Coverage Report

Created: 2026-03-26 13:18

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
985k
        : Scanner(params.state, parent, params.limit, params.profile),
72
985k
          _key_ranges(std::move(params.key_ranges)),
73
985k
          _tablet_reader_params({.tablet = std::move(params.tablet),
74
985k
                                 .tablet_schema {},
75
985k
                                 .aggregation = params.aggregation,
76
985k
                                 .version = {0, params.version},
77
985k
                                 .start_key {},
78
985k
                                 .end_key {},
79
985k
                                 .predicates {},
80
985k
                                 .function_filters {},
81
985k
                                 .delete_predicates {},
82
985k
                                 .target_cast_type_for_variants {},
83
985k
                                 .all_access_paths {},
84
985k
                                 .predicate_access_paths {},
85
985k
                                 .rs_splits {},
86
985k
                                 .return_columns {},
87
985k
                                 .output_columns {},
88
985k
                                 .remaining_conjunct_roots {},
89
985k
                                 .common_expr_ctxs_push_down {},
90
985k
                                 .topn_filter_source_node_ids {},
91
985k
                                 .filter_block_conjuncts {},
92
985k
                                 .key_group_cluster_key_idxes {},
93
985k
                                 .virtual_column_exprs {},
94
985k
                                 .vir_cid_to_idx_in_block {},
95
985k
                                 .vir_col_idx_to_type {},
96
985k
                                 .score_runtime {},
97
985k
                                 .collection_statistics {},
98
985k
                                 .ann_topn_runtime {},
99
985k
                                 .condition_cache_digest = parent->get_condition_cache_digest()}) {
100
985k
    _tablet_reader_params.set_read_source(std::move(params.read_source),
101
985k
                                          _state->skip_delete_bitmap());
102
985k
    _has_prepared = false;
103
985k
    _vector_search_params = params.state->get_vector_search_params();
104
985k
}
105
106
static std::string read_columns_to_string(TabletSchemaSPtr tablet_schema,
107
2.86k
                                          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
2.86k
    const int col_per_line = 10;
111
2.86k
    int i = 0;
112
2.86k
    std::string read_columns_string;
113
2.86k
    read_columns_string += "[";
114
12.9k
    for (auto it = read_columns.cbegin(); it != read_columns.cend(); it++) {
115
10.1k
        if (it != read_columns.cbegin()) {
116
7.32k
            read_columns_string += ", ";
117
7.32k
        }
118
10.1k
        read_columns_string += tablet_schema->columns().at(*it)->name();
119
10.1k
        if (i >= col_per_line) {
120
13
            read_columns_string += "\n";
121
13
            i = 0;
122
10.1k
        } else {
123
10.1k
            ++i;
124
10.1k
        }
125
10.1k
    }
126
2.86k
    read_columns_string += "]";
127
2.86k
    return read_columns_string;
128
2.86k
}
129
130
985k
Status OlapScanner::prepare() {
131
985k
    auto* local_state = static_cast<OlapScanLocalState*>(_local_state);
132
985k
    auto& tablet = _tablet_reader_params.tablet;
133
985k
    auto& tablet_schema = _tablet_reader_params.tablet_schema;
134
985k
    DBUG_EXECUTE_IF("CloudTablet.capture_rs_readers.return.e-230", {
135
985k
        LOG_WARNING("CloudTablet.capture_rs_readers.return e-230 init")
136
985k
                .tag("tablet_id", tablet->tablet_id());
137
985k
        return Status::Error<false>(-230, "injected error");
138
985k
    });
139
140
985k
    for (auto& ctx : local_state->_common_expr_ctxs_push_down) {
141
13.7k
        VExprContextSPtr context;
142
13.7k
        RETURN_IF_ERROR(ctx->clone(_state, context));
143
13.7k
        _common_expr_ctxs_push_down.emplace_back(context);
144
13.7k
        context->prepare_ann_range_search(_vector_search_params);
145
13.7k
    }
146
147
985k
    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
368
        VExprContextSPtr context;
150
368
        RETURN_IF_ERROR(pair.second->clone(_state, context));
151
368
        _slot_id_to_virtual_column_expr[pair.first] = context;
152
368
    }
153
154
985k
    _slot_id_to_index_in_block = local_state->_slot_id_to_index_in_block;
155
985k
    _slot_id_to_col_type = local_state->_slot_id_to_col_type;
156
985k
    _score_runtime = local_state->_score_runtime;
157
    // All scanners share the same ann_topn_runtime.
158
985k
    _ann_topn_runtime = local_state->_ann_topn_runtime;
159
160
    // set limit to reduce end of rowset and segment mem use
161
985k
    _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
985k
    _tablet_reader->set_batch_size(_state->batch_size());
167
985k
    TabletSchemaSPtr cached_schema;
168
985k
    std::string schema_key;
169
985k
    {
170
985k
        TOlapScanNode& olap_scan_node = local_state->olap_scan_node();
171
172
985k
        const auto check_can_use_cache = [&]() {
173
984k
            if (!(olap_scan_node.__isset.schema_version && olap_scan_node.__isset.columns_desc &&
174
984k
                  !olap_scan_node.columns_desc.empty() &&
175
984k
                  olap_scan_node.columns_desc[0].col_unique_id >= 0 && // Why check first column?
176
984k
                  tablet->tablet_schema()->num_variant_columns() == 0 &&
177
984k
                  tablet->tablet_schema()->num_virtual_columns() == 0)) {
178
6.23k
                return false;
179
6.23k
            }
180
181
978k
            const bool has_pruned_column =
182
8.46M
                    std::ranges::any_of(_output_tuple_desc->slots(), [](const auto& slot) {
183
8.46M
                        if ((slot->type()->get_primitive_type() == PrimitiveType::TYPE_STRUCT ||
184
8.46M
                             slot->type()->get_primitive_type() == PrimitiveType::TYPE_MAP ||
185
8.46M
                             slot->type()->get_primitive_type() == PrimitiveType::TYPE_ARRAY) &&
186
8.46M
                            !slot->all_access_paths().empty()) {
187
23.4k
                            return true;
188
23.4k
                        }
189
8.44M
                        return false;
190
8.46M
                    });
191
978k
            return !has_pruned_column;
192
984k
        }();
193
194
985k
        if (check_can_use_cache) {
195
955k
            schema_key =
196
955k
                    SchemaCache::get_schema_key(tablet->tablet_id(), olap_scan_node.columns_desc,
197
955k
                                                olap_scan_node.schema_version);
198
955k
            cached_schema = SchemaCache::instance()->get_schema(schema_key);
199
955k
        }
200
985k
        if (cached_schema && cached_schema->num_virtual_columns() == 0) {
201
779k
            tablet_schema = cached_schema;
202
779k
        } else {
203
            // If schema is not cached or cached schema has virtual columns,
204
            // we need to create a new TabletSchema.
205
205k
            tablet_schema = std::make_shared<TabletSchema>();
206
205k
            tablet_schema->copy_from(*tablet->tablet_schema());
207
205k
            if (olap_scan_node.__isset.columns_desc && !olap_scan_node.columns_desc.empty() &&
208
205k
                olap_scan_node.columns_desc[0].col_unique_id >= 0) {
209
                // Originally scanner get TabletSchema from tablet object in BE.
210
                // To support lightweight schema change for adding / dropping columns,
211
                // tabletschema is bounded to rowset and tablet's schema maybe outdated,
212
                //  so we have to use schema from a query plan witch FE puts it in query plans.
213
204k
                tablet_schema->clear_columns();
214
2.89M
                for (const auto& column_desc : olap_scan_node.columns_desc) {
215
2.89M
                    tablet_schema->append_column(TabletColumn(column_desc));
216
2.89M
                }
217
204k
                if (olap_scan_node.__isset.schema_version) {
218
204k
                    tablet_schema->set_schema_version(olap_scan_node.schema_version);
219
204k
                }
220
204k
            }
221
205k
            if (olap_scan_node.__isset.indexes_desc) {
222
204k
                tablet_schema->update_indexes_from_thrift(olap_scan_node.indexes_desc);
223
204k
            }
224
205k
        }
225
226
985k
        if (_tablet_reader_params.rs_splits.empty()) {
227
            // Non-pipeline mode, Tablet : Scanner = 1 : 1
228
            // acquire tablet rowset readers at the beginning of the scan node
229
            // to prevent this case: when there are lots of olap scanners to run for example 10000
230
            // the rowsets maybe compacted when the last olap scanner starts
231
0
            ReadSource read_source;
232
233
0
            if (config::is_cloud_mode()) {
234
                // FIXME(plat1ko): Avoid pointer cast
235
0
                ExecEnv::GetInstance()->storage_engine().to_cloud().tablet_hotspot().count(*tablet);
236
0
            }
237
238
0
            auto maybe_read_source = tablet->capture_read_source(
239
0
                    _tablet_reader_params.version,
240
0
                    {
241
0
                            .skip_missing_versions = _state->skip_missing_version(),
242
0
                            .enable_fetch_rowsets_from_peers =
243
0
                                    config::enable_fetch_rowsets_from_peer_replicas,
244
0
                            .enable_prefer_cached_rowset =
245
0
                                    config::is_cloud_mode() ? _state->enable_prefer_cached_rowset()
246
0
                                                            : false,
247
0
                            .query_freshness_tolerance_ms =
248
0
                                    config::is_cloud_mode() ? _state->query_freshness_tolerance_ms()
249
0
                                                            : -1,
250
0
                    });
251
0
            if (!maybe_read_source) {
252
0
                LOG(WARNING) << "fail to init reader. res=" << maybe_read_source.error();
253
0
                return maybe_read_source.error();
254
0
            }
255
256
0
            read_source = std::move(maybe_read_source.value());
257
258
0
            if (config::enable_mow_verbose_log && tablet->enable_unique_key_merge_on_write()) {
259
0
                LOG_INFO("finish capture_rs_readers for tablet={}, query_id={}",
260
0
                         tablet->tablet_id(), print_id(_state->query_id()));
261
0
            }
262
263
0
            if (!_state->skip_delete_predicate()) {
264
0
                read_source.fill_delete_predicates();
265
0
            }
266
0
            _tablet_reader_params.set_read_source(std::move(read_source));
267
0
        }
268
269
        // Initialize tablet_reader_params
270
985k
        RETURN_IF_ERROR(_init_tablet_reader_params(
271
985k
                local_state->_parent->cast<OlapScanOperatorX>()._slot_id_to_slot_desc, _key_ranges,
272
985k
                local_state->_slot_id_to_predicates, local_state->_push_down_functions));
273
985k
    }
274
275
    // add read columns in profile
276
985k
    if (_state->enable_profile()) {
277
2.83k
        _profile->add_info_string("ReadColumns",
278
2.83k
                                  read_columns_to_string(tablet_schema, _return_columns));
279
2.83k
    }
280
281
    // Add newly created tablet schema to schema cache if it does not have virtual columns.
282
985k
    if (cached_schema == nullptr && !schema_key.empty() &&
283
985k
        tablet_schema->num_virtual_columns() == 0 && !tablet_schema->has_pruned_columns()) {
284
174k
        SchemaCache::instance()->insert_schema(schema_key, tablet_schema);
285
174k
    }
286
287
985k
    if (_tablet_reader_params.score_runtime) {
288
14
        SCOPED_TIMER(local_state->_statistics_collect_timer);
289
14
        _tablet_reader_params.collection_statistics = std::make_shared<CollectionStatistics>();
290
291
14
        io::IOContext io_ctx {
292
14
                .reader_type = ReaderType::READER_QUERY,
293
14
                .expiration_time = tablet->ttl_seconds(),
294
14
                .query_id = &_state->query_id(),
295
14
                .file_cache_stats = &_tablet_reader->mutable_stats()->file_cache_stats,
296
14
                .is_inverted_index = true,
297
14
        };
298
299
14
        RETURN_IF_ERROR(_tablet_reader_params.collection_statistics->collect(
300
14
                _state, _tablet_reader_params.rs_splits, _tablet_reader_params.tablet_schema,
301
14
                _tablet_reader_params.common_expr_ctxs_push_down, &io_ctx));
302
14
    }
303
304
985k
    _has_prepared = true;
305
985k
    return Status::OK();
306
985k
}
307
308
981k
Status OlapScanner::_open_impl(RuntimeState* state) {
309
981k
    RETURN_IF_ERROR(Scanner::_open_impl(state));
310
981k
    SCOPED_TIMER(_local_state->cast<OlapScanLocalState>()._reader_init_timer);
311
312
981k
    auto res = _tablet_reader->init(_tablet_reader_params);
313
981k
    if (!res.ok()) {
314
53
        res.append("failed to initialize storage reader. tablet=" +
315
53
                   std::to_string(_tablet_reader_params.tablet->tablet_id()) +
316
53
                   ", backend=" + BackendOptions::get_localhost());
317
53
        return res;
318
53
    }
319
320
    // Do not hold rs_splits any more to release memory.
321
981k
    _tablet_reader_params.rs_splits.clear();
322
323
981k
    return Status::OK();
324
981k
}
325
326
// it will be called under tablet read lock because capture rs readers need
327
Status OlapScanner::_init_tablet_reader_params(
328
        const phmap::flat_hash_map<int, SlotDescriptor*>& slot_id_to_slot_desc,
329
        const std::vector<OlapScanRange*>& key_ranges,
330
        const phmap::flat_hash_map<int, std::vector<std::shared_ptr<ColumnPredicate>>>&
331
                slot_to_predicates,
332
983k
        const std::vector<FunctionFilter>& function_filters) {
333
    // if the table with rowset [0-x] or [0-1] [2-y], and [0-1] is empty
334
983k
    const bool single_version = _tablet_reader_params.has_single_version();
335
336
983k
    auto* olap_local_state = static_cast<OlapScanLocalState*>(_local_state);
337
983k
    bool read_mor_as_dup = olap_local_state->olap_scan_node().__isset.read_mor_as_dup &&
338
983k
                           olap_local_state->olap_scan_node().read_mor_as_dup;
339
984k
    if (_state->skip_storage_engine_merge() || read_mor_as_dup) {
340
49
        _tablet_reader_params.direct_mode = true;
341
49
        _tablet_reader_params.aggregation = true;
342
983k
    } else {
343
983k
        auto push_down_agg_type = _local_state->get_push_down_agg_type();
344
983k
        _tablet_reader_params.direct_mode = _tablet_reader_params.aggregation || single_version ||
345
983k
                                            (push_down_agg_type != TPushAggOp::NONE &&
346
10.1k
                                             push_down_agg_type != TPushAggOp::COUNT_ON_INDEX);
347
983k
    }
348
349
983k
    RETURN_IF_ERROR(_init_variant_columns());
350
983k
    RETURN_IF_ERROR(_init_return_columns());
351
352
983k
    _tablet_reader_params.reader_type = ReaderType::READER_QUERY;
353
983k
    _tablet_reader_params.push_down_agg_type_opt = _local_state->get_push_down_agg_type();
354
355
    // TODO: If a new runtime filter arrives after `_conjuncts` move to `_common_expr_ctxs_push_down`,
356
983k
    if (_common_expr_ctxs_push_down.empty()) {
357
972k
        for (auto& conjunct : _conjuncts) {
358
11.1k
            _tablet_reader_params.remaining_conjunct_roots.emplace_back(conjunct->root());
359
11.1k
        }
360
972k
    } else {
361
13.5k
        for (auto& ctx : _common_expr_ctxs_push_down) {
362
13.5k
            _tablet_reader_params.remaining_conjunct_roots.emplace_back(ctx->root());
363
13.5k
        }
364
10.9k
    }
365
366
983k
    _tablet_reader_params.common_expr_ctxs_push_down = _common_expr_ctxs_push_down;
367
983k
    _tablet_reader_params.virtual_column_exprs = _virtual_column_exprs;
368
983k
    _tablet_reader_params.vir_cid_to_idx_in_block = _vir_cid_to_idx_in_block;
369
983k
    _tablet_reader_params.vir_col_idx_to_type = _vir_col_idx_to_type;
370
983k
    _tablet_reader_params.score_runtime = _score_runtime;
371
983k
    _tablet_reader_params.output_columns = ((OlapScanLocalState*)_local_state)->_output_column_ids;
372
983k
    _tablet_reader_params.ann_topn_runtime = _ann_topn_runtime;
373
983k
    for (const auto& ele : ((OlapScanLocalState*)_local_state)->_cast_types_for_variants) {
374
1.45k
        _tablet_reader_params.target_cast_type_for_variants[ele.first] = ele.second;
375
1.45k
    };
376
983k
    auto& tablet_schema = _tablet_reader_params.tablet_schema;
377
6.48M
    for (auto& predicates : slot_to_predicates) {
378
6.48M
        const int sid = predicates.first;
379
6.48M
        DCHECK(slot_id_to_slot_desc.contains(sid));
380
6.48M
        int32_t index =
381
6.48M
                tablet_schema->field_index(slot_id_to_slot_desc.find(sid)->second->col_name());
382
6.48M
        if (index < 0) {
383
0
            throw Exception(
384
0
                    Status::InternalError("Column {} not found in tablet schema",
385
0
                                          slot_id_to_slot_desc.find(sid)->second->col_name()));
386
0
        }
387
6.48M
        for (auto& predicate : predicates.second) {
388
719k
            _tablet_reader_params.predicates.push_back(predicate->clone(index));
389
719k
        }
390
6.48M
    }
391
392
983k
    std::copy(function_filters.cbegin(), function_filters.cend(),
393
983k
              std::inserter(_tablet_reader_params.function_filters,
394
983k
                            _tablet_reader_params.function_filters.begin()));
395
396
    // Merge the columns in delete predicate that not in latest schema in to current tablet schema
397
983k
    for (auto& del_pred : _tablet_reader_params.delete_predicates) {
398
6.47k
        tablet_schema->merge_dropped_columns(*del_pred->tablet_schema());
399
6.47k
    }
400
401
    // Push key ranges to the tablet reader.
402
    // Skip the "full scan" placeholder (has_lower_bound == false) — when no key
403
    // predicates exist, start_key/end_key remain empty and the reader does a full scan.
404
1.65M
    for (auto* key_range : key_ranges) {
405
1.65M
        if (!key_range->has_lower_bound) {
406
193k
            continue;
407
193k
        }
408
409
1.46M
        _tablet_reader_params.start_key_include = key_range->begin_include;
410
1.46M
        _tablet_reader_params.end_key_include = key_range->end_include;
411
412
1.46M
        _tablet_reader_params.start_key.push_back(key_range->begin_scan_range);
413
1.46M
        _tablet_reader_params.end_key.push_back(key_range->end_scan_range);
414
1.46M
    }
415
416
983k
    _tablet_reader_params.profile = _local_state->custom_profile();
417
983k
    _tablet_reader_params.runtime_state = _state;
418
419
983k
    _tablet_reader_params.origin_return_columns = &_return_columns;
420
983k
    _tablet_reader_params.tablet_columns_convert_to_null_set = &_tablet_columns_convert_to_null_set;
421
422
983k
    if (_tablet_reader_params.direct_mode) {
423
972k
        _tablet_reader_params.return_columns = _return_columns;
424
972k
    } else {
425
        // we need to fetch all key columns to do the right aggregation on storage engine side.
426
37.3k
        for (size_t i = 0; i < tablet_schema->num_key_columns(); ++i) {
427
27.0k
            _tablet_reader_params.return_columns.push_back(i);
428
27.0k
        }
429
49.5k
        for (auto index : _return_columns) {
430
49.5k
            if (tablet_schema->column(index).is_key()) {
431
18.3k
                continue;
432
18.3k
            }
433
31.1k
            _tablet_reader_params.return_columns.push_back(index);
434
31.1k
        }
435
        // expand the sequence column
436
10.2k
        if (tablet_schema->has_sequence_col() || tablet_schema->has_seq_map()) {
437
39
            bool has_replace_col = false;
438
90
            for (auto col : _return_columns) {
439
90
                if (tablet_schema->column(col).aggregation() ==
440
90
                    FieldAggregationMethod::OLAP_FIELD_AGGREGATION_REPLACE) {
441
39
                    has_replace_col = true;
442
39
                    break;
443
39
                }
444
90
            }
445
39
            if (auto sequence_col_idx = tablet_schema->sequence_col_idx();
446
39
                has_replace_col && tablet_schema->has_sequence_col() &&
447
39
                std::find(_return_columns.begin(), _return_columns.end(), sequence_col_idx) ==
448
27
                        _return_columns.end()) {
449
16
                _tablet_reader_params.return_columns.push_back(sequence_col_idx);
450
16
            }
451
40
            if (has_replace_col) {
452
40
                const auto& val_to_seq = tablet_schema->value_col_idx_to_seq_col_idx();
453
40
                std::set<uint32_t> return_seq_columns;
454
455
240
                for (auto col : _tablet_reader_params.return_columns) {
456
                    // we need to add the necessary sequence column in _return_columns, and
457
                    // Avoid adding the same seq column twice
458
240
                    const auto val_iter = val_to_seq.find(col);
459
240
                    if (val_iter != val_to_seq.end()) {
460
42
                        auto seq = val_iter->second;
461
42
                        if (std::find(_tablet_reader_params.return_columns.begin(),
462
42
                                      _tablet_reader_params.return_columns.end(),
463
42
                                      seq) == _tablet_reader_params.return_columns.end()) {
464
4
                            return_seq_columns.insert(seq);
465
4
                        }
466
42
                    }
467
240
                }
468
40
                _tablet_reader_params.return_columns.insert(
469
40
                        std::end(_tablet_reader_params.return_columns),
470
40
                        std::begin(return_seq_columns), std::end(return_seq_columns));
471
40
            }
472
39
        }
473
10.2k
    }
474
475
983k
    _tablet_reader_params.use_page_cache = _state->enable_page_cache();
476
477
983k
    DBUG_EXECUTE_IF("NewOlapScanner::_init_tablet_reader_params.block", DBUG_BLOCK);
478
479
983k
    if (!_state->skip_storage_engine_merge()) {
480
982k
        auto* olap_scan_local_state = (OlapScanLocalState*)_local_state;
481
982k
        TOlapScanNode& olap_scan_node = olap_scan_local_state->olap_scan_node();
482
483
        // Set MOR value predicate pushdown flag
484
982k
        if (olap_scan_node.__isset.enable_mor_value_predicate_pushdown &&
485
982k
            olap_scan_node.enable_mor_value_predicate_pushdown) {
486
25
            _tablet_reader_params.enable_mor_value_predicate_pushdown = true;
487
25
        }
488
489
        // order by table keys optimization for topn
490
        // will only read head/tail of data file since it's already sorted by keys
491
982k
        if (olap_scan_node.__isset.sort_info && !olap_scan_node.sort_info.is_asc_order.empty()) {
492
1.77k
            _limit = _local_state->limit_per_scanner();
493
1.77k
            _tablet_reader_params.read_orderby_key = true;
494
1.77k
            if (!olap_scan_node.sort_info.is_asc_order[0]) {
495
368
                _tablet_reader_params.read_orderby_key_reverse = true;
496
368
            }
497
1.77k
            _tablet_reader_params.read_orderby_key_num_prefix_columns =
498
1.77k
                    olap_scan_node.sort_info.is_asc_order.size();
499
1.77k
            _tablet_reader_params.read_orderby_key_limit = _limit;
500
501
1.77k
            if (_tablet_reader_params.read_orderby_key_limit > 0 &&
502
1.77k
                olap_scan_local_state->_storage_no_merge()) {
503
1.66k
                _tablet_reader_params.filter_block_conjuncts = _conjuncts;
504
1.66k
                _conjuncts.clear();
505
1.66k
            }
506
1.77k
        }
507
508
        // set push down topn filter
509
982k
        _tablet_reader_params.topn_filter_source_node_ids =
510
982k
                olap_scan_local_state->get_topn_filter_source_node_ids(_state, true);
511
982k
        if (!_tablet_reader_params.topn_filter_source_node_ids.empty()) {
512
4.37k
            _tablet_reader_params.topn_filter_target_node_id =
513
4.37k
                    olap_scan_local_state->parent()->node_id();
514
4.37k
        }
515
982k
    }
516
517
    // If this is a Two-Phase read query, and we need to delay the release of Rowset
518
    // by rowset->update_delayed_expired_timestamp().This could expand the lifespan of Rowset
519
983k
    if (tablet_schema->field_index(BeConsts::ROWID_COL) >= 0) {
520
9
        constexpr static int delayed_s = 60;
521
13
        for (auto rs_reader : _tablet_reader_params.rs_splits) {
522
13
            uint64_t delayed_expired_timestamp =
523
13
                    UnixSeconds() + _tablet_reader_params.runtime_state->execution_timeout() +
524
13
                    delayed_s;
525
13
            rs_reader.rs_reader->rowset()->update_delayed_expired_timestamp(
526
13
                    delayed_expired_timestamp);
527
13
            ExecEnv::GetInstance()->storage_engine().add_quering_rowset(
528
13
                    rs_reader.rs_reader->rowset());
529
13
        }
530
9
    }
531
532
983k
    if (tablet_schema->has_global_row_id()) {
533
7.15k
        auto& id_file_map = _state->get_id_file_map();
534
14.7k
        for (auto rs_reader : _tablet_reader_params.rs_splits) {
535
14.7k
            id_file_map->add_temp_rowset(rs_reader.rs_reader->rowset());
536
14.7k
        }
537
7.15k
    }
538
539
983k
    return Status::OK();
540
983k
}
541
542
982k
Status OlapScanner::_init_variant_columns() {
543
982k
    auto& tablet_schema = _tablet_reader_params.tablet_schema;
544
982k
    if (tablet_schema->num_variant_columns() == 0) {
545
977k
        return Status::OK();
546
977k
    }
547
    // Parent column has path info to distinction from each other
548
15.2k
    for (auto* slot : _output_tuple_desc->slots()) {
549
15.2k
        if (slot->type()->get_primitive_type() == PrimitiveType::TYPE_VARIANT) {
550
            // Such columns are not exist in frontend schema info, so we need to
551
            // add them into tablet_schema for later column indexing.
552
7.59k
            TabletColumn subcol = TabletColumn::create_materialized_variant_column(
553
7.59k
                    tablet_schema->column_by_uid(slot->col_unique_id()).name_lower_case(),
554
7.59k
                    slot->column_paths(), slot->col_unique_id(),
555
7.59k
                    assert_cast<const DataTypeVariant&>(*remove_nullable(slot->type()))
556
7.59k
                            .variant_max_subcolumns_count());
557
7.59k
            if (tablet_schema->field_index(*subcol.path_info_ptr()) < 0) {
558
5.79k
                tablet_schema->append_column(subcol, TabletSchema::ColumnType::VARIANT);
559
5.79k
            }
560
7.59k
        }
561
15.2k
    }
562
5.47k
    variant_util::inherit_column_attributes(tablet_schema);
563
5.47k
    return Status::OK();
564
982k
}
565
566
982k
Status OlapScanner::_init_return_columns() {
567
8.49M
    for (auto* slot : _output_tuple_desc->slots()) {
568
        // variant column using path to index a column
569
8.49M
        int32_t index = 0;
570
8.49M
        auto& tablet_schema = _tablet_reader_params.tablet_schema;
571
8.49M
        if (slot->type()->get_primitive_type() == PrimitiveType::TYPE_VARIANT) {
572
7.61k
            index = tablet_schema->field_index(PathInData(
573
7.61k
                    tablet_schema->column_by_uid(slot->col_unique_id()).name_lower_case(),
574
7.61k
                    slot->column_paths()));
575
8.49M
        } else {
576
8.50M
            index = slot->col_unique_id() >= 0 ? tablet_schema->field_index(slot->col_unique_id())
577
18.4E
                                               : tablet_schema->field_index(slot->col_name());
578
8.49M
        }
579
580
8.49M
        if (index < 0) {
581
0
            return Status::InternalError(
582
0
                    "field name is invalid. field={}, field_name_to_index={}, col_unique_id={}",
583
0
                    slot->col_name(), tablet_schema->get_all_field_names(), slot->col_unique_id());
584
0
        }
585
586
8.49M
        if (slot->get_virtual_column_expr()) {
587
365
            ColumnId virtual_column_cid = index;
588
365
            _virtual_column_exprs[virtual_column_cid] = _slot_id_to_virtual_column_expr[slot->id()];
589
365
            size_t idx_in_block = _slot_id_to_index_in_block[slot->id()];
590
365
            _vir_cid_to_idx_in_block[virtual_column_cid] = idx_in_block;
591
365
            _vir_col_idx_to_type[idx_in_block] = _slot_id_to_col_type[slot->id()];
592
593
365
            VLOG_DEBUG << fmt::format(
594
1
                    "Virtual column, slot id: {}, cid {}, column index: {}, type: {}", slot->id(),
595
1
                    virtual_column_cid, _vir_cid_to_idx_in_block[virtual_column_cid],
596
1
                    _vir_col_idx_to_type[idx_in_block]->get_name());
597
365
        }
598
599
8.49M
        const auto& column = tablet_schema->column(index);
600
8.49M
        int32_t unique_id = column.unique_id() > 0 ? column.unique_id() : column.parent_unique_id();
601
8.49M
        if (!slot->all_access_paths().empty()) {
602
56.1k
            _tablet_reader_params.all_access_paths.insert({unique_id, slot->all_access_paths()});
603
56.1k
        }
604
605
8.49M
        if (!slot->predicate_access_paths().empty()) {
606
4.86k
            _tablet_reader_params.predicate_access_paths.insert(
607
4.86k
                    {unique_id, slot->predicate_access_paths()});
608
4.86k
        }
609
610
8.49M
        if ((slot->type()->get_primitive_type() == PrimitiveType::TYPE_STRUCT ||
611
8.51M
             slot->type()->get_primitive_type() == PrimitiveType::TYPE_MAP ||
612
8.50M
             slot->type()->get_primitive_type() == PrimitiveType::TYPE_ARRAY) &&
613
8.49M
            !slot->all_access_paths().empty()) {
614
50.6k
            tablet_schema->add_pruned_columns_data_type(column.unique_id(), slot->type());
615
50.6k
        }
616
617
8.49M
        _return_columns.push_back(index);
618
8.49M
        if (slot->is_nullable() && !tablet_schema->column(index).is_nullable()) {
619
0
            _tablet_columns_convert_to_null_set.emplace(index);
620
8.49M
        } else if (!slot->is_nullable() && tablet_schema->column(index).is_nullable()) {
621
0
            return Status::Error<ErrorCode::INVALID_SCHEMA>(
622
0
                    "slot(id: {}, name: {})'s nullable does not match "
623
0
                    "column(tablet id: {}, index: {}, name: {}) ",
624
0
                    slot->id(), slot->col_name(), tablet_schema->table_id(), index,
625
0
                    tablet_schema->column(index).name());
626
0
        }
627
8.49M
    }
628
629
982k
    if (_return_columns.empty()) {
630
0
        return Status::InternalError("failed to build storage scanner, no materialized slot!");
631
0
    }
632
633
982k
    return Status::OK();
634
982k
}
635
636
988k
doris::TabletStorageType OlapScanner::get_storage_type() {
637
988k
    if (config::is_cloud_mode()) {
638
        // we don't have cold storage in cloud mode, all storage is treated as local
639
978k
        return doris::TabletStorageType::STORAGE_TYPE_LOCAL;
640
978k
    }
641
9.75k
    int local_reader = 0;
642
56.5k
    for (const auto& reader : _tablet_reader_params.rs_splits) {
643
56.5k
        local_reader += reader.rs_reader->rowset()->is_local();
644
56.5k
    }
645
9.75k
    int total_reader = _tablet_reader_params.rs_splits.size();
646
647
9.75k
    if (local_reader == total_reader) {
648
9.75k
        return doris::TabletStorageType::STORAGE_TYPE_LOCAL;
649
9.75k
    } else if (local_reader == 0) {
650
0
        return doris::TabletStorageType::STORAGE_TYPE_REMOTE;
651
0
    }
652
1
    return doris::TabletStorageType::STORAGE_TYPE_REMOTE_AND_LOCAL;
653
9.75k
}
654
655
1.24M
Status OlapScanner::_get_block_impl(RuntimeState* state, Block* block, bool* eof) {
656
    // Read one block from block reader
657
    // ATTN: Here we need to let the _get_block_impl method guarantee the semantics of the interface,
658
    // that is, eof can be set to true only when the returned block is empty.
659
1.24M
    RETURN_IF_ERROR(_tablet_reader->next_block_with_aggregation(block, eof));
660
1.24M
    if (block->rows() > 0) {
661
263k
        _tablet_reader_params.tablet->read_block_count.fetch_add(1, std::memory_order_relaxed);
662
263k
        *eof = false;
663
263k
    }
664
1.24M
    return Status::OK();
665
1.24M
}
666
667
988k
Status OlapScanner::close(RuntimeState* state) {
668
988k
    if (!_try_close()) {
669
134
        return Status::OK();
670
134
    }
671
988k
    RETURN_IF_ERROR(Scanner::close(state));
672
988k
    return Status::OK();
673
988k
}
674
675
981k
void OlapScanner::update_realtime_counters() {
676
981k
    if (!_has_prepared) {
677
        // Counter update need prepare successfully, or it maybe core. For example, olap scanner
678
        // will open tablet reader during prepare, if not prepare successfully, tablet reader == nullptr.
679
0
        return;
680
0
    }
681
981k
    OlapScanLocalState* local_state = static_cast<OlapScanLocalState*>(_local_state);
682
981k
    const OlapReaderStatistics& stats = _tablet_reader->stats();
683
981k
    COUNTER_UPDATE(local_state->_read_compressed_counter, stats.compressed_bytes_read);
684
981k
    COUNTER_UPDATE(local_state->_read_uncompressed_counter, stats.uncompressed_bytes_read);
685
981k
    COUNTER_UPDATE(local_state->_scan_bytes, stats.uncompressed_bytes_read);
686
981k
    COUNTER_UPDATE(local_state->_scan_rows, stats.raw_rows_read);
687
688
    // Make sure the scan bytes and scan rows counter in audit log is the same as the counter in
689
    // doris metrics.
690
    // ScanBytes is the uncompressed bytes read from local + remote
691
    // bytes_read_from_local is the compressed bytes read from local
692
    // bytes_read_from_remote is the compressed bytes read from remote
693
    // scan bytes > bytes_read_from_local + bytes_read_from_remote
694
981k
    _state->get_query_ctx()->resource_ctx()->io_context()->update_scan_rows(stats.raw_rows_read);
695
981k
    _state->get_query_ctx()->resource_ctx()->io_context()->update_scan_bytes(
696
981k
            stats.uncompressed_bytes_read);
697
698
    // In case of no cache, we still need to update the IO stats. uncompressed bytes read == local + remote
699
981k
    if (stats.file_cache_stats.bytes_read_from_local == 0 &&
700
981k
        stats.file_cache_stats.bytes_read_from_remote == 0) {
701
893k
        _state->get_query_ctx()->resource_ctx()->io_context()->update_scan_bytes_from_local_storage(
702
893k
                stats.compressed_bytes_read);
703
893k
        DorisMetrics::instance()->query_scan_bytes_from_local->increment(
704
893k
                stats.compressed_bytes_read);
705
893k
    } else {
706
88.0k
        _state->get_query_ctx()->resource_ctx()->io_context()->update_scan_bytes_from_local_storage(
707
88.0k
                stats.file_cache_stats.bytes_read_from_local - _bytes_read_from_local);
708
88.0k
        _state->get_query_ctx()
709
88.0k
                ->resource_ctx()
710
88.0k
                ->io_context()
711
88.0k
                ->update_scan_bytes_from_remote_storage(
712
88.0k
                        stats.file_cache_stats.bytes_read_from_remote - _bytes_read_from_remote);
713
714
88.0k
        DorisMetrics::instance()->query_scan_bytes_from_local->increment(
715
88.0k
                stats.file_cache_stats.bytes_read_from_local - _bytes_read_from_local);
716
88.0k
        DorisMetrics::instance()->query_scan_bytes_from_remote->increment(
717
88.0k
                stats.file_cache_stats.bytes_read_from_remote - _bytes_read_from_remote);
718
88.0k
    }
719
720
981k
    _tablet_reader->mutable_stats()->compressed_bytes_read = 0;
721
981k
    _tablet_reader->mutable_stats()->uncompressed_bytes_read = 0;
722
981k
    _tablet_reader->mutable_stats()->raw_rows_read = 0;
723
724
981k
    _bytes_read_from_local = _tablet_reader->stats().file_cache_stats.bytes_read_from_local;
725
981k
    _bytes_read_from_remote = _tablet_reader->stats().file_cache_stats.bytes_read_from_remote;
726
981k
}
727
728
980k
void OlapScanner::_collect_profile_before_close() {
729
    //  Please don't directly enable the profile here, we need to set QueryStatistics using the counter inside.
730
980k
    if (_has_updated_counter) {
731
0
        return;
732
0
    }
733
980k
    _has_updated_counter = true;
734
980k
    _tablet_reader->update_profile(_profile);
735
736
980k
    Scanner::_collect_profile_before_close();
737
738
    // Update counters for OlapScanner
739
    // Update counters from tablet reader's stats
740
980k
    auto& stats = _tablet_reader->stats();
741
980k
    auto* local_state = (OlapScanLocalState*)_local_state;
742
980k
    COUNTER_UPDATE(local_state->_io_timer, stats.io_ns);
743
980k
    COUNTER_UPDATE(local_state->_read_compressed_counter, stats.compressed_bytes_read);
744
980k
    COUNTER_UPDATE(local_state->_scan_bytes, stats.uncompressed_bytes_read);
745
980k
    COUNTER_UPDATE(local_state->_decompressor_timer, stats.decompress_ns);
746
980k
    COUNTER_UPDATE(local_state->_read_uncompressed_counter, stats.uncompressed_bytes_read);
747
980k
    COUNTER_UPDATE(local_state->_block_load_timer, stats.block_load_ns);
748
980k
    COUNTER_UPDATE(local_state->_block_load_counter, stats.blocks_load);
749
980k
    COUNTER_UPDATE(local_state->_block_fetch_timer, stats.block_fetch_ns);
750
980k
    COUNTER_UPDATE(local_state->_delete_bitmap_get_agg_timer, stats.delete_bitmap_get_agg_ns);
751
980k
    COUNTER_UPDATE(local_state->_scan_rows, stats.raw_rows_read);
752
980k
    COUNTER_UPDATE(local_state->_vec_cond_timer, stats.vec_cond_ns);
753
980k
    COUNTER_UPDATE(local_state->_short_cond_timer, stats.short_cond_ns);
754
980k
    COUNTER_UPDATE(local_state->_expr_filter_timer, stats.expr_filter_ns);
755
980k
    COUNTER_UPDATE(local_state->_block_init_timer, stats.block_init_ns);
756
980k
    COUNTER_UPDATE(local_state->_block_init_seek_timer, stats.block_init_seek_ns);
757
980k
    COUNTER_UPDATE(local_state->_block_init_seek_counter, stats.block_init_seek_num);
758
980k
    COUNTER_UPDATE(local_state->_segment_generate_row_range_by_keys_timer,
759
980k
                   stats.generate_row_ranges_by_keys_ns);
760
980k
    COUNTER_UPDATE(local_state->_segment_generate_row_range_by_column_conditions_timer,
761
980k
                   stats.generate_row_ranges_by_column_conditions_ns);
762
980k
    COUNTER_UPDATE(local_state->_segment_generate_row_range_by_bf_timer,
763
980k
                   stats.generate_row_ranges_by_bf_ns);
764
980k
    COUNTER_UPDATE(local_state->_collect_iterator_merge_next_timer,
765
980k
                   stats.collect_iterator_merge_next_timer);
766
980k
    COUNTER_UPDATE(local_state->_segment_generate_row_range_by_zonemap_timer,
767
980k
                   stats.generate_row_ranges_by_zonemap_ns);
768
980k
    COUNTER_UPDATE(local_state->_segment_generate_row_range_by_dict_timer,
769
980k
                   stats.generate_row_ranges_by_dict_ns);
770
980k
    COUNTER_UPDATE(local_state->_predicate_column_read_timer, stats.predicate_column_read_ns);
771
980k
    COUNTER_UPDATE(local_state->_non_predicate_column_read_timer, stats.non_predicate_read_ns);
772
980k
    COUNTER_UPDATE(local_state->_predicate_column_read_seek_timer,
773
980k
                   stats.predicate_column_read_seek_ns);
774
980k
    COUNTER_UPDATE(local_state->_predicate_column_read_seek_counter,
775
980k
                   stats.predicate_column_read_seek_num);
776
980k
    COUNTER_UPDATE(local_state->_lazy_read_timer, stats.lazy_read_ns);
777
980k
    COUNTER_UPDATE(local_state->_lazy_read_seek_timer, stats.block_lazy_read_seek_ns);
778
980k
    COUNTER_UPDATE(local_state->_lazy_read_seek_counter, stats.block_lazy_read_seek_num);
779
980k
    COUNTER_UPDATE(local_state->_output_col_timer, stats.output_col_ns);
780
980k
    COUNTER_UPDATE(local_state->_rows_vec_cond_filtered_counter, stats.rows_vec_cond_filtered);
781
980k
    COUNTER_UPDATE(local_state->_rows_short_circuit_cond_filtered_counter,
782
980k
                   stats.rows_short_circuit_cond_filtered);
783
980k
    COUNTER_UPDATE(local_state->_rows_expr_cond_filtered_counter, stats.rows_expr_cond_filtered);
784
980k
    COUNTER_UPDATE(local_state->_rows_vec_cond_input_counter, stats.vec_cond_input_rows);
785
980k
    COUNTER_UPDATE(local_state->_rows_short_circuit_cond_input_counter,
786
980k
                   stats.short_circuit_cond_input_rows);
787
980k
    COUNTER_UPDATE(local_state->_rows_expr_cond_input_counter, stats.expr_cond_input_rows);
788
980k
    COUNTER_UPDATE(local_state->_stats_filtered_counter, stats.rows_stats_filtered);
789
980k
    COUNTER_UPDATE(local_state->_stats_rp_filtered_counter, stats.rows_stats_rp_filtered);
790
980k
    COUNTER_UPDATE(local_state->_dict_filtered_counter, stats.segment_dict_filtered);
791
980k
    COUNTER_UPDATE(local_state->_bf_filtered_counter, stats.rows_bf_filtered);
792
980k
    COUNTER_UPDATE(local_state->_del_filtered_counter, stats.rows_del_filtered);
793
980k
    COUNTER_UPDATE(local_state->_del_filtered_counter, stats.rows_del_by_bitmap);
794
980k
    COUNTER_UPDATE(local_state->_del_filtered_counter, stats.rows_vec_del_cond_filtered);
795
980k
    COUNTER_UPDATE(local_state->_conditions_filtered_counter, stats.rows_conditions_filtered);
796
980k
    COUNTER_UPDATE(local_state->_key_range_filtered_counter, stats.rows_key_range_filtered);
797
980k
    COUNTER_UPDATE(local_state->_total_pages_num_counter, stats.total_pages_num);
798
980k
    COUNTER_UPDATE(local_state->_cached_pages_num_counter, stats.cached_pages_num);
799
980k
    COUNTER_UPDATE(local_state->_inverted_index_filter_counter, stats.rows_inverted_index_filtered);
800
980k
    COUNTER_UPDATE(local_state->_inverted_index_filter_timer, stats.inverted_index_filter_timer);
801
980k
    COUNTER_UPDATE(local_state->_inverted_index_query_cache_hit_counter,
802
980k
                   stats.inverted_index_query_cache_hit);
803
980k
    COUNTER_UPDATE(local_state->_inverted_index_query_cache_miss_counter,
804
980k
                   stats.inverted_index_query_cache_miss);
805
980k
    COUNTER_UPDATE(local_state->_inverted_index_query_timer, stats.inverted_index_query_timer);
806
980k
    COUNTER_UPDATE(local_state->_inverted_index_query_null_bitmap_timer,
807
980k
                   stats.inverted_index_query_null_bitmap_timer);
808
980k
    COUNTER_UPDATE(local_state->_inverted_index_query_bitmap_copy_timer,
809
980k
                   stats.inverted_index_query_bitmap_copy_timer);
810
980k
    COUNTER_UPDATE(local_state->_inverted_index_searcher_open_timer,
811
980k
                   stats.inverted_index_searcher_open_timer);
812
980k
    COUNTER_UPDATE(local_state->_inverted_index_searcher_search_timer,
813
980k
                   stats.inverted_index_searcher_search_timer);
814
980k
    COUNTER_UPDATE(local_state->_inverted_index_searcher_search_init_timer,
815
980k
                   stats.inverted_index_searcher_search_init_timer);
816
980k
    COUNTER_UPDATE(local_state->_inverted_index_searcher_search_exec_timer,
817
980k
                   stats.inverted_index_searcher_search_exec_timer);
818
980k
    COUNTER_UPDATE(local_state->_inverted_index_searcher_cache_hit_counter,
819
980k
                   stats.inverted_index_searcher_cache_hit);
820
980k
    COUNTER_UPDATE(local_state->_inverted_index_searcher_cache_miss_counter,
821
980k
                   stats.inverted_index_searcher_cache_miss);
822
980k
    COUNTER_UPDATE(local_state->_inverted_index_downgrade_count_counter,
823
980k
                   stats.inverted_index_downgrade_count);
824
980k
    COUNTER_UPDATE(local_state->_inverted_index_analyzer_timer,
825
980k
                   stats.inverted_index_analyzer_timer);
826
980k
    COUNTER_UPDATE(local_state->_inverted_index_lookup_timer, stats.inverted_index_lookup_timer);
827
980k
    COUNTER_UPDATE(local_state->_variant_scan_sparse_column_timer,
828
980k
                   stats.variant_scan_sparse_column_timer_ns);
829
980k
    COUNTER_UPDATE(local_state->_variant_scan_sparse_column_bytes,
830
980k
                   stats.variant_scan_sparse_column_bytes);
831
980k
    COUNTER_UPDATE(local_state->_variant_fill_path_from_sparse_column_timer,
832
980k
                   stats.variant_fill_path_from_sparse_column_timer_ns);
833
980k
    COUNTER_UPDATE(local_state->_variant_subtree_default_iter_count,
834
980k
                   stats.variant_subtree_default_iter_count);
835
980k
    COUNTER_UPDATE(local_state->_variant_subtree_leaf_iter_count,
836
980k
                   stats.variant_subtree_leaf_iter_count);
837
980k
    COUNTER_UPDATE(local_state->_variant_subtree_hierarchical_iter_count,
838
980k
                   stats.variant_subtree_hierarchical_iter_count);
839
980k
    COUNTER_UPDATE(local_state->_variant_subtree_sparse_iter_count,
840
980k
                   stats.variant_subtree_sparse_iter_count);
841
980k
    COUNTER_UPDATE(local_state->_variant_doc_value_column_iter_count,
842
980k
                   stats.variant_doc_value_column_iter_count);
843
844
980k
    InvertedIndexProfileReporter inverted_index_profile;
845
980k
    inverted_index_profile.update(local_state->_index_filter_profile.get(),
846
980k
                                  &stats.inverted_index_stats);
847
848
    // only cloud deploy mode will use file cache.
849
980k
    if (config::is_cloud_mode() && config::enable_file_cache) {
850
972k
        io::FileCacheProfileReporter cache_profile(local_state->_segment_profile.get());
851
972k
        cache_profile.update(&stats.file_cache_stats);
852
972k
        _state->get_query_ctx()->resource_ctx()->io_context()->update_bytes_write_into_cache(
853
972k
                stats.file_cache_stats.bytes_write_into_cache);
854
972k
    }
855
980k
    COUNTER_UPDATE(local_state->_output_index_result_column_timer,
856
980k
                   stats.output_index_result_column_timer);
857
980k
    COUNTER_UPDATE(local_state->_filtered_segment_counter, stats.filtered_segment_number);
858
980k
    COUNTER_UPDATE(local_state->_total_segment_counter, stats.total_segment_number);
859
980k
    COUNTER_UPDATE(local_state->_condition_cache_hit_counter, stats.condition_cache_hit_seg_nums);
860
980k
    COUNTER_UPDATE(local_state->_condition_cache_filtered_rows_counter,
861
980k
                   stats.condition_cache_filtered_rows);
862
863
980k
    COUNTER_UPDATE(local_state->_tablet_reader_init_timer, stats.tablet_reader_init_timer_ns);
864
980k
    COUNTER_UPDATE(local_state->_tablet_reader_capture_rs_readers_timer,
865
980k
                   stats.tablet_reader_capture_rs_readers_timer_ns);
866
980k
    COUNTER_UPDATE(local_state->_tablet_reader_init_return_columns_timer,
867
980k
                   stats.tablet_reader_init_return_columns_timer_ns);
868
980k
    COUNTER_UPDATE(local_state->_tablet_reader_init_keys_param_timer,
869
980k
                   stats.tablet_reader_init_keys_param_timer_ns);
870
980k
    COUNTER_UPDATE(local_state->_tablet_reader_init_orderby_keys_param_timer,
871
980k
                   stats.tablet_reader_init_orderby_keys_param_timer_ns);
872
980k
    COUNTER_UPDATE(local_state->_tablet_reader_init_conditions_param_timer,
873
980k
                   stats.tablet_reader_init_conditions_param_timer_ns);
874
980k
    COUNTER_UPDATE(local_state->_tablet_reader_init_delete_condition_param_timer,
875
980k
                   stats.tablet_reader_init_delete_condition_param_timer_ns);
876
980k
    COUNTER_UPDATE(local_state->_block_reader_vcollect_iter_init_timer,
877
980k
                   stats.block_reader_vcollect_iter_init_timer_ns);
878
980k
    COUNTER_UPDATE(local_state->_block_reader_rs_readers_init_timer,
879
980k
                   stats.block_reader_rs_readers_init_timer_ns);
880
980k
    COUNTER_UPDATE(local_state->_block_reader_build_heap_init_timer,
881
980k
                   stats.block_reader_build_heap_init_timer_ns);
882
883
980k
    COUNTER_UPDATE(local_state->_rowset_reader_get_segment_iterators_timer,
884
980k
                   stats.rowset_reader_get_segment_iterators_timer_ns);
885
980k
    COUNTER_UPDATE(local_state->_rowset_reader_create_iterators_timer,
886
980k
                   stats.rowset_reader_create_iterators_timer_ns);
887
980k
    COUNTER_UPDATE(local_state->_rowset_reader_init_iterators_timer,
888
980k
                   stats.rowset_reader_init_iterators_timer_ns);
889
980k
    COUNTER_UPDATE(local_state->_rowset_reader_load_segments_timer,
890
980k
                   stats.rowset_reader_load_segments_timer_ns);
891
892
980k
    COUNTER_UPDATE(local_state->_segment_iterator_init_timer, stats.segment_iterator_init_timer_ns);
893
980k
    COUNTER_UPDATE(local_state->_segment_iterator_init_return_column_iterators_timer,
894
980k
                   stats.segment_iterator_init_return_column_iterators_timer_ns);
895
980k
    COUNTER_UPDATE(local_state->_segment_iterator_init_index_iterators_timer,
896
980k
                   stats.segment_iterator_init_index_iterators_timer_ns);
897
898
980k
    COUNTER_UPDATE(local_state->_segment_create_column_readers_timer,
899
980k
                   stats.segment_create_column_readers_timer_ns);
900
980k
    COUNTER_UPDATE(local_state->_segment_load_index_timer, stats.segment_load_index_timer_ns);
901
902
    // Update metrics
903
980k
    DorisMetrics::instance()->query_scan_bytes->increment(
904
980k
            local_state->_read_uncompressed_counter->value());
905
980k
    DorisMetrics::instance()->query_scan_rows->increment(local_state->_scan_rows->value());
906
980k
    auto& tablet = _tablet_reader_params.tablet;
907
980k
    tablet->query_scan_bytes->increment(local_state->_read_uncompressed_counter->value());
908
980k
    tablet->query_scan_rows->increment(local_state->_scan_rows->value());
909
980k
    tablet->query_scan_count->increment(1);
910
911
980k
    COUNTER_UPDATE(local_state->_ann_range_search_filter_counter,
912
980k
                   stats.rows_ann_index_range_filtered);
913
980k
    COUNTER_UPDATE(local_state->_ann_topn_filter_counter, stats.rows_ann_index_topn_filtered);
914
980k
    COUNTER_UPDATE(local_state->_ann_index_load_costs, stats.ann_index_load_ns);
915
980k
    COUNTER_UPDATE(local_state->_ann_range_search_costs, stats.ann_index_range_search_ns);
916
980k
    COUNTER_UPDATE(local_state->_ann_range_search_cnt, stats.ann_index_range_search_cnt);
917
980k
    COUNTER_UPDATE(local_state->_ann_range_engine_search_costs, stats.ann_range_engine_search_ns);
918
    // Engine prepare before search
919
980k
    COUNTER_UPDATE(local_state->_ann_range_pre_process_costs, stats.ann_range_pre_process_ns);
920
    // Post process parent: Doris result process + engine convert
921
980k
    COUNTER_UPDATE(local_state->_ann_range_post_process_costs,
922
980k
                   stats.ann_range_result_convert_ns + stats.ann_range_engine_convert_ns);
923
    // Engine convert (child under post-process)
924
980k
    COUNTER_UPDATE(local_state->_ann_range_engine_convert_costs, stats.ann_range_engine_convert_ns);
925
    // Doris-side result convert (child under post-process)
926
980k
    COUNTER_UPDATE(local_state->_ann_range_result_convert_costs, stats.ann_range_result_convert_ns);
927
928
980k
    COUNTER_UPDATE(local_state->_ann_topn_search_costs, stats.ann_topn_search_ns);
929
980k
    COUNTER_UPDATE(local_state->_ann_topn_search_cnt, stats.ann_index_topn_search_cnt);
930
931
    // Detailed ANN timers
932
    // ANN TopN timers with hierarchy
933
    // Engine search time (FAISS)
934
980k
    COUNTER_UPDATE(local_state->_ann_topn_engine_search_costs,
935
980k
                   stats.ann_index_topn_engine_search_ns);
936
    // Engine prepare time (allocations/buffer setup before search)
937
980k
    COUNTER_UPDATE(local_state->_ann_topn_pre_process_costs,
938
980k
                   stats.ann_index_topn_engine_prepare_ns);
939
    // Post process parent includes Doris result processing + engine convert
940
980k
    COUNTER_UPDATE(local_state->_ann_topn_post_process_costs,
941
980k
                   stats.ann_index_topn_result_process_ns + stats.ann_index_topn_engine_convert_ns);
942
    // Engine-side conversion time inside FAISS wrappers (child under post-process)
943
980k
    COUNTER_UPDATE(local_state->_ann_topn_engine_convert_costs,
944
980k
                   stats.ann_index_topn_engine_convert_ns);
945
946
    // Doris-side result convert costs (show separately as another child counter); use pure process time
947
980k
    COUNTER_UPDATE(local_state->_ann_topn_result_convert_costs,
948
980k
                   stats.ann_index_topn_result_process_ns);
949
950
980k
    COUNTER_UPDATE(local_state->_ann_fallback_brute_force_cnt, stats.ann_fall_back_brute_force_cnt);
951
952
    // Overhead counter removed; precise instrumentation is reported via engine_prepare above.
953
980k
}
954
955
#include "common/compile_check_avoid_end.h"
956
} // namespace doris