be/src/storage/tablet/tablet_reader.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 "storage/tablet/tablet_reader.h" |
19 | | |
20 | | #include <gen_cpp/olap_file.pb.h> |
21 | | #include <gen_cpp/segment_v2.pb.h> |
22 | | #include <thrift/protocol/TDebugProtocol.h> |
23 | | |
24 | | #include <algorithm> |
25 | | #include <functional> |
26 | | #include <iterator> |
27 | | #include <memory> |
28 | | #include <numeric> |
29 | | #include <ostream> |
30 | | #include <set> |
31 | | #include <shared_mutex> |
32 | | #include <unordered_map> |
33 | | |
34 | | #include "common/compiler_util.h" // IWYU pragma: keep |
35 | | #include "common/config.h" |
36 | | #include "common/exception.h" |
37 | | #include "common/logging.h" |
38 | | #include "common/status.h" |
39 | | #include "core/arena.h" |
40 | | #include "core/block/block.h" |
41 | | #include "exec/common/variant_util.h" |
42 | | #include "exprs/bloom_filter_func.h" |
43 | | #include "exprs/create_predicate_function.h" |
44 | | #include "exprs/hybrid_set.h" |
45 | | #include "runtime/query_context.h" |
46 | | #include "runtime/runtime_predicate.h" |
47 | | #include "runtime/runtime_state.h" |
48 | | #include "storage/delete/delete_handler.h" |
49 | | #include "storage/index/bloom_filter/bloom_filter.h" |
50 | | #include "storage/itoken_extractor.h" |
51 | | #include "storage/olap_common.h" |
52 | | #include "storage/olap_define.h" |
53 | | #include "storage/predicate/block_column_predicate.h" |
54 | | #include "storage/predicate/column_predicate.h" |
55 | | #include "storage/predicate/like_column_predicate.h" |
56 | | #include "storage/predicate/predicate_creator.h" |
57 | | #include "storage/row_cursor.h" |
58 | | #include "storage/schema.h" |
59 | | #include "storage/tablet/tablet.h" |
60 | | #include "storage/tablet/tablet_meta.h" |
61 | | #include "storage/tablet/tablet_schema.h" |
62 | | |
63 | | namespace doris { |
64 | | using namespace ErrorCode; |
65 | | |
66 | 408 | void TabletReader::ReaderParams::check_validation() const { |
67 | 408 | if (UNLIKELY(version.first == -1 && is_segcompaction == false)) { |
68 | 0 | throw Exception(Status::FatalError("version is not set. tablet={}", tablet->tablet_id())); |
69 | 0 | } |
70 | 408 | } |
71 | | |
72 | 408 | Status TabletReader::init(const ReaderParams& read_params) { |
73 | 408 | Status res = _init_params(read_params); |
74 | 408 | if (!res.ok()) { |
75 | 0 | LOG(WARNING) << "fail to init reader when init params. res:" << res |
76 | 0 | << ", tablet_id:" << read_params.tablet->tablet_id() |
77 | 0 | << ", schema_hash:" << read_params.tablet->schema_hash() |
78 | 0 | << ", reader type:" << int(read_params.reader_type) |
79 | 0 | << ", version:" << read_params.version; |
80 | 0 | } |
81 | 408 | return res; |
82 | 408 | } |
83 | | |
84 | | void TabletReader::remove_delete_columns_from_access_paths( |
85 | | const DeleteHandler& delete_handler, const TabletSchema& tablet_schema, |
86 | 2 | std::map<int32_t, TColumnAccessPaths>& all_access_paths) { |
87 | 2 | auto delete_predicates = AndBlockColumnPredicate::create_shared(); |
88 | 2 | std::unordered_map<int32_t, std::vector<std::shared_ptr<const ColumnPredicate>>> |
89 | 2 | del_predicates_for_zone_map; |
90 | 2 | delete_handler.get_delete_conditions_after_version(0, delete_predicates.get(), |
91 | 2 | &del_predicates_for_zone_map); |
92 | 2 | std::set<ColumnId> delete_column_ids; |
93 | 2 | delete_predicates->get_all_column_ids(delete_column_ids); |
94 | 4 | for (auto cid : delete_column_ids) { |
95 | 4 | all_access_paths.erase(tablet_schema.column(cid).unique_id()); |
96 | 4 | } |
97 | 2 | } |
98 | | |
99 | 386 | Status TabletReader::_capture_rs_readers(const ReaderParams& read_params) { |
100 | 386 | SCOPED_RAW_TIMER(&_stats.tablet_reader_capture_rs_readers_timer_ns); |
101 | 386 | if (read_params.rs_splits.empty()) { |
102 | 0 | return Status::InternalError("fail to acquire data sources. tablet={}", |
103 | 0 | _tablet->tablet_id()); |
104 | 0 | } |
105 | | |
106 | 386 | bool eof = false; |
107 | 386 | bool is_lower_key_included = _keys_param.start_key_include; |
108 | 386 | bool is_upper_key_included = _keys_param.end_key_include; |
109 | | |
110 | 386 | for (int i = 0; i < _keys_param.start_keys.size(); ++i) { |
111 | | // lower bound |
112 | 0 | RowCursor& start_key = _keys_param.start_keys[i]; |
113 | 0 | RowCursor& end_key = _keys_param.end_keys[i]; |
114 | |
|
115 | 0 | if (!is_lower_key_included) { |
116 | 0 | if (compare_row_key(start_key, end_key) >= 0) { |
117 | 0 | VLOG_NOTICE << "return EOF when lower key not include" |
118 | 0 | << ", start_key=" << start_key.to_string() |
119 | 0 | << ", end_key=" << end_key.to_string(); |
120 | 0 | eof = true; |
121 | 0 | break; |
122 | 0 | } |
123 | 0 | } else { |
124 | 0 | if (compare_row_key(start_key, end_key) > 0) { |
125 | 0 | VLOG_NOTICE << "return EOF when lower key include=" |
126 | 0 | << ", start_key=" << start_key.to_string() |
127 | 0 | << ", end_key=" << end_key.to_string(); |
128 | 0 | eof = true; |
129 | 0 | break; |
130 | 0 | } |
131 | 0 | } |
132 | | |
133 | 0 | _is_lower_keys_included.push_back(is_lower_key_included); |
134 | 0 | _is_upper_keys_included.push_back(is_upper_key_included); |
135 | 0 | } |
136 | | |
137 | 386 | if (eof) { |
138 | 0 | return Status::EndOfFile("reach end of scan range. tablet={}", _tablet->tablet_id()); |
139 | 0 | } |
140 | | |
141 | 386 | bool need_ordered_result = true; |
142 | 386 | if (read_params.reader_type == ReaderType::READER_QUERY || |
143 | 386 | read_params.reader_type == ReaderType::READER_BINLOG) { |
144 | 0 | if (_tablet_schema->keys_type() == DUP_KEYS) { |
145 | | // duplicated keys are allowed, no need to merge sort keys in rowset |
146 | 0 | need_ordered_result = false; |
147 | 0 | } |
148 | 0 | if (_tablet_schema->keys_type() == UNIQUE_KEYS && |
149 | 0 | _tablet->enable_unique_key_merge_on_write()) { |
150 | | // unique keys with merge on write, no need to merge sort keys in rowset |
151 | 0 | need_ordered_result = false; |
152 | 0 | } |
153 | 0 | if (_aggregation) { |
154 | | // compute engine will aggregate rows with the same key, |
155 | | // it's ok for rowset to return unordered result |
156 | 0 | need_ordered_result = false; |
157 | 0 | } |
158 | |
|
159 | 0 | if (_direct_mode) { |
160 | | // direct mode indicates that the storage layer does not need to merge, |
161 | | // it's ok for rowset to return unordered result |
162 | 0 | need_ordered_result = false; |
163 | 0 | } |
164 | |
|
165 | 0 | if (read_params.read_orderby_key) { |
166 | 0 | need_ordered_result = true; |
167 | 0 | } |
168 | 0 | } |
169 | | |
170 | 386 | _reader_context.reader_type = read_params.reader_type; |
171 | 386 | _reader_context.version = read_params.version; |
172 | 386 | _reader_context.tablet_schema = _tablet_schema; |
173 | 386 | _reader_context.need_ordered_result = need_ordered_result; |
174 | 386 | _reader_context.topn_filter_source_node_ids = read_params.topn_filter_source_node_ids; |
175 | 386 | _reader_context.topn_filter_target_node_id = read_params.topn_filter_target_node_id; |
176 | 386 | _reader_context.read_orderby_key_reverse = read_params.read_orderby_key_reverse; |
177 | 386 | _reader_context.use_insert_order_when_same = |
178 | 386 | read_params.use_insert_order_when_same || |
179 | 386 | read_params.reader_type == ReaderType::READER_BINLOG || |
180 | 386 | read_params.reader_type == ReaderType::READER_BINLOG_COMPACTION; |
181 | 386 | _reader_context.force_key_ordered_read = read_params.force_key_ordered_read; |
182 | 386 | _reader_context.read_orderby_key_limit = read_params.read_orderby_key_limit; |
183 | 386 | _reader_context.return_columns = &_return_columns; |
184 | 386 | _reader_context.tso_predicate_column_id = read_params.tso_predicate_column_id; |
185 | 386 | _reader_context.start_tso = read_params.start_tso; |
186 | 386 | _reader_context.end_tso = read_params.end_tso; |
187 | 386 | _reader_context.read_orderby_key_columns = |
188 | 386 | !_orderby_key_columns.empty() ? &_orderby_key_columns : nullptr; |
189 | 386 | _reader_context.predicates = &_col_predicates; |
190 | 386 | _reader_context.value_predicates = &_value_col_predicates; |
191 | 386 | _reader_context.lower_bound_keys = &_keys_param.start_keys; |
192 | 386 | _reader_context.is_lower_keys_included = &_is_lower_keys_included; |
193 | 386 | _reader_context.upper_bound_keys = &_keys_param.end_keys; |
194 | 386 | _reader_context.is_upper_keys_included = &_is_upper_keys_included; |
195 | 386 | _reader_context.delete_handler = &_delete_handler; |
196 | 386 | _reader_context.stats = &_stats; |
197 | 386 | _reader_context.use_page_cache = read_params.use_page_cache; |
198 | 386 | _reader_context.sequence_id_idx = _sequence_col_idx; |
199 | 386 | _reader_context.is_unique = tablet()->keys_type() == UNIQUE_KEYS; |
200 | 386 | _reader_context.merged_rows = &_merged_rows; |
201 | 386 | _reader_context.delete_bitmap = read_params.delete_bitmap; |
202 | 386 | _reader_context.enable_unique_key_merge_on_write = tablet()->enable_unique_key_merge_on_write(); |
203 | 386 | _reader_context.enable_mor_value_predicate_pushdown = |
204 | 386 | read_params.enable_mor_value_predicate_pushdown; |
205 | 386 | _reader_context.record_rowids = read_params.record_rowids; |
206 | 386 | _reader_context.rowid_conversion = read_params.rowid_conversion; |
207 | 386 | _reader_context.is_key_column_group = read_params.is_key_column_group; |
208 | 386 | _reader_context.common_expr_ctxs_push_down = read_params.common_expr_ctxs_push_down; |
209 | 386 | _reader_context.output_columns = &read_params.output_columns; |
210 | 386 | _reader_context.extra_columns = read_params.extra_columns; |
211 | 386 | _reader_context.push_down_agg_type_opt = read_params.push_down_agg_type_opt; |
212 | 386 | _reader_context.ttl_seconds = _tablet->ttl_seconds(); |
213 | 386 | _reader_context.score_runtime = read_params.score_runtime; |
214 | 386 | _reader_context.collection_statistics = read_params.collection_statistics; |
215 | | |
216 | 386 | _reader_context.virtual_column_exprs = read_params.virtual_column_exprs; |
217 | 386 | _reader_context.ann_topn_runtime = read_params.ann_topn_runtime; |
218 | | |
219 | 386 | _reader_context.condition_cache_digest = read_params.condition_cache_digest; |
220 | 386 | _reader_context.all_access_paths = read_params.all_access_paths; |
221 | 386 | _reader_context.predicate_access_paths = read_params.predicate_access_paths; |
222 | | |
223 | | // Force a full read of delete-condition columns: the FE can't see storage deletes and may |
224 | | // mark them meta-only (OFFSET/NULL), whose content-less read makes the delete predicate |
225 | | // match nothing and leak deleted rows. |
226 | 386 | if (!_delete_handler.empty() && !_reader_context.all_access_paths.empty()) { |
227 | 0 | remove_delete_columns_from_access_paths(_delete_handler, *_tablet_schema, |
228 | 0 | _reader_context.all_access_paths); |
229 | 0 | } |
230 | | |
231 | | // Propagate general read limit for DUP_KEYS and UNIQUE_KEYS with MOW |
232 | 386 | _reader_context.general_read_limit = read_params.general_read_limit; |
233 | | |
234 | 386 | return Status::OK(); |
235 | 386 | } |
236 | | |
237 | 0 | TabletColumn TabletReader::materialize_column(const TabletColumn& orig) { |
238 | 0 | if (!orig.is_variant_type()) { |
239 | 0 | return orig; |
240 | 0 | } |
241 | 0 | TabletColumn column_with_cast_type = orig; |
242 | 0 | auto cast_type = _reader_context.target_cast_type_for_variants.at(orig.name()); |
243 | 0 | return variant_util::get_column_by_type(cast_type, orig.name(), |
244 | 0 | { |
245 | 0 | .unique_id = orig.unique_id(), |
246 | 0 | .parent_unique_id = orig.parent_unique_id(), |
247 | 0 | .path_info = *orig.path_info_ptr(), |
248 | 0 | }); |
249 | 0 | } |
250 | | |
251 | 408 | Status TabletReader::_init_params(const ReaderParams& read_params) { |
252 | 408 | read_params.check_validation(); |
253 | | |
254 | 408 | _direct_mode = read_params.direct_mode; |
255 | 408 | _aggregation = read_params.aggregation; |
256 | 408 | _reader_type = read_params.reader_type; |
257 | 408 | _tablet = read_params.tablet; |
258 | 408 | _tablet_schema = read_params.tablet_schema; |
259 | 408 | _reader_context.runtime_state = read_params.runtime_state; |
260 | 408 | _reader_context.target_cast_type_for_variants = read_params.target_cast_type_for_variants; |
261 | | |
262 | 408 | RETURN_IF_ERROR(_init_conditions_param(read_params)); |
263 | | |
264 | 408 | Status res = _init_delete_condition(read_params); |
265 | 408 | if (!res.ok()) { |
266 | 0 | LOG(WARNING) << "fail to init delete param. res = " << res; |
267 | 0 | return res; |
268 | 0 | } |
269 | | |
270 | 408 | res = _init_return_columns(read_params); |
271 | 408 | if (!res.ok()) { |
272 | 0 | LOG(WARNING) << "fail to init return columns. res = " << res; |
273 | 0 | return res; |
274 | 0 | } |
275 | | |
276 | 408 | res = _init_keys_param(read_params); |
277 | 408 | if (!res.ok()) { |
278 | 0 | LOG(WARNING) << "fail to init keys param. res=" << res; |
279 | 0 | return res; |
280 | 0 | } |
281 | 408 | res = _init_orderby_keys_param(read_params); |
282 | 408 | if (!res.ok()) { |
283 | 0 | LOG(WARNING) << "fail to init orderby keys param. res=" << res; |
284 | 0 | return res; |
285 | 0 | } |
286 | 408 | if (_tablet_schema->has_sequence_col()) { |
287 | 8 | auto sequence_col_idx = _tablet_schema->sequence_col_idx(); |
288 | 8 | DCHECK_NE(sequence_col_idx, -1); |
289 | 16 | for (auto col : _return_columns) { |
290 | | // query has sequence col |
291 | 16 | if (col == sequence_col_idx) { |
292 | 4 | _sequence_col_idx = sequence_col_idx; |
293 | 4 | break; |
294 | 4 | } |
295 | 16 | } |
296 | 8 | } |
297 | | |
298 | 408 | return res; |
299 | 408 | } |
300 | | |
301 | 408 | Status TabletReader::_init_return_columns(const ReaderParams& read_params) { |
302 | 408 | SCOPED_RAW_TIMER(&_stats.tablet_reader_init_return_columns_timer_ns); |
303 | 408 | if (read_params.reader_type == ReaderType::READER_QUERY || |
304 | 408 | read_params.reader_type == ReaderType::READER_BINLOG) { |
305 | 22 | _return_columns = read_params.return_columns; |
306 | 22 | _tablet_columns_convert_to_null_set = read_params.tablet_columns_convert_to_null_set; |
307 | 37 | for (auto id : read_params.return_columns) { |
308 | 37 | if (_tablet_schema->column(id).is_key()) { |
309 | 22 | _key_cids.push_back(id); |
310 | 22 | } else { |
311 | 15 | _value_cids.push_back(id); |
312 | 15 | } |
313 | 37 | } |
314 | 386 | } else if (read_params.return_columns.empty()) { |
315 | 0 | for (uint32_t i = 0; i < _tablet_schema->num_columns(); ++i) { |
316 | 0 | _return_columns.push_back(i); |
317 | 0 | if (_tablet_schema->column(i).is_key()) { |
318 | 0 | _key_cids.push_back(i); |
319 | 0 | } else { |
320 | 0 | _value_cids.push_back(i); |
321 | 0 | } |
322 | 0 | } |
323 | 0 | VLOG_NOTICE << "return column is empty, using full column as default."; |
324 | 386 | } else if ((read_params.reader_type == ReaderType::READER_CUMULATIVE_COMPACTION || |
325 | 386 | read_params.reader_type == ReaderType::READER_SEGMENT_COMPACTION || |
326 | 386 | read_params.reader_type == ReaderType::READER_BASE_COMPACTION || |
327 | 386 | read_params.reader_type == ReaderType::READER_FULL_COMPACTION || |
328 | 386 | read_params.reader_type == ReaderType::READER_BINLOG_COMPACTION || |
329 | 386 | read_params.reader_type == ReaderType::READER_COLD_DATA_COMPACTION || |
330 | 386 | read_params.reader_type == ReaderType::READER_ALTER_TABLE) && |
331 | 386 | !read_params.return_columns.empty()) { |
332 | 386 | _return_columns = read_params.return_columns; |
333 | 1.18k | for (auto id : read_params.return_columns) { |
334 | 1.18k | if (_tablet_schema->column(id).is_key()) { |
335 | 136 | _key_cids.push_back(id); |
336 | 1.04k | } else { |
337 | 1.04k | _value_cids.push_back(id); |
338 | 1.04k | } |
339 | 1.18k | } |
340 | 386 | } else if (read_params.reader_type == ReaderType::READER_CHECKSUM) { |
341 | 0 | _return_columns = read_params.return_columns; |
342 | 0 | for (auto id : read_params.return_columns) { |
343 | 0 | if (_tablet_schema->column(id).is_key()) { |
344 | 0 | _key_cids.push_back(id); |
345 | 0 | } else { |
346 | 0 | _value_cids.push_back(id); |
347 | 0 | } |
348 | 0 | } |
349 | 0 | } else { |
350 | 0 | return Status::Error<INVALID_ARGUMENT>( |
351 | 0 | "fail to init return columns. reader_type={}, return_columns_size={}", |
352 | 0 | int(read_params.reader_type), read_params.return_columns.size()); |
353 | 0 | } |
354 | | |
355 | 408 | std::sort(_key_cids.begin(), _key_cids.end(), std::greater<>()); |
356 | | |
357 | 408 | return Status::OK(); |
358 | 408 | } |
359 | | |
360 | 408 | Status TabletReader::_init_keys_param(const ReaderParams& read_params) { |
361 | 408 | SCOPED_RAW_TIMER(&_stats.tablet_reader_init_keys_param_timer_ns); |
362 | 408 | if (read_params.start_key.empty()) { |
363 | 408 | return Status::OK(); |
364 | 408 | } |
365 | | |
366 | 0 | _keys_param.start_key_include = read_params.start_key_include; |
367 | 0 | _keys_param.end_key_include = read_params.end_key_include; |
368 | |
|
369 | 0 | size_t start_key_size = read_params.start_key.size(); |
370 | | //_keys_param.start_keys.resize(start_key_size); |
371 | 0 | std::vector<RowCursor>(start_key_size).swap(_keys_param.start_keys); |
372 | |
|
373 | 0 | size_t scan_key_size = read_params.start_key.front().size(); |
374 | 0 | if (scan_key_size > _tablet_schema->num_columns()) { |
375 | 0 | return Status::Error<INVALID_ARGUMENT>( |
376 | 0 | "Input param are invalid. Column count is bigger than num_columns of schema. " |
377 | 0 | "column_count={}, schema.num_columns={}", |
378 | 0 | scan_key_size, _tablet_schema->num_columns()); |
379 | 0 | } |
380 | | |
381 | 0 | for (size_t i = 0; i < start_key_size; ++i) { |
382 | 0 | if (read_params.start_key[i].size() != scan_key_size) { |
383 | 0 | return Status::Error<INVALID_ARGUMENT>( |
384 | 0 | "The start_key.at({}).size={}, not equals the scan_key_size={}", i, |
385 | 0 | read_params.start_key[i].size(), scan_key_size); |
386 | 0 | } |
387 | | |
388 | 0 | Status res = _keys_param.start_keys[i].init(_tablet_schema, read_params.start_key[i]); |
389 | 0 | if (!res.ok()) { |
390 | 0 | LOG(WARNING) << "fail to init row cursor. res = " << res; |
391 | 0 | return res; |
392 | 0 | } |
393 | 0 | } |
394 | | |
395 | 0 | size_t end_key_size = read_params.end_key.size(); |
396 | | //_keys_param.end_keys.resize(end_key_size); |
397 | 0 | std::vector<RowCursor>(end_key_size).swap(_keys_param.end_keys); |
398 | 0 | for (size_t i = 0; i < end_key_size; ++i) { |
399 | 0 | if (read_params.end_key[i].size() != scan_key_size) { |
400 | 0 | return Status::Error<INVALID_ARGUMENT>( |
401 | 0 | "The end_key.at({}).size={}, not equals the scan_key_size={}", i, |
402 | 0 | read_params.end_key[i].size(), scan_key_size); |
403 | 0 | } |
404 | | |
405 | 0 | Status res = _keys_param.end_keys[i].init(_tablet_schema, read_params.end_key[i]); |
406 | 0 | if (!res.ok()) { |
407 | 0 | LOG(WARNING) << "fail to init row cursor. res = " << res; |
408 | 0 | return res; |
409 | 0 | } |
410 | 0 | } |
411 | | |
412 | | //TODO:check the valid of start_key and end_key.(eg. start_key <= end_key) |
413 | | |
414 | 0 | return Status::OK(); |
415 | 0 | } |
416 | | |
417 | 408 | Status TabletReader::_init_orderby_keys_param(const ReaderParams& read_params) { |
418 | 408 | SCOPED_RAW_TIMER(&_stats.tablet_reader_init_orderby_keys_param_timer_ns); |
419 | | // UNIQUE_KEYS will compare all keys as before |
420 | 408 | if (_tablet_schema->keys_type() == DUP_KEYS || (_tablet_schema->keys_type() == UNIQUE_KEYS && |
421 | 346 | _tablet->enable_unique_key_merge_on_write())) { |
422 | 346 | if (!_tablet_schema->cluster_key_uids().empty()) { |
423 | 0 | if (read_params.read_orderby_key_num_prefix_columns > |
424 | 0 | _tablet_schema->cluster_key_uids().size()) { |
425 | 0 | return Status::Error<ErrorCode::INTERNAL_ERROR>( |
426 | 0 | "read_orderby_key_num_prefix_columns={} > cluster_keys.size()={}", |
427 | 0 | read_params.read_orderby_key_num_prefix_columns, |
428 | 0 | _tablet_schema->cluster_key_uids().size()); |
429 | 0 | } |
430 | 0 | for (uint32_t i = 0; i < read_params.read_orderby_key_num_prefix_columns; i++) { |
431 | 0 | auto cid = _tablet_schema->cluster_key_uids()[i]; |
432 | 0 | auto index = _tablet_schema->field_index(cid); |
433 | 0 | if (index < 0) { |
434 | 0 | return Status::Error<ErrorCode::INTERNAL_ERROR>( |
435 | 0 | "could not find cluster key column with unique_id=" + |
436 | 0 | std::to_string(cid) + |
437 | 0 | " in tablet schema, tablet_id=" + std::to_string(_tablet->tablet_id())); |
438 | 0 | } |
439 | 0 | for (uint32_t idx = 0; idx < _return_columns.size(); idx++) { |
440 | 0 | if (_return_columns[idx] == index) { |
441 | 0 | _orderby_key_columns.push_back(idx); |
442 | 0 | break; |
443 | 0 | } |
444 | 0 | } |
445 | 0 | } |
446 | 346 | } else { |
447 | | // find index in vector _return_columns |
448 | | // for the read_orderby_key_num_prefix_columns orderby keys |
449 | 346 | for (uint32_t i = 0; i < read_params.read_orderby_key_num_prefix_columns; i++) { |
450 | 0 | for (uint32_t idx = 0; idx < _return_columns.size(); idx++) { |
451 | 0 | if (_return_columns[idx] == i) { |
452 | 0 | _orderby_key_columns.push_back(idx); |
453 | 0 | break; |
454 | 0 | } |
455 | 0 | } |
456 | 0 | } |
457 | 346 | } |
458 | 346 | if (read_params.read_orderby_key_num_prefix_columns != _orderby_key_columns.size()) { |
459 | 0 | return Status::Error<ErrorCode::INTERNAL_ERROR>( |
460 | 0 | "read_orderby_key_num_prefix_columns != _orderby_key_columns.size, " |
461 | 0 | "read_params.read_orderby_key_num_prefix_columns={}, " |
462 | 0 | "_orderby_key_columns.size()={}", |
463 | 0 | read_params.read_orderby_key_num_prefix_columns, _orderby_key_columns.size()); |
464 | 0 | } |
465 | 346 | } |
466 | | |
467 | 408 | return Status::OK(); |
468 | 408 | } |
469 | | |
470 | 408 | Status TabletReader::_init_conditions_param(const ReaderParams& read_params) { |
471 | 408 | SCOPED_RAW_TIMER(&_stats.tablet_reader_init_conditions_param_timer_ns); |
472 | 408 | std::vector<std::shared_ptr<ColumnPredicate>> predicates; |
473 | 408 | std::copy(read_params.predicates.cbegin(), read_params.predicates.cend(), |
474 | 408 | std::inserter(predicates, predicates.begin())); |
475 | | // Function filter push down to storage engine |
476 | 408 | auto is_like_predicate = [](std::shared_ptr<ColumnPredicate> _pred) { |
477 | 0 | return dynamic_cast<LikeColumnPredicate*>(_pred.get()) != nullptr; |
478 | 0 | }; |
479 | | |
480 | 408 | for (const auto& filter : read_params.function_filters) { |
481 | 0 | predicates.emplace_back(_parse_to_predicate(filter)); |
482 | 0 | auto pred = predicates.back(); |
483 | |
|
484 | 0 | const auto& col = _tablet_schema->column(pred->column_id()); |
485 | 0 | const auto* tablet_index = _tablet_schema->get_ngram_bf_index(col.unique_id()); |
486 | 0 | if (is_like_predicate(pred) && tablet_index && config::enable_query_like_bloom_filter) { |
487 | 0 | std::unique_ptr<segment_v2::BloomFilter> ng_bf; |
488 | 0 | std::string pattern = pred->get_search_str(); |
489 | 0 | auto gram_bf_size = tablet_index->get_gram_bf_size(); |
490 | 0 | auto gram_size = tablet_index->get_gram_size(); |
491 | |
|
492 | 0 | RETURN_IF_ERROR(segment_v2::BloomFilter::create(segment_v2::NGRAM_BLOOM_FILTER, &ng_bf, |
493 | 0 | gram_bf_size)); |
494 | 0 | NgramTokenExtractor _token_extractor(gram_size); |
495 | |
|
496 | 0 | if (_token_extractor.string_like_to_bloom_filter(pattern.data(), pattern.length(), |
497 | 0 | *ng_bf)) { |
498 | 0 | pred->set_page_ng_bf(std::move(ng_bf)); |
499 | 0 | } |
500 | 0 | } |
501 | 0 | } |
502 | | |
503 | 408 | int32_t delete_sign_idx = _tablet_schema->delete_sign_idx(); |
504 | 408 | for (auto predicate : predicates) { |
505 | 0 | auto column = _tablet_schema->column(predicate->column_id()); |
506 | 0 | if (column.aggregation() != FieldAggregationMethod::OLAP_FIELD_AGGREGATION_NONE) { |
507 | | // When MOR value predicate pushdown is enabled, drop __DORIS_DELETE_SIGN__ |
508 | | // from storage-layer predicates entirely. Delete sign must only be evaluated |
509 | | // post-merge via VExpr to prevent deleted rows from reappearing. |
510 | 0 | if (read_params.enable_mor_value_predicate_pushdown && delete_sign_idx >= 0 && |
511 | 0 | predicate->column_id() == static_cast<uint32_t>(delete_sign_idx)) { |
512 | 0 | continue; |
513 | 0 | } |
514 | 0 | _value_col_predicates.push_back(predicate); |
515 | 0 | } else { |
516 | 0 | _col_predicates.push_back(predicate); |
517 | 0 | } |
518 | 0 | } |
519 | | |
520 | 408 | return Status::OK(); |
521 | 408 | } |
522 | | |
523 | | std::shared_ptr<ColumnPredicate> TabletReader::_parse_to_predicate( |
524 | 0 | const FunctionFilter& function_filter) { |
525 | 0 | int32_t index = _tablet_schema->field_index(function_filter._col_name); |
526 | 0 | if (index < 0) { |
527 | 0 | throw Exception(Status::InternalError("Column {} not found in tablet schema", |
528 | 0 | function_filter._col_name)); |
529 | 0 | return nullptr; |
530 | 0 | } |
531 | 0 | const TabletColumn& column = materialize_column(_tablet_schema->column(index)); |
532 | 0 | return create_column_predicate(index, std::make_shared<FunctionFilter>(function_filter), |
533 | 0 | column.type(), &column); |
534 | 0 | } |
535 | | |
536 | 408 | Status TabletReader::_init_delete_condition(const ReaderParams& read_params) { |
537 | 408 | SCOPED_RAW_TIMER(&_stats.tablet_reader_init_delete_condition_param_timer_ns); |
538 | | // If it's cumu and not allow do delete when cumu |
539 | 408 | if (read_params.reader_type == ReaderType::READER_SEGMENT_COMPACTION || |
540 | 408 | (read_params.reader_type == ReaderType::READER_CUMULATIVE_COMPACTION && |
541 | 408 | !config::enable_delete_when_cumu_compaction)) { |
542 | 40 | return Status::OK(); |
543 | 40 | } |
544 | 368 | bool cumu_delete = read_params.reader_type == ReaderType::READER_CUMULATIVE_COMPACTION && |
545 | 368 | config::enable_delete_when_cumu_compaction; |
546 | | // Delete sign could not be applied when delete on cumu compaction is enabled, bucause it is meant for delete with predicates. |
547 | | // If delete design is applied on cumu compaction, it will lose effect when doing base compaction. |
548 | | // `_delete_sign_available` indicates the condition where we could apply delete signs to data. |
549 | 368 | _delete_sign_available = (((read_params.reader_type == ReaderType::READER_BASE_COMPACTION || |
550 | 368 | read_params.reader_type == ReaderType::READER_FULL_COMPACTION) && |
551 | 368 | config::enable_prune_delete_sign_when_base_compaction) || |
552 | 368 | read_params.reader_type == ReaderType::READER_COLD_DATA_COMPACTION || |
553 | 368 | read_params.reader_type == ReaderType::READER_CHECKSUM); |
554 | | |
555 | | // `_filter_delete` indicates the condition where we should execlude deleted tuples when reading data. |
556 | | // However, queries will not use this condition but generate special where predicates to filter data. |
557 | | // (Though a lille bit confused, it is how the current logic working...) |
558 | 368 | _filter_delete = _delete_sign_available || cumu_delete; |
559 | 368 | return _delete_handler.init(_tablet_schema, read_params.delete_predicates, |
560 | 368 | read_params.version.second); |
561 | 408 | } |
562 | | |
563 | | Status TabletReader::init_reader_params_and_create_block( |
564 | | TabletSharedPtr tablet, ReaderType reader_type, |
565 | | const std::vector<RowsetSharedPtr>& input_rowsets, |
566 | 0 | TabletReader::ReaderParams* reader_params, Block* block) { |
567 | 0 | reader_params->tablet = tablet; |
568 | 0 | reader_params->reader_type = reader_type; |
569 | 0 | reader_params->version = |
570 | 0 | Version(input_rowsets.front()->start_version(), input_rowsets.back()->end_version()); |
571 | |
|
572 | 0 | TabletReadSource read_source; |
573 | 0 | for (const auto& rowset : input_rowsets) { |
574 | 0 | RowsetReaderSharedPtr rs_reader; |
575 | 0 | RETURN_IF_ERROR(rowset->create_reader(&rs_reader)); |
576 | 0 | read_source.rs_splits.emplace_back(std::move(rs_reader)); |
577 | 0 | } |
578 | 0 | read_source.fill_delete_predicates(); |
579 | 0 | reader_params->set_read_source(std::move(read_source)); |
580 | |
|
581 | 0 | std::vector<RowsetMetaSharedPtr> rowset_metas(input_rowsets.size()); |
582 | 0 | std::transform(input_rowsets.begin(), input_rowsets.end(), rowset_metas.begin(), |
583 | 0 | [](const RowsetSharedPtr& rowset) { return rowset->rowset_meta(); }); |
584 | 0 | TabletSchemaSPtr read_tablet_schema = |
585 | 0 | tablet->tablet_schema_with_merged_max_schema_version(rowset_metas); |
586 | 0 | TabletSchemaSPtr merge_tablet_schema = std::make_shared<TabletSchema>(); |
587 | 0 | merge_tablet_schema->copy_from(*read_tablet_schema); |
588 | | |
589 | | // Merge the columns in delete predicate that not in latest schema in to current tablet schema |
590 | 0 | for (auto& del_pred : reader_params->delete_predicates) { |
591 | 0 | merge_tablet_schema->merge_dropped_columns(*del_pred->tablet_schema()); |
592 | 0 | } |
593 | 0 | reader_params->tablet_schema = merge_tablet_schema; |
594 | |
|
595 | 0 | reader_params->return_columns.resize(read_tablet_schema->num_columns()); |
596 | 0 | std::iota(reader_params->return_columns.begin(), reader_params->return_columns.end(), 0); |
597 | 0 | reader_params->origin_return_columns = &reader_params->return_columns; |
598 | |
|
599 | 0 | *block = read_tablet_schema->create_block(); |
600 | |
|
601 | 0 | return Status::OK(); |
602 | 0 | } |
603 | | |
604 | | } // namespace doris |