Coverage Report

Created: 2026-05-18 13:00

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
337k
        : Scanner(params.state, parent, params.limit, params.profile),
74
337k
          _key_ranges(std::move(params.key_ranges)),
75
337k
          _tablet_reader_params({.tablet = std::move(params.tablet),
76
337k
                                 .tablet_schema {},
77
337k
                                 .reader_type = params.read_row_binlog ? ReaderType::READER_BINLOG
78
337k
                                                                       : ReaderType::READER_QUERY,
79
337k
                                 .aggregation = params.aggregation,
80
337k
                                 .version = {0, params.version},
81
337k
                                 .start_key {},
82
337k
                                 .end_key {},
83
337k
                                 .predicates {},
84
337k
                                 .function_filters {},
85
337k
                                 .delete_predicates {},
86
337k
                                 .target_cast_type_for_variants {},
87
337k
                                 .all_access_paths {},
88
337k
                                 .predicate_access_paths {},
89
337k
                                 .rs_splits {},
90
337k
                                 .return_columns {},
91
337k
                                 .output_columns {},
92
337k
                                 .remaining_conjunct_roots {},
93
337k
                                 .common_expr_ctxs_push_down {},
94
337k
                                 .topn_filter_source_node_ids {},
95
337k
                                 .key_group_cluster_key_idxes {},
96
337k
                                 .virtual_column_exprs {},
97
337k
                                 .vir_cid_to_idx_in_block {},
98
337k
                                 .vir_col_idx_to_type {},
99
337k
                                 .score_runtime {},
100
337k
                                 .collection_statistics {},
101
337k
                                 .ann_topn_runtime {},
102
337k
                                 .condition_cache_digest = parent->get_condition_cache_digest()}) {
103
337k
    _tablet_reader_params.set_read_source(std::move(params.read_source),
104
337k
                                          _state->skip_delete_bitmap());
105
337k
    _has_prepared = false;
106
337k
    _vector_search_params = params.state->get_vector_search_params();
107
337k
}
108
109
static std::string read_columns_to_string(TabletSchemaSPtr tablet_schema,
110
0
                                          const std::vector<uint32_t>& read_columns) {
111
    // avoid too long for one line,
112
    // it is hard to display in `show profile` stmt if one line is too long.
113
0
    const int col_per_line = 10;
114
0
    int i = 0;
115
0
    std::string read_columns_string;
116
0
    read_columns_string += "[";
117
0
    for (auto it = read_columns.cbegin(); it != read_columns.cend(); it++) {
118
0
        if (it != read_columns.cbegin()) {
119
0
            read_columns_string += ", ";
120
0
        }
121
0
        read_columns_string += tablet_schema->columns().at(*it)->name();
122
0
        if (i >= col_per_line) {
123
0
            read_columns_string += "\n";
124
0
            i = 0;
125
0
        } else {
126
0
            ++i;
127
0
        }
128
0
    }
129
0
    read_columns_string += "]";
130
0
    return read_columns_string;
131
0
}
132
133
337k
Status OlapScanner::_prepare_impl() {
134
337k
    auto* local_state = static_cast<OlapScanLocalState*>(_local_state);
135
337k
    auto& tablet = _tablet_reader_params.tablet;
136
337k
    auto& tablet_schema = _tablet_reader_params.tablet_schema;
137
337k
    DBUG_EXECUTE_IF("CloudTablet.capture_rs_readers.return.e-230", {
138
337k
        LOG_WARNING("CloudTablet.capture_rs_readers.return e-230 init")
139
337k
                .tag("tablet_id", tablet->tablet_id());
140
337k
        return Status::Error<false>(-230, "injected error");
141
337k
    });
142
143
337k
    for (auto& ctx : local_state->_common_expr_ctxs_push_down) {
144
131
        VExprContextSPtr context;
145
131
        RETURN_IF_ERROR(ctx->clone(_state, context));
146
131
        _common_expr_ctxs_push_down.emplace_back(context);
147
131
        context->prepare_ann_range_search(_vector_search_params);
148
131
    }
149
150
337k
    for (auto pair : local_state->_slot_id_to_virtual_column_expr) {
151
        // Scanner will be executed in a different thread, so we need to clone the context.
152
0
        VExprContextSPtr context;
153
0
        RETURN_IF_ERROR(pair.second->clone(_state, context));
154
0
        _slot_id_to_virtual_column_expr[pair.first] = context;
155
0
    }
156
157
337k
    _slot_id_to_index_in_block = local_state->_slot_id_to_index_in_block;
158
337k
    _slot_id_to_col_type = local_state->_slot_id_to_col_type;
159
337k
    _score_runtime = local_state->_score_runtime;
160
    // All scanners share the same ann_topn_runtime.
161
337k
    _ann_topn_runtime = local_state->_ann_topn_runtime;
162
163
    // set limit to reduce end of rowset and segment mem use
164
337k
    _tablet_reader = std::make_unique<BlockReader>();
165
    // batch size is passed down to segment iterator, use _state->batch_size()
166
    // instead of _parent->limit(), because if _parent->limit() is a very small
167
    // value (e.g. select a from t where a .. and b ... limit 1),
168
    // it will be very slow when reading data in segment iterator
169
337k
    _tablet_reader->set_batch_size(_state->batch_size());
170
    // Adaptive batch size: pass byte-budget settings to the storage reader.
171
    // The reader still uses batch_size() as the row ceiling.
172
337k
    _tablet_reader->set_preferred_block_size_bytes(_state->preferred_block_size_bytes());
173
337k
    {
174
337k
        TOlapScanNode& olap_scan_node = local_state->olap_scan_node();
175
337k
        TabletSchemaSPtr source_tablet_schema =
176
337k
                _tablet_reader_params.reader_type == ReaderType::READER_BINLOG
177
337k
                        ? tablet->row_binlog_tablet_schema()
178
337k
                        : tablet->tablet_schema();
179
180
337k
        tablet_schema = std::make_shared<TabletSchema>();
181
337k
        tablet_schema->copy_from(*source_tablet_schema);
182
337k
        if (olap_scan_node.__isset.columns_desc && !olap_scan_node.columns_desc.empty() &&
183
337k
            olap_scan_node.columns_desc[0].col_unique_id >= 0) {
184
337k
            tablet_schema->clear_columns();
185
5.70M
            for (const auto& column_desc : olap_scan_node.columns_desc) {
186
5.70M
                tablet_schema->append_column(TabletColumn(column_desc));
187
5.70M
            }
188
337k
            if (olap_scan_node.__isset.schema_version) {
189
337k
                tablet_schema->set_schema_version(olap_scan_node.schema_version);
190
337k
            }
191
337k
        }
192
337k
        if (olap_scan_node.__isset.indexes_desc) {
193
337k
            tablet_schema->update_indexes_from_thrift(olap_scan_node.indexes_desc);
194
337k
        }
195
196
337k
        if (_tablet_reader_params.rs_splits.empty()) {
197
            // Non-pipeline mode, Tablet : Scanner = 1 : 1
198
            // acquire tablet rowset readers at the beginning of the scan node
199
            // to prevent this case: when there are lots of olap scanners to run for example 10000
200
            // the rowsets maybe compacted when the last olap scanner starts
201
0
            ReadSource read_source;
202
203
0
            if (config::is_cloud_mode()) {
204
                // FIXME(plat1ko): Avoid pointer cast
205
0
                ExecEnv::GetInstance()->storage_engine().to_cloud().tablet_hotspot().count(*tablet);
206
0
            }
207
208
0
            auto maybe_read_source = tablet->capture_read_source(
209
0
                    _tablet_reader_params.version,
210
0
                    {
211
0
                            .skip_missing_versions = _state->skip_missing_version(),
212
0
                            .enable_fetch_rowsets_from_peers =
213
0
                                    config::enable_fetch_rowsets_from_peer_replicas,
214
0
                            .capture_row_binlog =
215
0
                                    _tablet_reader_params.reader_type == ReaderType::READER_BINLOG,
216
0
                            .enable_prefer_cached_rowset =
217
0
                                    config::is_cloud_mode() ? _state->enable_prefer_cached_rowset()
218
0
                                                            : false,
219
0
                            .query_freshness_tolerance_ms =
220
0
                                    config::is_cloud_mode() ? _state->query_freshness_tolerance_ms()
221
0
                                                            : -1,
222
0
                    });
223
0
            if (!maybe_read_source) {
224
0
                LOG(WARNING) << "fail to init reader. res=" << maybe_read_source.error();
225
0
                return maybe_read_source.error();
226
0
            }
227
0
            read_source = std::move(maybe_read_source.value());
228
229
0
            if (config::enable_mow_verbose_log && tablet->enable_unique_key_merge_on_write()) {
230
0
                LOG_INFO("finish capture_rs_readers for tablet={}, query_id={}",
231
0
                         tablet->tablet_id(), print_id(_state->query_id()));
232
0
            }
233
234
0
            if (!_state->skip_delete_predicate()) {
235
0
                read_source.fill_delete_predicates();
236
0
            }
237
0
            _tablet_reader_params.set_read_source(std::move(read_source));
238
0
        }
239
240
        // Initialize tablet_reader_params
241
337k
        RETURN_IF_ERROR(_init_tablet_reader_params(
242
337k
                local_state->_parent->cast<OlapScanOperatorX>()._slot_id_to_slot_desc, _key_ranges,
243
337k
                local_state->_slot_id_to_predicates, local_state->_push_down_functions));
244
337k
    }
245
246
    // add read columns in profile
247
337k
    if (_state->enable_profile()) {
248
0
        _profile->add_info_string("ReadColumns",
249
0
                                  read_columns_to_string(tablet_schema, _return_columns));
250
0
    }
251
252
337k
    if (_tablet_reader_params.score_runtime) {
253
0
        SCOPED_TIMER(local_state->_statistics_collect_timer);
254
0
        _tablet_reader_params.collection_statistics = std::make_shared<CollectionStatistics>();
255
256
0
        io::IOContext io_ctx {
257
0
                .reader_type = _tablet_reader_params.reader_type,
258
0
                .expiration_time = tablet->ttl_seconds(),
259
0
                .query_id = &_state->query_id(),
260
0
                .file_cache_stats = &_tablet_reader->mutable_stats()->file_cache_stats,
261
0
                .is_inverted_index = true,
262
0
        };
263
264
0
        RETURN_IF_ERROR(_tablet_reader_params.collection_statistics->collect(
265
0
                _state, _tablet_reader_params.rs_splits, _tablet_reader_params.tablet_schema,
266
0
                _tablet_reader_params.common_expr_ctxs_push_down, &io_ctx));
267
0
    }
268
269
337k
    _has_prepared = true;
270
337k
    return Status::OK();
271
337k
}
272
273
337k
Status OlapScanner::_open_impl(RuntimeState* state) {
274
337k
    RETURN_IF_ERROR(Scanner::_open_impl(state));
275
337k
    SCOPED_TIMER(_local_state->cast<OlapScanLocalState>()._reader_init_timer);
276
277
337k
    auto res = _tablet_reader->init(_tablet_reader_params);
278
337k
    if (!res.ok()) {
279
0
        res.append("failed to initialize storage reader. tablet=" +
280
0
                   std::to_string(_tablet_reader_params.tablet->tablet_id()) +
281
0
                   ", backend=" + BackendOptions::get_localhost());
282
0
        return res;
283
0
    }
284
285
    // Do not hold rs_splits any more to release memory.
286
337k
    _tablet_reader_params.rs_splits.clear();
287
288
337k
    return Status::OK();
289
337k
}
290
291
// it will be called under tablet read lock because capture rs readers need
292
Status OlapScanner::_init_tablet_reader_params(
293
        const phmap::flat_hash_map<int, SlotDescriptor*>& slot_id_to_slot_desc,
294
        const std::vector<OlapScanRange*>& key_ranges,
295
        const phmap::flat_hash_map<int, std::vector<std::shared_ptr<ColumnPredicate>>>&
296
                slot_to_predicates,
297
337k
        const std::vector<FunctionFilter>& function_filters) {
298
    // if the table with rowset [0-x] or [0-1] [2-y], and [0-1] is empty
299
337k
    const bool single_version = _tablet_reader_params.has_single_version();
300
301
337k
    auto* olap_local_state = static_cast<OlapScanLocalState*>(_local_state);
302
337k
    bool read_mor_as_dup = olap_local_state->olap_scan_node().__isset.read_mor_as_dup &&
303
337k
                           olap_local_state->olap_scan_node().read_mor_as_dup;
304
337k
    if (_state->skip_storage_engine_merge() || read_mor_as_dup) {
305
0
        _tablet_reader_params.direct_mode = true;
306
0
        _tablet_reader_params.aggregation = true;
307
337k
    } else {
308
337k
        auto push_down_agg_type = _local_state->get_push_down_agg_type();
309
337k
        _tablet_reader_params.direct_mode = _tablet_reader_params.aggregation || single_version ||
310
337k
                                            (push_down_agg_type != TPushAggOp::NONE &&
311
598
                                             push_down_agg_type != TPushAggOp::COUNT_ON_INDEX);
312
337k
    }
313
314
337k
    RETURN_IF_ERROR(_init_variant_columns());
315
337k
    RETURN_IF_ERROR(_init_return_columns());
316
317
337k
    _tablet_reader_params.push_down_agg_type_opt = _local_state->get_push_down_agg_type();
318
319
    // TODO: If a new runtime filter arrives after `_conjuncts` move to `_common_expr_ctxs_push_down`,
320
337k
    if (_common_expr_ctxs_push_down.empty()) {
321
337k
        for (auto& conjunct : _conjuncts) {
322
0
            _tablet_reader_params.remaining_conjunct_roots.emplace_back(conjunct->root());
323
0
        }
324
337k
    } else {
325
157
        for (auto& ctx : _common_expr_ctxs_push_down) {
326
129
            _tablet_reader_params.remaining_conjunct_roots.emplace_back(ctx->root());
327
129
        }
328
157
    }
329
330
337k
    _tablet_reader_params.common_expr_ctxs_push_down = _common_expr_ctxs_push_down;
331
337k
    _tablet_reader_params.virtual_column_exprs = _virtual_column_exprs;
332
337k
    _tablet_reader_params.vir_cid_to_idx_in_block = _vir_cid_to_idx_in_block;
333
337k
    _tablet_reader_params.vir_col_idx_to_type = _vir_col_idx_to_type;
334
337k
    _tablet_reader_params.score_runtime = _score_runtime;
335
337k
    _tablet_reader_params.output_columns = ((OlapScanLocalState*)_local_state)->_output_column_ids;
336
337k
    _tablet_reader_params.ann_topn_runtime = _ann_topn_runtime;
337
337k
    for (const auto& ele : ((OlapScanLocalState*)_local_state)->_cast_types_for_variants) {
338
4
        _tablet_reader_params.target_cast_type_for_variants[ele.first] = ele.second;
339
4
    };
340
337k
    auto& tablet_schema = _tablet_reader_params.tablet_schema;
341
3.69M
    for (auto& predicates : slot_to_predicates) {
342
3.69M
        const int sid = predicates.first;
343
3.69M
        DCHECK(slot_id_to_slot_desc.contains(sid));
344
3.69M
        int32_t index =
345
3.69M
                tablet_schema->field_index(slot_id_to_slot_desc.find(sid)->second->col_name());
346
3.69M
        if (index < 0) {
347
0
            throw Exception(
348
0
                    Status::InternalError("Column {} not found in tablet schema",
349
0
                                          slot_id_to_slot_desc.find(sid)->second->col_name()));
350
0
        }
351
3.69M
        for (auto& predicate : predicates.second) {
352
313k
            _tablet_reader_params.predicates.push_back(predicate->clone(index));
353
313k
        }
354
3.69M
    }
355
356
337k
    std::copy(function_filters.cbegin(), function_filters.cend(),
357
337k
              std::inserter(_tablet_reader_params.function_filters,
358
337k
                            _tablet_reader_params.function_filters.begin()));
359
360
    // Merge the columns in delete predicate that not in latest schema in to current tablet schema
361
337k
    for (auto& del_pred : _tablet_reader_params.delete_predicates) {
362
18
        tablet_schema->merge_dropped_columns(*del_pred->tablet_schema());
363
18
    }
364
365
    // Push key ranges to the tablet reader.
366
    // Skip the "full scan" placeholder (has_lower_bound == false) — when no key
367
    // predicates exist, start_key/end_key remain empty and the reader does a full scan.
368
355k
    for (auto* key_range : key_ranges) {
369
355k
        if (!key_range->has_lower_bound) {
370
23.4k
            continue;
371
23.4k
        }
372
373
332k
        _tablet_reader_params.start_key_include = key_range->begin_include;
374
332k
        _tablet_reader_params.end_key_include = key_range->end_include;
375
376
332k
        _tablet_reader_params.start_key.push_back(key_range->begin_scan_range);
377
332k
        _tablet_reader_params.end_key.push_back(key_range->end_scan_range);
378
332k
    }
379
380
337k
    _tablet_reader_params.profile = _local_state->custom_profile();
381
337k
    _tablet_reader_params.runtime_state = _state;
382
383
337k
    _tablet_reader_params.origin_return_columns = &_return_columns;
384
337k
    _tablet_reader_params.tablet_columns_convert_to_null_set = &_tablet_columns_convert_to_null_set;
385
386
337k
    if (_tablet_reader_params.direct_mode) {
387
337k
        _tablet_reader_params.return_columns = _return_columns;
388
337k
    } else {
389
        // we need to fetch all key columns to do the right aggregation on storage engine side.
390
1.89k
        for (size_t i = 0; i < tablet_schema->num_key_columns(); ++i) {
391
1.24k
            _tablet_reader_params.return_columns.push_back(i);
392
1.24k
        }
393
1.83k
        for (auto index : _return_columns) {
394
1.83k
            if (tablet_schema->column(index).is_key()) {
395
1.23k
                continue;
396
1.23k
            }
397
596
            _tablet_reader_params.return_columns.push_back(index);
398
596
        }
399
        // expand the sequence column
400
656
        if (tablet_schema->has_sequence_col() || tablet_schema->has_seq_map()) {
401
0
            bool has_replace_col = false;
402
0
            for (auto col : _return_columns) {
403
0
                if (tablet_schema->column(col).aggregation() ==
404
0
                    FieldAggregationMethod::OLAP_FIELD_AGGREGATION_REPLACE) {
405
0
                    has_replace_col = true;
406
0
                    break;
407
0
                }
408
0
            }
409
0
            if (auto sequence_col_idx = tablet_schema->sequence_col_idx();
410
0
                has_replace_col && tablet_schema->has_sequence_col() &&
411
0
                std::find(_return_columns.begin(), _return_columns.end(), sequence_col_idx) ==
412
0
                        _return_columns.end()) {
413
0
                _tablet_reader_params.return_columns.push_back(sequence_col_idx);
414
0
            }
415
0
            if (has_replace_col) {
416
0
                const auto& val_to_seq = tablet_schema->value_col_idx_to_seq_col_idx();
417
0
                std::set<uint32_t> return_seq_columns;
418
419
0
                for (auto col : _tablet_reader_params.return_columns) {
420
                    // we need to add the necessary sequence column in _return_columns, and
421
                    // Avoid adding the same seq column twice
422
0
                    const auto val_iter = val_to_seq.find(col);
423
0
                    if (val_iter != val_to_seq.end()) {
424
0
                        auto seq = val_iter->second;
425
0
                        if (std::find(_tablet_reader_params.return_columns.begin(),
426
0
                                      _tablet_reader_params.return_columns.end(),
427
0
                                      seq) == _tablet_reader_params.return_columns.end()) {
428
0
                            return_seq_columns.insert(seq);
429
0
                        }
430
0
                    }
431
0
                }
432
0
                _tablet_reader_params.return_columns.insert(
433
0
                        std::end(_tablet_reader_params.return_columns),
434
0
                        std::begin(return_seq_columns), std::end(return_seq_columns));
435
0
            }
436
0
        }
437
656
    }
438
439
337k
    _tablet_reader_params.use_page_cache = _state->enable_page_cache();
440
441
337k
    DBUG_EXECUTE_IF("NewOlapScanner::_init_tablet_reader_params.block", DBUG_BLOCK);
442
443
337k
    if (!_state->skip_storage_engine_merge()) {
444
337k
        auto* olap_scan_local_state = (OlapScanLocalState*)_local_state;
445
337k
        TOlapScanNode& olap_scan_node = olap_scan_local_state->olap_scan_node();
446
447
        // Set MOR value predicate pushdown flag
448
337k
        if (olap_scan_node.__isset.enable_mor_value_predicate_pushdown &&
449
337k
            olap_scan_node.enable_mor_value_predicate_pushdown) {
450
0
            _tablet_reader_params.enable_mor_value_predicate_pushdown = true;
451
0
        }
452
453
337k
        const bool has_key_topn =
454
337k
                olap_scan_node.__isset.sort_info && !olap_scan_node.sort_info.is_asc_order.empty();
455
337k
        if (has_key_topn) {
456
0
            _limit = _local_state->limit_per_scanner();
457
0
        }
458
459
337k
        const bool no_runtime_filters = _total_rf_num == 0;
460
337k
        const bool segment_limit_enabled = _state->enable_segment_limit_pushdown();
461
337k
        const bool storage_no_merge = olap_scan_local_state->_storage_no_merge();
462
463
337k
        if (_limit > 0 && no_runtime_filters && segment_limit_enabled && storage_no_merge) {
464
40
            for (const auto& conjunct : _conjuncts) {
465
0
                DORIS_CHECK(!olap_scan_local_state->_check_expr_storage_filter(
466
0
                        conjunct->root(), OlapScanLocalState::ExprStorageFilterCheckMode::
467
0
                                                  HAS_SEGMENT_EVALUABLE_EXPR));
468
0
            }
469
40
        }
470
471
        // Segment LIMIT has only two legal states: completely disabled, or enabled after every
472
        // row-filtering conjunct has become a storage predicate or SegmentIterator common expr.
473
337k
        const bool can_push_down_segment_limit = _limit > 0 && no_runtime_filters &&
474
337k
                                                 _conjuncts.empty() && segment_limit_enabled &&
475
337k
                                                 storage_no_merge;
476
337k
        if (can_push_down_segment_limit) {
477
40
            if (has_key_topn) {
478
0
                _tablet_reader_params.read_orderby_key = true;
479
0
                if (!olap_scan_node.sort_info.is_asc_order[0]) {
480
0
                    _tablet_reader_params.read_orderby_key_reverse = true;
481
0
                }
482
0
                _tablet_reader_params.read_orderby_key_num_prefix_columns =
483
0
                        olap_scan_node.sort_info.is_asc_order.size();
484
0
                _tablet_reader_params.read_orderby_key_limit = _limit;
485
40
            } else {
486
40
                _tablet_reader_params.general_read_limit = _limit;
487
40
            }
488
40
        }
489
490
337k
        if (_tablet_reader_params.read_orderby_key_limit > 0 ||
491
337k
            _tablet_reader_params.general_read_limit > 0) {
492
40
            DORIS_CHECK(can_push_down_segment_limit);
493
40
            DORIS_CHECK(_conjuncts.empty());
494
40
        }
495
496
        // A key TopN scan cannot share the plain LIMIT early-stop counter. If
497
        // storage TopN is pushed down, each scanner must produce its full local
498
        // candidates. If it is not pushed down for any reason, the upper TopN
499
        // still needs all rows from the scan.
500
337k
        if (has_key_topn) {
501
0
            _shared_scan_limit = nullptr;
502
0
            if (_tablet_reader_params.read_orderby_key_limit == 0) {
503
0
                _limit = -1;
504
0
            }
505
0
        }
506
        // Note: _shared_scan_limit is intentionally not pushed into the
507
        // storage layer. SegmentIterator's _process_eof() is irreversible,
508
        // so a concurrently-decremented atomic could reach 0 while a segment
509
        // still has data needed by other scanners.
510
511
        // set push down topn filter
512
337k
        _tablet_reader_params.topn_filter_source_node_ids =
513
337k
                olap_scan_local_state->get_topn_filter_source_node_ids(_state, true);
514
337k
        if (!_tablet_reader_params.topn_filter_source_node_ids.empty()) {
515
0
            _tablet_reader_params.topn_filter_target_node_id =
516
0
                    olap_scan_local_state->parent()->node_id();
517
0
        }
518
337k
    }
519
520
    // If this is a Two-Phase read query, and we need to delay the release of Rowset
521
    // by rowset->update_delayed_expired_timestamp().This could expand the lifespan of Rowset
522
337k
    if (tablet_schema->field_index(BeConsts::ROWID_COL) >= 0) {
523
0
        constexpr static int delayed_s = 60;
524
0
        for (auto rs_reader : _tablet_reader_params.rs_splits) {
525
0
            uint64_t delayed_expired_timestamp =
526
0
                    UnixSeconds() + _tablet_reader_params.runtime_state->execution_timeout() +
527
0
                    delayed_s;
528
0
            rs_reader.rs_reader->rowset()->update_delayed_expired_timestamp(
529
0
                    delayed_expired_timestamp);
530
0
            ExecEnv::GetInstance()->storage_engine().add_quering_rowset(
531
0
                    rs_reader.rs_reader->rowset());
532
0
        }
533
0
    }
534
535
337k
    if (tablet_schema->has_global_row_id()) {
536
6
        auto& id_file_map = _state->get_id_file_map();
537
6
        for (auto rs_reader : _tablet_reader_params.rs_splits) {
538
6
            id_file_map->add_temp_rowset(rs_reader.rs_reader->rowset());
539
6
        }
540
6
    }
541
542
337k
    return Status::OK();
543
337k
}
544
545
337k
Status OlapScanner::_init_variant_columns() {
546
337k
    auto& tablet_schema = _tablet_reader_params.tablet_schema;
547
337k
    if (tablet_schema->num_variant_columns() == 0) {
548
337k
        return Status::OK();
549
337k
    }
550
    // Parent column has path info to distinction from each other
551
18.4E
    for (auto* slot : _output_tuple_desc->slots()) {
552
60
        if (slot->type()->get_primitive_type() == PrimitiveType::TYPE_VARIANT) {
553
            // Such columns are not exist in frontend schema info, so we need to
554
            // add them into tablet_schema for later column indexing.
555
30
            const auto& dt_variant =
556
30
                    assert_cast<const DataTypeVariant&>(*remove_nullable(slot->type()));
557
30
            TabletColumn subcol = TabletColumn::create_materialized_variant_column(
558
30
                    tablet_schema->column_by_uid(slot->col_unique_id()).name_lower_case(),
559
30
                    slot->column_paths(), slot->col_unique_id(),
560
30
                    dt_variant.variant_max_subcolumns_count(), dt_variant.enable_doc_mode());
561
30
            if (tablet_schema->field_index(*subcol.path_info_ptr()) < 0) {
562
18
                tablet_schema->append_column(subcol, TabletSchema::ColumnType::VARIANT);
563
18
            }
564
30
        }
565
60
    }
566
18.4E
    variant_util::inherit_column_attributes(tablet_schema);
567
18.4E
    return Status::OK();
568
337k
}
569
570
337k
Status OlapScanner::_init_return_columns() {
571
4.91M
    for (auto* slot : _output_tuple_desc->slots()) {
572
        // variant column using path to index a column
573
4.91M
        int32_t index = 0;
574
4.91M
        auto& tablet_schema = _tablet_reader_params.tablet_schema;
575
4.91M
        if (slot->type()->get_primitive_type() == PrimitiveType::TYPE_VARIANT) {
576
30
            index = tablet_schema->field_index(PathInData(
577
30
                    tablet_schema->column_by_uid(slot->col_unique_id()).name_lower_case(),
578
30
                    slot->column_paths()));
579
4.91M
        } else {
580
4.91M
            index = slot->col_unique_id() >= 0 ? tablet_schema->field_index(slot->col_unique_id())
581
18.4E
                                               : tablet_schema->field_index(slot->col_name());
582
4.91M
        }
583
584
4.91M
        if (index < 0) {
585
0
            return Status::InternalError(
586
0
                    "field name is invalid. field={}, field_name_to_index={}, col_unique_id={}",
587
0
                    slot->col_name(), tablet_schema->get_all_field_names(), slot->col_unique_id());
588
0
        }
589
590
4.91M
        if (slot->get_virtual_column_expr()) {
591
0
            ColumnId virtual_column_cid = index;
592
0
            _virtual_column_exprs[virtual_column_cid] = _slot_id_to_virtual_column_expr[slot->id()];
593
0
            size_t idx_in_block = _slot_id_to_index_in_block[slot->id()];
594
0
            _vir_cid_to_idx_in_block[virtual_column_cid] = idx_in_block;
595
0
            _vir_col_idx_to_type[idx_in_block] = _slot_id_to_col_type[slot->id()];
596
597
0
            VLOG_DEBUG << fmt::format(
598
0
                    "Virtual column, slot id: {}, cid {}, column index: {}, type: {}", slot->id(),
599
0
                    virtual_column_cid, _vir_cid_to_idx_in_block[virtual_column_cid],
600
0
                    _vir_col_idx_to_type[idx_in_block]->get_name());
601
0
        }
602
603
4.91M
        const auto& column = tablet_schema->column(index);
604
4.91M
        int32_t unique_id =
605
18.4E
                column.unique_id() >= 0 ? column.unique_id() : column.parent_unique_id();
606
4.91M
        if (!slot->all_access_paths().empty()) {
607
4.07k
            _tablet_reader_params.all_access_paths.insert({unique_id, slot->all_access_paths()});
608
4.07k
        }
609
610
4.91M
        if (!slot->predicate_access_paths().empty()) {
611
2.61k
            _tablet_reader_params.predicate_access_paths.insert(
612
2.61k
                    {unique_id, slot->predicate_access_paths()});
613
2.61k
        }
614
615
4.91M
        if ((slot->type()->get_primitive_type() == PrimitiveType::TYPE_STRUCT ||
616
4.91M
             slot->type()->get_primitive_type() == PrimitiveType::TYPE_MAP ||
617
4.91M
             slot->type()->get_primitive_type() == PrimitiveType::TYPE_ARRAY) &&
618
4.91M
            !slot->all_access_paths().empty()) {
619
1.42k
            tablet_schema->add_pruned_columns_data_type(column.unique_id(), slot->type());
620
1.42k
        }
621
622
4.91M
        _return_columns.push_back(index);
623
4.91M
        if (slot->is_nullable() && !tablet_schema->column(index).is_nullable()) {
624
0
            _tablet_columns_convert_to_null_set.emplace(index);
625
4.91M
        } else if (!slot->is_nullable() && tablet_schema->column(index).is_nullable()) {
626
0
            return Status::Error<ErrorCode::INVALID_SCHEMA>(
627
0
                    "slot(id: {}, name: {})'s nullable does not match "
628
0
                    "column(tablet id: {}, index: {}, name: {}) ",
629
0
                    slot->id(), slot->col_name(), tablet_schema->table_id(), index,
630
0
                    tablet_schema->column(index).name());
631
0
        }
632
4.91M
    }
633
634
337k
    if (_return_columns.empty()) {
635
0
        return Status::InternalError("failed to build storage scanner, no materialized slot!");
636
0
    }
637
638
337k
    return Status::OK();
639
337k
}
640
641
340k
doris::TabletStorageType OlapScanner::get_storage_type() {
642
340k
    if (config::is_cloud_mode()) {
643
        // we don't have cold storage in cloud mode, all storage is treated as local
644
23
        return doris::TabletStorageType::STORAGE_TYPE_LOCAL;
645
23
    }
646
340k
    int local_reader = 0;
647
600k
    for (const auto& reader : _tablet_reader_params.rs_splits) {
648
600k
        local_reader += reader.rs_reader->rowset()->is_local();
649
600k
    }
650
340k
    int total_reader = _tablet_reader_params.rs_splits.size();
651
652
340k
    if (local_reader == total_reader) {
653
340k
        return doris::TabletStorageType::STORAGE_TYPE_LOCAL;
654
340k
    } else if (local_reader == 0) {
655
0
        return doris::TabletStorageType::STORAGE_TYPE_REMOTE;
656
0
    }
657
0
    return doris::TabletStorageType::STORAGE_TYPE_REMOTE_AND_LOCAL;
658
340k
}
659
660
355k
Status OlapScanner::_get_block_impl(RuntimeState* state, Block* block, bool* eof) {
661
    // Read one block from block reader
662
    // ATTN: Here we need to let the _get_block_impl method guarantee the semantics of the interface,
663
    // that is, eof can be set to true only when the returned block is empty.
664
355k
    RETURN_IF_ERROR(_tablet_reader->next_block_with_aggregation(block, eof));
665
355k
    if (block->rows() > 0) {
666
17.6k
        _tablet_reader_params.tablet->read_block_count.fetch_add(1, std::memory_order_relaxed);
667
17.6k
        *eof = false;
668
17.6k
    }
669
355k
#ifndef NDEBUG
670
355k
    RETURN_IF_ERROR(_check_ann_cache_hit_debug_points(_tablet_reader->stats()));
671
355k
#endif
672
355k
    return Status::OK();
673
355k
}
674
675
338k
Status OlapScanner::close(RuntimeState* state) {
676
338k
    if (!_try_close()) {
677
150
        return Status::OK();
678
150
    }
679
338k
    RETURN_IF_ERROR(Scanner::close(state));
680
338k
    return Status::OK();
681
338k
}
682
683
340k
void OlapScanner::update_realtime_counters() {
684
340k
    if (!_has_prepared) {
685
        // Counter update need prepare successfully, or it maybe core. For example, olap scanner
686
        // will open tablet reader during prepare, if not prepare successfully, tablet reader == nullptr.
687
0
        return;
688
0
    }
689
340k
    OlapScanLocalState* local_state = static_cast<OlapScanLocalState*>(_local_state);
690
340k
    const OlapReaderStatistics& stats = _tablet_reader->stats();
691
340k
    COUNTER_UPDATE(local_state->_read_compressed_counter, stats.compressed_bytes_read);
692
340k
    COUNTER_UPDATE(local_state->_read_uncompressed_counter, stats.uncompressed_bytes_read);
693
340k
    COUNTER_UPDATE(local_state->_scan_bytes, stats.uncompressed_bytes_read);
694
340k
    COUNTER_UPDATE(local_state->_scan_rows, stats.raw_rows_read);
695
696
    // Make sure the scan bytes and scan rows counter in audit log is the same as the counter in
697
    // doris metrics.
698
    // ScanBytes is the uncompressed bytes read from local + remote
699
    // bytes_read_from_local is the compressed bytes read from local
700
    // bytes_read_from_remote is the compressed bytes read from remote
701
    // scan bytes > bytes_read_from_local + bytes_read_from_remote
702
340k
    _state->get_query_ctx()->resource_ctx()->io_context()->update_scan_rows(stats.raw_rows_read);
703
340k
    _state->get_query_ctx()->resource_ctx()->io_context()->update_scan_bytes(
704
340k
            stats.uncompressed_bytes_read);
705
706
    // In case of no cache, we still need to update the IO stats. uncompressed bytes read == local + remote
707
340k
    if (stats.file_cache_stats.bytes_read_from_local == 0 &&
708
340k
        stats.file_cache_stats.bytes_read_from_remote == 0) {
709
340k
        _state->get_query_ctx()->resource_ctx()->io_context()->update_scan_bytes_from_local_storage(
710
340k
                stats.compressed_bytes_read);
711
340k
        DorisMetrics::instance()->query_scan_bytes_from_local->increment(
712
340k
                stats.compressed_bytes_read);
713
18.4E
    } else {
714
18.4E
        _state->get_query_ctx()->resource_ctx()->io_context()->update_scan_bytes_from_local_storage(
715
18.4E
                stats.file_cache_stats.bytes_read_from_local - _bytes_read_from_local);
716
18.4E
        _state->get_query_ctx()
717
18.4E
                ->resource_ctx()
718
18.4E
                ->io_context()
719
18.4E
                ->update_scan_bytes_from_remote_storage(
720
18.4E
                        stats.file_cache_stats.bytes_read_from_remote - _bytes_read_from_remote);
721
722
18.4E
        DorisMetrics::instance()->query_scan_bytes_from_local->increment(
723
18.4E
                stats.file_cache_stats.bytes_read_from_local - _bytes_read_from_local);
724
18.4E
        DorisMetrics::instance()->query_scan_bytes_from_remote->increment(
725
18.4E
                stats.file_cache_stats.bytes_read_from_remote - _bytes_read_from_remote);
726
18.4E
    }
727
728
340k
    _tablet_reader->mutable_stats()->compressed_bytes_read = 0;
729
340k
    _tablet_reader->mutable_stats()->uncompressed_bytes_read = 0;
730
340k
    _tablet_reader->mutable_stats()->raw_rows_read = 0;
731
732
340k
    _bytes_read_from_local = _tablet_reader->stats().file_cache_stats.bytes_read_from_local;
733
340k
    _bytes_read_from_remote = _tablet_reader->stats().file_cache_stats.bytes_read_from_remote;
734
340k
}
735
736
337k
void OlapScanner::_collect_profile_before_close() {
737
    //  Please don't directly enable the profile here, we need to set QueryStatistics using the counter inside.
738
337k
    if (_has_updated_counter) {
739
0
        return;
740
0
    }
741
337k
    _has_updated_counter = true;
742
337k
    _tablet_reader->update_profile(_profile);
743
744
337k
    Scanner::_collect_profile_before_close();
745
746
    // Update counters for OlapScanner
747
    // Update counters from tablet reader's stats
748
337k
    auto& stats = _tablet_reader->stats();
749
337k
    auto* local_state = (OlapScanLocalState*)_local_state;
750
337k
    COUNTER_UPDATE(local_state->_io_timer, stats.io_ns);
751
337k
    COUNTER_UPDATE(local_state->_read_compressed_counter, stats.compressed_bytes_read);
752
337k
    COUNTER_UPDATE(local_state->_scan_bytes, stats.uncompressed_bytes_read);
753
337k
    COUNTER_UPDATE(local_state->_decompressor_timer, stats.decompress_ns);
754
337k
    COUNTER_UPDATE(local_state->_read_uncompressed_counter, stats.uncompressed_bytes_read);
755
337k
    COUNTER_UPDATE(local_state->_block_load_timer, stats.block_load_ns);
756
337k
    COUNTER_UPDATE(local_state->_block_load_counter, stats.blocks_load);
757
337k
    COUNTER_UPDATE(local_state->_block_fetch_timer, stats.block_fetch_ns);
758
337k
    COUNTER_UPDATE(local_state->_delete_bitmap_get_agg_timer, stats.delete_bitmap_get_agg_ns);
759
337k
    COUNTER_UPDATE(local_state->_scan_rows, stats.raw_rows_read);
760
337k
    COUNTER_UPDATE(local_state->_vec_cond_timer, stats.vec_cond_ns);
761
337k
    COUNTER_UPDATE(local_state->_short_cond_timer, stats.short_cond_ns);
762
337k
    COUNTER_UPDATE(local_state->_expr_filter_timer, stats.expr_filter_ns);
763
337k
    COUNTER_UPDATE(local_state->_block_init_timer, stats.block_init_ns);
764
337k
    COUNTER_UPDATE(local_state->_block_init_seek_timer, stats.block_init_seek_ns);
765
337k
    COUNTER_UPDATE(local_state->_block_init_seek_counter, stats.block_init_seek_num);
766
337k
    COUNTER_UPDATE(local_state->_segment_generate_row_range_by_keys_timer,
767
337k
                   stats.generate_row_ranges_by_keys_ns);
768
337k
    COUNTER_UPDATE(local_state->_segment_generate_row_range_by_column_conditions_timer,
769
337k
                   stats.generate_row_ranges_by_column_conditions_ns);
770
337k
    COUNTER_UPDATE(local_state->_segment_generate_row_range_by_bf_timer,
771
337k
                   stats.generate_row_ranges_by_bf_ns);
772
337k
    COUNTER_UPDATE(local_state->_collect_iterator_merge_next_timer,
773
337k
                   stats.collect_iterator_merge_next_timer);
774
337k
    COUNTER_UPDATE(local_state->_segment_generate_row_range_by_zonemap_timer,
775
337k
                   stats.generate_row_ranges_by_zonemap_ns);
776
337k
    COUNTER_UPDATE(local_state->_segment_generate_row_range_by_dict_timer,
777
337k
                   stats.generate_row_ranges_by_dict_ns);
778
337k
    COUNTER_UPDATE(local_state->_predicate_column_read_timer, stats.predicate_column_read_ns);
779
337k
    COUNTER_UPDATE(local_state->_non_predicate_column_read_timer, stats.non_predicate_read_ns);
780
337k
    COUNTER_UPDATE(local_state->_predicate_column_read_seek_timer,
781
337k
                   stats.predicate_column_read_seek_ns);
782
337k
    COUNTER_UPDATE(local_state->_predicate_column_read_seek_counter,
783
337k
                   stats.predicate_column_read_seek_num);
784
337k
    COUNTER_UPDATE(local_state->_lazy_read_timer, stats.lazy_read_ns);
785
337k
    COUNTER_UPDATE(local_state->_lazy_read_seek_timer, stats.block_lazy_read_seek_ns);
786
337k
    COUNTER_UPDATE(local_state->_lazy_read_seek_counter, stats.block_lazy_read_seek_num);
787
337k
    COUNTER_UPDATE(local_state->_output_col_timer, stats.output_col_ns);
788
337k
    COUNTER_UPDATE(local_state->_rows_vec_cond_filtered_counter, stats.rows_vec_cond_filtered);
789
337k
    COUNTER_UPDATE(local_state->_rows_short_circuit_cond_filtered_counter,
790
337k
                   stats.rows_short_circuit_cond_filtered);
791
337k
    COUNTER_UPDATE(local_state->_rows_expr_cond_filtered_counter, stats.rows_expr_cond_filtered);
792
337k
    COUNTER_UPDATE(local_state->_rows_vec_cond_input_counter, stats.vec_cond_input_rows);
793
337k
    COUNTER_UPDATE(local_state->_rows_short_circuit_cond_input_counter,
794
337k
                   stats.short_circuit_cond_input_rows);
795
337k
    COUNTER_UPDATE(local_state->_rows_expr_cond_input_counter, stats.expr_cond_input_rows);
796
337k
    COUNTER_UPDATE(local_state->_stats_filtered_counter, stats.rows_stats_filtered);
797
337k
    COUNTER_UPDATE(local_state->_stats_rp_filtered_counter, stats.rows_stats_rp_filtered);
798
337k
    COUNTER_UPDATE(local_state->_dict_filtered_counter, stats.segment_dict_filtered);
799
337k
    COUNTER_UPDATE(local_state->_bf_filtered_counter, stats.rows_bf_filtered);
800
337k
    COUNTER_UPDATE(local_state->_del_filtered_counter, stats.rows_del_filtered);
801
337k
    COUNTER_UPDATE(local_state->_del_filtered_counter, stats.rows_del_by_bitmap);
802
337k
    COUNTER_UPDATE(local_state->_del_filtered_counter, stats.rows_vec_del_cond_filtered);
803
337k
    COUNTER_UPDATE(local_state->_conditions_filtered_counter, stats.rows_conditions_filtered);
804
337k
    COUNTER_UPDATE(local_state->_key_range_filtered_counter, stats.rows_key_range_filtered);
805
337k
    COUNTER_UPDATE(local_state->_total_pages_num_counter, stats.total_pages_num);
806
337k
    COUNTER_UPDATE(local_state->_cached_pages_num_counter, stats.cached_pages_num);
807
337k
    COUNTER_UPDATE(local_state->_inverted_index_filter_counter, stats.rows_inverted_index_filtered);
808
337k
    COUNTER_UPDATE(local_state->_inverted_index_filter_timer, stats.inverted_index_filter_timer);
809
337k
    COUNTER_UPDATE(local_state->_inverted_index_query_cache_hit_counter,
810
337k
                   stats.inverted_index_query_cache_hit);
811
337k
    COUNTER_UPDATE(local_state->_inverted_index_query_cache_miss_counter,
812
337k
                   stats.inverted_index_query_cache_miss);
813
337k
    COUNTER_UPDATE(local_state->_inverted_index_query_timer, stats.inverted_index_query_timer);
814
337k
    COUNTER_UPDATE(local_state->_inverted_index_query_null_bitmap_timer,
815
337k
                   stats.inverted_index_query_null_bitmap_timer);
816
337k
    COUNTER_UPDATE(local_state->_inverted_index_query_bitmap_copy_timer,
817
337k
                   stats.inverted_index_query_bitmap_copy_timer);
818
337k
    COUNTER_UPDATE(local_state->_inverted_index_searcher_open_timer,
819
337k
                   stats.inverted_index_searcher_open_timer);
820
337k
    COUNTER_UPDATE(local_state->_inverted_index_searcher_search_timer,
821
337k
                   stats.inverted_index_searcher_search_timer);
822
337k
    COUNTER_UPDATE(local_state->_inverted_index_searcher_search_init_timer,
823
337k
                   stats.inverted_index_searcher_search_init_timer);
824
337k
    COUNTER_UPDATE(local_state->_inverted_index_searcher_search_exec_timer,
825
337k
                   stats.inverted_index_searcher_search_exec_timer);
826
337k
    COUNTER_UPDATE(local_state->_inverted_index_searcher_cache_hit_counter,
827
337k
                   stats.inverted_index_searcher_cache_hit);
828
337k
    COUNTER_UPDATE(local_state->_inverted_index_searcher_cache_miss_counter,
829
337k
                   stats.inverted_index_searcher_cache_miss);
830
337k
    COUNTER_UPDATE(local_state->_inverted_index_downgrade_count_counter,
831
337k
                   stats.inverted_index_downgrade_count);
832
337k
    COUNTER_UPDATE(local_state->_inverted_index_analyzer_timer,
833
337k
                   stats.inverted_index_analyzer_timer);
834
337k
    COUNTER_UPDATE(local_state->_inverted_index_lookup_timer, stats.inverted_index_lookup_timer);
835
337k
    COUNTER_UPDATE(local_state->_variant_scan_sparse_column_timer,
836
337k
                   stats.variant_scan_sparse_column_timer_ns);
837
337k
    COUNTER_UPDATE(local_state->_variant_scan_sparse_column_bytes,
838
337k
                   stats.variant_scan_sparse_column_bytes);
839
337k
    COUNTER_UPDATE(local_state->_variant_fill_path_from_sparse_column_timer,
840
337k
                   stats.variant_fill_path_from_sparse_column_timer_ns);
841
337k
    COUNTER_UPDATE(local_state->_variant_subtree_default_iter_count,
842
337k
                   stats.variant_subtree_default_iter_count);
843
337k
    COUNTER_UPDATE(local_state->_variant_subtree_leaf_iter_count,
844
337k
                   stats.variant_subtree_leaf_iter_count);
845
337k
    COUNTER_UPDATE(local_state->_variant_subtree_hierarchical_iter_count,
846
337k
                   stats.variant_subtree_hierarchical_iter_count);
847
337k
    COUNTER_UPDATE(local_state->_variant_subtree_sparse_iter_count,
848
337k
                   stats.variant_subtree_sparse_iter_count);
849
337k
    COUNTER_UPDATE(local_state->_variant_doc_value_column_iter_count,
850
337k
                   stats.variant_doc_value_column_iter_count);
851
852
337k
    if (stats.adaptive_batch_size_predict_max_rows > 0) {
853
314k
        local_state->_adaptive_batch_predict_min_rows_counter->set(
854
314k
                stats.adaptive_batch_size_predict_min_rows);
855
314k
        local_state->_adaptive_batch_predict_max_rows_counter->set(
856
314k
                stats.adaptive_batch_size_predict_max_rows);
857
314k
    }
858
859
337k
    InvertedIndexProfileReporter inverted_index_profile;
860
337k
    inverted_index_profile.update(local_state->_index_filter_profile.get(),
861
337k
                                  &stats.inverted_index_stats);
862
863
    // only cloud deploy mode will use file cache.
864
337k
    if (config::is_cloud_mode() && config::enable_file_cache) {
865
23
        io::FileCacheProfileReporter cache_profile(local_state->_segment_profile.get());
866
23
        cache_profile.update(&stats.file_cache_stats);
867
23
        _state->get_query_ctx()->resource_ctx()->io_context()->update_bytes_write_into_cache(
868
23
                stats.file_cache_stats.bytes_write_into_cache);
869
23
    }
870
337k
    COUNTER_UPDATE(local_state->_output_index_result_column_timer,
871
337k
                   stats.output_index_result_column_timer);
872
337k
    COUNTER_UPDATE(local_state->_filtered_segment_counter, stats.filtered_segment_number);
873
337k
    COUNTER_UPDATE(local_state->_total_segment_counter, stats.total_segment_number);
874
337k
    COUNTER_UPDATE(local_state->_condition_cache_hit_counter, stats.condition_cache_hit_seg_nums);
875
337k
    COUNTER_UPDATE(local_state->_condition_cache_filtered_rows_counter,
876
337k
                   stats.condition_cache_filtered_rows);
877
878
337k
    COUNTER_UPDATE(local_state->_tablet_reader_init_timer, stats.tablet_reader_init_timer_ns);
879
337k
    COUNTER_UPDATE(local_state->_tablet_reader_capture_rs_readers_timer,
880
337k
                   stats.tablet_reader_capture_rs_readers_timer_ns);
881
337k
    COUNTER_UPDATE(local_state->_tablet_reader_init_return_columns_timer,
882
337k
                   stats.tablet_reader_init_return_columns_timer_ns);
883
337k
    COUNTER_UPDATE(local_state->_tablet_reader_init_keys_param_timer,
884
337k
                   stats.tablet_reader_init_keys_param_timer_ns);
885
337k
    COUNTER_UPDATE(local_state->_tablet_reader_init_orderby_keys_param_timer,
886
337k
                   stats.tablet_reader_init_orderby_keys_param_timer_ns);
887
337k
    COUNTER_UPDATE(local_state->_tablet_reader_init_conditions_param_timer,
888
337k
                   stats.tablet_reader_init_conditions_param_timer_ns);
889
337k
    COUNTER_UPDATE(local_state->_tablet_reader_init_delete_condition_param_timer,
890
337k
                   stats.tablet_reader_init_delete_condition_param_timer_ns);
891
337k
    COUNTER_UPDATE(local_state->_block_reader_vcollect_iter_init_timer,
892
337k
                   stats.block_reader_vcollect_iter_init_timer_ns);
893
337k
    COUNTER_UPDATE(local_state->_block_reader_rs_readers_init_timer,
894
337k
                   stats.block_reader_rs_readers_init_timer_ns);
895
337k
    COUNTER_UPDATE(local_state->_block_reader_build_heap_init_timer,
896
337k
                   stats.block_reader_build_heap_init_timer_ns);
897
898
337k
    COUNTER_UPDATE(local_state->_rowset_reader_get_segment_iterators_timer,
899
337k
                   stats.rowset_reader_get_segment_iterators_timer_ns);
900
337k
    COUNTER_UPDATE(local_state->_rowset_reader_create_iterators_timer,
901
337k
                   stats.rowset_reader_create_iterators_timer_ns);
902
337k
    COUNTER_UPDATE(local_state->_rowset_reader_init_iterators_timer,
903
337k
                   stats.rowset_reader_init_iterators_timer_ns);
904
337k
    COUNTER_UPDATE(local_state->_rowset_reader_load_segments_timer,
905
337k
                   stats.rowset_reader_load_segments_timer_ns);
906
907
337k
    COUNTER_UPDATE(local_state->_segment_iterator_init_timer, stats.segment_iterator_init_timer_ns);
908
337k
    COUNTER_UPDATE(local_state->_segment_iterator_init_return_column_iterators_timer,
909
337k
                   stats.segment_iterator_init_return_column_iterators_timer_ns);
910
337k
    COUNTER_UPDATE(local_state->_segment_iterator_init_index_iterators_timer,
911
337k
                   stats.segment_iterator_init_index_iterators_timer_ns);
912
337k
    COUNTER_UPDATE(local_state->_segment_iterator_init_segment_prefetchers_timer,
913
337k
                   stats.segment_iterator_init_segment_prefetchers_timer_ns);
914
915
337k
    COUNTER_UPDATE(local_state->_segment_create_column_readers_timer,
916
337k
                   stats.segment_create_column_readers_timer_ns);
917
337k
    COUNTER_UPDATE(local_state->_segment_load_index_timer, stats.segment_load_index_timer_ns);
918
919
    // Update metrics
920
337k
    DorisMetrics::instance()->query_scan_bytes->increment(
921
337k
            local_state->_read_uncompressed_counter->value());
922
337k
    DorisMetrics::instance()->query_scan_rows->increment(local_state->_scan_rows->value());
923
337k
    auto& tablet = _tablet_reader_params.tablet;
924
337k
    tablet->query_scan_bytes->increment(local_state->_read_uncompressed_counter->value());
925
337k
    tablet->query_scan_rows->increment(local_state->_scan_rows->value());
926
337k
    tablet->query_scan_count->increment(1);
927
928
337k
    COUNTER_UPDATE(local_state->_ann_range_search_filter_counter,
929
337k
                   stats.rows_ann_index_range_filtered);
930
337k
    COUNTER_UPDATE(local_state->_ann_topn_filter_counter, stats.rows_ann_index_topn_filtered);
931
337k
    COUNTER_UPDATE(local_state->_ann_index_load_costs, stats.ann_index_load_ns);
932
337k
    COUNTER_UPDATE(local_state->_ann_ivf_on_disk_load_costs, stats.ann_ivf_on_disk_load_ns);
933
337k
    COUNTER_UPDATE(local_state->_ann_ivf_on_disk_cache_hit_cnt,
934
337k
                   stats.ann_ivf_on_disk_cache_hit_cnt);
935
337k
    COUNTER_UPDATE(local_state->_ann_ivf_on_disk_cache_miss_cnt,
936
337k
                   stats.ann_ivf_on_disk_cache_miss_cnt);
937
337k
    COUNTER_UPDATE(local_state->_ann_range_search_costs, stats.ann_index_range_search_ns);
938
337k
    COUNTER_UPDATE(local_state->_ann_range_search_cnt, stats.ann_index_range_search_cnt);
939
337k
    COUNTER_UPDATE(local_state->_ann_range_engine_search_costs, stats.ann_range_engine_search_ns);
940
    // Engine prepare before search
941
337k
    COUNTER_UPDATE(local_state->_ann_range_pre_process_costs, stats.ann_range_pre_process_ns);
942
    // Post process parent: Doris result process + engine convert
943
337k
    COUNTER_UPDATE(local_state->_ann_range_post_process_costs,
944
337k
                   stats.ann_range_result_convert_ns + stats.ann_range_engine_convert_ns);
945
    // Engine convert (child under post-process)
946
337k
    COUNTER_UPDATE(local_state->_ann_range_engine_convert_costs, stats.ann_range_engine_convert_ns);
947
    // Doris-side result convert (child under post-process)
948
337k
    COUNTER_UPDATE(local_state->_ann_range_result_convert_costs, stats.ann_range_result_convert_ns);
949
950
337k
    COUNTER_UPDATE(local_state->_ann_topn_search_costs, stats.ann_topn_search_ns);
951
337k
    COUNTER_UPDATE(local_state->_ann_topn_search_cnt, stats.ann_index_topn_search_cnt);
952
337k
    COUNTER_UPDATE(local_state->_ann_cache_hit_cnt, stats.ann_index_cache_hits);
953
337k
    COUNTER_UPDATE(local_state->_ann_range_cache_hit_cnt, stats.ann_index_range_cache_hits);
954
955
    // Detailed ANN timers
956
    // ANN TopN timers with hierarchy
957
    // Engine search time (FAISS)
958
337k
    COUNTER_UPDATE(local_state->_ann_topn_engine_search_costs,
959
337k
                   stats.ann_index_topn_engine_search_ns);
960
    // Engine prepare time (allocations/buffer setup before search)
961
337k
    COUNTER_UPDATE(local_state->_ann_topn_pre_process_costs,
962
337k
                   stats.ann_index_topn_engine_prepare_ns);
963
    // Post process parent includes Doris result processing + engine convert
964
337k
    COUNTER_UPDATE(local_state->_ann_topn_post_process_costs,
965
337k
                   stats.ann_index_topn_result_process_ns + stats.ann_index_topn_engine_convert_ns);
966
    // Engine-side conversion time inside FAISS wrappers (child under post-process)
967
337k
    COUNTER_UPDATE(local_state->_ann_topn_engine_convert_costs,
968
337k
                   stats.ann_index_topn_engine_convert_ns);
969
970
    // Doris-side result convert costs (show separately as another child counter); use pure process time
971
337k
    COUNTER_UPDATE(local_state->_ann_topn_result_convert_costs,
972
337k
                   stats.ann_index_topn_result_process_ns);
973
974
337k
    COUNTER_UPDATE(local_state->_ann_fallback_brute_force_cnt, stats.ann_fall_back_brute_force_cnt);
975
976
    // Overhead counter removed; precise instrumentation is reported via engine_prepare above.
977
337k
}
978
979
#ifndef NDEBUG
980
355k
Status OlapScanner::_check_ann_cache_hit_debug_points(const OlapReaderStatistics& stats) {
981
355k
    DBUG_EXECUTE_IF("olap_scanner.ann_topn_cache_hits", {
982
355k
        auto expected_hits = dp->param<int32_t>("expected_hits", -1);
983
355k
        auto min_hits = dp->param<int32_t>("min_hits", -1);
984
355k
        if (expected_hits >= 0 && stats.ann_index_cache_hits != expected_hits) {
985
355k
            return Status::Error<ErrorCode::INTERNAL_ERROR>(
986
355k
                    "ann_index_cache_hits: {} not equal to expected: {}",
987
355k
                    stats.ann_index_cache_hits, expected_hits);
988
355k
        }
989
355k
        if (min_hits >= 0 && stats.ann_index_cache_hits < min_hits) {
990
355k
            return Status::Error<ErrorCode::INTERNAL_ERROR>(
991
355k
                    "ann_index_cache_hits: {} less than expected min: {}",
992
355k
                    stats.ann_index_cache_hits, min_hits);
993
355k
        }
994
355k
    })
995
355k
    DBUG_EXECUTE_IF("olap_scanner.ann_range_cache_hits", {
996
355k
        auto expected_hits = dp->param<int32_t>("expected_hits", -1);
997
355k
        auto min_hits = dp->param<int32_t>("min_hits", -1);
998
355k
        if (expected_hits >= 0 && stats.ann_index_range_cache_hits != expected_hits) {
999
355k
            return Status::Error<ErrorCode::INTERNAL_ERROR>(
1000
355k
                    "ann_index_range_cache_hits: {} not equal to expected: {}",
1001
355k
                    stats.ann_index_range_cache_hits, expected_hits);
1002
355k
        }
1003
355k
        if (min_hits >= 0 && stats.ann_index_range_cache_hits < min_hits) {
1004
355k
            return Status::Error<ErrorCode::INTERNAL_ERROR>(
1005
355k
                    "ann_index_range_cache_hits: {} less than expected min: {}",
1006
355k
                    stats.ann_index_range_cache_hits, min_hits);
1007
355k
        }
1008
355k
    })
1009
355k
    return Status::OK();
1010
355k
}
1011
#endif
1012
1013
#include "common/compile_check_avoid_end.h"
1014
} // namespace doris