Coverage Report

Created: 2026-09-20 20:45

next uncovered line (L), next uncovered region (R), next uncovered branch (B)
be/src/service/point_query_executor.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 "service/point_query_executor.h"
19
20
#include <fmt/format.h>
21
#include <gen_cpp/Descriptors_types.h>
22
#include <gen_cpp/Exprs_types.h>
23
#include <gen_cpp/PaloInternalService_types.h>
24
#include <gen_cpp/internal_service.pb.h>
25
#include <glog/logging.h>
26
#include <google/protobuf/extension_set.h>
27
#include <stdlib.h>
28
29
#include <memory>
30
#include <unordered_map>
31
#include <vector>
32
33
#include "cloud/cloud_tablet.h"
34
#include "cloud/config.h"
35
#include "common/cast_set.h"
36
#include "common/consts.h"
37
#include "common/status.h"
38
#include "core/data_type/data_type_factory.hpp"
39
#include "core/data_type_serde/data_type_serde.h"
40
#include "exec/common/variant_util.h"
41
#include "exec/sink/writer/vmysql_result_writer.h"
42
#include "exprs/vexpr.h"
43
#include "exprs/vexpr_context.h"
44
#include "exprs/vexpr_fwd.h"
45
#include "exprs/vslot_ref.h"
46
#include "io/cache/remote_scan_cache_write_limiter.h"
47
#include "io/io_common.h"
48
#include "runtime/descriptors.h"
49
#include "runtime/exec_env.h"
50
#include "runtime/result_block_buffer.h"
51
#include "runtime/runtime_profile.h"
52
#include "runtime/runtime_state.h"
53
#include "runtime/thread_context.h"
54
#include "storage/row_cursor.h"
55
#include "storage/rowset/beta_rowset.h"
56
#include "storage/rowset/rowset_fwd.h"
57
#include "storage/segment/column_reader.h"
58
#include "storage/tablet/tablet_schema.h"
59
#include "storage/utils.h"
60
#include "util/defer_op.h"
61
#include "util/jsonb/serialize.h"
62
#include "util/lru_cache.h"
63
#include "util/simd/bits.h"
64
#include "util/thrift_util.h"
65
66
namespace doris {
67
68
class PointQueryResultBlockBuffer final : public MySQLResultBlockBuffer {
69
public:
70
0
    PointQueryResultBlockBuffer(RuntimeState* state) : MySQLResultBlockBuffer(state) {}
71
0
    ~PointQueryResultBlockBuffer() override = default;
72
0
    std::shared_ptr<TFetchDataResult> get_block() {
73
0
        std::lock_guard<std::mutex> l(_lock);
74
0
        DCHECK_EQ(_result_batch_queue.size(), 1);
75
0
        auto result = std::move(_result_batch_queue.front());
76
0
        _result_batch_queue.pop_front();
77
0
        return result;
78
0
    }
79
};
80
81
4.09k
Reusable::~Reusable() = default;
82
83
// get missing and include column ids
84
// input include_cids : the output expr slots columns unique ids
85
// missing_cids : the output expr columns that not in row columns cids
86
static void get_missing_and_include_cids(const TabletSchema& schema,
87
                                         const std::vector<SlotDescriptor*>& slots,
88
                                         int target_rs_column_id,
89
                                         std::unordered_set<int>& missing_cids,
90
4.04k
                                         std::unordered_set<int>& include_cids) {
91
4.04k
    missing_cids.clear();
92
4.04k
    include_cids.clear();
93
4.04k
    for (auto* slot : slots) {
94
4.04k
        missing_cids.insert(slot->col_unique_id());
95
4.04k
    }
96
    // insert delete sign column id
97
4.04k
    missing_cids.insert(schema.columns()[schema.delete_sign_idx()]->unique_id());
98
4.04k
    if (target_rs_column_id == -1) {
99
        // no row store columns
100
4.04k
        return;
101
4.04k
    }
102
1
    const TabletColumn& target_rs_column = schema.column_by_uid(target_rs_column_id);
103
1
    DCHECK(target_rs_column.is_row_store_column());
104
    // The full column group is considered a full match, thus no missing cids
105
1
    if (schema.row_columns_uids().empty()) {
106
0
        missing_cids.clear();
107
0
        return;
108
0
    }
109
1
    for (int cid : schema.row_columns_uids()) {
110
0
        missing_cids.erase(cid);
111
0
        include_cids.insert(cid);
112
0
    }
113
1
}
114
115
constexpr static int s_preallocted_blocks_num = 32;
116
117
static void extract_slot_ref(const VExprSPtr& expr, TupleDescriptor* tuple_desc,
118
4.04k
                             std::vector<SlotDescriptor*>& slots) {
119
4.04k
    const auto& children = expr->children();
120
4.04k
    for (const auto& i : children) {
121
0
        extract_slot_ref(i, tuple_desc, slots);
122
0
    }
123
124
4.04k
    auto node_type = expr->node_type();
125
4.04k
    if (node_type == TExprNodeType::SLOT_REF) {
126
4.04k
        int column_id = static_cast<const VSlotRef*>(expr.get())->column_id();
127
4.04k
        auto* slot_desc = tuple_desc->slots()[column_id];
128
4.04k
        slots.push_back(slot_desc);
129
4.04k
    }
130
4.04k
}
131
132
Status Reusable::init(const TDescriptorTable& t_desc_tbl, const std::vector<TExpr>& output_exprs,
133
                      const TQueryOptions& query_options, const TabletSchema& schema,
134
4.04k
                      size_t block_size) {
135
4.04k
    _runtime_state = RuntimeState::create_unique();
136
4.04k
    _runtime_state->set_query_options(query_options);
137
4.04k
    RETURN_IF_ERROR(DescriptorTbl::create(_runtime_state->obj_pool(), t_desc_tbl, &_desc_tbl));
138
4.04k
    _runtime_state->set_desc_tbl(_desc_tbl);
139
89.0k
    for (const auto* slot : tuple_desc()->slots()) {
140
89.0k
        if (!slot->all_access_paths().empty() || !slot->predicate_access_paths().empty()) {
141
0
            return Status::InternalError(
142
0
                    "Short-circuit point query does not support nested column access paths, "
143
0
                    "slot: {}. Please upgrade FE to disable nested column pruning for "
144
0
                    "short-circuit point queries.",
145
0
                    slot->col_name());
146
0
        }
147
89.0k
    }
148
4.04k
    _block_pool.resize(block_size);
149
8.09k
    for (auto& i : _block_pool) {
150
8.09k
        i = Block::create_unique(tuple_desc()->slots(), 2);
151
        // Name is useless but cost space
152
8.09k
        i->clear_names();
153
8.09k
    }
154
155
4.04k
    RETURN_IF_ERROR(VExpr::create_expr_trees(output_exprs, _output_exprs_ctxs));
156
4.04k
    RowDescriptor row_desc(tuple_desc());
157
    // Prepare the exprs to run.
158
4.04k
    RETURN_IF_ERROR(VExpr::prepare(_output_exprs_ctxs, _runtime_state.get(), row_desc));
159
4.04k
    RETURN_IF_ERROR(VExpr::open(_output_exprs_ctxs, _runtime_state.get()));
160
4.04k
    _create_timestamp = butil::gettimeofday_ms();
161
4.04k
    _data_type_serdes = create_data_type_serdes(tuple_desc()->slots());
162
4.04k
    _col_default_values.resize(tuple_desc()->slots().size());
163
4.04k
    bool has_delete_sign = false;
164
93.0k
    for (int i = 0; i < tuple_desc()->slots().size(); ++i) {
165
89.0k
        auto* slot = tuple_desc()->slots()[i];
166
89.0k
        if (slot->col_name() == DELETE_SIGN) {
167
0
            has_delete_sign = true;
168
0
        }
169
89.0k
        _col_uid_to_idx[slot->col_unique_id()] = i;
170
89.0k
        _col_default_values[i] = slot->col_default_value();
171
89.0k
    }
172
173
    // Get the output slot descriptors
174
4.04k
    std::vector<SlotDescriptor*> output_slot_descs;
175
4.04k
    for (const auto& expr : _output_exprs_ctxs) {
176
4.04k
        extract_slot_ref(expr->root(), tuple_desc(), output_slot_descs);
177
4.04k
    }
178
179
    // get the delete sign idx in block
180
4.04k
    if (has_delete_sign) {
181
0
        _delete_sign_idx = _col_uid_to_idx[schema.columns()[schema.delete_sign_idx()]->unique_id()];
182
0
    }
183
184
4.04k
    if (schema.have_column(BeConsts::ROW_STORE_COL)) {
185
0
        const auto& column = *DORIS_TRY(schema.column(BeConsts::ROW_STORE_COL));
186
0
        _row_store_column_ids = column.unique_id();
187
0
    }
188
4.04k
    get_missing_and_include_cids(schema, output_slot_descs, _row_store_column_ids,
189
4.04k
                                 _missing_col_uids, _include_col_uids);
190
191
4.04k
    return Status::OK();
192
4.04k
}
193
194
0
std::unique_ptr<Block> Reusable::get_block() {
195
0
    std::lock_guard lock(_block_mutex);
196
0
    if (_block_pool.empty()) {
197
0
        auto block = Block::create_unique(tuple_desc()->slots(), 2);
198
        // Name is useless but cost space
199
0
        block->clear_names();
200
0
        return block;
201
0
    }
202
0
    auto block = std::move(_block_pool.back());
203
0
    CHECK(block != nullptr);
204
0
    _block_pool.pop_back();
205
0
    return block;
206
0
}
207
208
0
void Reusable::return_block(std::unique_ptr<Block>& block) {
209
0
    std::lock_guard lock(_block_mutex);
210
0
    if (block == nullptr) {
211
0
        return;
212
0
    }
213
0
    block->clear_column_data();
214
0
    _block_pool.push_back(std::move(block));
215
0
    if (_block_pool.size() > s_preallocted_blocks_num) {
216
0
        _block_pool.resize(s_preallocted_blocks_num);
217
0
    }
218
0
}
219
220
0
LookupConnectionCache* LookupConnectionCache::create_global_instance(size_t capacity) {
221
0
    DCHECK(ExecEnv::GetInstance()->get_lookup_connection_cache() == nullptr);
222
0
    auto* res = new LookupConnectionCache(capacity);
223
0
    return res;
224
0
}
225
226
RowCache::RowCache(int64_t capacity, int num_shards)
227
5
        : LRUCachePolicy(CachePolicy::CacheType::POINT_QUERY_ROW_CACHE, capacity,
228
5
                         LRUCacheType::SIZE, config::point_query_row_cache_stale_sweep_time_sec,
229
5
                         num_shards, /*element count capacity */ 0,
230
5
                         /*enable prune*/ true, /*is lru-k*/ true) {}
231
232
// Create global instance of this class
233
0
RowCache* RowCache::create_global_cache(int64_t capacity, uint32_t num_shards) {
234
0
    DCHECK(ExecEnv::GetInstance()->get_row_cache() == nullptr);
235
0
    auto* res = new RowCache(capacity, num_shards);
236
0
    return res;
237
0
}
238
239
12
RowCache* RowCache::instance() {
240
12
    return ExecEnv::GetInstance()->get_row_cache();
241
12
}
242
243
10
bool RowCache::lookup(const RowCacheKey& key, CacheHandle* handle) {
244
10
    const std::string& encoded_key = key.encode();
245
10
    auto* lru_handle = LRUCachePolicy::lookup(encoded_key);
246
10
    if (!lru_handle) {
247
        // cache miss
248
5
        return false;
249
5
    }
250
5
    *handle = CacheHandle(this, lru_handle);
251
5
    return true;
252
10
}
253
254
15
void RowCache::insert(const RowCacheKey& key, const Slice& value) {
255
15
    char* cache_value = static_cast<char*>(malloc(value.size));
256
15
    memcpy(cache_value, value.data, value.size);
257
15
    auto* row_cache_value = new RowCacheValue;
258
15
    row_cache_value->cache_value = cache_value;
259
15
    const std::string& encoded_key = key.encode();
260
15
    auto* handle = LRUCachePolicy::insert(encoded_key, row_cache_value, value.size, value.size,
261
15
                                          CachePriority::NORMAL);
262
    // handle will released
263
15
    auto tmp = CacheHandle {this, handle};
264
15
}
265
266
3
void RowCache::erase(const RowCacheKey& key) {
267
3
    const std::string& encoded_key = key.encode();
268
3
    LRUCachePolicy::erase(encoded_key);
269
3
}
270
271
2.09k
LookupConnectionCache::CacheValue::~CacheValue() {
272
2.09k
    SCOPED_SWITCH_THREAD_MEM_TRACKER_LIMITER(
273
2.09k
            ExecEnv::GetInstance()->point_query_executor_mem_tracker());
274
2.09k
    item.reset();
275
2.09k
}
276
277
0
PointQueryExecutor::~PointQueryExecutor() {
278
0
    SCOPED_SWITCH_THREAD_MEM_TRACKER_LIMITER(
279
0
            ExecEnv::GetInstance()->point_query_executor_mem_tracker());
280
0
    _tablet.reset();
281
0
    _reusable.reset();
282
0
    _result_block.reset();
283
0
    _row_read_ctxs.clear();
284
0
}
285
286
0
void PointQueryExecutor::_init_remote_scan_cache_write_limiter() {
287
0
    const auto& query_options = _reusable->runtime_state()->query_options();
288
0
    const bool initialize_remote_scan_cache_write_limiter =
289
0
            config::is_cloud_mode() && config::enable_file_cache &&
290
0
            query_options.__isset.file_cache_query_limit_bytes &&
291
0
            query_options.file_cache_query_limit_bytes >= 0 &&
292
0
            query_options.query_type == TQueryType::SELECT;
293
0
    if (initialize_remote_scan_cache_write_limiter) {
294
0
        _remote_scan_cache_write_limiter = std::make_unique<io::RemoteScanCacheWriteLimiter>(
295
0
                _reusable->runtime_state()->query_id(), query_options.file_cache_query_limit_bytes);
296
0
    }
297
0
}
298
299
Status PointQueryExecutor::init(const PTabletKeyLookupRequest* request,
300
0
                                PTabletKeyLookupResponse* response) {
301
0
    SCOPED_TIMER(&_profile_metrics.init_ns);
302
0
    _response = response;
303
    // using cache
304
0
    __int128_t uuid =
305
0
            static_cast<__int128_t>(request->uuid().uuid_high()) << 64 | request->uuid().uuid_low();
306
0
    SCOPED_ATTACH_TASK(ExecEnv::GetInstance()->point_query_executor_mem_tracker());
307
0
    auto cache_handle = LookupConnectionCache::instance()->get(uuid);
308
0
    _binary_row_format = request->is_binary_row();
309
0
    _tablet = DORIS_TRY(ExecEnv::get_tablet(request->tablet_id()));
310
0
    if (cache_handle != nullptr) {
311
0
        _reusable = cache_handle;
312
0
        _profile_metrics.hit_lookup_cache = true;
313
0
    } else {
314
        // Lightweight request: FE may omit reusable query context and rely on uuid cache.
315
        // If cache miss and required parameters are absent, ask FE to resend a full request.
316
0
        if (uuid != 0 && (!request->has_desc_tbl() || request->desc_tbl().empty() ||
317
0
                          !request->has_output_expr() || request->output_expr().empty() ||
318
0
                          !request->has_query_options() || request->query_options().empty())) {
319
0
            if (VLOG_DEBUG_IS_ON) {
320
0
                VLOG_DEBUG << "lookup connection cache miss, ask FE to resend query context"
321
0
                           << ", tablet_id=" << request->tablet_id()
322
0
                           << ", uuid_high=" << request->uuid().uuid_high()
323
0
                           << ", uuid_low=" << request->uuid().uuid_low();
324
0
            }
325
0
            response->set_need_resend_query_context(true);
326
0
            return Status::OK();
327
0
        }
328
0
        if (uuid == 0 && (!request->has_desc_tbl() || request->desc_tbl().empty() ||
329
0
                          !request->has_output_expr() || request->output_expr().empty())) {
330
0
            return Status::InvalidArgument(
331
0
                    "tablet_fetch_data requires desc_tbl/output_expr when uuid is not set");
332
0
        }
333
        // init handle
334
0
        auto reusable_ptr = std::make_shared<Reusable>();
335
0
        TDescriptorTable t_desc_tbl;
336
0
        TExprList t_output_exprs;
337
0
        auto len = cast_set<uint32_t>(request->desc_tbl().size());
338
0
        RETURN_IF_ERROR(
339
0
                deserialize_thrift_msg(reinterpret_cast<const uint8_t*>(request->desc_tbl().data()),
340
0
                                       &len, false, &t_desc_tbl));
341
0
        len = cast_set<uint32_t>(request->output_expr().size());
342
0
        RETURN_IF_ERROR(deserialize_thrift_msg(
343
0
                reinterpret_cast<const uint8_t*>(request->output_expr().data()), &len, false,
344
0
                &t_output_exprs));
345
0
        _reusable = reusable_ptr;
346
0
        TQueryOptions t_query_options;
347
0
        len = cast_set<uint32_t>(request->query_options().size());
348
0
        if (request->has_query_options()) {
349
0
            RETURN_IF_ERROR(deserialize_thrift_msg(
350
0
                    reinterpret_cast<const uint8_t*>(request->query_options().data()), &len, false,
351
0
                    &t_query_options));
352
0
        }
353
0
        if (uuid != 0) {
354
            // could be reused by requests after, pre allocte more blocks
355
0
            RETURN_IF_ERROR(reusable_ptr->init(t_desc_tbl, t_output_exprs.exprs, t_query_options,
356
0
                                               *_tablet->tablet_schema(),
357
0
                                               s_preallocted_blocks_num));
358
0
            LookupConnectionCache::instance()->add(uuid, reusable_ptr);
359
0
        } else {
360
0
            RETURN_IF_ERROR(reusable_ptr->init(t_desc_tbl, t_output_exprs.exprs, t_query_options,
361
0
                                               *_tablet->tablet_schema(), 1));
362
0
        }
363
0
    }
364
0
    _init_remote_scan_cache_write_limiter();
365
    // Set timezone from request for functions like from_unixtime()
366
0
    if (request->has_time_zone() && !request->time_zone().empty()) {
367
0
        _reusable->runtime_state()->set_timezone(request->time_zone());
368
0
    }
369
0
    if (request->has_version() && request->version() >= 0) {
370
0
        _version = request->version();
371
0
    }
372
0
    RETURN_IF_ERROR(_init_keys(request));
373
0
    _result_block = _reusable->get_block();
374
0
    CHECK(_result_block != nullptr);
375
376
0
    return Status::OK();
377
0
}
378
379
0
Status PointQueryExecutor::lookup_up() {
380
0
    SCOPED_ATTACH_TASK(ExecEnv::GetInstance()->point_query_executor_mem_tracker());
381
0
    RETURN_IF_ERROR(_lookup_row_key());
382
0
    RETURN_IF_ERROR(_lookup_row_data());
383
0
    RETURN_IF_ERROR(_output_data());
384
0
    return Status::OK();
385
0
}
386
387
0
void PointQueryExecutor::print_profile() {
388
0
    auto init_us = _profile_metrics.init_ns.value() / 1000;
389
0
    auto init_key_us = _profile_metrics.init_key_ns.value() / 1000;
390
0
    auto lookup_key_us = _profile_metrics.lookup_key_ns.value() / 1000;
391
0
    auto lookup_data_us = _profile_metrics.lookup_data_ns.value() / 1000;
392
0
    auto output_data_us = _profile_metrics.output_data_ns.value() / 1000;
393
0
    auto load_segments_key_us = _profile_metrics.load_segment_key_stage_ns.value() / 1000;
394
0
    auto load_segments_data_us = _profile_metrics.load_segment_data_stage_ns.value() / 1000;
395
0
    auto total_us = init_us + lookup_key_us + lookup_data_us + output_data_us;
396
0
    auto read_stats = _profile_metrics.read_stats;
397
0
    const std::string stats_str = fmt::format(
398
0
            "[lookup profile:{}us] init:{}us, init_key:{}us,"
399
0
            " lookup_key:{}us, load_segments_key:{}us, lookup_data:{}us, load_segments_data:{}us,"
400
0
            " output_data:{}us, "
401
0
            "hit_lookup_cache:{}"
402
0
            ", is_binary_row:{}, output_columns:{}, total_keys:{}, row_cache_hits:{}"
403
0
            ", hit_cached_pages:{}, total_pages_read:{}, compressed_bytes_read:{}, "
404
0
            "io_latency:{}ns, "
405
0
            "uncompressed_bytes_read:{}, result_data_bytes:{}, row_hits:{}"
406
0
            ", rs_column_uid:{}, bytes_read_from_local:{}, bytes_read_from_remote:{}, "
407
0
            "local_io_timer:{}, remote_io_timer:{}, local_write_timer:{}",
408
0
            total_us, init_us, init_key_us, lookup_key_us, load_segments_key_us, lookup_data_us,
409
0
            load_segments_data_us, output_data_us, _profile_metrics.hit_lookup_cache,
410
0
            _binary_row_format, _reusable->output_exprs().size(), _row_read_ctxs.size(),
411
0
            _profile_metrics.row_cache_hits, read_stats.cached_pages_num,
412
0
            read_stats.total_pages_num, read_stats.compressed_bytes_read, read_stats.io_ns,
413
0
            read_stats.uncompressed_bytes_read, _profile_metrics.result_data_bytes, _row_hits,
414
0
            _reusable->rs_column_uid(),
415
0
            _profile_metrics.read_stats.file_cache_stats.bytes_read_from_local,
416
0
            _profile_metrics.read_stats.file_cache_stats.bytes_read_from_remote,
417
0
            _profile_metrics.read_stats.file_cache_stats.local_io_timer,
418
0
            _profile_metrics.read_stats.file_cache_stats.remote_io_timer,
419
0
            _profile_metrics.read_stats.file_cache_stats.write_cache_io_timer);
420
421
0
    constexpr static int kSlowThreholdUs = 50 * 1000; // 50ms
422
0
    if (total_us > kSlowThreholdUs) {
423
0
        LOG(WARNING) << "slow query, " << stats_str;
424
0
    } else if (VLOG_DEBUG_IS_ON) {
425
0
        VLOG_DEBUG << stats_str;
426
0
    } else {
427
0
        LOG_EVERY_N(INFO, 1000) << stats_str;
428
0
    }
429
0
}
430
431
0
Status PointQueryExecutor::_init_keys(const PTabletKeyLookupRequest* request) {
432
0
    SCOPED_TIMER(&_profile_metrics.init_key_ns);
433
0
    const auto& schema = _tablet->tablet_schema();
434
    // Point query is only supported on merge-on-write unique key tables.
435
0
    DCHECK(schema->keys_type() == UNIQUE_KEYS && _tablet->enable_unique_key_merge_on_write());
436
0
    if (schema->keys_type() != UNIQUE_KEYS || !_tablet->enable_unique_key_merge_on_write()) {
437
0
        return Status::InvalidArgument(
438
0
                "Point query is only supported on merge-on-write unique key tables, "
439
0
                "tablet_id={}",
440
0
                _tablet->tablet_id());
441
0
    }
442
    // 1. get primary key from conditions
443
0
    _row_read_ctxs.resize(request->key_tuples().size());
444
    // get row cursor and encode keys
445
0
    for (int i = 0; i < request->key_tuples().size(); ++i) {
446
0
        const KeyTuple& key_tuple = request->key_tuples(i);
447
0
        if (UNLIKELY(cast_set<size_t>(key_tuple.key_column_literals_size()) !=
448
0
                     schema->num_key_columns())) {
449
0
            return Status::InvalidArgument(
450
0
                    "Key column count mismatch. expected={}, actual={}, tablet_id={}",
451
0
                    schema->num_key_columns(), key_tuple.key_column_literals_size(),
452
0
                    _tablet->tablet_id());
453
0
        }
454
0
        RowCursor cursor;
455
0
        std::vector<Field> key_fields;
456
0
        key_fields.reserve(key_tuple.key_column_literals_size());
457
0
        for (int j = 0; j < key_tuple.key_column_literals_size(); ++j) {
458
0
            const auto& literal_bytes = key_tuple.key_column_literals(j);
459
0
            TExprNode expr_node;
460
0
            auto len = cast_set<uint32_t>(literal_bytes.size());
461
0
            RETURN_IF_ERROR(
462
0
                    deserialize_thrift_msg(reinterpret_cast<const uint8_t*>(literal_bytes.data()),
463
0
                                           &len, false, &expr_node));
464
0
            const auto& col = schema->column(j);
465
0
            auto data_type = DataTypeFactory::instance().create_data_type(
466
0
                    col.type(), col.precision(), col.frac(), col.length());
467
0
            key_fields.push_back(data_type->get_field(expr_node));
468
0
        }
469
0
        RETURN_IF_ERROR(cursor.init_scan_key(_tablet->tablet_schema(), std::move(key_fields)));
470
0
        cursor.encode_key_with_padding<true>(&_row_read_ctxs[i]._primary_key,
471
0
                                             _tablet->tablet_schema()->num_key_columns(), true);
472
0
    }
473
0
    return Status::OK();
474
0
}
475
476
0
Status PointQueryExecutor::_lookup_row_key() {
477
0
    SCOPED_TIMER(&_profile_metrics.lookup_key_ns);
478
    // 2. lookup row location
479
0
    Status st;
480
0
    if (_version >= 0) {
481
0
        CHECK(config::is_cloud_mode()) << "Only cloud mode support snapshot read at present";
482
0
        SyncOptions options;
483
0
        options.query_version = _version;
484
0
        RETURN_IF_ERROR(std::dynamic_pointer_cast<CloudTablet>(_tablet)->sync_rowsets(options));
485
0
    }
486
0
    std::vector<RowsetSharedPtr> specified_rowsets;
487
0
    {
488
0
        std::shared_lock rlock(_tablet->get_header_lock());
489
0
        specified_rowsets = _tablet->get_rowset_by_ids(nullptr);
490
0
    }
491
0
    io::IOContext io_ctx;
492
0
    io_ctx.reader_type = ReaderType::READER_QUERY;
493
0
    io_ctx.file_cache_stats = &_profile_metrics.read_stats.file_cache_stats;
494
0
    io_ctx.remote_scan_cache_write_limiter = _remote_scan_cache_write_limiter.get();
495
0
    std::vector<std::unique_ptr<SegmentCacheHandle>> segment_caches(specified_rowsets.size());
496
0
    for (size_t i = 0; i < _row_read_ctxs.size(); ++i) {
497
0
        RowLocation location;
498
0
        if (!config::disable_storage_row_cache) {
499
0
            RowCache::CacheHandle cache_handle;
500
0
            auto hit_cache = RowCache::instance()->lookup(
501
0
                    {_tablet->tablet_id(), _row_read_ctxs[i]._primary_key}, &cache_handle);
502
0
            if (hit_cache) {
503
0
                _row_read_ctxs[i]._cached_row_data = std::move(cache_handle);
504
0
                ++_profile_metrics.row_cache_hits;
505
0
                continue;
506
0
            }
507
0
        }
508
        // Get rowlocation and rowset, ctx._rowset_ptr will acquire wrap this ptr
509
0
        auto rowset_ptr = std::make_unique<RowsetSharedPtr>();
510
0
        st = (_tablet->lookup_row_key(_row_read_ctxs[i]._primary_key, nullptr, false,
511
0
                                      specified_rowsets, &location, INT32_MAX /*rethink?*/,
512
0
                                      segment_caches, rowset_ptr.get(), false, nullptr,
513
0
                                      &_profile_metrics.read_stats, nullptr, &io_ctx));
514
0
        if (st.is<ErrorCode::KEY_NOT_FOUND>()) {
515
0
            continue;
516
0
        }
517
0
        RETURN_IF_ERROR(st);
518
0
        _row_read_ctxs[i]._row_location = location;
519
        // acquire and wrap this rowset
520
0
        (*rowset_ptr)->acquire();
521
0
        VLOG_DEBUG << "aquire rowset " << (*rowset_ptr)->rowset_id();
522
0
        _row_read_ctxs[i]._rowset_ptr = std::unique_ptr<RowsetSharedPtr, decltype(&release_rowset)>(
523
0
                rowset_ptr.release(), &release_rowset);
524
0
        _row_hits++;
525
0
    }
526
0
    return Status::OK();
527
0
}
528
529
0
Status PointQueryExecutor::_lookup_row_data() {
530
    // 3. get values
531
0
    SCOPED_TIMER(&_profile_metrics.lookup_data_ns);
532
0
    {
533
0
        auto result_columns_guard = _result_block->mutate_columns_scoped();
534
0
        MutableColumns& result_columns = result_columns_guard.mutable_columns();
535
0
        for (size_t i = 0; i < _row_read_ctxs.size(); ++i) {
536
0
            if (_row_read_ctxs[i]._cached_row_data.valid()) {
537
0
                RETURN_IF_ERROR(JsonbSerializeUtil::jsonb_to_columns(
538
0
                        _reusable->get_data_type_serdes(),
539
0
                        _row_read_ctxs[i]._cached_row_data.data().data,
540
0
                        _row_read_ctxs[i]._cached_row_data.data().size,
541
0
                        _reusable->get_col_uid_to_idx(), result_columns,
542
0
                        _reusable->get_col_default_values(), _reusable->include_col_uids()));
543
0
                continue;
544
0
            }
545
0
            if (!_row_read_ctxs[i]._row_location.has_value()) {
546
0
                continue;
547
0
            }
548
0
            std::string value;
549
            // fill block by row store
550
0
            if (_reusable->rs_column_uid() != -1) {
551
0
                bool use_row_cache = !config::disable_storage_row_cache;
552
0
                io::IOContext io_ctx;
553
0
                io_ctx.reader_type = ReaderType::READER_QUERY;
554
0
                io_ctx.file_cache_stats = &_profile_metrics.read_stats.file_cache_stats;
555
0
                io_ctx.remote_scan_cache_write_limiter = _remote_scan_cache_write_limiter.get();
556
0
                RETURN_IF_ERROR(_tablet->lookup_row_data(
557
0
                        _row_read_ctxs[i]._primary_key, _row_read_ctxs[i]._row_location.value(),
558
0
                        *(_row_read_ctxs[i]._rowset_ptr), _profile_metrics.read_stats, value,
559
0
                        use_row_cache, &io_ctx));
560
                // serialize value to block, currently only jsonb row format
561
0
                RETURN_IF_ERROR(JsonbSerializeUtil::jsonb_to_columns(
562
0
                        _reusable->get_data_type_serdes(), value.data(), value.size(),
563
0
                        _reusable->get_col_uid_to_idx(), result_columns,
564
0
                        _reusable->get_col_default_values(), _reusable->include_col_uids()));
565
0
            }
566
0
            if (!_reusable->missing_col_uids().empty()) {
567
0
                if (!_reusable->runtime_state()->enable_short_circuit_query_access_column_store()) {
568
0
                    std::string missing_columns;
569
0
                    for (int cid : _reusable->missing_col_uids()) {
570
                        // Named from the query, not the tablet schema: a column a light schema
571
                        // change just added is not in the tablet schema yet, and column_by_uid
572
                        // throws on a uid it does not hold.
573
0
                        const int pos = _reusable->get_col_uid_to_idx().at(cid);
574
0
                        missing_columns += _reusable->tuple_desc()->slots()[pos]->col_name() + ",";
575
0
                    }
576
0
                    return Status::InternalError(
577
0
                            "Not support column store, set store_row_column=true or "
578
0
                            "row_store_columns in table properties, missing columns: " +
579
0
                            missing_columns + " should be added to row store");
580
0
                }
581
                // fill missing columns by column store
582
0
                RowLocation row_loc = _row_read_ctxs[i]._row_location.value();
583
0
                BetaRowsetSharedPtr rowset = std::static_pointer_cast<BetaRowset>(
584
0
                        _tablet->get_rowset(row_loc.rowset_id));
585
0
                SegmentCacheHandle segment_cache;
586
0
                io::IOContext io_ctx;
587
0
                io_ctx.reader_type = ReaderType::READER_QUERY;
588
0
                io_ctx.file_cache_stats = &_read_stats.file_cache_stats;
589
0
                io_ctx.remote_scan_cache_write_limiter = _remote_scan_cache_write_limiter.get();
590
0
                {
591
0
                    SCOPED_TIMER(&_profile_metrics.load_segment_data_stage_ns);
592
0
                    RETURN_IF_ERROR(SegmentLoader::instance()->load_segments(
593
0
                            rowset, &segment_cache, true, false, &_read_stats, &io_ctx));
594
0
                }
595
                // find segment
596
0
                auto it = std::find_if(segment_cache.get_segments().cbegin(),
597
0
                                       segment_cache.get_segments().cend(),
598
0
                                       [&](const segment_v2::SegmentSharedPtr& seg) {
599
0
                                           return seg->id() == row_loc.segment_id;
600
0
                                       });
601
0
                const auto& segment = *it;
602
0
                const auto tablet_schema = _tablet->tablet_schema();
603
0
                for (int cid : _reusable->missing_col_uids()) {
604
0
                    int pos = _reusable->get_col_uid_to_idx().at(cid);
605
0
                    std::vector<segment_v2::rowid_t> row_ids {
606
0
                            static_cast<segment_v2::rowid_t>(row_loc.row_id)};
607
0
                    auto& column = result_columns[pos];
608
0
                    std::unique_ptr<ColumnIterator> iter;
609
0
                    SlotDescriptor* slot = _reusable->tuple_desc()->slots()[pos];
610
0
                    int32_t index = slot->col_unique_id() >= 0
611
0
                                            ? tablet_schema->field_index(slot->col_unique_id())
612
0
                                            : tablet_schema->field_index(slot->col_name());
613
                    // A light schema change adds a column without writing anything, and
614
                    // BaseTablet::_max_version_schema only moves forward on write, so until the
615
                    // next load the tablet schema here does not have it. No segment holds it
616
                    // either, so describe it from the slot and let the segment answer with its
617
                    // default -- the same answer it gives once a load does bring the column in.
618
0
                    TabletColumn column_awaiting_a_load;
619
0
                    if (index < 0) {
620
0
                        column_awaiting_a_load = variant_util::get_column_by_type(
621
0
                                slot->type(), slot->col_name(),
622
0
                                variant_util::ExtraInfo {.unique_id = slot->col_unique_id()});
623
0
                        if (!slot->col_default_value().empty()) {
624
0
                            column_awaiting_a_load.set_default_value(slot->col_default_value());
625
0
                        }
626
0
                    }
627
0
                    const TabletColumn& read_column =
628
0
                            index >= 0 ? tablet_schema->column(index) : column_awaiting_a_load;
629
0
                    StorageReadOptions storage_read_options;
630
0
                    storage_read_options.stats = &_read_stats;
631
0
                    storage_read_options.io_ctx = io_ctx;
632
0
                    RETURN_IF_ERROR(segment->seek_and_read_by_rowid(
633
0
                            read_column, slot, row_ids, column, storage_read_options, iter));
634
0
                }
635
0
            }
636
0
        }
637
0
        if (result_columns.size() > _reusable->include_col_uids().size()) {
638
            // Padding rows for some columns that no need to output to mysql client
639
            // eg. SELECT k1,v1,v2 FROM TABLE WHERE k1 = 1, k1 is not in output slots, tuple as bellow
640
            // TupleDescriptor{id=1, tbl=table_with_column_group}
641
            // SlotDescriptor{id=8, col=v1, colUniqueId=1 ...}
642
            // SlotDescriptor{id=9, col=v2, colUniqueId=2 ...}
643
            // thus missing in include_col_uids and missing_col_uids
644
0
            for (auto& column : result_columns) {
645
0
                int padding_rows = _row_hits - cast_set<int>(column->size());
646
0
                if (padding_rows > 0) {
647
0
                    column->insert_many_defaults(padding_rows);
648
0
                }
649
0
            }
650
0
        }
651
0
    }
652
    // filter rows by delete sign
653
0
    if (_row_hits > 0 && _reusable->delete_sign_idx() != -1) {
654
0
        size_t filtered = 0;
655
0
        size_t total = 0;
656
0
        {
657
            // clear_column_data will check reference of ColumnPtr, so we need to release
658
            // reference before clear_column_data
659
0
            ColumnPtr delete_filter_columns =
660
0
                    _result_block->get_columns()[_reusable->delete_sign_idx()];
661
0
            const auto& filter =
662
0
                    assert_cast<const ColumnInt8*>(delete_filter_columns.get())->get_data();
663
0
            filtered = filter.size() - simd::count_zero_num((int8_t*)filter.data(), filter.size());
664
0
            total = filter.size();
665
0
        }
666
667
0
        if (filtered == total) {
668
0
            _result_block->clear_column_data();
669
0
        } else if (filtered > 0) {
670
0
            return Status::NotSupported("Not implemented since only single row at present");
671
0
        }
672
0
    }
673
0
    return Status::OK();
674
0
}
675
676
0
Status serialize_block(std::shared_ptr<TFetchDataResult> res, PTabletKeyLookupResponse* response) {
677
0
    uint8_t* buf = nullptr;
678
0
    uint32_t len = 0;
679
0
    ThriftSerializer ser(false, 4096);
680
0
    RETURN_IF_ERROR(ser.serialize(&(res->result_batch), &len, &buf));
681
0
    response->set_row_batch(std::string((const char*)buf, len));
682
0
    return Status::OK();
683
0
}
684
685
0
Status PointQueryExecutor::_output_data() {
686
    // 4. exprs exec and serialize to mysql row batches
687
0
    SCOPED_TIMER(&_profile_metrics.output_data_ns);
688
0
    if (_result_block->rows()) {
689
0
        RuntimeState state;
690
0
        auto buffer = std::make_shared<PointQueryResultBlockBuffer>(&state);
691
        // TODO reuse mysql_writer
692
0
        VMysqlResultWriter mysql_writer(buffer, _reusable->output_exprs(), nullptr,
693
0
                                        _binary_row_format);
694
0
        RETURN_IF_ERROR(mysql_writer.init(_reusable->runtime_state()));
695
0
        _result_block->clear_names();
696
0
        RETURN_IF_ERROR(mysql_writer.write(_reusable->runtime_state(), *_result_block));
697
0
        RETURN_IF_ERROR(serialize_block(buffer->get_block(), _response));
698
0
        VLOG_DEBUG << "dump block " << _result_block->dump_data();
699
0
    } else {
700
0
        _response->set_empty_batch(true);
701
0
    }
702
0
    _profile_metrics.result_data_bytes = _result_block->bytes();
703
0
    _reusable->return_block(_result_block);
704
0
    return Status::OK();
705
0
}
706
707
} // namespace doris