Coverage Report

Created: 2026-08-28 02:04

next uncovered line (L), next uncovered region (R), next uncovered branch (B)
/root/doris/be/src/exec/rowid_fetcher.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/rowid_fetcher.h"
19
20
#include <brpc/callback.h>
21
#include <butil/endpoint.h>
22
#include <fmt/format.h>
23
#include <gen_cpp/data.pb.h>
24
#include <gen_cpp/internal_service.pb.h>
25
#include <gen_cpp/olap_file.pb.h>
26
#include <gen_cpp/types.pb.h>
27
#include <glog/logging.h>
28
#include <stddef.h>
29
#include <stdint.h>
30
31
#include <algorithm>
32
#include <cstdint>
33
#include <limits>
34
#include <memory>
35
#include <ostream>
36
#include <string>
37
#include <unordered_map>
38
#include <utility>
39
#include <vector>
40
41
#include "bthread/countdown_event.h"
42
#include "cloud/cloud_storage_engine.h"
43
#include "cloud/cloud_tablet.h"
44
#include "cloud/cloud_tablet_mgr.h"
45
#include "cloud/config.h"
46
#include "common/config.h"
47
#include "common/consts.h"
48
#include "common/exception.h"
49
#include "common/signal_handler.h"
50
#include "core/assert_cast.h"
51
#include "core/block/block.h" // Block
52
#include "core/column/column.h"
53
#include "core/column/column_nullable.h"
54
#include "core/column/column_string.h"
55
#include "core/data_type/data_type_struct.h"
56
#include "core/data_type_serde/data_type_serde.h"
57
#include "core/string_ref.h"
58
#include "core/types.h"
59
#include "exec/scan/file_scanner.h"
60
#include "format/orc/vorc_reader.h"
61
#include "format/parquet/vparquet_reader.h"
62
#include "format_v2/table/lance_reader.h"
63
#include "io/io_common.h"
64
#include "runtime/descriptors.h"
65
#include "runtime/exec_env.h"      // ExecEnv
66
#include "runtime/fragment_mgr.h"  // FragmentMgr
67
#include "runtime/runtime_state.h" // RuntimeState
68
#include "runtime/workload_group/workload_group_manager.h"
69
#include "semaphore"
70
#include "storage/olap_common.h"
71
#include "storage/rowset/beta_rowset.h"
72
#include "storage/segment/column_reader.h"
73
#include "storage/storage_engine.h"
74
#include "storage/tablet/tablet_fwd.h"
75
#include "storage/tablet/tablet_schema.h"
76
#include "storage/tablet_info.h" // DorisNodesInfo
77
#include "storage/utils.h"
78
#include "util/brpc_client_cache.h" // BrpcClientCache
79
#include "util/defer_op.h"
80
#include "util/jsonb/serialize.h"
81
#include "util/pretty_printer.h"
82
83
namespace doris {
84
85
#include "common/compile_check_begin.h"
86
87
namespace {
88
89
void set_topn_lazy_materialization_file_cache_stats(
90
0
        const io::FileCacheStatistics& stats, PTopNLazyMaterializationFileCacheStats* pstats) {
91
0
    pstats->set_local_io_count(stats.num_local_io_total);
92
0
    pstats->set_local_io_bytes(stats.bytes_read_from_local);
93
0
    pstats->set_remote_io_count(stats.num_remote_io_total);
94
0
    pstats->set_remote_io_bytes(stats.bytes_read_from_remote);
95
0
    pstats->set_skip_cache_io_count(stats.num_skip_cache_io_total);
96
0
    pstats->set_write_cache_bytes(stats.bytes_write_into_cache);
97
0
    pstats->set_local_io_time(stats.local_io_timer);
98
0
    pstats->set_remote_io_time(stats.remote_io_timer);
99
0
    pstats->set_write_cache_io_time(stats.write_cache_io_timer);
100
0
}
101
102
} // namespace
103
104
0
Status RowIDFetcher::init() {
105
0
    DorisNodesInfo nodes_info;
106
0
    nodes_info.setNodes(_fetch_option.t_fetch_opt.nodes_info);
107
0
    for (auto [node_id, node_info] : nodes_info.nodes_info()) {
108
0
        auto client = ExecEnv::GetInstance()->brpc_internal_client_cache()->get_client(
109
0
                node_info.host, node_info.brpc_port);
110
0
        if (!client) {
111
0
            LOG(WARNING) << "Get rpc stub failed, host=" << node_info.host
112
0
                         << ", port=" << node_info.brpc_port;
113
0
            return Status::InternalError("RowIDFetcher failed to init rpc client, host={}, port={}",
114
0
                                         node_info.host, node_info.brpc_port);
115
0
        }
116
0
        _stubs.push_back(client);
117
0
    }
118
0
    return Status::OK();
119
0
}
120
121
0
PMultiGetRequest RowIDFetcher::_init_fetch_request(const ColumnString& row_locs) const {
122
0
    PMultiGetRequest mget_req;
123
0
    _fetch_option.desc->to_protobuf(mget_req.mutable_desc());
124
0
    for (SlotDescriptor* slot : _fetch_option.desc->slots()) {
125
        // ignore rowid
126
0
        if (slot->col_name() == BeConsts::ROWID_COL) {
127
0
            continue;
128
0
        }
129
0
        slot->to_protobuf(mget_req.add_slots());
130
0
    }
131
0
    for (size_t i = 0; i < row_locs.size(); ++i) {
132
0
        PRowLocation row_loc;
133
0
        StringRef row_id_rep = row_locs.get_data_at(i);
134
        // TODO: When transferring data between machines with different byte orders (endianness),
135
        // not performing proper handling may lead to issues in parsing and exchanging the data.
136
0
        auto location = reinterpret_cast<const GlobalRowLoacation*>(row_id_rep.data);
137
0
        row_loc.set_tablet_id(location->tablet_id);
138
0
        row_loc.set_rowset_id(location->row_location.rowset_id.to_string());
139
0
        row_loc.set_segment_id(location->row_location.segment_id);
140
0
        row_loc.set_ordinal_id(location->row_location.row_id);
141
0
        *mget_req.add_row_locs() = std::move(row_loc);
142
0
    }
143
    // Set column desc
144
0
    for (const TColumn& tcolumn : _fetch_option.t_fetch_opt.column_desc) {
145
0
        TabletColumn column(tcolumn);
146
0
        column.to_schema_pb(mget_req.add_column_desc());
147
0
    }
148
0
    PUniqueId& query_id = *mget_req.mutable_query_id();
149
0
    query_id.set_hi(_fetch_option.runtime_state->query_id().hi);
150
0
    query_id.set_lo(_fetch_option.runtime_state->query_id().lo);
151
0
    mget_req.set_be_exec_version(_fetch_option.runtime_state->be_exec_version());
152
0
    mget_req.set_fetch_row_store(_fetch_option.t_fetch_opt.fetch_row_store);
153
0
    return mget_req;
154
0
}
155
156
Status RowIDFetcher::_merge_rpc_results(const PMultiGetRequest& request,
157
                                        const std::vector<PMultiGetResponse>& rsps,
158
                                        const std::vector<brpc::Controller>& cntls,
159
                                        Block* output_block,
160
0
                                        std::vector<PRowLocation>* rows_id) const {
161
0
    output_block->clear();
162
0
    for (const auto& cntl : cntls) {
163
0
        if (cntl.Failed()) {
164
0
            LOG(WARNING) << "Failed to fetch meet rpc error:" << cntl.ErrorText()
165
0
                         << ", host:" << cntl.remote_side();
166
0
            return Status::InternalError(cntl.ErrorText());
167
0
        }
168
0
    }
169
0
    DataTypeSerDeSPtrs serdes;
170
0
    std::unordered_map<uint32_t, uint32_t> col_uid_to_idx;
171
0
    std::vector<std::string> default_values;
172
0
    default_values.resize(_fetch_option.desc->slots().size());
173
0
    auto merge_function = [&](const PMultiGetResponse& resp) {
174
0
        Status st(Status::create(resp.status()));
175
0
        if (!st.ok()) {
176
0
            LOG(WARNING) << "Failed to fetch " << st.to_string();
177
0
            return st;
178
0
        }
179
0
        for (const PRowLocation& row_id : resp.row_locs()) {
180
0
            rows_id->push_back(row_id);
181
0
        }
182
        // Merge binary rows
183
0
        if (request.fetch_row_store()) {
184
0
            CHECK(resp.row_locs().size() == resp.binary_row_data_size());
185
0
            if (output_block->is_empty_column()) {
186
0
                *output_block = Block(_fetch_option.desc->slots(), 1);
187
0
            }
188
0
            if (serdes.empty() && col_uid_to_idx.empty()) {
189
0
                serdes = create_data_type_serdes(_fetch_option.desc->slots());
190
0
                for (int i = 0; i < _fetch_option.desc->slots().size(); ++i) {
191
0
                    col_uid_to_idx[_fetch_option.desc->slots()[i]->col_unique_id()] = i;
192
0
                    default_values[i] = _fetch_option.desc->slots()[i]->col_default_value();
193
0
                }
194
0
            }
195
0
            auto output_columns_guard = output_block->mutate_columns_scoped();
196
0
            MutableColumns& output_columns = output_columns_guard.mutable_columns();
197
0
            for (int i = 0; i < resp.binary_row_data_size(); ++i) {
198
0
                RETURN_IF_ERROR(JsonbSerializeUtil::jsonb_to_columns(
199
0
                        serdes, resp.binary_row_data(i).data(), resp.binary_row_data(i).size(),
200
0
                        col_uid_to_idx, output_columns, default_values, {}));
201
0
            }
202
0
            return Status::OK();
203
0
        }
204
        // Merge partial blocks
205
0
        Block partial_block;
206
0
        [[maybe_unused]] size_t uncompressed_size = 0;
207
0
        [[maybe_unused]] int64_t uncompressed_time = 0;
208
209
0
        RETURN_IF_ERROR(
210
0
                partial_block.deserialize(resp.block(), &uncompressed_size, &uncompressed_time));
211
0
        if (partial_block.is_empty_column()) {
212
0
            return Status::OK();
213
0
        }
214
0
        CHECK(resp.row_locs().size() == partial_block.rows());
215
0
        if (output_block->is_empty_column()) {
216
0
            output_block->swap(partial_block);
217
0
        } else if (partial_block.columns() != output_block->columns()) {
218
0
            return Status::Error<ErrorCode::INTERNAL_ERROR>(
219
0
                    "Merge block not match, self:[{}], input:[{}], ", output_block->dump_types(),
220
0
                    partial_block.dump_types());
221
0
        } else {
222
0
            for (int i = 0; i < output_block->columns(); ++i) {
223
0
                auto column_guard = output_block->mutate_column_scoped(i);
224
0
                MutableColumnPtr& column = column_guard.mutable_column();
225
0
                column->insert_range_from(
226
0
                        *partial_block.get_by_position(i).column->convert_to_full_column_if_const(),
227
0
                        0, partial_block.rows());
228
0
            }
229
0
        }
230
0
        return Status::OK();
231
0
    };
232
233
0
    for (const auto& resp : rsps) {
234
0
        RETURN_IF_ERROR(merge_function(resp));
235
0
    }
236
0
    return Status::OK();
237
0
}
238
239
0
Status RowIDFetcher::fetch(const ColumnPtr& column_row_ids, Block* res_block) {
240
0
    CHECK(!_stubs.empty());
241
0
    PMultiGetRequest mget_req = _init_fetch_request(
242
0
            assert_cast<const ColumnString&>(*remove_nullable(column_row_ids).get()));
243
0
    std::vector<PMultiGetResponse> resps(_stubs.size());
244
0
    std::vector<brpc::Controller> cntls(_stubs.size());
245
0
    bthread::CountdownEvent counter(cast_set<int>(_stubs.size()));
246
0
    for (size_t i = 0; i < _stubs.size(); ++i) {
247
0
        cntls[i].set_timeout_ms(_fetch_option.runtime_state->execution_timeout() * 1000);
248
0
        auto callback = brpc::NewCallback(fetch_callback, &counter);
249
0
        _stubs[i]->multiget_data(&cntls[i], &mget_req, &resps[i], callback);
250
0
    }
251
0
    counter.wait();
252
253
    // Merge
254
0
    std::vector<PRowLocation> rows_locs;
255
0
    rows_locs.reserve(rows_locs.size());
256
0
    RETURN_IF_ERROR(_merge_rpc_results(mget_req, resps, cntls, res_block, &rows_locs));
257
0
    if (rows_locs.size() < column_row_ids->size()) {
258
0
        return Status::InternalError("Miss matched return row loc count {}, expected {}, input {}",
259
0
                                     rows_locs.size(), res_block->rows(), column_row_ids->size());
260
0
    }
261
    // Final sort by row_ids sequence, since row_ids is already sorted if need
262
0
    std::map<GlobalRowLoacation, size_t> positions;
263
0
    for (size_t i = 0; i < rows_locs.size(); ++i) {
264
0
        RowsetId rowset_id;
265
0
        rowset_id.init(rows_locs[i].rowset_id());
266
0
        GlobalRowLoacation grl(rows_locs[i].tablet_id(), rowset_id,
267
0
                               cast_set<uint32_t>(rows_locs[i].segment_id()),
268
0
                               cast_set<uint32_t>(rows_locs[i].ordinal_id()));
269
0
        positions[grl] = i;
270
0
    };
271
    // TODO remove this warning code
272
0
    if (positions.size() < rows_locs.size()) {
273
0
        LOG(WARNING) << "cwntains duplicated row entry";
274
0
    }
275
0
    IColumn::Permutation permutation;
276
0
    permutation.reserve(column_row_ids->size());
277
0
    for (size_t i = 0; i < column_row_ids->size(); ++i) {
278
0
        auto location =
279
0
                reinterpret_cast<const GlobalRowLoacation*>(column_row_ids->get_data_at(i).data);
280
0
        permutation.push_back(positions[*location]);
281
0
    }
282
0
    for (size_t i = 0; i < res_block->columns(); ++i) {
283
0
        res_block->get_by_position(i).column =
284
0
                res_block->get_by_position(i).column->permute(permutation, permutation.size());
285
0
    }
286
    // Check row consistency
287
0
    RETURN_IF_CATCH_EXCEPTION(res_block->check_number_of_rows());
288
0
    VLOG_DEBUG << "dump block:" << res_block->dump_data(0, 10);
289
0
    return Status::OK();
290
0
}
291
292
struct IteratorKey {
293
    int64_t tablet_id;
294
    RowsetId rowset_id;
295
    uint64_t segment_id;
296
    int slot_id;
297
298
    // unordered map std::equal_to
299
12
    bool operator==(const IteratorKey& rhs) const {
300
12
        return tablet_id == rhs.tablet_id && rowset_id == rhs.rowset_id &&
301
12
               segment_id == rhs.segment_id && slot_id == rhs.slot_id;
302
12
    }
303
};
304
305
struct SegKey {
306
    int64_t tablet_id;
307
    RowsetId rowset_id;
308
    uint64_t segment_id;
309
310
    // unordered map std::equal_to
311
0
    bool operator==(const SegKey& rhs) const {
312
0
        return tablet_id == rhs.tablet_id && rowset_id == rhs.rowset_id &&
313
0
               segment_id == rhs.segment_id;
314
0
    }
315
};
316
317
struct HashOfSegKey {
318
0
    size_t operator()(const SegKey& key) const {
319
0
        size_t seed = 0;
320
0
        seed = HashUtil::hash64(&key.tablet_id, sizeof(key.tablet_id), seed);
321
0
        seed = HashUtil::hash64(&key.rowset_id.hi, sizeof(key.rowset_id.hi), seed);
322
0
        seed = HashUtil::hash64(&key.rowset_id.mi, sizeof(key.rowset_id.mi), seed);
323
0
        seed = HashUtil::hash64(&key.rowset_id.lo, sizeof(key.rowset_id.lo), seed);
324
0
        seed = HashUtil::hash64(&key.segment_id, sizeof(key.segment_id), seed);
325
0
        return seed;
326
0
    }
327
};
328
329
struct HashOfIteratorKey {
330
16
    size_t operator()(const IteratorKey& key) const {
331
16
        size_t seed = 0;
332
16
        seed = HashUtil::hash64(&key.tablet_id, sizeof(key.tablet_id), seed);
333
16
        seed = HashUtil::hash64(&key.rowset_id.hi, sizeof(key.rowset_id.hi), seed);
334
16
        seed = HashUtil::hash64(&key.rowset_id.mi, sizeof(key.rowset_id.mi), seed);
335
16
        seed = HashUtil::hash64(&key.rowset_id.lo, sizeof(key.rowset_id.lo), seed);
336
16
        seed = HashUtil::hash64(&key.segment_id, sizeof(key.segment_id), seed);
337
16
        seed = HashUtil::hash64(&key.slot_id, sizeof(key.slot_id), seed);
338
16
        return seed;
339
16
    }
340
};
341
342
struct IteratorItem {
343
    std::unique_ptr<ColumnIterator> iterator;
344
    SegmentSharedPtr segment;
345
    // for holding the reference of storage read options to avoid use after release
346
    StorageReadOptions storage_read_options;
347
};
348
349
static void set_slot_access_paths(const SlotDescriptor& slot, const TabletSchema& schema,
350
12
                                  StorageReadOptions& storage_read_options) {
351
12
    int32_t unique_id = slot.col_unique_id();
352
12
    const int field_index =
353
12
            unique_id >= 0 ? schema.field_index(unique_id) : schema.field_index(slot.col_name());
354
12
    if (field_index >= 0) {
355
12
        const auto& column = schema.column(field_index);
356
12
        unique_id = column.unique_id() >= 0 ? column.unique_id() : column.parent_unique_id();
357
12
    }
358
12
    if (unique_id < 0) {
359
0
        return;
360
0
    }
361
362
12
    if (!slot.all_access_paths().empty()) {
363
0
        storage_read_options.all_access_paths[unique_id] = slot.all_access_paths();
364
0
    }
365
366
12
    if (!slot.predicate_access_paths().empty()) {
367
0
        storage_read_options.predicate_access_paths[unique_id] = slot.predicate_access_paths();
368
0
    }
369
12
}
370
371
struct SegItem {
372
    BaseTabletSPtr tablet;
373
    BetaRowsetSharedPtr rowset;
374
    // for holding the reference of segment to avoid use after release
375
    SegmentSharedPtr segment;
376
};
377
378
Status RowIdStorageReader::read_by_rowids(const PMultiGetRequest& request,
379
2
                                          PMultiGetResponse* response) {
380
    // read from storage engine row id by row id
381
2
    OlapReaderStatistics stats;
382
2
    Block result_block;
383
2
    int64_t acquire_tablet_ms = 0;
384
2
    int64_t acquire_rowsets_ms = 0;
385
2
    int64_t acquire_segments_ms = 0;
386
2
    int64_t lookup_row_data_ms = 0;
387
388
    // init desc
389
2
    std::vector<SlotDescriptor> slots;
390
2
    slots.reserve(request.slots().size());
391
4
    for (const auto& pslot : request.slots()) {
392
4
        slots.push_back(SlotDescriptor(pslot));
393
4
    }
394
395
    // init read schema
396
2
    TabletSchema full_read_schema;
397
2
    for (const ColumnPB& column_pb : request.column_desc()) {
398
2
        full_read_schema.append_column(TabletColumn(column_pb));
399
2
    }
400
401
2
    std::unordered_map<IteratorKey, IteratorItem, HashOfIteratorKey> iterator_map;
402
    // read row by row
403
8
    for (int i = 0; i < request.row_locs_size(); ++i) {
404
6
        const auto& row_loc = request.row_locs(i);
405
6
        MonotonicStopWatch watch;
406
6
        watch.start();
407
6
        BaseTabletSPtr tablet = scope_timer_run(
408
6
                [&]() {
409
6
                    auto res = ExecEnv::get_tablet(row_loc.tablet_id(), nullptr, true);
410
6
                    return !res.has_value() ? nullptr
411
6
                                            : std::dynamic_pointer_cast<BaseTablet>(res.value());
412
6
                },
413
6
                &acquire_tablet_ms);
414
6
        RowsetId rowset_id;
415
6
        rowset_id.init(row_loc.rowset_id());
416
6
        if (!tablet) {
417
0
            continue;
418
0
        }
419
        // We ensured it's rowset is not released when init Tablet reader param, rowset->update_delayed_expired_timestamp();
420
6
        BetaRowsetSharedPtr rowset = std::static_pointer_cast<BetaRowset>(scope_timer_run(
421
6
                [&]() {
422
6
                    return ExecEnv::GetInstance()->storage_engine().get_quering_rowset(rowset_id);
423
6
                },
424
6
                &acquire_rowsets_ms));
425
6
        if (!rowset) {
426
0
            LOG(INFO) << "no such rowset " << rowset_id;
427
0
            continue;
428
0
        }
429
6
        size_t row_size = 0;
430
6
        Defer _defer([&]() {
431
6
            LOG_EVERY_N(INFO, 100)
432
1
                    << "multiget_data single_row, cost(us):" << watch.elapsed_time() / 1000
433
1
                    << ", row_size:" << row_size;
434
6
            *response->add_row_locs() = row_loc;
435
6
        });
436
        // TODO: supoort session variable enable_page_cache and disable_file_cache if necessary.
437
6
        SegmentCacheHandle segment_cache;
438
6
        RETURN_IF_ERROR(scope_timer_run(
439
6
                [&]() {
440
6
                    return SegmentLoader::instance()->load_segments(rowset, &segment_cache, true);
441
6
                },
442
6
                &acquire_segments_ms));
443
        // find segment
444
6
        auto it = std::find_if(segment_cache.get_segments().cbegin(),
445
6
                               segment_cache.get_segments().cend(),
446
6
                               [&row_loc](const segment_v2::SegmentSharedPtr& seg) {
447
6
                                   return seg->id() == row_loc.segment_id();
448
6
                               });
449
6
        if (it == segment_cache.get_segments().end()) {
450
0
            continue;
451
0
        }
452
6
        segment_v2::SegmentSharedPtr segment = *it;
453
6
        GlobalRowLoacation row_location(row_loc.tablet_id(), rowset->rowset_id(),
454
6
                                        cast_set<uint32_t>(row_loc.segment_id()),
455
6
                                        cast_set<uint32_t>(row_loc.ordinal_id()));
456
        // fetch by row store, more effcient way
457
6
        if (request.fetch_row_store()) {
458
0
            if (!tablet->tablet_schema()->has_row_store_for_all_columns()) {
459
0
                return Status::InternalError("Tablet {} does not have row store for all columns",
460
0
                                             tablet->tablet_id());
461
0
            }
462
0
            RowLocation loc(rowset_id, segment->id(), cast_set<uint32_t>(row_loc.ordinal_id()));
463
0
            std::string* value = response->add_binary_row_data();
464
0
            RETURN_IF_ERROR(scope_timer_run(
465
0
                    [&]() { return tablet->lookup_row_data({}, loc, rowset, stats, *value); },
466
0
                    &lookup_row_data_ms));
467
0
            row_size = value->size();
468
0
            continue;
469
0
        }
470
471
        // fetch by column store
472
6
        if (result_block.is_empty_column()) {
473
2
            result_block = Block(slots, request.row_locs().size());
474
2
        }
475
6
        VLOG_DEBUG << "Read row location "
476
0
                   << fmt::format("{}, {}, {}, {}", row_location.tablet_id,
477
0
                                  row_location.row_location.rowset_id.to_string(),
478
0
                                  row_location.row_location.segment_id,
479
0
                                  row_location.row_location.row_id);
480
18
        for (int x = 0; x < slots.size(); ++x) {
481
12
            auto row_id = static_cast<segment_v2::rowid_t>(row_loc.ordinal_id());
482
            // The scoped mutation republishes a potentially replaced V2 column when the reader
483
            // transfers ownership of its first assembled batch.
484
12
            auto column_guard = result_block.mutate_column_scoped(x);
485
12
            MutableColumnPtr& column = column_guard.mutable_column();
486
12
            IteratorKey iterator_key {.tablet_id = tablet->tablet_id(),
487
12
                                      .rowset_id = rowset_id,
488
12
                                      .segment_id = row_loc.segment_id(),
489
12
                                      .slot_id = slots[x].id()};
490
12
            IteratorItem& iterator_item = iterator_map[iterator_key];
491
12
            if (iterator_item.segment == nullptr) {
492
                // hold the reference
493
4
                iterator_map[iterator_key].segment = segment;
494
4
                iterator_item.storage_read_options.stats = &stats;
495
4
                iterator_item.storage_read_options.io_ctx.reader_type = ReaderType::READER_QUERY;
496
4
            }
497
12
            segment = iterator_item.segment;
498
12
            set_slot_access_paths(slots[x], full_read_schema, iterator_item.storage_read_options);
499
12
            RETURN_IF_ERROR(segment->seek_and_read_by_rowid(
500
12
                    full_read_schema, &slots[x], row_id, column, iterator_item.storage_read_options,
501
12
                    iterator_item.iterator));
502
12
        }
503
6
    }
504
    // serialize block if not empty
505
2
    if (!result_block.is_empty_column()) {
506
2
        VLOG_DEBUG << "dump block:" << result_block.dump_data(0, 10)
507
0
                   << ", be_exec_version:" << request.be_exec_version();
508
2
        [[maybe_unused]] size_t compressed_size = 0;
509
2
        [[maybe_unused]] size_t uncompressed_size = 0;
510
2
        [[maybe_unused]] int64_t compress_time = 0;
511
2
        int be_exec_version = request.has_be_exec_version() ? request.be_exec_version() : 0;
512
2
        RETURN_IF_ERROR(result_block.serialize(be_exec_version, response->mutable_block(),
513
2
                                               &uncompressed_size, &compressed_size, &compress_time,
514
2
                                               segment_v2::CompressionTypePB::LZ4));
515
2
    }
516
517
2
    LOG(INFO) << "Query stats: "
518
2
              << fmt::format(
519
2
                         "query_id:{}, "
520
2
                         "hit_cached_pages:{}, total_pages_read:{}, compressed_bytes_read:{}, "
521
2
                         "io_latency:{}ns, "
522
2
                         "uncompressed_bytes_read:{},"
523
2
                         "bytes_read:{},"
524
2
                         "acquire_tablet_ms:{}, acquire_rowsets_ms:{}, acquire_segments_ms:{}, "
525
2
                         "lookup_row_data_ms:{}",
526
2
                         print_id(request.query_id()), stats.cached_pages_num,
527
2
                         stats.total_pages_num, stats.compressed_bytes_read, stats.io_ns,
528
2
                         stats.uncompressed_bytes_read, stats.bytes_read, acquire_tablet_ms,
529
2
                         acquire_rowsets_ms, acquire_segments_ms, lookup_row_data_ms);
530
2
    return Status::OK();
531
2
}
532
533
Status RowIdStorageReader::read_by_rowids(const PMultiGetRequestV2& request,
534
0
                                          PMultiGetResponseV2* response) {
535
0
    if (request.request_block_descs_size()) {
536
0
        auto tquery_id = ((UniqueId)request.query_id()).to_thrift();
537
        // todo: use mutableBlock instead of block
538
0
        std::vector<Block> result_blocks(request.request_block_descs_size());
539
540
0
        OlapReaderStatistics stats;
541
0
        int64_t acquire_tablet_ms = 0;
542
0
        int64_t acquire_rowsets_ms = 0;
543
0
        int64_t acquire_segments_ms = 0;
544
0
        int64_t lookup_row_data_ms = 0;
545
546
0
        int64_t external_init_reader_avg_ms = 0;
547
0
        int64_t external_get_block_avg_ms = 0;
548
0
        size_t external_scan_range_cnt = 0;
549
550
0
        const auto file_cache_miss_policy =
551
0
                request.file_cache_remote_only_on_miss()
552
0
                        ? io::FileCacheMissPolicy::REMOTE_ONLY_ON_MISS
553
0
                        : io::FileCacheMissPolicy::READ_THROUGH_AND_WRITE_BACK;
554
555
        // Add counters for different file mapping types
556
0
        std::unordered_map<FileMappingType, int64_t> file_type_counts;
557
558
0
        auto id_file_map =
559
0
                ExecEnv::GetInstance()->get_id_manager()->get_id_file_map(request.query_id());
560
        // if id_file_map is null, means the BE not have scan range, just return ok
561
0
        if (!id_file_map) {
562
            // padding empty block to response
563
0
            LOG(INFO) << "id_file_map not found for query_id: " << print_id(request.query_id());
564
0
            for (int i = 0; i < request.request_block_descs_size(); ++i) {
565
0
                response->add_blocks();
566
0
            }
567
0
            return Status::OK();
568
0
        }
569
570
0
        for (int i = 0; i < request.request_block_descs_size(); ++i) {
571
0
            const auto& request_block_desc = request.request_block_descs(i);
572
0
            PMultiGetBlockV2* pblock = response->add_blocks();
573
0
            if (request_block_desc.row_id_size() >= 1) {
574
0
                if (request_block_desc.file_id_size() != request_block_desc.row_id_size()) {
575
0
                    return Status::InvalidArgument(
576
0
                            "Row-id fetch has mismatched file_id and row_id counts: "
577
0
                            "file_ids={}, row_ids={}",
578
0
                            request_block_desc.file_id_size(), request_block_desc.row_id_size());
579
0
                }
580
0
                const auto row_location_version = request_block_desc.row_location_version();
581
0
                switch (row_location_version) {
582
0
                case static_cast<uint32_t>(ROW_VERSION::FILE_LOCAL_ROW_ID):
583
0
                case static_cast<uint32_t>(ROW_VERSION::LANCE_DATASET_ROW_ID):
584
0
                    break;
585
0
                default:
586
0
                    return Status::NotSupported("unsupported row location version in fetch: {}",
587
0
                                                row_location_version);
588
0
                }
589
                // Rows in this request block belong to the same relation. The first mapping is
590
                // sufficient to select the internal/external fetch path; individual file IDs
591
                // are still used below to group the actual reads.
592
0
                const auto first_file_id = request_block_desc.file_id(0);
593
0
                auto first_file_mapping = id_file_map->get_file_mapping(first_file_id);
594
0
                if (!first_file_mapping) {
595
0
                    return Status::InternalError(
596
0
                            "Backend:{} file mapping not found, query_id: {}, file_id: {}",
597
0
                            BackendOptions::get_localhost(), print_id(request.query_id()),
598
0
                            first_file_id);
599
0
                }
600
0
                if (first_file_mapping->type != FileMappingType::EXTERNAL &&
601
0
                    row_location_version ==
602
0
                            static_cast<uint32_t>(ROW_VERSION::LANCE_DATASET_ROW_ID)) {
603
0
                    return Status::InvalidArgument(
604
0
                            "Lance dataset row IDs require an external file mapping");
605
0
                }
606
0
                file_type_counts[first_file_mapping->type] += request_block_desc.row_id_size();
607
608
                // prepare slots to build block
609
0
                std::vector<SlotDescriptor> slots;
610
0
                slots.reserve(request_block_desc.slots_size());
611
0
                for (const auto& pslot : request_block_desc.slots()) {
612
0
                    slots.push_back(SlotDescriptor(pslot));
613
0
                }
614
0
                try {
615
0
                    if (first_file_mapping->type == FileMappingType::INTERNAL) {
616
0
                        RETURN_IF_ERROR(read_batch_doris_format_row(
617
0
                                request_block_desc, id_file_map, slots, tquery_id, result_blocks[i],
618
0
                                stats, &acquire_tablet_ms, &acquire_rowsets_ms,
619
0
                                &acquire_segments_ms, &lookup_row_data_ms, file_cache_miss_policy));
620
0
                    } else {
621
0
                        RETURN_IF_ERROR(read_batch_external_row(
622
0
                                request.wg_id(), request_block_desc, id_file_map, slots,
623
0
                                first_file_mapping, tquery_id, result_blocks[i],
624
0
                                pblock->mutable_profile(), &external_init_reader_avg_ms,
625
0
                                &external_get_block_avg_ms, &external_scan_range_cnt));
626
0
                    }
627
0
                } catch (const Exception& e) {
628
0
                    return Status::Error<false>(e.code(), "Row id fetch failed because {}",
629
0
                                                e.what());
630
0
                }
631
0
            }
632
633
0
            [[maybe_unused]] size_t compressed_size = 0;
634
0
            [[maybe_unused]] size_t uncompressed_size = 0;
635
0
            [[maybe_unused]] int64_t compress_time = 0;
636
0
            int be_exec_version = request.has_be_exec_version() ? request.be_exec_version() : 0;
637
0
            RETURN_IF_ERROR(result_blocks[i].serialize(
638
0
                    be_exec_version, pblock->mutable_block(), &uncompressed_size, &compressed_size,
639
0
                    &compress_time, segment_v2::CompressionTypePB::LZ4));
640
0
        }
641
642
        // Build file type statistics string
643
0
        std::string file_type_stats;
644
0
        for (const auto& [type, count] : file_type_counts) {
645
0
            if (!file_type_stats.empty()) {
646
0
                file_type_stats += ", ";
647
0
            }
648
0
            file_type_stats += fmt::format("{}:{}", type, count);
649
0
        }
650
651
0
        LOG(INFO) << "Query stats: "
652
0
                  << fmt::format(
653
0
                             "query_id:{}, "
654
0
                             "Internal table:"
655
0
                             "hit_cached_pages:{}, total_pages_read:{}, compressed_bytes_read:{}, "
656
0
                             "io_latency:{}ns, uncompressed_bytes_read:{}, bytes_read:{}, "
657
0
                             "acquire_tablet_ms:{}, acquire_rowsets_ms:{}, acquire_segments_ms:{}, "
658
0
                             "lookup_row_data_ms:{}, file_types:[{}]; "
659
0
                             "External table : init_reader_ms:{}, get_block_ms:{}, "
660
0
                             "external_scan_range_cnt:{}",
661
0
                             print_id(request.query_id()), stats.cached_pages_num,
662
0
                             stats.total_pages_num, stats.compressed_bytes_read, stats.io_ns,
663
0
                             stats.uncompressed_bytes_read, stats.bytes_read, acquire_tablet_ms,
664
0
                             acquire_rowsets_ms, acquire_segments_ms, lookup_row_data_ms,
665
0
                             file_type_stats, external_init_reader_avg_ms,
666
0
                             external_get_block_avg_ms, external_scan_range_cnt);
667
0
        set_topn_lazy_materialization_file_cache_stats(
668
0
                stats.file_cache_stats,
669
0
                response->mutable_topn_lazy_materialization_file_cache_stats());
670
0
    }
671
672
0
    if (request.has_gc_id_map() && request.gc_id_map()) {
673
0
        ExecEnv::GetInstance()->get_id_manager()->remove_id_file_map(request.query_id());
674
0
    }
675
676
0
    return Status::OK();
677
0
}
678
679
Status RowIdStorageReader::read_batch_doris_format_row(
680
        const PRequestBlockDesc& request_block_desc, std::shared_ptr<IdFileMap> id_file_map,
681
        std::vector<SlotDescriptor>& slots, const TUniqueId& query_id, Block& result_block,
682
        OlapReaderStatistics& stats, int64_t* acquire_tablet_ms, int64_t* acquire_rowsets_ms,
683
        int64_t* acquire_segments_ms, int64_t* lookup_row_data_ms,
684
0
        io::FileCacheMissPolicy file_cache_miss_policy) {
685
0
    if (result_block.is_empty_column()) [[likely]] {
686
0
        result_block = Block(slots, request_block_desc.row_id_size());
687
0
    }
688
0
    TabletSchema full_read_schema;
689
0
    for (const ColumnPB& column_pb : request_block_desc.column_descs()) {
690
0
        full_read_schema.append_column(TabletColumn(column_pb));
691
0
    }
692
693
0
    std::unordered_map<IteratorKey, IteratorItem, HashOfIteratorKey> iterator_map;
694
0
    std::unordered_map<SegKey, SegItem, HashOfSegKey> seg_map;
695
0
    std::string row_store_buffer;
696
0
    RowStoreReadStruct row_store_read_struct(row_store_buffer);
697
0
    if (request_block_desc.fetch_row_store()) {
698
0
        for (int i = 0; i < request_block_desc.slots_size(); ++i) {
699
0
            row_store_read_struct.serdes.emplace_back(slots[i].get_data_type_ptr()->get_serde());
700
0
            row_store_read_struct.col_uid_to_idx[slots[i].col_unique_id()] = i;
701
0
            row_store_read_struct.default_values.emplace_back(slots[i].col_default_value());
702
0
        }
703
0
    }
704
705
0
    std::vector<uint32_t> row_ids;
706
0
    int k = 1;
707
0
    auto max_k = 0;
708
0
    for (int j = 0; j < request_block_desc.row_id_size();) {
709
0
        auto file_id = request_block_desc.file_id(j);
710
0
        const auto row_id = request_block_desc.row_id(j);
711
0
        if (row_id > std::numeric_limits<uint32_t>::max()) {
712
0
            return Status::InvalidArgument("internal row id exceeds uint32 range: {}", row_id);
713
0
        }
714
0
        row_ids.emplace_back(static_cast<uint32_t>(row_id));
715
0
        auto file_mapping = id_file_map->get_file_mapping(file_id);
716
0
        if (!file_mapping) {
717
0
            return Status::InternalError(
718
0
                    "Backend:{} file_mapping not found, query_id: {}, file_id: {}",
719
0
                    BackendOptions::get_localhost(), print_id(query_id), file_id);
720
0
        }
721
0
        for (k = 1; j + k < request_block_desc.row_id_size(); ++k) {
722
0
            if (request_block_desc.file_id(j + k) == file_id) {
723
0
                const auto next_row_id = request_block_desc.row_id(j + k);
724
0
                if (next_row_id > std::numeric_limits<uint32_t>::max()) {
725
0
                    return Status::InvalidArgument("internal row id exceeds uint32 range: {}",
726
0
                                                   next_row_id);
727
0
                }
728
0
                row_ids.emplace_back(static_cast<uint32_t>(next_row_id));
729
0
            } else {
730
0
                break;
731
0
            }
732
0
        }
733
734
0
        RETURN_IF_ERROR(read_doris_format_row(
735
0
                id_file_map, file_mapping, row_ids, slots, full_read_schema, row_store_read_struct,
736
0
                stats, acquire_tablet_ms, acquire_rowsets_ms, acquire_segments_ms,
737
0
                lookup_row_data_ms, seg_map, iterator_map, file_cache_miss_policy, result_block));
738
739
0
        j += k;
740
0
        max_k = std::max(max_k, k);
741
0
        row_ids.clear();
742
0
    }
743
744
0
    return Status::OK();
745
0
}
746
747
const std::string RowIdStorageReader::ScannersRunningTimeProfile = "ScannersRunningTime";
748
const std::string RowIdStorageReader::InitReaderAvgTimeProfile = "InitReaderAvgTime";
749
const std::string RowIdStorageReader::GetBlockAvgTimeProfile = "GetBlockAvgTime";
750
const std::string RowIdStorageReader::FileReadLinesProfile = "FileReadLines";
751
const std::string RowIdStorageReader::LanceDatasetOpenTimeProfile = "LanceDatasetOpenTime";
752
const std::string RowIdStorageReader::LanceRowIdTakeReadTimeProfile = "LanceRowIdTakeReadTime";
753
const std::string RowIdStorageReader::LanceArrowToDorisBlockTimeProfile =
754
        "LanceArrowToDorisBlockTime";
755
const std::string RowIdStorageReader::LanceRowIdFetchTotalTimeProfile = "LanceRowIdFetchTotalTime";
756
const std::string RowIdStorageReader::TopNLazyMaterializationSecondPhaseLocalIOCount =
757
        "TopNLazyMaterializationSecondPhaseLocalIOCount";
758
const std::string RowIdStorageReader::TopNLazyMaterializationSecondPhaseLocalIOBytes =
759
        "TopNLazyMaterializationSecondPhaseLocalIOBytes";
760
const std::string RowIdStorageReader::TopNLazyMaterializationSecondPhaseRemoteIOCount =
761
        "TopNLazyMaterializationSecondPhaseRemoteIOCount";
762
const std::string RowIdStorageReader::TopNLazyMaterializationSecondPhaseRemoteIOBytes =
763
        "TopNLazyMaterializationSecondPhaseRemoteIOBytes";
764
const std::string RowIdStorageReader::TopNLazyMaterializationSecondPhaseSkipCacheIOCount =
765
        "TopNLazyMaterializationSecondPhaseSkipCacheIOCount";
766
const std::string RowIdStorageReader::TopNLazyMaterializationSecondPhaseWriteCacheBytes =
767
        "TopNLazyMaterializationSecondPhaseWriteCacheBytes";
768
const std::string RowIdStorageReader::TopNLazyMaterializationSecondPhaseLocalIOTime =
769
        "TopNLazyMaterializationSecondPhaseLocalIOTime";
770
const std::string RowIdStorageReader::TopNLazyMaterializationSecondPhaseRemoteIOTime =
771
        "TopNLazyMaterializationSecondPhaseRemoteIOTime";
772
const std::string RowIdStorageReader::TopNLazyMaterializationSecondPhaseWriteCacheIOTime =
773
        "TopNLazyMaterializationSecondPhaseWriteCacheIOTime";
774
const std::string RowIdStorageReader::TopNLazyMaterializationSecondPhaseRowsRead =
775
        "TopNLazyMaterializationSecondPhaseRowsRead";
776
const std::string RowIdStorageReader::TopNLazyMaterializationSecondPhaseSegmentsRead =
777
        "TopNLazyMaterializationSecondPhaseSegmentsRead";
778
779
Status RowIdStorageReader::read_lance_rows_by_row_ids(
780
        const TFileRangeDesc& scan_range_desc, const std::vector<uint64_t>& row_ids,
781
        const std::vector<SlotDescriptor>& slots, RuntimeState* runtime_state,
782
        RuntimeProfile* runtime_profile, const TFileScanRangeParams& scan_params, Block* block,
783
0
        RowIdStorageReader::ExternalFetchStatistics* fetch_statistics) {
784
    // Unlike Parquet/ORC, a Lance row ID is a native uint64 ID in a fixed dataset snapshot,
785
    // rather than a row ordinal interpreted by a reader for one physical file range. Phase-two
786
    // fetch therefore opens that dataset snapshot and uses dataset-level take_rows directly. It
787
    // does not create a FileScanner or rescan the fragment split that produced the row ID.
788
0
    std::vector<format::ColumnDefinition> projected_columns;
789
0
    projected_columns.reserve(slots.size());
790
0
    for (const auto& slot : slots) {
791
0
        format::ColumnDefinition column;
792
0
        column.identifier = Field::create_field<TYPE_STRING>(slot.col_name());
793
0
        column.name = slot.col_name();
794
0
        column.type = slot.get_data_type_ptr();
795
0
        projected_columns.emplace_back(std::move(column));
796
0
    }
797
798
0
    format::lance::LanceTableReader reader;
799
0
    auto lance_scan_params = scan_params;
800
0
    RETURN_IF_ERROR(scope_timer_run(
801
0
            [&]() {
802
0
                return reader.init({
803
0
                        .projected_columns = projected_columns,
804
0
                        .conjuncts = {},
805
0
                        .format = format::FileFormat::LANCE,
806
0
                        .scan_params = &lance_scan_params,
807
0
                        .io_ctx = nullptr,
808
0
                        .runtime_state = runtime_state,
809
0
                        .scanner_profile = runtime_profile,
810
0
                });
811
0
            },
812
0
            &fetch_statistics->init_reader_ms));
813
0
    RETURN_IF_ERROR(scope_timer_run(
814
0
            [&]() { return reader.read_by_row_ids(scan_range_desc, row_ids, block); },
815
0
            &fetch_statistics->get_block_ms));
816
0
    RETURN_IF_ERROR(reader.close());
817
818
0
    const auto collect_lance_fetch_time = [&](const std::string& timer_name) {
819
0
        if (const auto* counter = runtime_profile->get_counter(timer_name); counter != nullptr) {
820
0
            fetch_statistics->lance_fetch_times_ns.emplace(timer_name, counter->value());
821
0
        }
822
0
    };
823
0
    collect_lance_fetch_time(LanceDatasetOpenTimeProfile);
824
0
    collect_lance_fetch_time(LanceRowIdTakeReadTimeProfile);
825
0
    collect_lance_fetch_time(LanceArrowToDorisBlockTimeProfile);
826
0
    collect_lance_fetch_time(LanceRowIdFetchTotalTimeProfile);
827
0
    return Status::OK();
828
0
}
829
830
Status RowIdStorageReader::read_external_row_from_file_mapping(
831
        size_t idx, const std::multimap<uint64_t, size_t>& row_ids,
832
        const std::shared_ptr<FileMapping>& file_mapping, const std::vector<SlotDescriptor>& slots,
833
        const TUniqueId& query_id, const std::shared_ptr<RuntimeState>& runtime_state,
834
        std::vector<Block>& scan_blocks, std::vector<std::pair<size_t, size_t>>& row_id_block_idx,
835
        std::vector<RowIdStorageReader::ExternalFetchStatistics>& fetch_statistics,
836
        const TFileScanRangeParams& rpc_scan_params,
837
        const std::unordered_map<std::string, int>& colname_to_slot_id,
838
0
        TupleDescriptor& tuple_desc) {
839
0
    SCOPED_ATTACH_TASK(ExecEnv::GetInstance()->rowid_storage_reader_tracker());
840
0
    signal::set_signal_task_id(query_id);
841
842
0
    std::vector<uint64_t> read_ids;
843
    // Generate an ordered, deduplicated list with the help of the multimap ordering.
844
0
    for (const auto& [row_id, result_block_idx] : row_ids) {
845
0
        if (read_ids.empty() || read_ids.back() != row_id) {
846
0
            read_ids.emplace_back(row_id);
847
0
        }
848
0
        row_id_block_idx[result_block_idx] = std::make_pair(idx, read_ids.size() - 1);
849
0
    }
850
851
0
    scan_blocks[idx] = Block(slots, read_ids.size());
852
853
0
    auto& external_info = file_mapping->get_external_file_info();
854
0
    auto& scan_range_desc = external_info.scan_range_desc;
855
856
    // Clear to avoid reading iceberg position delete file...
857
0
    scan_range_desc.table_format_params.iceberg_params = TIcebergFileDesc {};
858
859
    // Clear to avoid reading hive transactional delete delta file...
860
0
    scan_range_desc.table_format_params.transactional_hive_params = TTransactionalHiveDesc {};
861
862
0
    std::unique_ptr<RuntimeProfile> sub_runtime_profile =
863
0
            std::make_unique<RuntimeProfile>("ExternalRowIDFetcher");
864
0
    const auto format_type = scan_range_desc.__isset.format_type ? scan_range_desc.format_type
865
0
                                                                 : rpc_scan_params.format_type;
866
0
    if (format_type == TFileFormatType::FORMAT_LANCE) {
867
0
        RETURN_IF_ERROR(read_lance_rows_by_row_ids(
868
0
                scan_range_desc, read_ids, slots, runtime_state.get(), sub_runtime_profile.get(),
869
0
                rpc_scan_params, &scan_blocks[idx], &fetch_statistics[idx]));
870
0
    } else {
871
        // Parquet/ORC row IDs are consumed as row ordinals within the exact physical file range
872
        // recorded by phase one. Keep using FileScanner so the format reader can resolve those
873
        // ordinals against that range; unlike Lance, ranges cannot be merged at dataset level.
874
0
        std::list<int64_t> legacy_read_ids;
875
0
        for (const auto row_id : read_ids) {
876
0
            if (row_id > static_cast<uint64_t>(std::numeric_limits<int64_t>::max())) {
877
0
                return Status::InvalidArgument("legacy external row id exceeds int64 range: {}",
878
0
                                               row_id);
879
0
            }
880
0
            legacy_read_ids.emplace_back(static_cast<int64_t>(row_id));
881
0
        }
882
0
        std::unique_ptr<FileScanner> vfile_scanner_ptr =
883
0
                FileScanner::create_unique(runtime_state.get(), sub_runtime_profile.get(),
884
0
                                           &rpc_scan_params, &colname_to_slot_id, &tuple_desc);
885
886
0
        RETURN_IF_ERROR(vfile_scanner_ptr->prepare_for_read_lines(scan_range_desc));
887
0
        RETURN_IF_ERROR(vfile_scanner_ptr->read_lines_from_range(
888
0
                scan_range_desc, legacy_read_ids, &scan_blocks[idx], external_info,
889
0
                &fetch_statistics[idx].init_reader_ms, &fetch_statistics[idx].get_block_ms));
890
0
    }
891
892
0
    auto file_read_bytes_counter =
893
0
            sub_runtime_profile->get_counter(FileScanner::FileReadBytesProfile);
894
895
0
    if (file_read_bytes_counter != nullptr) {
896
0
        fetch_statistics[idx].file_read_bytes = PrettyPrinter::print(
897
0
                file_read_bytes_counter->value(), file_read_bytes_counter->type());
898
0
    }
899
900
0
    auto file_read_times_counter =
901
0
            sub_runtime_profile->get_counter(FileScanner::FileReadTimeProfile);
902
0
    if (file_read_times_counter != nullptr) {
903
0
        fetch_statistics[idx].file_read_times = PrettyPrinter::print(
904
0
                file_read_times_counter->value(), file_read_times_counter->type());
905
0
    }
906
907
0
    return Status::OK();
908
0
}
909
910
Status RowIdStorageReader::read_batch_external_row(
911
        const uint64_t workload_group_id, const PRequestBlockDesc& request_block_desc,
912
        std::shared_ptr<IdFileMap> id_file_map, std::vector<SlotDescriptor>& slots,
913
        std::shared_ptr<FileMapping> first_file_mapping, const TUniqueId& query_id,
914
        Block& result_block, PRuntimeProfileTree* pprofile, int64_t* init_reader_avg_ms,
915
0
        int64_t* get_block_avg_ms, size_t* scan_range_cnt) {
916
0
    TFileScanRangeParams rpc_scan_params;
917
0
    TupleDescriptor tuple_desc(request_block_desc.desc(), false);
918
0
    std::unordered_map<std::string, int> colname_to_slot_id;
919
0
    std::shared_ptr<RuntimeState> runtime_state = nullptr;
920
921
0
    int max_file_scanners = 0;
922
0
    {
923
0
        if (result_block.is_empty_column()) [[likely]] {
924
0
            result_block = Block(slots, request_block_desc.row_id_size());
925
0
        }
926
927
0
        auto& external_info = first_file_mapping->get_external_file_info();
928
0
        int plan_node_id = external_info.plan_node_id;
929
0
        const auto& first_scan_range_desc = external_info.scan_range_desc;
930
931
0
        const auto& external_scan_params = id_file_map->get_external_scan_params();
932
0
        const auto scan_params = external_scan_params.find(plan_node_id);
933
0
        if (scan_params == external_scan_params.end()) {
934
0
            return Status::InternalError("External scan params not found for plan node id {}",
935
0
                                         plan_node_id);
936
0
        }
937
0
        rpc_scan_params = scan_params->second;
938
0
        if (request_block_desc.row_location_version() ==
939
0
            static_cast<uint32_t>(ROW_VERSION::LANCE_DATASET_ROW_ID)) {
940
0
            const auto format_type = first_scan_range_desc.__isset.format_type
941
0
                                             ? first_scan_range_desc.format_type
942
0
                                             : rpc_scan_params.format_type;
943
0
            if (format_type != TFileFormatType::FORMAT_LANCE) {
944
0
                return Status::InvalidArgument(
945
0
                        "Lance dataset row IDs cannot be fetched with file format {}",
946
0
                        static_cast<int>(format_type));
947
0
            }
948
0
        }
949
950
0
        rpc_scan_params.required_slots.clear();
951
0
        rpc_scan_params.column_idxs.clear();
952
0
        rpc_scan_params.slot_name_to_schema_pos.clear();
953
954
0
        std::set partition_name_set(first_scan_range_desc.columns_from_path_keys.begin(),
955
0
                                    first_scan_range_desc.columns_from_path_keys.end());
956
0
        for (auto slot_idx = 0; slot_idx < slots.size(); ++slot_idx) {
957
0
            auto& slot = slots[slot_idx];
958
0
            tuple_desc.add_slot(&slot);
959
0
            colname_to_slot_id.emplace(slot.col_name(), slot.id());
960
0
            TFileScanSlotInfo slot_info;
961
0
            slot_info.slot_id = slot.id();
962
0
            auto column_idx = request_block_desc.column_idxs(slot_idx);
963
0
            if (partition_name_set.contains(slot.col_name())) {
964
                //This is partition column.
965
0
                slot_info.is_file_slot = false;
966
0
            } else {
967
0
                rpc_scan_params.column_idxs.emplace_back(column_idx);
968
0
                slot_info.is_file_slot = true;
969
0
            }
970
0
            rpc_scan_params.default_value_of_src_slot.emplace(slot.id(), TExpr {});
971
0
            rpc_scan_params.required_slots.emplace_back(slot_info);
972
0
            rpc_scan_params.slot_name_to_schema_pos.emplace(slot.col_name(), column_idx);
973
0
        }
974
975
0
        const auto& query_options = id_file_map->get_query_options();
976
0
        const auto& query_globals = id_file_map->get_query_globals();
977
        /*
978
         * The scan stage needs the information in query_options to generate different behaviors according to the specific variables:
979
         *  query_options.hive_parquet_use_column_names, query_options.truncate_char_or_varchar_columns,query_globals.time_zone ...
980
         *
981
         * To ensure the same behavior as the scan stage, I get query_options query_globals from id_file_map, then create runtime_state
982
         * and pass it to vfile_scanner so that the runtime_state information is the same as the scan stage and the behavior is also consistent.
983
         */
984
0
        runtime_state = RuntimeState::create_shared(
985
0
                query_id, -1, query_options, query_globals, ExecEnv::GetInstance(),
986
0
                ExecEnv::GetInstance()->rowid_storage_reader_tracker());
987
988
0
        max_file_scanners = id_file_map->get_max_file_scanners();
989
0
    }
990
991
    // Hash(TFileRangeDesc) => { all the rows that need to be read and their positions in the result block. } +  file mapping
992
    // The multimap retains duplicate row IDs because the same source row can appear more than once
993
    // in the materialized result.
994
0
    std::map<std::string, std::pair<std::multimap<uint64_t, size_t>, std::shared_ptr<FileMapping>>>
995
0
            scan_rows;
996
997
    // Block corresponding to the order of `scan_rows` map.
998
0
    std::vector<Block> scan_blocks;
999
1000
    // row_id (Indexing of vectors) => < In which block, which line in the block >
1001
0
    std::vector<std::pair<size_t, size_t>> row_id_block_idx;
1002
1003
    // Count the time/bytes it takes to read each TFileRangeDesc. (for profile)
1004
0
    std::vector<ExternalFetchStatistics> fetch_statistics;
1005
1006
0
    auto hash_file_range = [&rpc_scan_params](const ExternalFileMappingInfo& external_info) {
1007
0
        const auto& file_range_desc = external_info.scan_range_desc;
1008
0
        const auto format_type = file_range_desc.__isset.format_type ? file_range_desc.format_type
1009
0
                                                                     : rpc_scan_params.format_type;
1010
0
        if (format_type == TFileFormatType::FORMAT_LANCE) {
1011
            // Parquet and ORC row IDs are offsets within a physical file range, so their fetch
1012
            // path must keep each path/start_offset pair separate. Lance row IDs instead belong
1013
            // to a fixed dataset snapshot. Although phase one registers one file mapping per
1014
            // fragment split, phase two uses dataset-level take_rows and can fetch row IDs from
1015
            // all of those fragments together. Group the mappings by scan node, snapshot, and
1016
            // dataset URI; the plan node keeps independent scans of the same snapshot isolated.
1017
0
            const auto& lance_params = file_range_desc.table_format_params.lance_params;
1018
0
            return fmt::format("lance:{}:{}:{}", external_info.plan_node_id, lance_params.version,
1019
0
                               lance_params.dataset_uri);
1020
0
        }
1021
0
        std::string value;
1022
0
        value.resize(file_range_desc.path.size() + sizeof(file_range_desc.start_offset));
1023
0
        auto* ptr = value.data();
1024
1025
0
        memcpy(ptr, &file_range_desc.start_offset, sizeof(file_range_desc.start_offset));
1026
0
        ptr += sizeof(file_range_desc.start_offset);
1027
0
        memcpy(ptr, file_range_desc.path.data(), file_range_desc.path.size());
1028
0
        return value;
1029
0
    };
1030
1031
0
    for (int j = 0; j < request_block_desc.row_id_size(); ++j) {
1032
0
        auto file_id = request_block_desc.file_id(j);
1033
0
        auto file_mapping = id_file_map->get_file_mapping(file_id);
1034
0
        if (!file_mapping) {
1035
0
            return Status::InternalError(
1036
0
                    "Backend:{} file_mapping not found, query_id: {}, file_id: {}",
1037
0
                    BackendOptions::get_localhost(), print_id(query_id), file_id);
1038
0
        }
1039
1040
0
        const auto& external_info = file_mapping->get_external_file_info();
1041
0
        const auto& scan_range_hash = hash_file_range(external_info);
1042
0
        if (scan_rows.contains(scan_range_hash)) {
1043
0
            scan_rows.at(scan_range_hash).first.emplace(request_block_desc.row_id(j), j);
1044
0
        } else {
1045
0
            std::multimap<uint64_t, size_t> rows {{request_block_desc.row_id(j), j}};
1046
0
            scan_rows.emplace(scan_range_hash, std::make_pair(std::move(rows), file_mapping));
1047
0
        }
1048
0
    }
1049
1050
0
    scan_blocks.resize(scan_rows.size());
1051
0
    row_id_block_idx.resize(request_block_desc.row_id_size());
1052
0
    fetch_statistics.resize(scan_rows.size());
1053
1054
    // Get the workload group for subsequent scan task submission.
1055
0
    std::vector<uint64_t> workload_group_ids;
1056
0
    workload_group_ids.emplace_back(workload_group_id);
1057
0
    auto wg = ExecEnv::GetInstance()->workload_group_mgr()->get_group(workload_group_ids);
1058
0
    doris::TaskScheduler* exec_sched = nullptr;
1059
0
    ScannerScheduler* scan_sched = nullptr;
1060
0
    ScannerScheduler* remote_scan_sched = nullptr;
1061
0
    wg->get_query_scheduler(&exec_sched, &scan_sched, &remote_scan_sched);
1062
0
    DCHECK(remote_scan_sched);
1063
1064
0
    int64_t scan_running_time = 0;
1065
0
    RETURN_IF_ERROR(scope_timer_run(
1066
0
            [&]() -> Status {
1067
                // Make sure to insert data into result_block only after all scan tasks have been executed.
1068
0
                std::atomic<int> producer_count {0};
1069
0
                std::condition_variable cv;
1070
0
                std::mutex mtx;
1071
0
                Status scan_status = Status::OK();
1072
1073
                //semaphore: Limit the number of scan tasks submitted at one time
1074
0
                std::counting_semaphore semaphore {max_file_scanners};
1075
1076
0
                size_t idx = 0;
1077
0
                for (const auto& [_, scan_info] : scan_rows) {
1078
0
                    semaphore.acquire();
1079
0
                    RETURN_IF_ERROR(remote_scan_sched->submit_scan_task(
1080
0
                            SimplifiedScanTask(
1081
0
                                    [&, idx, scan_info]() -> bool {
1082
0
                                        Defer complete_task {[&]() {
1083
0
                                            semaphore.release();
1084
0
                                            if (++producer_count == scan_rows.size()) {
1085
0
                                                std::lock_guard<std::mutex> lock(mtx);
1086
0
                                                cv.notify_one();
1087
0
                                            }
1088
0
                                        }};
1089
0
                                        const auto& [row_ids, file_mapping] = scan_info;
1090
0
                                        auto status = read_external_row_from_file_mapping(
1091
0
                                                idx, row_ids, file_mapping, slots, query_id,
1092
0
                                                runtime_state, scan_blocks, row_id_block_idx,
1093
0
                                                fetch_statistics, rpc_scan_params,
1094
0
                                                colname_to_slot_id, tuple_desc);
1095
0
                                        if (!status.ok()) {
1096
0
                                            std::lock_guard<std::mutex> lock(mtx);
1097
0
                                            if (scan_status.ok()) {
1098
0
                                                scan_status = status;
1099
0
                                            }
1100
0
                                        }
1101
                                        // The return value indicates whether this one-shot
1102
                                        // scheduler task has completed, not whether the fetch
1103
                                        // succeeded. The fetch status is propagated by scan_status.
1104
0
                                        return true;
1105
0
                                    },
1106
0
                                    nullptr, nullptr),
1107
0
                            fmt::format("{}-read_batch_external_row-{}", print_id(query_id), idx)));
1108
0
                    idx++;
1109
0
                }
1110
1111
0
                {
1112
0
                    std::unique_lock<std::mutex> lock(mtx);
1113
0
                    cv.wait(lock, [&] { return producer_count == scan_rows.size(); });
1114
0
                }
1115
0
                RETURN_IF_ERROR(scan_status);
1116
0
                return Status::OK();
1117
0
            },
1118
0
            &scan_running_time));
1119
1120
    // Insert the read data into result_block.
1121
0
    for (size_t column_id = 0; column_id < result_block.columns(); column_id++) {
1122
0
        auto dst_col_guard = result_block.mutate_column_scoped(column_id);
1123
0
        MutableColumnPtr& dst_col = dst_col_guard.mutable_column();
1124
1125
0
        std::vector<const IColumn*> scan_src_columns;
1126
0
        scan_src_columns.reserve(row_id_block_idx.size());
1127
0
        std::vector<size_t> scan_positions;
1128
0
        scan_positions.reserve(row_id_block_idx.size());
1129
0
        for (const auto& [pos_block, block_idx] : row_id_block_idx) {
1130
0
            DCHECK(scan_blocks.size() > pos_block);
1131
0
            DCHECK(scan_blocks[pos_block].get_columns().size() > column_id);
1132
0
            scan_src_columns.emplace_back(scan_blocks[pos_block].get_columns()[column_id].get());
1133
0
            scan_positions.emplace_back(block_idx);
1134
0
        }
1135
0
        dst_col->insert_from_multi_column(scan_src_columns, scan_positions);
1136
0
    }
1137
1138
    // Statistical runtime profile information.
1139
0
    std::unique_ptr<RuntimeProfile> runtime_profile =
1140
0
            std::make_unique<RuntimeProfile>("ExternalRowIDFetcher");
1141
0
    {
1142
0
        runtime_profile->add_info_string(ScannersRunningTimeProfile,
1143
0
                                         std::to_string(scan_running_time) + "ms");
1144
0
        fmt::memory_buffer file_read_lines_buffer;
1145
0
        format_to(file_read_lines_buffer, "[");
1146
0
        fmt::memory_buffer file_read_bytes_buffer;
1147
0
        format_to(file_read_bytes_buffer, "[");
1148
0
        fmt::memory_buffer file_read_times_buffer;
1149
0
        format_to(file_read_times_buffer, "[");
1150
1151
0
        std::map<std::string, int64_t> lance_fetch_times_ns;
1152
0
        size_t idx = 0;
1153
0
        for (const auto& [_, scan_info] : scan_rows) {
1154
0
            format_to(file_read_lines_buffer, "{}, ", scan_info.first.size());
1155
0
            *init_reader_avg_ms += fetch_statistics[idx].init_reader_ms;
1156
0
            *get_block_avg_ms += fetch_statistics[idx].get_block_ms;
1157
0
            format_to(file_read_bytes_buffer, "{}, ", fetch_statistics[idx].file_read_bytes);
1158
0
            format_to(file_read_times_buffer, "{}, ", fetch_statistics[idx].file_read_times);
1159
0
            for (const auto& [time_name, time_value] : fetch_statistics[idx].lance_fetch_times_ns) {
1160
0
                lance_fetch_times_ns[time_name] += time_value;
1161
0
            }
1162
0
            idx++;
1163
0
        }
1164
1165
0
        format_to(file_read_lines_buffer, "]");
1166
0
        format_to(file_read_bytes_buffer, "]");
1167
0
        format_to(file_read_times_buffer, "]");
1168
1169
0
        *init_reader_avg_ms /= fetch_statistics.size();
1170
0
        *get_block_avg_ms /= fetch_statistics.size();
1171
0
        runtime_profile->add_info_string(InitReaderAvgTimeProfile,
1172
0
                                         std::to_string(*init_reader_avg_ms) + "ms");
1173
0
        runtime_profile->add_info_string(GetBlockAvgTimeProfile,
1174
0
                                         std::to_string(*get_block_avg_ms) + "ms");
1175
0
        runtime_profile->add_info_string(FileReadLinesProfile,
1176
0
                                         fmt::to_string(file_read_lines_buffer));
1177
0
        runtime_profile->add_info_string(FileScanner::FileReadBytesProfile,
1178
0
                                         fmt::to_string(file_read_bytes_buffer));
1179
0
        runtime_profile->add_info_string(FileScanner::FileReadTimeProfile,
1180
0
                                         fmt::to_string(file_read_times_buffer));
1181
0
        for (const auto& [time_name, time_value] : lance_fetch_times_ns) {
1182
0
            runtime_profile->add_info_string(time_name,
1183
0
                                             PrettyPrinter::print(time_value, TUnit::TIME_NS));
1184
0
        }
1185
0
    }
1186
1187
0
    runtime_profile->to_proto(pprofile, 2);
1188
1189
0
    *scan_range_cnt = scan_rows.size();
1190
1191
0
    return Status::OK();
1192
0
}
1193
1194
Status RowIdStorageReader::read_doris_format_row(
1195
        const std::shared_ptr<IdFileMap>& id_file_map,
1196
        const std::shared_ptr<FileMapping>& file_mapping, const std::vector<uint32_t>& row_ids,
1197
        std::vector<SlotDescriptor>& slots, const TabletSchema& full_read_schema,
1198
        RowStoreReadStruct& row_store_read_struct, OlapReaderStatistics& stats,
1199
        int64_t* acquire_tablet_ms, int64_t* acquire_rowsets_ms, int64_t* acquire_segments_ms,
1200
        int64_t* lookup_row_data_ms, std::unordered_map<SegKey, SegItem, HashOfSegKey>& seg_map,
1201
        std::unordered_map<IteratorKey, IteratorItem, HashOfIteratorKey>& iterator_map,
1202
0
        io::FileCacheMissPolicy file_cache_miss_policy, Block& result_block) {
1203
0
    auto [tablet_id, rowset_id, segment_id] = file_mapping->get_doris_format_info();
1204
0
    SegKey seg_key {.tablet_id = tablet_id, .rowset_id = rowset_id, .segment_id = segment_id};
1205
1206
0
    BaseTabletSPtr tablet;
1207
0
    BetaRowsetSharedPtr rowset;
1208
0
    SegmentSharedPtr segment;
1209
0
    if (seg_map.find(seg_key) == seg_map.end()) {
1210
0
        tablet = scope_timer_run(
1211
0
                [&]() {
1212
0
                    auto res = ExecEnv::get_tablet(tablet_id);
1213
0
                    return !res.has_value() ? nullptr
1214
0
                                            : std::dynamic_pointer_cast<BaseTablet>(res.value());
1215
0
                },
1216
0
                acquire_tablet_ms);
1217
0
        if (!tablet) {
1218
0
            return Status::InternalError(
1219
0
                    "Backend:{} tablet not found, tablet_id: {}, rowset_id: {}, segment_id: {}, "
1220
0
                    "row_id: {}",
1221
0
                    BackendOptions::get_localhost(), tablet_id, rowset_id.to_string(), segment_id,
1222
0
                    row_ids[0]);
1223
0
        }
1224
1225
0
        rowset = std::static_pointer_cast<BetaRowset>(scope_timer_run(
1226
0
                [&]() { return id_file_map->get_temp_rowset(tablet_id, rowset_id); },
1227
0
                acquire_rowsets_ms));
1228
0
        if (!rowset) {
1229
0
            return Status::InternalError(
1230
0
                    "Backend:{} rowset_id not found, tablet_id: {}, rowset_id: {}, segment_id: {}, "
1231
0
                    "row_id: {}",
1232
0
                    BackendOptions::get_localhost(), tablet_id, rowset_id.to_string(), segment_id,
1233
0
                    row_ids[0]);
1234
0
        }
1235
1236
0
        SegmentCacheHandle segment_cache;
1237
0
        RETURN_IF_ERROR(scope_timer_run(
1238
0
                [&]() {
1239
0
                    return SegmentLoader::instance()->load_segments(rowset, &segment_cache, true);
1240
0
                },
1241
0
                acquire_segments_ms));
1242
1243
0
        auto it = std::find_if(segment_cache.get_segments().cbegin(),
1244
0
                               segment_cache.get_segments().cend(),
1245
0
                               [segment_id](const segment_v2::SegmentSharedPtr& seg) {
1246
0
                                   return seg->id() == segment_id;
1247
0
                               });
1248
0
        if (it == segment_cache.get_segments().end()) {
1249
0
            return Status::InternalError(
1250
0
                    "Backend:{} segment not found, tablet_id: {}, rowset_id: {}, segment_id: {}, "
1251
0
                    "row_id: {}",
1252
0
                    BackendOptions::get_localhost(), tablet_id, rowset_id.to_string(), segment_id,
1253
0
                    row_ids[0]);
1254
0
        }
1255
0
        segment = *it;
1256
0
        seg_map[seg_key] = SegItem {.tablet = tablet, .rowset = rowset, .segment = segment};
1257
0
    } else {
1258
0
        auto& seg_item = seg_map[seg_key];
1259
0
        tablet = seg_item.tablet;
1260
0
        rowset = seg_item.rowset;
1261
0
        segment = seg_item.segment;
1262
0
    }
1263
1264
    // if row_store_read_struct not empty, means the line we should read from row_store
1265
0
    if (!row_store_read_struct.default_values.empty()) {
1266
0
        if (!tablet->tablet_schema()->has_row_store_for_all_columns()) {
1267
0
            return Status::InternalError("Tablet {} does not have row store for all columns",
1268
0
                                         tablet->tablet_id());
1269
0
        }
1270
0
        auto result_columns_guard = result_block.mutate_columns_scoped();
1271
0
        MutableColumns& result_columns = result_columns_guard.mutable_columns();
1272
0
        io::IOContext io_ctx;
1273
0
        io_ctx.reader_type = ReaderType::READER_QUERY;
1274
0
        io_ctx.file_cache_stats = &stats.file_cache_stats;
1275
0
        io_ctx.file_cache_miss_policy = file_cache_miss_policy;
1276
0
        for (auto row_id : row_ids) {
1277
0
            RowLocation loc(rowset_id, segment->id(), cast_set<uint32_t>(row_id));
1278
0
            row_store_read_struct.row_store_buffer.clear();
1279
0
            RETURN_IF_ERROR(scope_timer_run(
1280
0
                    [&]() {
1281
0
                        return tablet->lookup_row_data({}, loc, rowset, stats,
1282
0
                                                       row_store_read_struct.row_store_buffer,
1283
0
                                                       false, &io_ctx);
1284
0
                    },
1285
0
                    lookup_row_data_ms));
1286
1287
0
            RETURN_IF_ERROR(JsonbSerializeUtil::jsonb_to_columns(
1288
0
                    row_store_read_struct.serdes, row_store_read_struct.row_store_buffer.data(),
1289
0
                    row_store_read_struct.row_store_buffer.size(),
1290
0
                    row_store_read_struct.col_uid_to_idx, result_columns,
1291
0
                    row_store_read_struct.default_values, {}));
1292
0
        }
1293
0
    } else {
1294
0
        for (int x = 0; x < slots.size(); ++x) {
1295
0
            auto column_guard = result_block.mutate_column_scoped(x);
1296
0
            MutableColumnPtr& column = column_guard.mutable_column();
1297
0
            IteratorKey iterator_key {.tablet_id = tablet_id,
1298
0
                                      .rowset_id = rowset_id,
1299
0
                                      .segment_id = segment_id,
1300
0
                                      .slot_id = slots[x].id()};
1301
0
            IteratorItem& iterator_item = iterator_map[iterator_key];
1302
0
            if (iterator_item.segment == nullptr) {
1303
0
                iterator_map[iterator_key].segment = segment;
1304
0
                iterator_item.storage_read_options.stats = &stats;
1305
0
                iterator_item.storage_read_options.io_ctx.reader_type = ReaderType::READER_QUERY;
1306
0
                iterator_item.storage_read_options.io_ctx.file_cache_miss_policy =
1307
0
                        file_cache_miss_policy;
1308
0
            }
1309
0
            set_slot_access_paths(slots[x], full_read_schema, iterator_item.storage_read_options);
1310
0
            for (auto row_id : row_ids) {
1311
0
                RETURN_IF_ERROR(segment->seek_and_read_by_rowid(
1312
0
                        full_read_schema, &slots[x], row_id, column,
1313
0
                        iterator_item.storage_read_options, iterator_item.iterator));
1314
0
            }
1315
0
        }
1316
0
    }
1317
0
    return Status::OK();
1318
0
}
1319
1320
#include "common/compile_check_end.h"
1321
1322
} // namespace doris