Coverage Report

Created: 2026-03-30 07:45

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/cache/schema_cache.h"
55
#include "storage/id_manager.h"
56
#include "storage/index/inverted/inverted_index_profile.h"
57
#include "storage/iterator/block_reader.h"
58
#include "storage/olap_common.h"
59
#include "storage/olap_tuple.h"
60
#include "storage/olap_utils.h"
61
#include "storage/storage_engine.h"
62
#include "storage/tablet/tablet_schema.h"
63
#include "util/json/path_in_data.h"
64
65
namespace doris {
66
#include "common/compile_check_avoid_begin.h"
67
68
using ReadSource = TabletReadSource;
69
70
OlapScanner::OlapScanner(ScanLocalStateBase* parent, OlapScanner::Params&& params)
71
1.26M
        : Scanner(params.state, parent, params.limit, params.profile),
72
1.26M
          _key_ranges(std::move(params.key_ranges)),
73
1.26M
          _tablet_reader_params({.tablet = std::move(params.tablet),
74
1.26M
                                 .tablet_schema {},
75
1.26M
                                 .aggregation = params.aggregation,
76
1.26M
                                 .version = {0, params.version},
77
1.26M
                                 .start_key {},
78
1.26M
                                 .end_key {},
79
1.26M
                                 .predicates {},
80
1.26M
                                 .function_filters {},
81
1.26M
                                 .delete_predicates {},
82
1.26M
                                 .target_cast_type_for_variants {},
83
1.26M
                                 .all_access_paths {},
84
1.26M
                                 .predicate_access_paths {},
85
1.26M
                                 .rs_splits {},
86
1.26M
                                 .return_columns {},
87
1.26M
                                 .output_columns {},
88
1.26M
                                 .remaining_conjunct_roots {},
89
1.26M
                                 .common_expr_ctxs_push_down {},
90
1.26M
                                 .topn_filter_source_node_ids {},
91
1.26M
                                 .filter_block_conjuncts {},
92
1.26M
                                 .key_group_cluster_key_idxes {},
93
1.26M
                                 .virtual_column_exprs {},
94
1.26M
                                 .vir_cid_to_idx_in_block {},
95
1.26M
                                 .vir_col_idx_to_type {},
96
1.26M
                                 .score_runtime {},
97
1.26M
                                 .collection_statistics {},
98
1.26M
                                 .ann_topn_runtime {},
99
1.26M
                                 .condition_cache_digest = parent->get_condition_cache_digest()}) {
100
1.26M
    _tablet_reader_params.set_read_source(std::move(params.read_source),
101
1.26M
                                          _state->skip_delete_bitmap());
102
1.26M
    _has_prepared = false;
103
1.26M
    _vector_search_params = params.state->get_vector_search_params();
104
1.26M
}
105
106
static std::string read_columns_to_string(TabletSchemaSPtr tablet_schema,
107
3.36k
                                          const std::vector<uint32_t>& read_columns) {
108
    // avoid too long for one line,
109
    // it is hard to display in `show profile` stmt if one line is too long.
110
3.36k
    const int col_per_line = 10;
111
3.36k
    int i = 0;
112
3.36k
    std::string read_columns_string;
113
3.36k
    read_columns_string += "[";
114
17.3k
    for (auto it = read_columns.cbegin(); it != read_columns.cend(); it++) {
115
14.0k
        if (it != read_columns.cbegin()) {
116
10.6k
            read_columns_string += ", ";
117
10.6k
        }
118
14.0k
        read_columns_string += tablet_schema->columns().at(*it)->name();
119
14.0k
        if (i >= col_per_line) {
120
13
            read_columns_string += "\n";
121
13
            i = 0;
122
14.0k
        } else {
123
14.0k
            ++i;
124
14.0k
        }
125
14.0k
    }
126
3.36k
    read_columns_string += "]";
127
3.36k
    return read_columns_string;
128
3.36k
}
129
130
1.26M
Status OlapScanner::prepare() {
131
1.26M
    auto* local_state = static_cast<OlapScanLocalState*>(_local_state);
132
1.26M
    auto& tablet = _tablet_reader_params.tablet;
133
1.26M
    auto& tablet_schema = _tablet_reader_params.tablet_schema;
134
1.26M
    DBUG_EXECUTE_IF("CloudTablet.capture_rs_readers.return.e-230", {
135
1.26M
        LOG_WARNING("CloudTablet.capture_rs_readers.return e-230 init")
136
1.26M
                .tag("tablet_id", tablet->tablet_id());
137
1.26M
        return Status::Error<false>(-230, "injected error");
138
1.26M
    });
139
140
1.26M
    for (auto& ctx : local_state->_common_expr_ctxs_push_down) {
141
13.8k
        VExprContextSPtr context;
142
13.8k
        RETURN_IF_ERROR(ctx->clone(_state, context));
143
13.8k
        _common_expr_ctxs_push_down.emplace_back(context);
144
13.8k
        context->prepare_ann_range_search(_vector_search_params);
145
13.8k
    }
146
147
1.26M
    for (auto pair : local_state->_slot_id_to_virtual_column_expr) {
148
        // Scanner will be executed in a different thread, so we need to clone the context.
149
387
        VExprContextSPtr context;
150
387
        RETURN_IF_ERROR(pair.second->clone(_state, context));
151
387
        _slot_id_to_virtual_column_expr[pair.first] = context;
152
387
    }
153
154
1.26M
    _slot_id_to_index_in_block = local_state->_slot_id_to_index_in_block;
155
1.26M
    _slot_id_to_col_type = local_state->_slot_id_to_col_type;
156
1.26M
    _score_runtime = local_state->_score_runtime;
157
    // All scanners share the same ann_topn_runtime.
158
1.26M
    _ann_topn_runtime = local_state->_ann_topn_runtime;
159
160
    // set limit to reduce end of rowset and segment mem use
161
1.26M
    _tablet_reader = std::make_unique<BlockReader>();
162
    // batch size is passed down to segment iterator, use _state->batch_size()
163
    // instead of _parent->limit(), because if _parent->limit() is a very small
164
    // value (e.g. select a from t where a .. and b ... limit 1),
165
    // it will be very slow when reading data in segment iterator
166
1.26M
    _tablet_reader->set_batch_size(_state->batch_size());
167
1.26M
    TabletSchemaSPtr cached_schema;
168
1.26M
    std::string schema_key;
169
1.26M
    {
170
1.26M
        TOlapScanNode& olap_scan_node = local_state->olap_scan_node();
171
172
1.26M
        const auto check_can_use_cache = [&]() {
173
1.26M
            if (!(olap_scan_node.__isset.schema_version && olap_scan_node.__isset.columns_desc &&
174
1.26M
                  !olap_scan_node.columns_desc.empty() &&
175
1.26M
                  olap_scan_node.columns_desc[0].col_unique_id >= 0 && // Why check first column?
176
1.26M
                  tablet->tablet_schema()->num_variant_columns() == 0 &&
177
1.26M
                  tablet->tablet_schema()->num_virtual_columns() == 0)) {
178
6.78k
                return false;
179
6.78k
            }
180
181
            // If `delete_predicates` is not empty, will merge the columns in delete predicate into current tablet schema
182
1.25M
            if (!_tablet_reader_params.delete_predicates.empty()) {
183
3.40k
                return false;
184
3.40k
            }
185
186
1.25M
            const bool has_pruned_column =
187
12.6M
                    std::ranges::any_of(_output_tuple_desc->slots(), [](const auto& slot) {
188
12.6M
                        if ((slot->type()->get_primitive_type() == PrimitiveType::TYPE_STRUCT ||
189
12.6M
                             slot->type()->get_primitive_type() == PrimitiveType::TYPE_MAP ||
190
12.6M
                             slot->type()->get_primitive_type() == PrimitiveType::TYPE_ARRAY) &&
191
12.6M
                            !slot->all_access_paths().empty()) {
192
24.9k
                            return true;
193
24.9k
                        }
194
12.6M
                        return false;
195
12.6M
                    });
196
1.25M
            return !has_pruned_column;
197
1.25M
        }();
198
199
1.26M
        if (check_can_use_cache) {
200
1.23M
            schema_key =
201
1.23M
                    SchemaCache::get_schema_key(tablet->tablet_id(), olap_scan_node.columns_desc,
202
1.23M
                                                olap_scan_node.schema_version);
203
1.23M
            cached_schema = SchemaCache::instance()->get_schema(schema_key);
204
1.23M
        }
205
1.26M
        if (cached_schema && cached_schema->num_virtual_columns() == 0) {
206
1.04M
            tablet_schema = cached_schema;
207
1.04M
        } else {
208
            // If schema is not cached or cached schema has virtual columns,
209
            // we need to create a new TabletSchema.
210
219k
            tablet_schema = std::make_shared<TabletSchema>();
211
219k
            tablet_schema->copy_from(*tablet->tablet_schema());
212
219k
            if (olap_scan_node.__isset.columns_desc && !olap_scan_node.columns_desc.empty() &&
213
219k
                olap_scan_node.columns_desc[0].col_unique_id >= 0) {
214
                // Originally scanner get TabletSchema from tablet object in BE.
215
                // To support lightweight schema change for adding / dropping columns,
216
                // tabletschema is bounded to rowset and tablet's schema maybe outdated,
217
                //  so we have to use schema from a query plan witch FE puts it in query plans.
218
218k
                tablet_schema->clear_columns();
219
3.15M
                for (const auto& column_desc : olap_scan_node.columns_desc) {
220
3.15M
                    tablet_schema->append_column(TabletColumn(column_desc));
221
3.15M
                }
222
219k
                if (olap_scan_node.__isset.schema_version) {
223
219k
                    tablet_schema->set_schema_version(olap_scan_node.schema_version);
224
219k
                }
225
218k
            }
226
219k
            if (olap_scan_node.__isset.indexes_desc) {
227
218k
                tablet_schema->update_indexes_from_thrift(olap_scan_node.indexes_desc);
228
218k
            }
229
219k
        }
230
231
1.26M
        if (_tablet_reader_params.rs_splits.empty()) {
232
            // Non-pipeline mode, Tablet : Scanner = 1 : 1
233
            // acquire tablet rowset readers at the beginning of the scan node
234
            // to prevent this case: when there are lots of olap scanners to run for example 10000
235
            // the rowsets maybe compacted when the last olap scanner starts
236
0
            ReadSource read_source;
237
238
0
            if (config::is_cloud_mode()) {
239
                // FIXME(plat1ko): Avoid pointer cast
240
0
                ExecEnv::GetInstance()->storage_engine().to_cloud().tablet_hotspot().count(*tablet);
241
0
            }
242
243
0
            auto maybe_read_source = tablet->capture_read_source(
244
0
                    _tablet_reader_params.version,
245
0
                    {
246
0
                            .skip_missing_versions = _state->skip_missing_version(),
247
0
                            .enable_fetch_rowsets_from_peers =
248
0
                                    config::enable_fetch_rowsets_from_peer_replicas,
249
0
                            .enable_prefer_cached_rowset =
250
0
                                    config::is_cloud_mode() ? _state->enable_prefer_cached_rowset()
251
0
                                                            : false,
252
0
                            .query_freshness_tolerance_ms =
253
0
                                    config::is_cloud_mode() ? _state->query_freshness_tolerance_ms()
254
0
                                                            : -1,
255
0
                    });
256
0
            if (!maybe_read_source) {
257
0
                LOG(WARNING) << "fail to init reader. res=" << maybe_read_source.error();
258
0
                return maybe_read_source.error();
259
0
            }
260
261
0
            read_source = std::move(maybe_read_source.value());
262
263
0
            if (config::enable_mow_verbose_log && tablet->enable_unique_key_merge_on_write()) {
264
0
                LOG_INFO("finish capture_rs_readers for tablet={}, query_id={}",
265
0
                         tablet->tablet_id(), print_id(_state->query_id()));
266
0
            }
267
268
0
            if (!_state->skip_delete_predicate()) {
269
0
                read_source.fill_delete_predicates();
270
0
            }
271
0
            _tablet_reader_params.set_read_source(std::move(read_source));
272
0
        }
273
274
        // Initialize tablet_reader_params
275
1.26M
        RETURN_IF_ERROR(_init_tablet_reader_params(
276
1.26M
                local_state->_parent->cast<OlapScanOperatorX>()._slot_id_to_slot_desc, _key_ranges,
277
1.26M
                local_state->_slot_id_to_predicates, local_state->_push_down_functions));
278
1.26M
    }
279
280
    // add read columns in profile
281
1.26M
    if (_state->enable_profile()) {
282
3.36k
        _profile->add_info_string("ReadColumns",
283
3.36k
                                  read_columns_to_string(tablet_schema, _return_columns));
284
3.36k
    }
285
286
    // Add newly created tablet schema to schema cache if it does not have virtual columns.
287
1.26M
    if (cached_schema == nullptr && !schema_key.empty() &&
288
1.26M
        tablet_schema->num_virtual_columns() == 0 && !tablet_schema->has_pruned_columns()) {
289
183k
        SchemaCache::instance()->insert_schema(schema_key, tablet_schema);
290
183k
    }
291
292
1.26M
    if (_tablet_reader_params.score_runtime) {
293
14
        SCOPED_TIMER(local_state->_statistics_collect_timer);
294
14
        _tablet_reader_params.collection_statistics = std::make_shared<CollectionStatistics>();
295
296
14
        io::IOContext io_ctx {
297
14
                .reader_type = ReaderType::READER_QUERY,
298
14
                .expiration_time = tablet->ttl_seconds(),
299
14
                .query_id = &_state->query_id(),
300
14
                .file_cache_stats = &_tablet_reader->mutable_stats()->file_cache_stats,
301
14
                .is_inverted_index = true,
302
14
        };
303
304
14
        RETURN_IF_ERROR(_tablet_reader_params.collection_statistics->collect(
305
14
                _state, _tablet_reader_params.rs_splits, _tablet_reader_params.tablet_schema,
306
14
                _tablet_reader_params.common_expr_ctxs_push_down, &io_ctx));
307
14
    }
308
309
1.26M
    _has_prepared = true;
310
1.26M
    return Status::OK();
311
1.26M
}
312
313
1.26M
Status OlapScanner::_open_impl(RuntimeState* state) {
314
1.26M
    RETURN_IF_ERROR(Scanner::_open_impl(state));
315
1.26M
    SCOPED_TIMER(_local_state->cast<OlapScanLocalState>()._reader_init_timer);
316
317
1.26M
    auto res = _tablet_reader->init(_tablet_reader_params);
318
1.26M
    if (!res.ok()) {
319
99
        res.append("failed to initialize storage reader. tablet=" +
320
99
                   std::to_string(_tablet_reader_params.tablet->tablet_id()) +
321
99
                   ", backend=" + BackendOptions::get_localhost());
322
99
        return res;
323
99
    }
324
325
    // Do not hold rs_splits any more to release memory.
326
1.26M
    _tablet_reader_params.rs_splits.clear();
327
328
1.26M
    return Status::OK();
329
1.26M
}
330
331
// it will be called under tablet read lock because capture rs readers need
332
Status OlapScanner::_init_tablet_reader_params(
333
        const phmap::flat_hash_map<int, SlotDescriptor*>& slot_id_to_slot_desc,
334
        const std::vector<OlapScanRange*>& key_ranges,
335
        const phmap::flat_hash_map<int, std::vector<std::shared_ptr<ColumnPredicate>>>&
336
                slot_to_predicates,
337
1.26M
        const std::vector<FunctionFilter>& function_filters) {
338
    // if the table with rowset [0-x] or [0-1] [2-y], and [0-1] is empty
339
1.26M
    const bool single_version = _tablet_reader_params.has_single_version();
340
341
1.26M
    auto* olap_local_state = static_cast<OlapScanLocalState*>(_local_state);
342
1.26M
    bool read_mor_as_dup = olap_local_state->olap_scan_node().__isset.read_mor_as_dup &&
343
1.26M
                           olap_local_state->olap_scan_node().read_mor_as_dup;
344
1.26M
    if (_state->skip_storage_engine_merge() || read_mor_as_dup) {
345
43
        _tablet_reader_params.direct_mode = true;
346
43
        _tablet_reader_params.aggregation = true;
347
1.26M
    } else {
348
1.26M
        auto push_down_agg_type = _local_state->get_push_down_agg_type();
349
1.26M
        _tablet_reader_params.direct_mode = _tablet_reader_params.aggregation || single_version ||
350
1.26M
                                            (push_down_agg_type != TPushAggOp::NONE &&
351
10.6k
                                             push_down_agg_type != TPushAggOp::COUNT_ON_INDEX);
352
1.26M
    }
353
354
1.26M
    RETURN_IF_ERROR(_init_variant_columns());
355
1.26M
    RETURN_IF_ERROR(_init_return_columns());
356
357
1.26M
    _tablet_reader_params.reader_type = ReaderType::READER_QUERY;
358
1.26M
    _tablet_reader_params.push_down_agg_type_opt = _local_state->get_push_down_agg_type();
359
360
    // TODO: If a new runtime filter arrives after `_conjuncts` move to `_common_expr_ctxs_push_down`,
361
1.26M
    if (_common_expr_ctxs_push_down.empty()) {
362
1.25M
        for (auto& conjunct : _conjuncts) {
363
13.2k
            _tablet_reader_params.remaining_conjunct_roots.emplace_back(conjunct->root());
364
13.2k
        }
365
1.25M
    } else {
366
13.7k
        for (auto& ctx : _common_expr_ctxs_push_down) {
367
13.7k
            _tablet_reader_params.remaining_conjunct_roots.emplace_back(ctx->root());
368
13.7k
        }
369
12.6k
    }
370
371
1.26M
    _tablet_reader_params.common_expr_ctxs_push_down = _common_expr_ctxs_push_down;
372
1.26M
    _tablet_reader_params.virtual_column_exprs = _virtual_column_exprs;
373
1.26M
    _tablet_reader_params.vir_cid_to_idx_in_block = _vir_cid_to_idx_in_block;
374
1.26M
    _tablet_reader_params.vir_col_idx_to_type = _vir_col_idx_to_type;
375
1.26M
    _tablet_reader_params.score_runtime = _score_runtime;
376
1.26M
    _tablet_reader_params.output_columns = ((OlapScanLocalState*)_local_state)->_output_column_ids;
377
1.26M
    _tablet_reader_params.ann_topn_runtime = _ann_topn_runtime;
378
1.26M
    for (const auto& ele : ((OlapScanLocalState*)_local_state)->_cast_types_for_variants) {
379
1.95k
        _tablet_reader_params.target_cast_type_for_variants[ele.first] = ele.second;
380
1.95k
    };
381
1.26M
    auto& tablet_schema = _tablet_reader_params.tablet_schema;
382
9.72M
    for (auto& predicates : slot_to_predicates) {
383
9.72M
        const int sid = predicates.first;
384
9.72M
        DCHECK(slot_id_to_slot_desc.contains(sid));
385
9.72M
        int32_t index =
386
9.72M
                tablet_schema->field_index(slot_id_to_slot_desc.find(sid)->second->col_name());
387
9.72M
        if (index < 0) {
388
0
            throw Exception(
389
0
                    Status::InternalError("Column {} not found in tablet schema",
390
0
                                          slot_id_to_slot_desc.find(sid)->second->col_name()));
391
0
        }
392
9.72M
        for (auto& predicate : predicates.second) {
393
982k
            _tablet_reader_params.predicates.push_back(predicate->clone(index));
394
982k
        }
395
9.72M
    }
396
397
1.26M
    std::copy(function_filters.cbegin(), function_filters.cend(),
398
1.26M
              std::inserter(_tablet_reader_params.function_filters,
399
1.26M
                            _tablet_reader_params.function_filters.begin()));
400
401
    // Merge the columns in delete predicate that not in latest schema in to current tablet schema
402
1.26M
    for (auto& del_pred : _tablet_reader_params.delete_predicates) {
403
7.22k
        tablet_schema->merge_dropped_columns(*del_pred->tablet_schema());
404
7.22k
    }
405
406
    // Push key ranges to the tablet reader.
407
    // Skip the "full scan" placeholder (has_lower_bound == false) — when no key
408
    // predicates exist, start_key/end_key remain empty and the reader does a full scan.
409
2.03M
    for (auto* key_range : key_ranges) {
410
2.03M
        if (!key_range->has_lower_bound) {
411
206k
            continue;
412
206k
        }
413
414
1.83M
        _tablet_reader_params.start_key_include = key_range->begin_include;
415
1.83M
        _tablet_reader_params.end_key_include = key_range->end_include;
416
417
1.83M
        _tablet_reader_params.start_key.push_back(key_range->begin_scan_range);
418
1.83M
        _tablet_reader_params.end_key.push_back(key_range->end_scan_range);
419
1.83M
    }
420
421
1.26M
    _tablet_reader_params.profile = _local_state->custom_profile();
422
1.26M
    _tablet_reader_params.runtime_state = _state;
423
424
1.26M
    _tablet_reader_params.origin_return_columns = &_return_columns;
425
1.26M
    _tablet_reader_params.tablet_columns_convert_to_null_set = &_tablet_columns_convert_to_null_set;
426
427
1.26M
    if (_tablet_reader_params.direct_mode) {
428
1.25M
        _tablet_reader_params.return_columns = _return_columns;
429
1.25M
    } else {
430
        // we need to fetch all key columns to do the right aggregation on storage engine side.
431
40.3k
        for (size_t i = 0; i < tablet_schema->num_key_columns(); ++i) {
432
27.8k
            _tablet_reader_params.return_columns.push_back(i);
433
27.8k
        }
434
51.6k
        for (auto index : _return_columns) {
435
51.6k
            if (tablet_schema->column(index).is_key()) {
436
19.0k
                continue;
437
19.0k
            }
438
32.5k
            _tablet_reader_params.return_columns.push_back(index);
439
32.5k
        }
440
        // expand the sequence column
441
12.5k
        if (tablet_schema->has_sequence_col() || tablet_schema->has_seq_map()) {
442
40
            bool has_replace_col = false;
443
91
            for (auto col : _return_columns) {
444
91
                if (tablet_schema->column(col).aggregation() ==
445
91
                    FieldAggregationMethod::OLAP_FIELD_AGGREGATION_REPLACE) {
446
40
                    has_replace_col = true;
447
40
                    break;
448
40
                }
449
91
            }
450
40
            if (auto sequence_col_idx = tablet_schema->sequence_col_idx();
451
40
                has_replace_col && tablet_schema->has_sequence_col() &&
452
40
                std::find(_return_columns.begin(), _return_columns.end(), sequence_col_idx) ==
453
28
                        _return_columns.end()) {
454
16
                _tablet_reader_params.return_columns.push_back(sequence_col_idx);
455
16
            }
456
40
            if (has_replace_col) {
457
39
                const auto& val_to_seq = tablet_schema->value_col_idx_to_seq_col_idx();
458
39
                std::set<uint32_t> return_seq_columns;
459
460
243
                for (auto col : _tablet_reader_params.return_columns) {
461
                    // we need to add the necessary sequence column in _return_columns, and
462
                    // Avoid adding the same seq column twice
463
243
                    const auto val_iter = val_to_seq.find(col);
464
243
                    if (val_iter != val_to_seq.end()) {
465
42
                        auto seq = val_iter->second;
466
42
                        if (std::find(_tablet_reader_params.return_columns.begin(),
467
42
                                      _tablet_reader_params.return_columns.end(),
468
42
                                      seq) == _tablet_reader_params.return_columns.end()) {
469
4
                            return_seq_columns.insert(seq);
470
4
                        }
471
42
                    }
472
243
                }
473
39
                _tablet_reader_params.return_columns.insert(
474
39
                        std::end(_tablet_reader_params.return_columns),
475
39
                        std::begin(return_seq_columns), std::end(return_seq_columns));
476
39
            }
477
40
        }
478
12.5k
    }
479
480
1.26M
    _tablet_reader_params.use_page_cache = _state->enable_page_cache();
481
482
1.26M
    DBUG_EXECUTE_IF("NewOlapScanner::_init_tablet_reader_params.block", DBUG_BLOCK);
483
484
1.26M
    if (!_state->skip_storage_engine_merge()) {
485
1.26M
        auto* olap_scan_local_state = (OlapScanLocalState*)_local_state;
486
1.26M
        TOlapScanNode& olap_scan_node = olap_scan_local_state->olap_scan_node();
487
488
        // Set MOR value predicate pushdown flag
489
1.26M
        if (olap_scan_node.__isset.enable_mor_value_predicate_pushdown &&
490
1.26M
            olap_scan_node.enable_mor_value_predicate_pushdown) {
491
25
            _tablet_reader_params.enable_mor_value_predicate_pushdown = true;
492
25
        }
493
494
        // order by table keys optimization for topn
495
        // will only read head/tail of data file since it's already sorted by keys
496
1.26M
        if (olap_scan_node.__isset.sort_info && !olap_scan_node.sort_info.is_asc_order.empty()) {
497
2.04k
            _limit = _local_state->limit_per_scanner();
498
2.04k
            _tablet_reader_params.read_orderby_key = true;
499
2.04k
            if (!olap_scan_node.sort_info.is_asc_order[0]) {
500
312
                _tablet_reader_params.read_orderby_key_reverse = true;
501
312
            }
502
2.04k
            _tablet_reader_params.read_orderby_key_num_prefix_columns =
503
2.04k
                    olap_scan_node.sort_info.is_asc_order.size();
504
2.04k
            _tablet_reader_params.read_orderby_key_limit = _limit;
505
506
2.04k
            if (_tablet_reader_params.read_orderby_key_limit > 0 &&
507
2.05k
                olap_scan_local_state->_storage_no_merge()) {
508
1.93k
                _tablet_reader_params.filter_block_conjuncts = _conjuncts;
509
1.93k
                _conjuncts.clear();
510
1.93k
            }
511
2.04k
        }
512
513
        // set push down topn filter
514
1.26M
        _tablet_reader_params.topn_filter_source_node_ids =
515
1.26M
                olap_scan_local_state->get_topn_filter_source_node_ids(_state, true);
516
1.26M
        if (!_tablet_reader_params.topn_filter_source_node_ids.empty()) {
517
4.42k
            _tablet_reader_params.topn_filter_target_node_id =
518
4.42k
                    olap_scan_local_state->parent()->node_id();
519
4.42k
        }
520
1.26M
    }
521
522
    // If this is a Two-Phase read query, and we need to delay the release of Rowset
523
    // by rowset->update_delayed_expired_timestamp().This could expand the lifespan of Rowset
524
1.26M
    if (tablet_schema->field_index(BeConsts::ROWID_COL) >= 0) {
525
9
        constexpr static int delayed_s = 60;
526
12
        for (auto rs_reader : _tablet_reader_params.rs_splits) {
527
12
            uint64_t delayed_expired_timestamp =
528
12
                    UnixSeconds() + _tablet_reader_params.runtime_state->execution_timeout() +
529
12
                    delayed_s;
530
12
            rs_reader.rs_reader->rowset()->update_delayed_expired_timestamp(
531
12
                    delayed_expired_timestamp);
532
12
            ExecEnv::GetInstance()->storage_engine().add_quering_rowset(
533
12
                    rs_reader.rs_reader->rowset());
534
12
        }
535
9
    }
536
537
1.26M
    if (tablet_schema->has_global_row_id()) {
538
6.98k
        auto& id_file_map = _state->get_id_file_map();
539
14.5k
        for (auto rs_reader : _tablet_reader_params.rs_splits) {
540
14.5k
            id_file_map->add_temp_rowset(rs_reader.rs_reader->rowset());
541
14.5k
        }
542
6.98k
    }
543
544
1.26M
    return Status::OK();
545
1.26M
}
546
547
1.26M
Status OlapScanner::_init_variant_columns() {
548
1.26M
    auto& tablet_schema = _tablet_reader_params.tablet_schema;
549
1.26M
    if (tablet_schema->num_variant_columns() == 0) {
550
1.25M
        return Status::OK();
551
1.25M
    }
552
    // Parent column has path info to distinction from each other
553
17.9k
    for (auto* slot : _output_tuple_desc->slots()) {
554
17.9k
        if (slot->type()->get_primitive_type() == PrimitiveType::TYPE_VARIANT) {
555
            // Such columns are not exist in frontend schema info, so we need to
556
            // add them into tablet_schema for later column indexing.
557
9.17k
            TabletColumn subcol = TabletColumn::create_materialized_variant_column(
558
9.17k
                    tablet_schema->column_by_uid(slot->col_unique_id()).name_lower_case(),
559
9.17k
                    slot->column_paths(), slot->col_unique_id(),
560
9.17k
                    assert_cast<const DataTypeVariant&>(*remove_nullable(slot->type()))
561
9.17k
                            .variant_max_subcolumns_count());
562
9.17k
            if (tablet_schema->field_index(*subcol.path_info_ptr()) < 0) {
563
7.24k
                tablet_schema->append_column(subcol, TabletSchema::ColumnType::VARIANT);
564
7.24k
            }
565
9.17k
        }
566
17.9k
    }
567
6.14k
    variant_util::inherit_column_attributes(tablet_schema);
568
6.14k
    return Status::OK();
569
1.26M
}
570
571
1.26M
Status OlapScanner::_init_return_columns() {
572
12.7M
    for (auto* slot : _output_tuple_desc->slots()) {
573
        // variant column using path to index a column
574
12.7M
        int32_t index = 0;
575
12.7M
        auto& tablet_schema = _tablet_reader_params.tablet_schema;
576
12.7M
        if (slot->type()->get_primitive_type() == PrimitiveType::TYPE_VARIANT) {
577
9.17k
            index = tablet_schema->field_index(PathInData(
578
9.17k
                    tablet_schema->column_by_uid(slot->col_unique_id()).name_lower_case(),
579
9.17k
                    slot->column_paths()));
580
12.7M
        } else {
581
12.7M
            index = slot->col_unique_id() >= 0 ? tablet_schema->field_index(slot->col_unique_id())
582
18.4E
                                               : tablet_schema->field_index(slot->col_name());
583
12.7M
        }
584
585
12.7M
        if (index < 0) {
586
0
            return Status::InternalError(
587
0
                    "field name is invalid. field={}, field_name_to_index={}, col_unique_id={}",
588
0
                    slot->col_name(), tablet_schema->get_all_field_names(), slot->col_unique_id());
589
0
        }
590
591
12.7M
        if (slot->get_virtual_column_expr()) {
592
386
            ColumnId virtual_column_cid = index;
593
386
            _virtual_column_exprs[virtual_column_cid] = _slot_id_to_virtual_column_expr[slot->id()];
594
386
            size_t idx_in_block = _slot_id_to_index_in_block[slot->id()];
595
386
            _vir_cid_to_idx_in_block[virtual_column_cid] = idx_in_block;
596
386
            _vir_col_idx_to_type[idx_in_block] = _slot_id_to_col_type[slot->id()];
597
598
386
            VLOG_DEBUG << fmt::format(
599
4
                    "Virtual column, slot id: {}, cid {}, column index: {}, type: {}", slot->id(),
600
4
                    virtual_column_cid, _vir_cid_to_idx_in_block[virtual_column_cid],
601
4
                    _vir_col_idx_to_type[idx_in_block]->get_name());
602
386
        }
603
604
12.7M
        const auto& column = tablet_schema->column(index);
605
12.7M
        int32_t unique_id = column.unique_id() > 0 ? column.unique_id() : column.parent_unique_id();
606
12.7M
        if (!slot->all_access_paths().empty()) {
607
84.5k
            _tablet_reader_params.all_access_paths.insert({unique_id, slot->all_access_paths()});
608
84.5k
        }
609
610
12.7M
        if (!slot->predicate_access_paths().empty()) {
611
6.01k
            _tablet_reader_params.predicate_access_paths.insert(
612
6.01k
                    {unique_id, slot->predicate_access_paths()});
613
6.01k
        }
614
615
12.7M
        if ((slot->type()->get_primitive_type() == PrimitiveType::TYPE_STRUCT ||
616
12.7M
             slot->type()->get_primitive_type() == PrimitiveType::TYPE_MAP ||
617
12.7M
             slot->type()->get_primitive_type() == PrimitiveType::TYPE_ARRAY) &&
618
12.7M
            !slot->all_access_paths().empty()) {
619
77.4k
            tablet_schema->add_pruned_columns_data_type(column.unique_id(), slot->type());
620
77.4k
        }
621
622
12.7M
        _return_columns.push_back(index);
623
12.7M
        if (slot->is_nullable() && !tablet_schema->column(index).is_nullable()) {
624
0
            _tablet_columns_convert_to_null_set.emplace(index);
625
12.7M
        } else if (!slot->is_nullable() && tablet_schema->column(index).is_nullable()) {
626
0
            return Status::Error<ErrorCode::INVALID_SCHEMA>(
627
0
                    "slot(id: {}, name: {})'s nullable does not match "
628
0
                    "column(tablet id: {}, index: {}, name: {}) ",
629
0
                    slot->id(), slot->col_name(), tablet_schema->table_id(), index,
630
0
                    tablet_schema->column(index).name());
631
0
        }
632
12.7M
    }
633
634
1.26M
    if (_return_columns.empty()) {
635
0
        return Status::InternalError("failed to build storage scanner, no materialized slot!");
636
0
    }
637
638
1.26M
    return Status::OK();
639
1.26M
}
640
641
1.27M
doris::TabletStorageType OlapScanner::get_storage_type() {
642
1.27M
    if (config::is_cloud_mode()) {
643
        // we don't have cold storage in cloud mode, all storage is treated as local
644
985k
        return doris::TabletStorageType::STORAGE_TYPE_LOCAL;
645
985k
    }
646
285k
    int local_reader = 0;
647
510k
    for (const auto& reader : _tablet_reader_params.rs_splits) {
648
510k
        local_reader += reader.rs_reader->rowset()->is_local();
649
510k
    }
650
285k
    int total_reader = _tablet_reader_params.rs_splits.size();
651
652
285k
    if (local_reader == total_reader) {
653
285k
        return doris::TabletStorageType::STORAGE_TYPE_LOCAL;
654
285k
    } 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
285k
}
659
660
1.55M
Status OlapScanner::_get_block_impl(RuntimeState* state, Block* block, bool* eof) {
661
    // Read one block from block reader
662
    // ATTN: Here we need to let the _get_block_impl method guarantee the semantics of the interface,
663
    // that is, eof can be set to true only when the returned block is empty.
664
1.55M
    RETURN_IF_ERROR(_tablet_reader->next_block_with_aggregation(block, eof));
665
1.55M
    if (block->rows() > 0) {
666
290k
        _tablet_reader_params.tablet->read_block_count.fetch_add(1, std::memory_order_relaxed);
667
290k
        *eof = false;
668
290k
    }
669
1.55M
    return Status::OK();
670
1.55M
}
671
672
1.27M
Status OlapScanner::close(RuntimeState* state) {
673
1.27M
    if (!_try_close()) {
674
134
        return Status::OK();
675
134
    }
676
1.27M
    RETURN_IF_ERROR(Scanner::close(state));
677
1.27M
    return Status::OK();
678
1.27M
}
679
680
1.26M
void OlapScanner::update_realtime_counters() {
681
1.26M
    if (!_has_prepared) {
682
        // Counter update need prepare successfully, or it maybe core. For example, olap scanner
683
        // will open tablet reader during prepare, if not prepare successfully, tablet reader == nullptr.
684
0
        return;
685
0
    }
686
1.26M
    OlapScanLocalState* local_state = static_cast<OlapScanLocalState*>(_local_state);
687
1.26M
    const OlapReaderStatistics& stats = _tablet_reader->stats();
688
1.26M
    COUNTER_UPDATE(local_state->_read_compressed_counter, stats.compressed_bytes_read);
689
1.26M
    COUNTER_UPDATE(local_state->_read_uncompressed_counter, stats.uncompressed_bytes_read);
690
1.26M
    COUNTER_UPDATE(local_state->_scan_bytes, stats.uncompressed_bytes_read);
691
1.26M
    COUNTER_UPDATE(local_state->_scan_rows, stats.raw_rows_read);
692
693
    // Make sure the scan bytes and scan rows counter in audit log is the same as the counter in
694
    // doris metrics.
695
    // ScanBytes is the uncompressed bytes read from local + remote
696
    // bytes_read_from_local is the compressed bytes read from local
697
    // bytes_read_from_remote is the compressed bytes read from remote
698
    // scan bytes > bytes_read_from_local + bytes_read_from_remote
699
1.26M
    _state->get_query_ctx()->resource_ctx()->io_context()->update_scan_rows(stats.raw_rows_read);
700
1.26M
    _state->get_query_ctx()->resource_ctx()->io_context()->update_scan_bytes(
701
1.26M
            stats.uncompressed_bytes_read);
702
703
    // In case of no cache, we still need to update the IO stats. uncompressed bytes read == local + remote
704
1.26M
    if (stats.file_cache_stats.bytes_read_from_local == 0 &&
705
1.26M
        stats.file_cache_stats.bytes_read_from_remote == 0) {
706
1.04M
        _state->get_query_ctx()->resource_ctx()->io_context()->update_scan_bytes_from_local_storage(
707
1.04M
                stats.compressed_bytes_read);
708
1.04M
        DorisMetrics::instance()->query_scan_bytes_from_local->increment(
709
1.04M
                stats.compressed_bytes_read);
710
1.04M
    } else {
711
217k
        _state->get_query_ctx()->resource_ctx()->io_context()->update_scan_bytes_from_local_storage(
712
217k
                stats.file_cache_stats.bytes_read_from_local - _bytes_read_from_local);
713
217k
        _state->get_query_ctx()
714
217k
                ->resource_ctx()
715
217k
                ->io_context()
716
217k
                ->update_scan_bytes_from_remote_storage(
717
217k
                        stats.file_cache_stats.bytes_read_from_remote - _bytes_read_from_remote);
718
719
217k
        DorisMetrics::instance()->query_scan_bytes_from_local->increment(
720
217k
                stats.file_cache_stats.bytes_read_from_local - _bytes_read_from_local);
721
217k
        DorisMetrics::instance()->query_scan_bytes_from_remote->increment(
722
217k
                stats.file_cache_stats.bytes_read_from_remote - _bytes_read_from_remote);
723
217k
    }
724
725
1.26M
    _tablet_reader->mutable_stats()->compressed_bytes_read = 0;
726
1.26M
    _tablet_reader->mutable_stats()->uncompressed_bytes_read = 0;
727
1.26M
    _tablet_reader->mutable_stats()->raw_rows_read = 0;
728
729
1.26M
    _bytes_read_from_local = _tablet_reader->stats().file_cache_stats.bytes_read_from_local;
730
1.26M
    _bytes_read_from_remote = _tablet_reader->stats().file_cache_stats.bytes_read_from_remote;
731
1.26M
}
732
733
1.26M
void OlapScanner::_collect_profile_before_close() {
734
    //  Please don't directly enable the profile here, we need to set QueryStatistics using the counter inside.
735
1.26M
    if (_has_updated_counter) {
736
0
        return;
737
0
    }
738
1.26M
    _has_updated_counter = true;
739
1.26M
    _tablet_reader->update_profile(_profile);
740
741
1.26M
    Scanner::_collect_profile_before_close();
742
743
    // Update counters for OlapScanner
744
    // Update counters from tablet reader's stats
745
1.26M
    auto& stats = _tablet_reader->stats();
746
1.26M
    auto* local_state = (OlapScanLocalState*)_local_state;
747
1.26M
    COUNTER_UPDATE(local_state->_io_timer, stats.io_ns);
748
1.26M
    COUNTER_UPDATE(local_state->_read_compressed_counter, stats.compressed_bytes_read);
749
1.26M
    COUNTER_UPDATE(local_state->_scan_bytes, stats.uncompressed_bytes_read);
750
1.26M
    COUNTER_UPDATE(local_state->_decompressor_timer, stats.decompress_ns);
751
1.26M
    COUNTER_UPDATE(local_state->_read_uncompressed_counter, stats.uncompressed_bytes_read);
752
1.26M
    COUNTER_UPDATE(local_state->_block_load_timer, stats.block_load_ns);
753
1.26M
    COUNTER_UPDATE(local_state->_block_load_counter, stats.blocks_load);
754
1.26M
    COUNTER_UPDATE(local_state->_block_fetch_timer, stats.block_fetch_ns);
755
1.26M
    COUNTER_UPDATE(local_state->_delete_bitmap_get_agg_timer, stats.delete_bitmap_get_agg_ns);
756
1.26M
    COUNTER_UPDATE(local_state->_scan_rows, stats.raw_rows_read);
757
1.26M
    COUNTER_UPDATE(local_state->_vec_cond_timer, stats.vec_cond_ns);
758
1.26M
    COUNTER_UPDATE(local_state->_short_cond_timer, stats.short_cond_ns);
759
1.26M
    COUNTER_UPDATE(local_state->_expr_filter_timer, stats.expr_filter_ns);
760
1.26M
    COUNTER_UPDATE(local_state->_block_init_timer, stats.block_init_ns);
761
1.26M
    COUNTER_UPDATE(local_state->_block_init_seek_timer, stats.block_init_seek_ns);
762
1.26M
    COUNTER_UPDATE(local_state->_block_init_seek_counter, stats.block_init_seek_num);
763
1.26M
    COUNTER_UPDATE(local_state->_segment_generate_row_range_by_keys_timer,
764
1.26M
                   stats.generate_row_ranges_by_keys_ns);
765
1.26M
    COUNTER_UPDATE(local_state->_segment_generate_row_range_by_column_conditions_timer,
766
1.26M
                   stats.generate_row_ranges_by_column_conditions_ns);
767
1.26M
    COUNTER_UPDATE(local_state->_segment_generate_row_range_by_bf_timer,
768
1.26M
                   stats.generate_row_ranges_by_bf_ns);
769
1.26M
    COUNTER_UPDATE(local_state->_collect_iterator_merge_next_timer,
770
1.26M
                   stats.collect_iterator_merge_next_timer);
771
1.26M
    COUNTER_UPDATE(local_state->_segment_generate_row_range_by_zonemap_timer,
772
1.26M
                   stats.generate_row_ranges_by_zonemap_ns);
773
1.26M
    COUNTER_UPDATE(local_state->_segment_generate_row_range_by_dict_timer,
774
1.26M
                   stats.generate_row_ranges_by_dict_ns);
775
1.26M
    COUNTER_UPDATE(local_state->_predicate_column_read_timer, stats.predicate_column_read_ns);
776
1.26M
    COUNTER_UPDATE(local_state->_non_predicate_column_read_timer, stats.non_predicate_read_ns);
777
1.26M
    COUNTER_UPDATE(local_state->_predicate_column_read_seek_timer,
778
1.26M
                   stats.predicate_column_read_seek_ns);
779
1.26M
    COUNTER_UPDATE(local_state->_predicate_column_read_seek_counter,
780
1.26M
                   stats.predicate_column_read_seek_num);
781
1.26M
    COUNTER_UPDATE(local_state->_lazy_read_timer, stats.lazy_read_ns);
782
1.26M
    COUNTER_UPDATE(local_state->_lazy_read_seek_timer, stats.block_lazy_read_seek_ns);
783
1.26M
    COUNTER_UPDATE(local_state->_lazy_read_seek_counter, stats.block_lazy_read_seek_num);
784
1.26M
    COUNTER_UPDATE(local_state->_output_col_timer, stats.output_col_ns);
785
1.26M
    COUNTER_UPDATE(local_state->_rows_vec_cond_filtered_counter, stats.rows_vec_cond_filtered);
786
1.26M
    COUNTER_UPDATE(local_state->_rows_short_circuit_cond_filtered_counter,
787
1.26M
                   stats.rows_short_circuit_cond_filtered);
788
1.26M
    COUNTER_UPDATE(local_state->_rows_expr_cond_filtered_counter, stats.rows_expr_cond_filtered);
789
1.26M
    COUNTER_UPDATE(local_state->_rows_vec_cond_input_counter, stats.vec_cond_input_rows);
790
1.26M
    COUNTER_UPDATE(local_state->_rows_short_circuit_cond_input_counter,
791
1.26M
                   stats.short_circuit_cond_input_rows);
792
1.26M
    COUNTER_UPDATE(local_state->_rows_expr_cond_input_counter, stats.expr_cond_input_rows);
793
1.26M
    COUNTER_UPDATE(local_state->_stats_filtered_counter, stats.rows_stats_filtered);
794
1.26M
    COUNTER_UPDATE(local_state->_stats_rp_filtered_counter, stats.rows_stats_rp_filtered);
795
1.26M
    COUNTER_UPDATE(local_state->_dict_filtered_counter, stats.segment_dict_filtered);
796
1.26M
    COUNTER_UPDATE(local_state->_bf_filtered_counter, stats.rows_bf_filtered);
797
1.26M
    COUNTER_UPDATE(local_state->_del_filtered_counter, stats.rows_del_filtered);
798
1.26M
    COUNTER_UPDATE(local_state->_del_filtered_counter, stats.rows_del_by_bitmap);
799
1.26M
    COUNTER_UPDATE(local_state->_del_filtered_counter, stats.rows_vec_del_cond_filtered);
800
1.26M
    COUNTER_UPDATE(local_state->_conditions_filtered_counter, stats.rows_conditions_filtered);
801
1.26M
    COUNTER_UPDATE(local_state->_key_range_filtered_counter, stats.rows_key_range_filtered);
802
1.26M
    COUNTER_UPDATE(local_state->_total_pages_num_counter, stats.total_pages_num);
803
1.26M
    COUNTER_UPDATE(local_state->_cached_pages_num_counter, stats.cached_pages_num);
804
1.26M
    COUNTER_UPDATE(local_state->_inverted_index_filter_counter, stats.rows_inverted_index_filtered);
805
1.26M
    COUNTER_UPDATE(local_state->_inverted_index_filter_timer, stats.inverted_index_filter_timer);
806
1.26M
    COUNTER_UPDATE(local_state->_inverted_index_query_cache_hit_counter,
807
1.26M
                   stats.inverted_index_query_cache_hit);
808
1.26M
    COUNTER_UPDATE(local_state->_inverted_index_query_cache_miss_counter,
809
1.26M
                   stats.inverted_index_query_cache_miss);
810
1.26M
    COUNTER_UPDATE(local_state->_inverted_index_query_timer, stats.inverted_index_query_timer);
811
1.26M
    COUNTER_UPDATE(local_state->_inverted_index_query_null_bitmap_timer,
812
1.26M
                   stats.inverted_index_query_null_bitmap_timer);
813
1.26M
    COUNTER_UPDATE(local_state->_inverted_index_query_bitmap_copy_timer,
814
1.26M
                   stats.inverted_index_query_bitmap_copy_timer);
815
1.26M
    COUNTER_UPDATE(local_state->_inverted_index_searcher_open_timer,
816
1.26M
                   stats.inverted_index_searcher_open_timer);
817
1.26M
    COUNTER_UPDATE(local_state->_inverted_index_searcher_search_timer,
818
1.26M
                   stats.inverted_index_searcher_search_timer);
819
1.26M
    COUNTER_UPDATE(local_state->_inverted_index_searcher_search_init_timer,
820
1.26M
                   stats.inverted_index_searcher_search_init_timer);
821
1.26M
    COUNTER_UPDATE(local_state->_inverted_index_searcher_search_exec_timer,
822
1.26M
                   stats.inverted_index_searcher_search_exec_timer);
823
1.26M
    COUNTER_UPDATE(local_state->_inverted_index_searcher_cache_hit_counter,
824
1.26M
                   stats.inverted_index_searcher_cache_hit);
825
1.26M
    COUNTER_UPDATE(local_state->_inverted_index_searcher_cache_miss_counter,
826
1.26M
                   stats.inverted_index_searcher_cache_miss);
827
1.26M
    COUNTER_UPDATE(local_state->_inverted_index_downgrade_count_counter,
828
1.26M
                   stats.inverted_index_downgrade_count);
829
1.26M
    COUNTER_UPDATE(local_state->_inverted_index_analyzer_timer,
830
1.26M
                   stats.inverted_index_analyzer_timer);
831
1.26M
    COUNTER_UPDATE(local_state->_inverted_index_lookup_timer, stats.inverted_index_lookup_timer);
832
1.26M
    COUNTER_UPDATE(local_state->_variant_scan_sparse_column_timer,
833
1.26M
                   stats.variant_scan_sparse_column_timer_ns);
834
1.26M
    COUNTER_UPDATE(local_state->_variant_scan_sparse_column_bytes,
835
1.26M
                   stats.variant_scan_sparse_column_bytes);
836
1.26M
    COUNTER_UPDATE(local_state->_variant_fill_path_from_sparse_column_timer,
837
1.26M
                   stats.variant_fill_path_from_sparse_column_timer_ns);
838
1.26M
    COUNTER_UPDATE(local_state->_variant_subtree_default_iter_count,
839
1.26M
                   stats.variant_subtree_default_iter_count);
840
1.26M
    COUNTER_UPDATE(local_state->_variant_subtree_leaf_iter_count,
841
1.26M
                   stats.variant_subtree_leaf_iter_count);
842
1.26M
    COUNTER_UPDATE(local_state->_variant_subtree_hierarchical_iter_count,
843
1.26M
                   stats.variant_subtree_hierarchical_iter_count);
844
1.26M
    COUNTER_UPDATE(local_state->_variant_subtree_sparse_iter_count,
845
1.26M
                   stats.variant_subtree_sparse_iter_count);
846
1.26M
    COUNTER_UPDATE(local_state->_variant_doc_value_column_iter_count,
847
1.26M
                   stats.variant_doc_value_column_iter_count);
848
849
1.26M
    InvertedIndexProfileReporter inverted_index_profile;
850
1.26M
    inverted_index_profile.update(local_state->_index_filter_profile.get(),
851
1.26M
                                  &stats.inverted_index_stats);
852
853
    // only cloud deploy mode will use file cache.
854
1.26M
    if (config::is_cloud_mode() && config::enable_file_cache) {
855
979k
        io::FileCacheProfileReporter cache_profile(local_state->_segment_profile.get());
856
979k
        cache_profile.update(&stats.file_cache_stats);
857
979k
        _state->get_query_ctx()->resource_ctx()->io_context()->update_bytes_write_into_cache(
858
979k
                stats.file_cache_stats.bytes_write_into_cache);
859
979k
    }
860
1.26M
    COUNTER_UPDATE(local_state->_output_index_result_column_timer,
861
1.26M
                   stats.output_index_result_column_timer);
862
1.26M
    COUNTER_UPDATE(local_state->_filtered_segment_counter, stats.filtered_segment_number);
863
1.26M
    COUNTER_UPDATE(local_state->_total_segment_counter, stats.total_segment_number);
864
1.26M
    COUNTER_UPDATE(local_state->_condition_cache_hit_counter, stats.condition_cache_hit_seg_nums);
865
1.26M
    COUNTER_UPDATE(local_state->_condition_cache_filtered_rows_counter,
866
1.26M
                   stats.condition_cache_filtered_rows);
867
868
1.26M
    COUNTER_UPDATE(local_state->_tablet_reader_init_timer, stats.tablet_reader_init_timer_ns);
869
1.26M
    COUNTER_UPDATE(local_state->_tablet_reader_capture_rs_readers_timer,
870
1.26M
                   stats.tablet_reader_capture_rs_readers_timer_ns);
871
1.26M
    COUNTER_UPDATE(local_state->_tablet_reader_init_return_columns_timer,
872
1.26M
                   stats.tablet_reader_init_return_columns_timer_ns);
873
1.26M
    COUNTER_UPDATE(local_state->_tablet_reader_init_keys_param_timer,
874
1.26M
                   stats.tablet_reader_init_keys_param_timer_ns);
875
1.26M
    COUNTER_UPDATE(local_state->_tablet_reader_init_orderby_keys_param_timer,
876
1.26M
                   stats.tablet_reader_init_orderby_keys_param_timer_ns);
877
1.26M
    COUNTER_UPDATE(local_state->_tablet_reader_init_conditions_param_timer,
878
1.26M
                   stats.tablet_reader_init_conditions_param_timer_ns);
879
1.26M
    COUNTER_UPDATE(local_state->_tablet_reader_init_delete_condition_param_timer,
880
1.26M
                   stats.tablet_reader_init_delete_condition_param_timer_ns);
881
1.26M
    COUNTER_UPDATE(local_state->_block_reader_vcollect_iter_init_timer,
882
1.26M
                   stats.block_reader_vcollect_iter_init_timer_ns);
883
1.26M
    COUNTER_UPDATE(local_state->_block_reader_rs_readers_init_timer,
884
1.26M
                   stats.block_reader_rs_readers_init_timer_ns);
885
1.26M
    COUNTER_UPDATE(local_state->_block_reader_build_heap_init_timer,
886
1.26M
                   stats.block_reader_build_heap_init_timer_ns);
887
888
1.26M
    COUNTER_UPDATE(local_state->_rowset_reader_get_segment_iterators_timer,
889
1.26M
                   stats.rowset_reader_get_segment_iterators_timer_ns);
890
1.26M
    COUNTER_UPDATE(local_state->_rowset_reader_create_iterators_timer,
891
1.26M
                   stats.rowset_reader_create_iterators_timer_ns);
892
1.26M
    COUNTER_UPDATE(local_state->_rowset_reader_init_iterators_timer,
893
1.26M
                   stats.rowset_reader_init_iterators_timer_ns);
894
1.26M
    COUNTER_UPDATE(local_state->_rowset_reader_load_segments_timer,
895
1.26M
                   stats.rowset_reader_load_segments_timer_ns);
896
897
1.26M
    COUNTER_UPDATE(local_state->_segment_iterator_init_timer, stats.segment_iterator_init_timer_ns);
898
1.26M
    COUNTER_UPDATE(local_state->_segment_iterator_init_return_column_iterators_timer,
899
1.26M
                   stats.segment_iterator_init_return_column_iterators_timer_ns);
900
1.26M
    COUNTER_UPDATE(local_state->_segment_iterator_init_index_iterators_timer,
901
1.26M
                   stats.segment_iterator_init_index_iterators_timer_ns);
902
903
1.26M
    COUNTER_UPDATE(local_state->_segment_create_column_readers_timer,
904
1.26M
                   stats.segment_create_column_readers_timer_ns);
905
1.26M
    COUNTER_UPDATE(local_state->_segment_load_index_timer, stats.segment_load_index_timer_ns);
906
907
    // Update metrics
908
1.26M
    DorisMetrics::instance()->query_scan_bytes->increment(
909
1.26M
            local_state->_read_uncompressed_counter->value());
910
1.26M
    DorisMetrics::instance()->query_scan_rows->increment(local_state->_scan_rows->value());
911
1.26M
    auto& tablet = _tablet_reader_params.tablet;
912
1.26M
    tablet->query_scan_bytes->increment(local_state->_read_uncompressed_counter->value());
913
1.26M
    tablet->query_scan_rows->increment(local_state->_scan_rows->value());
914
1.26M
    tablet->query_scan_count->increment(1);
915
916
1.26M
    COUNTER_UPDATE(local_state->_ann_range_search_filter_counter,
917
1.26M
                   stats.rows_ann_index_range_filtered);
918
1.26M
    COUNTER_UPDATE(local_state->_ann_topn_filter_counter, stats.rows_ann_index_topn_filtered);
919
1.26M
    COUNTER_UPDATE(local_state->_ann_index_load_costs, stats.ann_index_load_ns);
920
1.26M
    COUNTER_UPDATE(local_state->_ann_range_search_costs, stats.ann_index_range_search_ns);
921
1.26M
    COUNTER_UPDATE(local_state->_ann_range_search_cnt, stats.ann_index_range_search_cnt);
922
1.26M
    COUNTER_UPDATE(local_state->_ann_range_engine_search_costs, stats.ann_range_engine_search_ns);
923
    // Engine prepare before search
924
1.26M
    COUNTER_UPDATE(local_state->_ann_range_pre_process_costs, stats.ann_range_pre_process_ns);
925
    // Post process parent: Doris result process + engine convert
926
1.26M
    COUNTER_UPDATE(local_state->_ann_range_post_process_costs,
927
1.26M
                   stats.ann_range_result_convert_ns + stats.ann_range_engine_convert_ns);
928
    // Engine convert (child under post-process)
929
1.26M
    COUNTER_UPDATE(local_state->_ann_range_engine_convert_costs, stats.ann_range_engine_convert_ns);
930
    // Doris-side result convert (child under post-process)
931
1.26M
    COUNTER_UPDATE(local_state->_ann_range_result_convert_costs, stats.ann_range_result_convert_ns);
932
933
1.26M
    COUNTER_UPDATE(local_state->_ann_topn_search_costs, stats.ann_topn_search_ns);
934
1.26M
    COUNTER_UPDATE(local_state->_ann_topn_search_cnt, stats.ann_index_topn_search_cnt);
935
936
    // Detailed ANN timers
937
    // ANN TopN timers with hierarchy
938
    // Engine search time (FAISS)
939
1.26M
    COUNTER_UPDATE(local_state->_ann_topn_engine_search_costs,
940
1.26M
                   stats.ann_index_topn_engine_search_ns);
941
    // Engine prepare time (allocations/buffer setup before search)
942
1.26M
    COUNTER_UPDATE(local_state->_ann_topn_pre_process_costs,
943
1.26M
                   stats.ann_index_topn_engine_prepare_ns);
944
    // Post process parent includes Doris result processing + engine convert
945
1.26M
    COUNTER_UPDATE(local_state->_ann_topn_post_process_costs,
946
1.26M
                   stats.ann_index_topn_result_process_ns + stats.ann_index_topn_engine_convert_ns);
947
    // Engine-side conversion time inside FAISS wrappers (child under post-process)
948
1.26M
    COUNTER_UPDATE(local_state->_ann_topn_engine_convert_costs,
949
1.26M
                   stats.ann_index_topn_engine_convert_ns);
950
951
    // Doris-side result convert costs (show separately as another child counter); use pure process time
952
1.26M
    COUNTER_UPDATE(local_state->_ann_topn_result_convert_costs,
953
1.26M
                   stats.ann_index_topn_result_process_ns);
954
955
1.26M
    COUNTER_UPDATE(local_state->_ann_fallback_brute_force_cnt, stats.ann_fall_back_brute_force_cnt);
956
957
    // Overhead counter removed; precise instrumentation is reported via engine_prepare above.
958
1.26M
}
959
960
#include "common/compile_check_avoid_end.h"
961
} // namespace doris