version) {
184
1.18M
                tablet_schema->set_schema_version(olap_scan_node.schema_version);
185
1.18M
            }
186
1.18M
        }
187
1.18M
        if (olap_scan_node.__isset.indexes_desc) {
188
1.18M
            tablet_schema->update_indexes_from_thrift(olap_scan_node.indexes_desc);
189
1.18M
        }
190
191
1.18M
        if (_tablet_reader_params.rs_splits.empty()) {
192
            // Non-pipeline mode, Tablet : Scanner = 1 : 1
193
            // acquire tablet rowset readers at the beginning of the scan node
194
            // to prevent this case: when there are lots of olap scanners to run for example 10000
195
            // the rowsets maybe compacted when the last olap scanner starts
196
0
            ReadSource read_source;
197
198
0
            if (config::is_cloud_mode()) {
199
                // FIXME(plat1ko): Avoid pointer cast
200
0
                ExecEnv::GetInstance()->storage_engine().to_cloud().tablet_hotspot().count(*tablet);
201
0
            }
202
203
0
            auto maybe_read_source = tablet->capture_read_source(
204
0
                    _tablet_reader_params.version,
205
0
                    {
206
0
                            .skip_missing_versions = _state->skip_missing_version(),
207
0
                            .enable_fetch_rowsets_from_peers =
208
0
                                    config::enable_fetch_rowsets_from_peer_replicas,
209
0
                            .enable_prefer_cached_rowset =
210
0
                                    config::is_cloud_mode() ? _state->enable_prefer_cached_rowset()
211
0
                                                            : false,
212
0
                            .query_freshness_tolerance_ms =
213
0
                                    config::is_cloud_mode() ? _state->query_freshness_tolerance_ms()
214
0
                                                            : -1,
215
0
                    });
216
0
            if (!maybe_read_source) {
217
0
                LOG(WARNING) << "fail to init reader. res=" << maybe_read_source.error();
218
0
                return maybe_read_source.error();
219
0
            }
220
221
0
            read_source = std::move(maybe_read_source.value());
222
223
0
            if (config::enable_mow_verbose_log && tablet->enable_unique_key_merge_on_write()) {
224
0
                LOG_INFO("finish capture_rs_readers for tablet={}, query_id={}",
225
0
                         tablet->tablet_id(), print_id(_state->query_id()));
226
0
            }
227
228
0
            if (!_state->skip_delete_predicate()) {
229
0
                read_source.fill_delete_predicates();
230
0
            }
231
0
            _tablet_reader_params.set_read_source(std::move(read_source));
232
0
        }
233
234
        // Initialize tablet_reader_params
235
1.18M
        RETURN_IF_ERROR(_init_tablet_reader_params(
236
1.18M
                local_state->_parent->cast<OlapScanOperatorX>()._slot_id_to_slot_desc, _key_ranges,
237
1.18M
                local_state->_slot_id_to_predicates, local_state->_push_down_functions));
238
1.18M
    }
239
240
    // add read columns in profile
241
1.18M
    if (_state->enable_profile()) {
242
2.69k
        _profile->add_info_string("ReadColumns",
243
2.69k
                                  read_columns_to_string(tablet_schema, _return_columns));
244
2.69k
    }
245
246
1.18M
    if (_tablet_reader_params.score_runtime) {
247
14
        SCOPED_TIMER(local_state->_statistics_collect_timer);
248
14
        _tablet_reader_params.collection_statistics = std::make_shared<CollectionStatistics>();
249
250
14
        io::IOContext io_ctx {
251
14
                .reader_type = ReaderType::READER_QUERY,
252
14
                .expiration_time = tablet->ttl_seconds(),
253
14
                .query_id = &_state->query_id(),
254
14
                .file_cache_stats = &_tablet_reader->mutable_stats()->file_cache_stats,
255
14
                .is_inverted_index = true,
256
14
        };
257
258
14
        RETURN_IF_ERROR(_tablet_reader_params.collection_statistics->collect(
259
14
                _state, _tablet_reader_params.rs_splits, _tablet_reader_params.tablet_schema,
260
14
                _tablet_reader_params.common_expr_ctxs_push_down, &io_ctx));
261
14
    }
262
263
1.18M
    _has_prepared = true;
264
1.18M
    return Status::OK();
265
1.18M
}
266
267
1.18M
Status OlapScanner::_open_impl(RuntimeState* state) {
268
1.18M
    RETURN_IF_ERROR(Scanner::_open_impl(state));
269
1.18M
    SCOPED_TIMER(_local_state->cast<OlapScanLocalState>()._reader_init_timer);
270
271
1.18M
    auto res = _tablet_reader->init(_tablet_reader_params);
272
1.18M
    if (!res.ok()) {
273
72
        res.append("failed to initialize storage reader. tablet=" +
274
72
                   std::to_string(_tablet_reader_params.tablet->tablet_id()) +
275
72
                   ", backend=" + BackendOptions::get_localhost());
276
72
        return res;
277
72
    }
278
279
    // Do not hold rs_splits any more to release memory.
280
1.18M
    _tablet_reader_params.rs_splits.clear();
281
282
1.18M
    return Status::OK();
283
1.18M
}
284
285
// it will be called under tablet read lock because capture rs readers need
286
Status OlapScanner::_init_tablet_reader_params(
287
        const phmap::flat_hash_map<int, SlotDescriptor*>& slot_id_to_slot_desc,
288
        const std::vector<OlapScanRange*>& key_ranges,
289
        const phmap::flat_hash_map<int, std::vector<std::shared_ptr<ColumnPredicate>>>&
290
                slot_to_predicates,
291
1.18M
        const std::vector<FunctionFilter>& function_filters) {
292
    // if the table with rowset [0-x] or [0-1] [2-y], and [0-1] is empty
293
1.18M
    const bool single_version = _tablet_reader_params.has_single_version();
294
295
1.18M
    auto* olap_local_state = static_cast<OlapScanLocalState*>(_local_state);
296
1.18M
    bool read_mor_as_dup = olap_local_state->olap_scan_node().__isset.read_mor_as_dup &&
297
1.18M
                           olap_local_state->olap_scan_node().read_mor_as_dup;
298
1.18M
    if (_state->skip_storage_engine_merge() || read_mor_as_dup) {
299
49
        _tablet_reader_params.direct_mode = true;
300
49
        _tablet_reader_params.aggregation = true;
301
1.18M
    } else {
302
1.18M
        auto push_down_agg_type = _local_state->get_push_down_agg_type();
303
1.18M
        _tablet_reader_params.direct_mode = _tablet_reader_params.aggregation || single_version ||
304
1.18M
                                            (push_down_agg_type != TPushAggOp::NONE &&
305
11.0k
                                             push_down_agg_type != TPushAggOp::COUNT_ON_INDEX);
306
1.18M
    }
307
308
1.18M
    RETURN_IF_ERROR(_init_variant_columns());
309
1.18M
    RETURN_IF_ERROR(_init_return_columns());
310
311
1.18M
    _tablet_reader_params.reader_type = ReaderType::READER_QUERY;
312
1.18M
    _tablet_reader_params.push_down_agg_type_opt = _local_state->get_push_down_agg_type();
313
314
    // TODO: If a new runtime filter arrives after `_conjuncts` move to `_common_expr_ctxs_push_down`,
315
1.18M
    if (_common_expr_ctxs_push_down.empty()) {
316
1.16M
        for (auto& conjunct : _conjuncts) {
317
3.78k
            _tablet_reader_params.remaining_conjunct_roots.emplace_back(conjunct->root());
318
3.78k
        }
319
1.16M
    } else {
320
21.8k
        for (auto& ctx : _common_expr_ctxs_push_down) {
321
21.8k
            _tablet_reader_params.remaining_conjunct_roots.emplace_back(ctx->root());
322
21.8k
        }
323
18.5k
    }
324
325
1.18M
    _tablet_reader_params.common_expr_ctxs_push_down = _common_expr_ctxs_push_down;
326
1.18M
    _tablet_reader_params.virtual_column_exprs = _virtual_column_exprs;
327
1.18M
    _tablet_reader_params.vir_cid_to_idx_in_block = _vir_cid_to_idx_in_block;
328
1.18M
    _tablet_reader_params.vir_col_idx_to_type = _vir_col_idx_to_type;
329
1.18M
    _tablet_reader_params.score_runtime = _score_runtime;
330
1.18M
    _tablet_reader_params.output_columns = ((OlapScanLocalState*)_local_state)->_output_column_ids;
331
1.18M
    _tablet_reader_params.ann_topn_runtime = _ann_topn_runtime;
332
1.18M
    for (const auto& ele : ((OlapScanLocalState*)_local_state)->_cast_types_for_variants) {
333
1.67k
        _tablet_reader_params.target_cast_type_for_variants[ele.first] = ele.second;
334
1.67k
    };
335
1.18M
    auto& tablet_schema = _tablet_reader_params.tablet_schema;
336
9.71M
    for (auto& predicates : slot_to_predicates) {
337
9.71M
        const int sid = predicates.first;
338
9.71M
        DCHECK(slot_id_to_slot_desc.contains(sid));
339
9.71M
        int32_t index =
340
9.71M
                tablet_schema->field_index(slot_id_to_slot_desc.find(sid)->second->col_name());
341
9.71M
        if (index < 0) {
342
0
            throw Exception(
343
0
                    Status::InternalError("Column {} not found in tablet schema",
344
0
                                          slot_id_to_slot_desc.find(sid)->second->col_name()));
345
0
        }
346
9.71M
        for (auto& predicate : predicates.second) {
347
958k
            _tablet_reader_params.predicates.push_back(predicate->clone(index));
348
958k
        }
349
9.71M
    }
350
351
1.18M
    std::copy(function_filters.cbegin(), function_filters.cend(),
352
1.18M
              std::inserter(_tablet_reader_params.function_filters,
353
1.18M
                            _tablet_reader_params.function_filters.begin()));
354
355
    // Merge the columns in delete predicate that not in latest schema in to current tablet schema
356
1.18M
    for (auto& del_pred : _tablet_reader_params.delete_predicates) {
357
6.31k
        tablet_schema->merge_dropped_columns(*del_pred->tablet_schema());
358
6.31k
    }
359
360
    // Push key ranges to the tablet reader.
361
    // Skip the "full scan" placeholder (has_lower_bound == false) — when no key
362
    // predicates exist, start_key/end_key remain empty and the reader does a full scan.
363
1.92M
    for (auto* key_range : key_ranges) {
364
1.92M
        if (!key_range->has_lower_bound) {
365
151k
            continue;
366
151k
        }
367
368
1.77M
        _tablet_reader_params.start_key_include = key_range->begin_include;
369
1.77M
        _tablet_reader_params.end_key_include = key_range->end_include;
370
371
1.77M
        _tablet_reader_params.start_key.push_back(key_range->begin_scan_range);
372
1.77M
        _tablet_reader_params.end_key.push_back(key_range->end_scan_range);
373
1.77M
    }
374
375
1.18M
    _tablet_reader_params.profile = _local_state->custom_profile();
376
1.18M
    _tablet_reader_params.runtime_state = _state;
377
378
1.18M
    _tablet_reader_params.origin_return_columns = &_return_columns;
379
1.18M
    _tablet_reader_params.tablet_columns_convert_to_null_set = &_tablet_columns_convert_to_null_set;
380
381
1.18M
    if (_tablet_reader_params.direct_mode) {
382
1.16M
        _tablet_reader_params.return_columns = _return_columns;
383
1.16M
    } else {
384
        // we need to fetch all key columns to do the right aggregation on storage engine side.
385
41.6k
        for (size_t i = 0; i < tablet_schema->num_key_columns(); ++i) {
386
30.1k
            _tablet_reader_params.return_columns.push_back(i);
387
30.1k
        }
388
54.6k
        for (auto index : _return_columns) {
389
54.6k
            if (tablet_schema->column(index).is_key()) {
390
21.3k
                continue;
391
21.3k
            }
392
33.3k
            _tablet_reader_params.return_columns.push_back(index);
393
33.3k
        }
394
        // expand the sequence column
395
11.4k
        if (tablet_schema->has_sequence_col() || tablet_schema->has_seq_map()) {
396
40
            bool has_replace_col = false;
397
90
            for (auto col : _return_columns) {
398
90
                if (tablet_schema->column(col).aggregation() ==
399
90
                    FieldAggregationMethod::OLAP_FIELD_AGGREGATION_REPLACE) {
400
39
                    has_replace_col = true;
401
39
                    break;
402
39
                }
403
90
            }
404
40
            if (auto sequence_col_idx = tablet_schema->sequence_col_idx();
405
40
                has_replace_col && tablet_schema->has_sequence_col() &&
406
40
                std::find(_return_columns.begin(), _return_columns.end(), sequence_col_idx) ==
407
27
                        _return_columns.end()) {
408
16
                _tablet_reader_params.return_columns.push_back(sequence_col_idx);
409
16
            }
410
40
            if (has_replace_col) {
411
39
                const auto& val_to_seq = tablet_schema->value_col_idx_to_seq_col_idx();
412
39
                std::set<uint32_t> return_seq_columns;
413
414
234
                for (auto col : _tablet_reader_params.return_columns) {
415
                    // we need to add the necessary sequence column in _return_columns, and
416
                    // Avoid adding the same seq column twice
417
234
                    const auto val_iter = val_to_seq.find(col);
418
234
                    if (val_iter != val_to_seq.end()) {
419
42
                        auto seq = val_iter->second;
420
42
                        if (std::find(_tablet_reader_params.return_columns.begin(),
421
42
                                      _tablet_reader_params.return_columns.end(),
422
42
                                      seq) == _tablet_reader_params.return_columns.end()) {
423
4
                            return_seq_columns.insert(seq);
424
4
                        }
425
42
                    }
426
234
                }
427
39
                _tablet_reader_params.return_columns.insert(
428
39
                        std::end(_tablet_reader_params.return_columns),
429
39
                        std::begin(return_seq_columns), std::end(return_seq_columns));
430
39
            }
431
40
        }
432
11.4k
    }
433
434
1.18M
    _tablet_reader_params.use_page_cache = _state->enable_page_cache();
435
436
1.18M
    DBUG_EXECUTE_IF("NewOlapScanner::_init_tablet_reader_params.block", DBUG_BLOCK);
437
438
1.18M
    if (!_state->skip_storage_engine_merge()) {
439
1.18M
        auto* olap_scan_local_state = (OlapScanLocalState*)_local_state;
440
1.18M
        TOlapScanNode& olap_scan_node = olap_scan_local_state->olap_scan_node();
441
442
        // Set MOR value predicate pushdown flag
443
1.18M
        if (olap_scan_node.__isset.enable_mor_value_predicate_pushdown &&
444
1.18M
            olap_scan_node.enable_mor_value_predicate_pushdown) {
445
25
            _tablet_reader_params.enable_mor_value_predicate_pushdown = true;
446
25
        }
447
448
1.18M
        const bool has_key_topn =
449
1.18M
                olap_scan_node.__isset.sort_info && !olap_scan_node.sort_info.is_asc_order.empty();
450
1.18M
        if (has_key_topn) {
451
2.13k
            _limit = _local_state->limit_per_scanner();
452
2.13k
        }
453
454
1.18M
        const bool no_runtime_filters = _total_rf_num == 0;
455
1.18M
        const bool segment_limit_enabled = _state->enable_segment_limit_pushdown();
456
1.18M
        const bool storage_no_merge = olap_scan_local_state->_storage_no_merge();
457
458
1.18M
        if (_limit > 0 && no_runtime_filters && segment_limit_enabled && storage_no_merge) {
459
3.87k
            for (const auto& conjunct : _conjuncts) {
460
0
                DORIS_CHECK(!olap_scan_local_state->_check_expr_storage_filter(
461
0
                        conjunct->root(), OlapScanLocalState::ExprStorageFilterCheckMode::
462
0
                                                  HAS_SEGMENT_EVALUABLE_EXPR));
463
0
            }
464
3.87k
        }
465
466
        // Segment LIMIT has only two legal states: completely disabled, or enabled after every
467
        // row-filtering conjunct has become a storage predicate or SegmentIterator common expr.
468
1.18M
        const bool can_push_down_segment_limit = _limit > 0 && no_runtime_filters &&
469
1.18M
                                                 _conjuncts.empty() && segment_limit_enabled &&
470
1.18M
                                                 storage_no_merge;
471
1.18M
        if (can_push_down_segment_limit) {
472
3.87k
            if (has_key_topn) {
473
2.03k
                _tablet_reader_params.read_orderby_key = true;
474
2.03k
                if (!olap_scan_node.sort_info.is_asc_order[0]) {
475
247
                    _tablet_reader_params.read_orderby_key_reverse = true;
476
247
                }
477
2.03k
                _tablet_reader_params.read_orderby_key_num_prefix_columns =
478
2.03k
                        olap_scan_node.sort_info.is_asc_order.size();
479
2.03k
                _tablet_reader_params.read_orderby_key_limit = _limit;
480
2.03k
            } else {
481
1.83k
                _tablet_reader_params.general_read_limit = _limit;
482
1.83k
            }
483
3.87k
        }
484
485
1.18M
        if (_tablet_reader_params.read_orderby_key_limit > 0 ||
486
1.18M
            _tablet_reader_params.general_read_limit > 0) {
487
3.87k
            DORIS_CHECK(can_push_down_segment_limit);
488
3.87k
            DORIS_CHECK(_conjuncts.empty());
489
3.87k
        }
490
491
        // A key TopN scan cannot share the plain LIMIT early-stop counter. If
492
        // storage TopN is pushed down, each scanner must produce its full local
493
        // candidates. If it is not pushed down for any reason, the upper TopN
494
        // still needs all rows from the scan.
495
1.18M
        if (has_key_topn) {
496
2.13k
            _shared_scan_limit = nullptr;
497
2.13k
            if (_tablet_reader_params.read_orderby_key_limit == 0) {
498
91
                _limit = -1;
499
91
            }
500
2.13k
        }
501
        // Note: _shared_scan_limit is intentionally not pushed into the
502
        // storage layer. SegmentIterator's _process_eof() is irreversible,
503
        // so a concurrently-decremented atomic could reach 0 while a segment
504
        // still has data needed by other scanners.
505
506
        // set push down topn filter
507
1.18M
        _tablet_reader_params.topn_filter_source_node_ids =
508
1.18M
                olap_scan_local_state->get_topn_filter_source_node_ids(_state, true);
509
1.18M
        if (!_tablet_reader_params.topn_filter_source_node_ids.empty()) {
510
4.82k
            _tablet_reader_params.topn_filter_target_node_id =
511
4.82k
                    olap_scan_local_state->parent()->node_id();
512
4.82k
        }
513
1.18M
    }
514
515
    // If this is a Two-Phase read query, and we need to delay the release of Rowset
516
    // by rowset->update_delayed_expired_timestamp().This could expand the lifespan of Rowset
517
1.18M
    if (tablet_schema->field_index(BeConsts::ROWID_COL) >= 0) {
518
0
        constexpr static int delayed_s = 60;
519
0
        for (auto rs_reader : _tablet_reader_params.rs_splits) {
520
0
            uint64_t delayed_expired_timestamp =
521
0
                    UnixSeconds() + _tablet_reader_params.runtime_state->execution_timeout() +
522
0
                    delayed_s;
523
0
            rs_reader.rs_reader->rowset()->update_delayed_expired_timestamp(
524
0
                    delayed_expired_timestamp);
525
0
            ExecEnv::GetInstance()->storage_engine().add_quering_rowset(
526
0
                    rs_reader.rs_reader->rowset());
527
0
        }
528
0
    }
529
530
1.18M
    if (tablet_schema->has_global_row_id()) {
531
6.78k
        auto& id_file_map = _state->get_id_file_map();
532
12.8k
        for (auto rs_reader : _tablet_reader_params.rs_splits) {
533
12.8k
            id_file_map->add_temp_rowset(rs_reader.rs_reader->rowset());
534
12.8k
        }
535
6.78k
    }
536
537
1.18M
    return Status::OK();
538
1.18M
}
539
540
1.18M
Status OlapScanner::_init_variant_columns() {
541
1.18M
    auto& tablet_schema = _tablet_reader_params.tablet_schema;
542
1.18M
    if (tablet_schema->num_variant_columns() == 0) {
543
1.17M
        return Status::OK();
544
1.17M
    }
545
    // Parent column has path info to distinction from each other
546
16.3k
    for (auto* slot : _output_tuple_desc->slots()) {
547
16.3k
        if (slot->type()->get_primitive_type() == PrimitiveType::TYPE_VARIANT) {
548
            // Such columns are not exist in frontend schema info, so we need to
549
            // add them into tablet_schema for later column indexing.
550
8.06k
            const auto& dt_variant =
551
8.06k
                    assert_cast<const DataTypeVariant&>(*remove_nullable(slot->type()));
552
8.06k
            TabletColumn subcol = TabletColumn::create_materialized_variant_column(
553
8.06k
                    tablet_schema->column_by_uid(slot->col_unique_id()).name_lower_case(),
554
8.06k
                    slot->column_paths(), slot->col_unique_id(),
555
8.06k
                    dt_variant.variant_max_subcolumns_count(), dt_variant.enable_doc_mode());
556
8.06k
            if (tablet_schema->field_index(*subcol.path_info_ptr()) < 0) {
557
6.20k
                tablet_schema->append_column(subcol, TabletSchema::ColumnType::VARIANT);
558
6.20k
            }
559
8.06k
        }
560
16.3k
    }
561
6.22k
    variant_util::inherit_column_attributes(tablet_schema);
562
6.22k
    return Status::OK();
563
1.18M
}
564
565
1.18M
Status OlapScanner::_init_return_columns() {
566
12.8M
    for (auto* slot : _output_tuple_desc->slots()) {
567
        // variant column using path to index a column
568
12.8M
        int32_t index = 0;
569
12.8M
        auto& tablet_schema = _tablet_reader_params.tablet_schema;
570
12.8M
        if (slot->type()->get_primitive_type() == PrimitiveType::TYPE_VARIANT) {
571
8.06k
            index = tablet_schema->field_index(PathInData(
572
8.06k
                    tablet_schema->column_by_uid(slot->col_unique_id()).name_lower_case(),
573
8.06k
                    slot->column_paths()));
574
12.7M
        } else {
575
12.8M
            index = slot->col_unique_id() >= 0 ? tablet_schema->field_index(slot->col_unique_id())
576
18.4E
                                               : tablet_schema->field_index(slot->col_name());
577
12.7M
        }
578
579
12.8M
        if (index < 0) {
580
0
            return Status::InternalError(
581
0
                    "field name is invalid. field={}, field_name_to_index={}, col_unique_id={}",
582
0
                    slot->col_name(), tablet_schema->get_all_field_names(), slot->col_unique_id());
583
0
        }
584
585
12.8M
        if (slot->get_virtual_column_expr()) {
586
404
            ColumnId virtual_column_cid = index;
587
404
            _virtual_column_exprs[virtual_column_cid] = _slot_id_to_virtual_column_expr[slot->id()];
588
404
            size_t idx_in_block = _slot_id_to_index_in_block[slot->id()];
589
404
            _vir_cid_to_idx_in_block[virtual_column_cid] = idx_in_block;
590
404
            _vir_col_idx_to_type[idx_in_block] = _slot_id_to_col_type[slot->id()];
591
592
404
            VLOG_DEBUG << fmt::format(
593
2
                    "Virtual column, slot id: {}, cid {}, column index: {}, type: {}", slot->id(),
594
2
                    virtual_column_cid, _vir_cid_to_idx_in_block[virtual_column_cid],
595
2
                    _vir_col_idx_to_type[idx_in_block]->get_name());
596
404
        }
597
598
12.8M
        const auto& column = tablet_schema->column(index);
599
12.8M
        int32_t unique_id =
600
12.8M
                column.unique_id() >= 0 ? column.unique_id() : column.parent_unique_id();
601
12.8M
        if (!slot->all_access_paths().empty()) {
602
76.9k
            _tablet_reader_params.all_access_paths.insert({unique_id, slot->all_access_paths()});
603
76.9k
        }
604
605
12.8M
        if (!slot->predicate_access_paths().empty()) {
606
13.5k
            _tablet_reader_params.predicate_access_paths.insert(
607
13.5k
                    {unique_id, slot->predicate_access_paths()});
608
13.5k
        }
609
610
12.8M
        if ((slot->type()->get_primitive_type() == PrimitiveType::TYPE_STRUCT ||
611
12.8M
             slot->type()->get_primitive_type() == PrimitiveType::TYPE_MAP ||
612
12.8M
             slot->type()->get_primitive_type() == PrimitiveType::TYPE_ARRAY) &&
613
12.8M
            !slot->all_access_paths().empty()) {
614
62.3k
            tablet_schema->add_pruned_columns_data_type(column.unique_id(), slot->type());
615
62.3k
        }
616
617
12.8M
        _return_columns.push_back(index);
618
12.8M
        if (slot->is_nullable() && !tablet_schema->column(index).is_nullable()) {
619
0
            _tablet_columns_convert_to_null_set.emplace(index);
620
12.8M
        } else if (!slot->is_nullable() && tablet_schema->column(index).is_nullable()) {
621
0
            return Status::Error<ErrorCode::INVALID_SCHEMA>(
622
0
                    "slot(id: {}, name: {})'s nullable does not match "
623
0
                    "column(tablet id: {}, index: {}, name: {}) ",
624
0
                    slot->id(), slot->col_name(), tablet_schema->table_id(), index,
625
0
                    tablet_schema->column(index).name());
626
0
        }
627
12.8M
    }
628
629
1.18M
    if (_return_columns.empty()) {
630
0
        return Status::InternalError("failed to build storage scanner, no materialized slot!");
631
0
    }
632
633
1.18M
    return Status::OK();
634
1.18M
}
635
636
1.22M
doris::TabletStorageType OlapScanner::get_storage_type() {
637
1.22M
    if (config::is_cloud_mode()) {
638
        // we don't have cold storage in cloud mode, all storage is treated as local
639
886k
        return doris::TabletStorageType::STORAGE_TYPE_LOCAL;
640
886k
    }
641
340k
    int local_reader = 0;
642
520k
    for (const auto& reader : _tablet_reader_params.rs_splits) {
643
520k
        local_reader += reader.rs_reader->rowset()->is_local();
644
520k
    }
645
340k
    int total_reader = _tablet_reader_params.rs_splits.size();
646
647
340k
    if (local_reader == total_reader) {
648
340k
        return doris::TabletStorageType::STORAGE_TYPE_LOCAL;
649
340k
    } else if (local_reader == 0) {
650
0
        return doris::TabletStorageType::STORAGE_TYPE_REMOTE;
651
0
    }
652
1
    return doris::TabletStorageType::STORAGE_TYPE_REMOTE_AND_LOCAL;
653
340k
}
654
655
1.44M
Status OlapScanner::_get_block_impl(RuntimeState* state, Block* block, bool* eof) {
656
    // Read one block from block reader
657
    // ATTN: Here we need to let the _get_block_impl method guarantee the semantics of the interface,
658
    // that is, eof can be set to true only when the returned block is empty.
659
1.44M
    RETURN_IF_ERROR(_tablet_reader->next_block_with_aggregation(block, eof));
660
1.44M
    if (block->rows() > 0) {
661
262k
        _tablet_reader_params.tablet->read_block_count.fetch_add(1, std::memory_order_relaxed);
662
262k
        *eof = false;
663
262k
    }
664
1.44M
#ifndef NDEBUG
665
1.44M
    RETURN_IF_ERROR(_check_ann_cache_hit_debug_points(_tablet_reader->stats()));
666
1.44M
#endif
667
1.44M
    return Status::OK();
668
1.44M
}
669
670
1.18M
Status OlapScanner::close(RuntimeState* state) {
671
1.18M
    if (!_try_close()) {
672
150
        return Status::OK();
673
150
    }
674
1.18M
    RETURN_IF_ERROR(Scanner::close(state));
675
1.18M
    return Status::OK();
676
1.18M
}
677
678
1.22M
void OlapScanner::update_realtime_counters() {
679
1.22M
    if (!_has_prepared) {
680
        // Counter update need prepare successfully, or it maybe core. For example, olap scanner
681
        // will open tablet reader during prepare, if not prepare successfully, tablet reader == nullptr.
682
0
        return;
683
0
    }
684
1.22M
    OlapScanLocalState* local_state = static_cast<OlapScanLocalState*>(_local_state);
685
1.22M
    const OlapReaderStatistics& stats = _tablet_reader->stats();
686
1.22M
    COUNTER_UPDATE(local_state->_read_compressed_counter, stats.compressed_bytes_read);
687
1.22M
    COUNTER_UPDATE(local_state->_read_uncompressed_counter, stats.uncompressed_bytes_read);
688
1.22M
    COUNTER_UPDATE(local_state->_scan_bytes, stats.uncompressed_bytes_read);
689
1.22M
    COUNTER_UPDATE(local_state->_scan_rows, stats.raw_rows_read);
690
691
    // Make sure the scan bytes and scan rows counter in audit log is the same as the counter in
692
    // doris metrics.
693
    // ScanBytes is the uncompressed bytes read from local + remote
694
    // bytes_read_from_local is the compressed bytes read from local
695
    // bytes_read_from_remote is the compressed bytes read from remote
696
    // scan bytes > bytes_read_from_local + bytes_read_from_remote
697
1.22M
    _state->get_query_ctx()->resource_ctx()->io_context()->update_scan_rows(stats.raw_rows_read);
698
1.22M
    _state->get_query_ctx()->resource_ctx()->io_context()->update_scan_bytes(
699
1.22M
            stats.uncompressed_bytes_read);
700
701
    // In case of no cache, we still need to update the IO stats. uncompressed bytes read == local + remote
702
1.22M
    if (stats.file_cache_stats.bytes_read_from_local == 0 &&
703
1.22M
        stats.file_cache_stats.bytes_read_from_remote == 0) {
704
974k
        _state->get_query_ctx()->resource_ctx()->io_context()->update_scan_bytes_from_local_storage(
705
974k
                stats.compressed_bytes_read);
706
974k
        DorisMetrics::instance()->query_scan_bytes_from_local->increment(
707
974k
                stats.compressed_bytes_read);
708
974k
    } else {
709
251k
        _state->get_query_ctx()->resource_ctx()->io_context()->update_scan_bytes_from_local_storage(
710
251k
                stats.file_cache_stats.bytes_read_from_local - _bytes_read_from_local);
711
251k
        _state->get_query_ctx()
712
251k
                ->resource_ctx()
713
251k
                ->io_context()
714
251k
                ->update_scan_bytes_from_remote_storage(
715
251k
                        stats.file_cache_stats.bytes_read_from_remote - _bytes_read_from_remote);
716
717
251k
        DorisMetrics::instance()->query_scan_bytes_from_local->increment(
718
251k
                stats.file_cache_stats.bytes_read_from_local - _bytes_read_from_local);
719
251k
        DorisMetrics::instance()->query_scan_bytes_from_remote->increment(
720
251k
                stats.file_cache_stats.bytes_read_from_remote - _bytes_read_from_remote);
721
251k
    }
722
723
1.22M
    _tablet_reader->mutable_stats()->compressed_bytes_read = 0;
724
1.22M
    _tablet_reader->mutable_stats()->uncompressed_bytes_read = 0;
725
1.22M
    _tablet_reader->mutable_stats()->raw_rows_read = 0;
726
727
1.22M
    _bytes_read_from_local = _tablet_reader->stats().file_cache_stats.bytes_read_from_local;
728
1.22M
    _bytes_read_from_remote = _tablet_reader->stats().file_cache_stats.bytes_read_from_remote;
729
1.22M
}
730
731
1.18M
void OlapScanner::_collect_profile_before_close() {
732
    //  Please don't directly enable the profile here, we need to set QueryStatistics using the counter inside.
733
1.18M
    if (_has_updated_counter) {
734
0
        return;
735
0
    }
736
1.18M
    _has_updated_counter = true;
737
1.18M
    _tablet_reader->update_profile(_profile);
738
739
1.18M
    Scanner::_collect_profile_before_close();
740
741
    // Update counters for OlapScanner
742
    // Update counters from tablet reader's stats
743
1.18M
    auto& stats = _tablet_reader->stats();
744
1.18M
    auto* local_state = (OlapScanLocalState*)_local_state;
745
1.18M
    COUNTER_UPDATE(local_state->_io_timer, stats.io_ns);
746
1.18M
    COUNTER_UPDATE(local_state->_read_compressed_counter, stats.compressed_bytes_read);
747
1.18M
    COUNTER_UPDATE(local_state->_scan_bytes, stats.uncompressed_bytes_read);
748
1.18M
    COUNTER_UPDATE(local_state->_decompressor_timer, stats.decompress_ns);
749
1.18M
    COUNTER_UPDATE(local_state->_read_uncompressed_counter, stats.uncompressed_bytes_read);
750
1.18M
    COUNTER_UPDATE(local_state->_block_load_timer, stats.block_load_ns);
751
1.18M
    COUNTER_UPDATE(local_state->_block_load_counter, stats.blocks_load);
752
1.18M
    COUNTER_UPDATE(local_state->_block_fetch_timer, stats.block_fetch_ns);
753
1.18M
    COUNTER_UPDATE(local_state->_delete_bitmap_get_agg_timer, stats.delete_bitmap_get_agg_ns);
754
1.18M
    COUNTER_UPDATE(local_state->_scan_rows, stats.raw_rows_read);
755
1.18M
    COUNTER_UPDATE(local_state->_vec_cond_timer, stats.vec_cond_ns);
756
1.18M
    COUNTER_UPDATE(local_state->_short_cond_timer, stats.short_cond_ns);
757
1.18M
    COUNTER_UPDATE(local_state->_expr_filter_timer, stats.expr_filter_ns);
758
1.18M
    COUNTER_UPDATE(local_state->_block_init_timer, stats.block_init_ns);
759
1.18M
    COUNTER_UPDATE(local_state->_block_init_seek_timer, stats.block_init_seek_ns);
760
1.18M
    COUNTER_UPDATE(local_state->_block_init_seek_counter, stats.block_init_seek_num);
761
1.18M
    COUNTER_UPDATE(local_state->_segment_generate_row_range_by_keys_timer,
762
1.18M
                   stats.generate_row_ranges_by_keys_ns);
763
1.18M
    COUNTER_UPDATE(local_state->_segment_generate_row_range_by_column_conditions_timer,
764
1.18M
                   stats.generate_row_ranges_by_column_conditions_ns);
765
1.18M
    COUNTER_UPDATE(local_state->_segment_generate_row_range_by_bf_timer,
766
1.18M
                   stats.generate_row_ranges_by_bf_ns);
767
1.18M
    COUNTER_UPDATE(local_state->_collect_iterator_merge_next_timer,
768
1.18M
                   stats.collect_iterator_merge_next_timer);
769
1.18M
    COUNTER_UPDATE(local_state->_segment_generate_row_range_by_zonemap_timer,
770
1.18M
                   stats.generate_row_ranges_by_zonemap_ns);
771
1.18M
    COUNTER_UPDATE(local_state->_segment_generate_row_range_by_dict_timer,
772
1.18M
                   stats.generate_row_ranges_by_dict_ns);
773
1.18M
    COUNTER_UPDATE(local_state->_predicate_column_read_timer, stats.predicate_column_read_ns);
774
1.18M
    COUNTER_UPDATE(local_state->_non_predicate_column_read_timer, stats.non_predicate_read_ns);
775
1.18M
    COUNTER_UPDATE(local_state->_predicate_column_read_seek_timer,
776
1.18M
                   stats.predicate_column_read_seek_ns);
777
1.18M
    COUNTER_UPDATE(local_state->_predicate_column_read_seek_counter,
778
1.18M
                   stats.predicate_column_read_seek_num);
779
1.18M
    COUNTER_UPDATE(local_state->_lazy_read_timer, stats.lazy_read_ns);
780
1.18M
    COUNTER_UPDATE(local_state->_lazy_read_seek_timer, stats.block_lazy_read_seek_ns);
781
1.18M
    COUNTER_UPDATE(local_state->_lazy_read_seek_counter, stats.block_lazy_read_seek_num);
782
1.18M
    COUNTER_UPDATE(local_state->_output_col_timer, stats.output_col_ns);
783
1.18M
    COUNTER_UPDATE(local_state->_rows_vec_cond_filtered_counter, stats.rows_vec_cond_filtered);
784
1.18M
    COUNTER_UPDATE(local_state->_rows_short_circuit_cond_filtered_counter,
785
1.18M
                   stats.rows_short_circuit_cond_filtered);
786
1.18M
    COUNTER_UPDATE(local_state->_rows_expr_cond_filtered_counter, stats.rows_expr_cond_filtered);
787
1.18M
    COUNTER_UPDATE(local_state->_rows_vec_cond_input_counter, stats.vec_cond_input_rows);
788
1.18M
    COUNTER_UPDATE(local_state->_rows_short_circuit_cond_input_counter,
789
1.18M
                   stats.short_circuit_cond_input_rows);
790
1.18M
    COUNTER_UPDATE(local_state->_rows_expr_cond_input_counter, stats.expr_cond_input_rows);
791
1.18M
    COUNTER_UPDATE(local_state->_stats_filtered_counter, stats.rows_stats_filtered);
792
1.18M
    COUNTER_UPDATE(local_state->_stats_rp_filtered_counter, stats.rows_stats_rp_filtered);
793
1.18M
    COUNTER_UPDATE(local_state->_dict_filtered_counter, stats.segment_dict_filtered);
794
1.18M
    COUNTER_UPDATE(local_state->_bf_filtered_counter, stats.rows_bf_filtered);
795
1.18M
    COUNTER_UPDATE(local_state->_del_filtered_counter, stats.rows_del_filtered);
796
1.18M
    COUNTER_UPDATE(local_state->_del_filtered_counter, stats.rows_del_by_bitmap);
797
1.18M
    COUNTER_UPDATE(local_state->_del_filtered_counter, stats.rows_vec_del_cond_filtered);
798
1.18M
    COUNTER_UPDATE(local_state->_conditions_filtered_counter, stats.rows_conditions_filtered);
799
1.18M
    COUNTER_UPDATE(local_state->_key_range_filtered_counter, stats.rows_key_range_filtered);
800
1.18M
    COUNTER_UPDATE(local_state->_total_pages_num_counter, stats.total_pages_num);
801
1.18M
    COUNTER_UPDATE(local_state->_cached_pages_num_counter, stats.cached_pages_num);
802
1.18M
    COUNTER_UPDATE(local_state->_inverted_index_filter_counter, stats.rows_inverted_index_filtered);
803
1.18M
    COUNTER_UPDATE(local_state->_inverted_index_filter_timer, stats.inverted_index_filter_timer);
804
1.18M
    COUNTER_UPDATE(local_state->_inverted_index_query_cache_hit_counter,
805
1.18M
                   stats.inverted_index_query_cache_hit);
806
1.18M
    COUNTER_UPDATE(local_state->_inverted_index_query_cache_miss_counter,
807
1.18M
                   stats.inverted_index_query_cache_miss);
808
1.18M
    COUNTER_UPDATE(local_state->_inverted_index_query_timer, stats.inverted_index_query_timer);
809
1.18M
    COUNTER_UPDATE(local_state->_inverted_index_query_null_bitmap_timer,
810
1.18M
                   stats.inverted_index_query_null_bitmap_timer);
811
1.18M
    COUNTER_UPDATE(local_state->_inverted_index_query_bitmap_copy_timer,
812
1.18M
                   stats.inverted_index_query_bitmap_copy_timer);
813
1.18M
    COUNTER_UPDATE(local_state->_inverted_index_searcher_open_timer,
814
1.18M
                   stats.inverted_index_searcher_open_timer);
815
1.18M
    COUNTER_UPDATE(local_state->_inverted_index_searcher_search_timer,
816
1.18M
                   stats.inverted_index_searcher_search_timer);
817
1.18M
    COUNTER_UPDATE(local_state->_inverted_index_searcher_search_init_timer,
818
1.18M
                   stats.inverted_index_searcher_search_init_timer);
819
1.18M
    COUNTER_UPDATE(local_state->_inverted_index_searcher_search_exec_timer,
820
1.18M
                   stats.inverted_index_searcher_search_exec_timer);
821
1.18M
    COUNTER_UPDATE(local_state->_inverted_index_searcher_cache_hit_counter,
822
1.18M
                   stats.inverted_index_searcher_cache_hit);
823
1.18M
    COUNTER_UPDATE(local_state->_inverted_index_searcher_cache_miss_counter,
824
1.18M
                   stats.inverted_index_searcher_cache_miss);
825
1.18M
    COUNTER_UPDATE(local_state->_inverted_index_downgrade_count_counter,
826
1.18M
                   stats.inverted_index_downgrade_count);
827
1.18M
    COUNTER_UPDATE(local_state->_inverted_index_analyzer_timer,
828
1.18M
                   stats.inverted_index_analyzer_timer);
829
1.18M
    COUNTER_UPDATE(local_state->_inverted_index_lookup_timer, stats.inverted_index_lookup_timer);
830
1.18M
    COUNTER_UPDATE(local_state->_variant_scan_sparse_column_timer,
831
1.18M
                   stats.variant_scan_sparse_column_timer_ns);
832
1.18M
    COUNTER_UPDATE(local_state->_variant_scan_sparse_column_bytes,
833
1.18M
                   stats.variant_scan_sparse_column_bytes);
834
1.18M
    COUNTER_UPDATE(local_state->_variant_fill_path_from_sparse_column_timer,
835
1.18M
                   stats.variant_fill_path_from_sparse_column_timer_ns);
836
1.18M
    COUNTER_UPDATE(local_state->_variant_subtree_default_iter_count,
837
1.18M
                   stats.variant_subtree_default_iter_count);
838
1.18M
    COUNTER_UPDATE(local_state->_variant_subtree_leaf_iter_count,
839
1.18M
                   stats.variant_subtree_leaf_iter_count);
840
1.18M
    COUNTER_UPDATE(local_state->_variant_subtree_hierarchical_iter_count,
841
1.18M
                   stats.variant_subtree_hierarchical_iter_count);
842
1.18M
    COUNTER_UPDATE(local_state->_variant_subtree_sparse_iter_count,
843
1.18M
                   stats.variant_subtree_sparse_iter_count);
844
1.18M
    COUNTER_UPDATE(local_state->_variant_doc_value_column_iter_count,
845
1.18M
                   stats.variant_doc_value_column_iter_count);
846
847
1.18M
    if (stats.adaptive_batch_size_predict_max_rows > 0) {
848
922k
        local_state->_adaptive_batch_predict_min_rows_counter->set(
849
922k
                stats.adaptive_batch_size_predict_min_rows);
850
922k
        local_state->_adaptive_batch_predict_max_rows_counter->set(
851
922k
                stats.adaptive_batch_size_predict_max_rows);
852
922k
    }
853
854
1.18M
    InvertedIndexProfileReporter inverted_index_profile;
855
1.18M
    inverted_index_profile.update(local_state->_index_filter_profile.get(),
856
1.18M
                                  &stats.inverted_index_stats);
857
858
    // only cloud deploy mode will use file cache.
859
1.18M
    if (config::is_cloud_mode() && config::enable_file_cache) {
860
843k
        io::FileCacheProfileReporter cache_profile(local_state->_segment_profile.get());
861
843k
        cache_profile.update(&stats.file_cache_stats);
862
843k
        _state->get_query_ctx()->resource_ctx()->io_context()->update_bytes_write_into_cache(
863
843k
                stats.file_cache_stats.bytes_write_into_cache);
864
843k
    }
865
1.18M
    COUNTER_UPDATE(local_state->_output_index_result_column_timer,
866
1.18M
                   stats.output_index_result_column_timer);
867
1.18M
    COUNTER_UPDATE(local_state->_filtered_segment_counter, stats.filtered_segment_number);
868
1.18M
    COUNTER_UPDATE(local_state->_total_segment_counter, stats.total_segment_number);
869
1.18M
    COUNTER_UPDATE(local_state->_condition_cache_hit_counter, stats.condition_cache_hit_seg_nums);
870
1.18M
    COUNTER_UPDATE(local_state->_condition_cache_filtered_rows_counter,
871
1.18M
                   stats.condition_cache_filtered_rows);
872
873
1.18M
    COUNTER_UPDATE(local_state->_tablet_reader_init_timer, stats.tablet_reader_init_timer_ns);
874
1.18M
    COUNTER_UPDATE(local_state->_tablet_reader_capture_rs_readers_timer,
875
1.18M
                   stats.tablet_reader_capture_rs_readers_timer_ns);
876
1.18M
    COUNTER_UPDATE(local_state->_tablet_reader_init_return_columns_timer,
877
1.18M
                   stats.tablet_reader_init_return_columns_timer_ns);
878
1.18M
    COUNTER_UPDATE(local_state->_tablet_reader_init_keys_param_timer,
879
1.18M
                   stats.tablet_reader_init_keys_param_timer_ns);
880
1.18M
    COUNTER_UPDATE(local_state->_tablet_reader_init_orderby_keys_param_timer,
881
1.18M
                   stats.tablet_reader_init_orderby_keys_param_timer_ns);
882
1.18M
    COUNTER_UPDATE(local_state->_tablet_reader_init_conditions_param_timer,
883
1.18M
                   stats.tablet_reader_init_conditions_param_timer_ns);
884
1.18M
    COUNTER_UPDATE(local_state->_tablet_reader_init_delete_condition_param_timer,
885
1.18M
                   stats.tablet_reader_init_delete_condition_param_timer_ns);
886
1.18M
    COUNTER_UPDATE(local_state->_block_reader_vcollect_iter_init_timer,
887
1.18M
                   stats.block_reader_vcollect_iter_init_timer_ns);
888
1.18M
    COUNTER_UPDATE(local_state->_block_reader_rs_readers_init_timer,
889
1.18M
                   stats.block_reader_rs_readers_init_timer_ns);
890
1.18M
    COUNTER_UPDATE(local_state->_block_reader_build_heap_init_timer,
891
1.18M
                   stats.block_reader_build_heap_init_timer_ns);
892
893
1.18M
    COUNTER_UPDATE(local_state->_rowset_reader_get_segment_iterators_timer,
894
1.18M
                   stats.rowset_reader_get_segment_iterators_timer_ns);
895
1.18M
    COUNTER_UPDATE(local_state->_rowset_reader_create_iterators_timer,
896
1.18M
                   stats.rowset_reader_create_iterators_timer_ns);
897
1.18M
    COUNTER_UPDATE(local_state->_rowset_reader_init_iterators_timer,
898
1.18M
                   stats.rowset_reader_init_iterators_timer_ns);
899
1.18M
    COUNTER_UPDATE(local_state->_rowset_reader_load_segments_timer,
900
1.18M
                   stats.rowset_reader_load_segments_timer_ns);
901
902
1.18M
    COUNTER_UPDATE(local_state->_segment_iterator_init_timer, stats.segment_iterator_init_timer_ns);
903
1.18M
    COUNTER_UPDATE(local_state->_segment_iterator_init_return_column_iterators_timer,
904
1.18M
                   stats.segment_iterator_init_return_column_iterators_timer_ns);
905
1.18M
    COUNTER_UPDATE(local_state->_segment_iterator_init_index_iterators_timer,
906
1.18M
                   stats.segment_iterator_init_index_iterators_timer_ns);
907
1.18M
    COUNTER_UPDATE(local_state->_segment_iterator_init_segment_prefetchers_timer,
908
1.18M
                   stats.segment_iterator_init_segment_prefetchers_timer_ns);
909
910
1.18M
    COUNTER_UPDATE(local_state->_segment_create_column_readers_timer,
911
1.18M
                   stats.segment_create_column_readers_timer_ns);
912
1.18M
    COUNTER_UPDATE(local_state->_segment_load_index_timer, stats.segment_load_index_timer_ns);
913
914
    // Update metrics
915
1.18M
    DorisMetrics::instance()->query_scan_bytes->increment(
916
1.18M
            local_state->_read_uncompressed_counter->value());
917
1.18M
    DorisMetrics::instance()->query_scan_rows->increment(local_state->_scan_rows->value());
918
1.18M
    auto& tablet = _tablet_reader_params.tablet;
919
1.18M
    tablet->query_scan_bytes->increment(local_state->_read_uncompressed_counter->value());
920
1.18M
    tablet->query_scan_rows->increment(local_state->_scan_rows->value());
921
1.18M
    tablet->query_scan_count->increment(1);
922
923
1.18M
    COUNTER_UPDATE(local_state->_ann_range_search_filter_counter,
924
1.18M
                   stats.rows_ann_index_range_filtered);
925
1.18M
    COUNTER_UPDATE(local_state->_ann_topn_filter_counter, stats.rows_ann_index_topn_filtered);
926
1.18M
    COUNTER_UPDATE(local_state->_ann_index_load_costs, stats.ann_index_load_ns);
927
1.18M
    COUNTER_UPDATE(local_state->_ann_ivf_on_disk_load_costs, stats.ann_ivf_on_disk_load_ns);
928
1.18M
    COUNTER_UPDATE(local_state->_ann_ivf_on_disk_cache_hit_cnt,
929
1.18M
                   stats.ann_ivf_on_disk_cache_hit_cnt);
930
1.18M
    COUNTER_UPDATE(local_state->_ann_ivf_on_disk_cache_miss_cnt,
931
1.18M
                   stats.ann_ivf_on_disk_cache_miss_cnt);
932
1.18M
    COUNTER_UPDATE(local_state->_ann_range_search_costs, stats.ann_index_range_search_ns);
933
1.18M
    COUNTER_UPDATE(local_state->_ann_range_search_cnt, stats.ann_index_range_search_cnt);
934
1.18M
    COUNTER_UPDATE(local_state->_ann_range_engine_search_costs, stats.ann_range_engine_search_ns);
935
    // Engine prepare before search
936
1.18M
    COUNTER_UPDATE(local_state->_ann_range_pre_process_costs, stats.ann_range_pre_process_ns);
937
    // Post process parent: Doris result process + engine convert
938
1.18M
    COUNTER_UPDATE(local_state->_ann_range_post_process_costs,
939
1.18M
                   stats.ann_range_result_convert_ns + stats.ann_range_engine_convert_ns);
940
    // Engine convert (child under post-process)
941
1.18M
    COUNTER_UPDATE(local_state->_ann_range_engine_convert_costs, stats.ann_range_engine_convert_ns);
942
    // Doris-side result convert (child under post-process)
943
1.18M
    COUNTER_UPDATE(local_state->_ann_range_result_convert_costs, stats.ann_range_result_convert_ns);
944
945
1.18M
    COUNTER_UPDATE(local_state->_ann_topn_search_costs, stats.ann_topn_search_ns);
946
1.18M
    COUNTER_UPDATE(local_state->_ann_topn_search_cnt, stats.ann_index_topn_search_cnt);
947
1.18M
    COUNTER_UPDATE(local_state->_ann_cache_hit_cnt, stats.ann_index_cache_hits);
948
1.18M
    COUNTER_UPDATE(local_state->_ann_range_cache_hit_cnt, stats.ann_index_range_cache_hits);
949
950
    // Detailed ANN timers
951
    // ANN TopN timers with hierarchy
952
    // Engine search time (FAISS)
953
1.18M
    COUNTER_UPDATE(local_state->_ann_topn_engine_search_costs,
954
1.18M
                   stats.ann_index_topn_engine_search_ns);
955
    // Engine prepare time (allocations/buffer setup before search)
956
1.18M
    COUNTER_UPDATE(local_state->_ann_topn_pre_process_costs,
957
1.18M
                   stats.ann_index_topn_engine_prepare_ns);
958
    // Post process parent includes Doris result processing + engine convert
959
1.18M
    COUNTER_UPDATE(local_state->_ann_topn_post_process_costs,
960
1.18M
                   stats.ann_index_topn_result_process_ns + stats.ann_index_topn_engine_convert_ns);
961
    // Engine-side conversion time inside FAISS wrappers (child under post-process)
962
1.18M
    COUNTER_UPDATE(local_state->_ann_topn_engine_convert_costs,
963
1.18M
                   stats.ann_index_topn_engine_convert_ns);
964
965
    // Doris-side result convert costs (show separately as another child counter); use pure process time
966
1.18M
    COUNTER_UPDATE(local_state->_ann_topn_result_convert_costs,
967
1.18M
                   stats.ann_index_topn_result_process_ns);
968
969
1.18M
    COUNTER_UPDATE(local_state->_ann_fallback_brute_force_cnt, stats.ann_fall_back_brute_force_cnt);
970
971
    // Overhead counter removed; precise instrumentation is reported via engine_prepare above.
972
1.18M
}
973
974
#ifndef NDEBUG
975
1.44M
Status OlapScanner::_check_ann_cache_hit_debug_points(const OlapReaderStatistics& stats) {
976
1.44M
    DBUG_EXECUTE_IF("olap_scanner.ann_topn_cache_hits", {
977
1.44M
        auto expected_hits = dp->param<int32_t>("expected_hits", -1);
978
1.44M
        auto min_hits = dp->param<int32_t>("min_hits", -1);
979
1.44M
        if (expected_hits >= 0 && stats.ann_index_cache_hits != expected_hits) {
980
1.44M
            return Status::Error<ErrorCode::INTERNAL_ERROR>(
981
1.44M
                    "ann_index_cache_hits: {} not equal to expected: {}",
982
1.44M
                    stats.ann_index_cache_hits, expected_hits);
983
1.44M
        }
984
1.44M
        if (min_hits >= 0 && stats.ann_index_cache_hits < min_hits) {
985
1.44M
            return Status::Error<ErrorCode::INTERNAL_ERROR>(
986
1.44M
                    "ann_index_cache_hits: {} less than expected min: {}",
987
1.44M
                    stats.ann_index_cache_hits, min_hits);
988
1.44M
        }
989
1.44M
    })
990
1.44M
    DBUG_EXECUTE_IF("olap_scanner.ann_range_cache_hits", {
991
1.44M
        auto expected_hits = dp->param<int32_t>("expected_hits", -1);
992
1.44M
        auto min_hits = dp->param<int32_t>("min_hits", -1);
993
1.44M
        if (expected_hits >= 0 && stats.ann_index_range_cache_hits != expected_hits) {
994
1.44M
            return Status::Error<ErrorCode::INTERNAL_ERROR>(
995
1.44M
                    "ann_index_range_cache_hits: {} not equal to expected: {}",
996
1.44M
                    stats.ann_index_range_cache_hits, expected_hits);
997
1.44M
        }
998
1.44M
        if (min_hits >= 0 && stats.ann_index_range_cache_hits < min_hits) {
999
1.44M
            return Status::Error<ErrorCode::INTERNAL_ERROR>(
1000
1.44M
                    "ann_index_range_cache_hits: {} less than expected min: {}",
1001
1.44M
                    stats.ann_index_range_cache_hits, min_hits);
1002
1.44M
        }
1003
1.44M
    })
1004
1.44M
    return Status::OK();
1005
1.44M
}
1006
#endif
1007
1008
#include "common/compile_check_avoid_end.h"
1009
} // namespace doris