be/src/format/parquet/vparquet_group_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 "format/parquet/vparquet_group_reader.h" |
19 | | |
20 | | #include <gen_cpp/Exprs_types.h> |
21 | | #include <gen_cpp/Opcodes_types.h> |
22 | | #include <gen_cpp/Types_types.h> |
23 | | #include <gen_cpp/parquet_types.h> |
24 | | #include <string.h> |
25 | | |
26 | | #include <algorithm> |
27 | | #include <boost/iterator/iterator_facade.hpp> |
28 | | #include <memory> |
29 | | #include <ostream> |
30 | | |
31 | | #include "common/config.h" |
32 | | #include "common/logging.h" |
33 | | #include "common/object_pool.h" |
34 | | #include "common/status.h" |
35 | | #include "core/assert_cast.h" |
36 | | #include "core/block/block.h" |
37 | | #include "core/block/column_with_type_and_name.h" |
38 | | #include "core/column/column_const.h" |
39 | | #include "core/column/column_nullable.h" |
40 | | #include "core/column/column_string.h" |
41 | | #include "core/column/column_vector.h" |
42 | | #include "core/custom_allocator.h" |
43 | | #include "core/data_type/data_type.h" |
44 | | #include "core/data_type/data_type_string.h" |
45 | | #include "core/data_type/define_primitive_type.h" |
46 | | #include "core/pod_array.h" |
47 | | #include "core/types.h" |
48 | | #include "exprs/create_predicate_function.h" |
49 | | #include "exprs/hybrid_set.h" |
50 | | #include "exprs/vdirect_in_predicate.h" |
51 | | #include "exprs/vectorized_fn_call.h" |
52 | | #include "exprs/vexpr.h" |
53 | | #include "exprs/vexpr_context.h" |
54 | | #include "exprs/vliteral.h" |
55 | | #include "exprs/vslot_ref.h" |
56 | | #include "format/parquet/schema_desc.h" |
57 | | #include "format/parquet/vparquet_column_reader.h" |
58 | | #include "runtime/descriptors.h" |
59 | | #include "runtime/runtime_state.h" |
60 | | #include "runtime/thread_context.h" |
61 | | #include "storage/segment/column_reader.h" |
62 | | |
63 | | namespace cctz { |
64 | | class time_zone; |
65 | | } // namespace cctz |
66 | | namespace doris { |
67 | | class RuntimeState; |
68 | | |
69 | | namespace io { |
70 | | struct IOContext; |
71 | | } // namespace io |
72 | | } // namespace doris |
73 | | |
74 | | namespace doris { |
75 | | #include "common/compile_check_begin.h" |
76 | | const std::vector<int64_t> RowGroupReader::NO_DELETE = {}; |
77 | | static constexpr uint32_t MAX_DICT_CODE_PREDICATE_TO_REWRITE = std::numeric_limits<uint32_t>::max(); |
78 | | |
79 | | RowGroupReader::RowGroupReader(io::FileReaderSPtr file_reader, |
80 | | const std::vector<std::string>& read_columns, |
81 | | const int32_t row_group_id, const tparquet::RowGroup& row_group, |
82 | | const cctz::time_zone* ctz, io::IOContext* io_ctx, |
83 | | const PositionDeleteContext& position_delete_ctx, |
84 | | const LazyReadContext& lazy_read_ctx, RuntimeState* state, |
85 | | const std::set<uint64_t>& column_ids, |
86 | | const std::set<uint64_t>& filter_column_ids) |
87 | 35 | : _file_reader(file_reader), |
88 | 35 | _read_table_columns(read_columns), |
89 | 35 | _row_group_id(row_group_id), |
90 | 35 | _row_group_meta(row_group), |
91 | 35 | _remaining_rows(row_group.num_rows), |
92 | 35 | _ctz(ctz), |
93 | 35 | _io_ctx(io_ctx), |
94 | 35 | _position_delete_ctx(position_delete_ctx), |
95 | 35 | _lazy_read_ctx(lazy_read_ctx), |
96 | 35 | _state(state), |
97 | 35 | _obj_pool(new ObjectPool()), |
98 | 35 | _column_ids(column_ids), |
99 | 35 | _filter_column_ids(filter_column_ids) {} |
100 | | |
101 | 35 | RowGroupReader::~RowGroupReader() { |
102 | 35 | _column_readers.clear(); |
103 | 35 | _obj_pool->clear(); |
104 | 35 | } |
105 | | |
106 | | Status RowGroupReader::init( |
107 | | const FieldDescriptor& schema, RowRanges& row_ranges, |
108 | | std::unordered_map<int, tparquet::OffsetIndex>& col_offsets, |
109 | | const TupleDescriptor* tuple_descriptor, const RowDescriptor* row_descriptor, |
110 | | const std::unordered_map<std::string, int>* colname_to_slot_id, |
111 | | const VExprContextSPtrs* not_single_slot_filter_conjuncts, |
112 | 35 | const std::unordered_map<int, VExprContextSPtrs>* slot_id_to_filter_conjuncts) { |
113 | 35 | _tuple_descriptor = tuple_descriptor; |
114 | 35 | _row_descriptor = row_descriptor; |
115 | 35 | _col_name_to_slot_id = colname_to_slot_id; |
116 | 35 | _slot_id_to_filter_conjuncts = slot_id_to_filter_conjuncts; |
117 | 35 | _read_ranges = row_ranges; |
118 | 35 | _remaining_rows = row_ranges.count(); |
119 | | |
120 | 35 | if (_read_table_columns.empty()) { |
121 | | // Query task that only select columns in path. |
122 | 1 | return Status::OK(); |
123 | 1 | } |
124 | 34 | const size_t MAX_GROUP_BUF_SIZE = config::parquet_rowgroup_max_buffer_mb << 20; |
125 | 34 | const size_t MAX_COLUMN_BUF_SIZE = config::parquet_column_max_buffer_mb << 20; |
126 | 34 | size_t max_buf_size = |
127 | 34 | std::min(MAX_COLUMN_BUF_SIZE, MAX_GROUP_BUF_SIZE / _read_table_columns.size()); |
128 | 104 | for (const auto& read_table_col : _read_table_columns) { |
129 | 104 | auto read_file_col = _table_info_node_ptr->children_file_column_name(read_table_col); |
130 | 104 | auto* field = schema.get_column(read_file_col); |
131 | 104 | std::unique_ptr<ParquetColumnReader> reader; |
132 | 104 | RETURN_IF_ERROR(ParquetColumnReader::create( |
133 | 104 | _file_reader, field, _row_group_meta, _read_ranges, _ctz, _io_ctx, reader, |
134 | 104 | max_buf_size, col_offsets, _state, false, _column_ids, _filter_column_ids)); |
135 | 104 | if (reader == nullptr) { |
136 | 0 | VLOG_DEBUG << "Init row group(" << _row_group_id << ") reader failed"; |
137 | 0 | return Status::Corruption("Init row group reader failed"); |
138 | 0 | } |
139 | 104 | _column_readers[read_table_col] = std::move(reader); |
140 | 104 | } |
141 | | |
142 | 34 | bool disable_dict_filter = false; |
143 | 34 | if (not_single_slot_filter_conjuncts != nullptr && !not_single_slot_filter_conjuncts->empty()) { |
144 | 0 | disable_dict_filter = true; |
145 | 0 | _filter_conjuncts.insert(_filter_conjuncts.end(), not_single_slot_filter_conjuncts->begin(), |
146 | 0 | not_single_slot_filter_conjuncts->end()); |
147 | 0 | } |
148 | | |
149 | | // Check if single slot can be filtered by dict. |
150 | 34 | if (_slot_id_to_filter_conjuncts && !_slot_id_to_filter_conjuncts->empty()) { |
151 | 6 | const std::vector<std::string>& predicate_col_names = |
152 | 6 | _lazy_read_ctx.predicate_columns.first; |
153 | 6 | const std::vector<int>& predicate_col_slot_ids = _lazy_read_ctx.predicate_columns.second; |
154 | 14 | for (size_t i = 0; i < predicate_col_names.size(); ++i) { |
155 | 8 | const std::string& predicate_col_name = predicate_col_names[i]; |
156 | 8 | int slot_id = predicate_col_slot_ids[i]; |
157 | 8 | auto predicate_file_col_name = |
158 | 8 | _table_info_node_ptr->children_file_column_name(predicate_col_name); |
159 | 8 | auto field = schema.get_column(predicate_file_col_name); |
160 | 8 | if (!disable_dict_filter && !_lazy_read_ctx.has_complex_type && |
161 | 8 | _can_filter_by_dict( |
162 | 8 | slot_id, _row_group_meta.columns[field->physical_column_index].meta_data)) { |
163 | 2 | _dict_filter_cols.emplace_back(std::make_pair(predicate_col_name, slot_id)); |
164 | 6 | } else { |
165 | 6 | if (_slot_id_to_filter_conjuncts->find(slot_id) != |
166 | 6 | _slot_id_to_filter_conjuncts->end()) { |
167 | 6 | for (auto& ctx : _slot_id_to_filter_conjuncts->at(slot_id)) { |
168 | 6 | _filter_conjuncts.push_back(ctx); |
169 | 6 | } |
170 | 6 | } |
171 | 6 | } |
172 | 8 | } |
173 | | // Add predicate_partition_columns in _slot_id_to_filter_conjuncts(single slot conjuncts) |
174 | | // to _filter_conjuncts, others should be added from not_single_slot_filter_conjuncts. |
175 | 6 | for (auto& kv : _lazy_read_ctx.predicate_partition_columns) { |
176 | 4 | auto& [value, slot_desc] = kv.second; |
177 | 4 | auto iter = _slot_id_to_filter_conjuncts->find(slot_desc->id()); |
178 | 4 | if (iter != _slot_id_to_filter_conjuncts->end()) { |
179 | 4 | for (auto& ctx : iter->second) { |
180 | 4 | _filter_conjuncts.push_back(ctx); |
181 | 4 | } |
182 | 4 | } |
183 | 4 | } |
184 | | //For check missing column : missing column == xx, missing column is null,missing column is not null. |
185 | 6 | _filter_conjuncts.insert(_filter_conjuncts.end(), |
186 | 6 | _lazy_read_ctx.missing_columns_conjuncts.begin(), |
187 | 6 | _lazy_read_ctx.missing_columns_conjuncts.end()); |
188 | 6 | RETURN_IF_ERROR(_rewrite_dict_predicates()); |
189 | 6 | } |
190 | | // _state is nullptr in some ut. |
191 | 34 | if (_state && _state->enable_adjust_conjunct_order_by_cost()) { |
192 | 8 | std::ranges::sort(_filter_conjuncts, [](const auto& a, const auto& b) { |
193 | 8 | return a->execute_cost() < b->execute_cost(); |
194 | 8 | }); |
195 | 8 | } |
196 | 34 | return Status::OK(); |
197 | 34 | } |
198 | | |
199 | | bool RowGroupReader::_can_filter_by_dict(int slot_id, |
200 | 8 | const tparquet::ColumnMetaData& column_metadata) { |
201 | 8 | SlotDescriptor* slot = nullptr; |
202 | 8 | const std::vector<SlotDescriptor*>& slots = _tuple_descriptor->slots(); |
203 | 14 | for (auto each : slots) { |
204 | 14 | if (each->id() == slot_id) { |
205 | 8 | slot = each; |
206 | 8 | break; |
207 | 8 | } |
208 | 14 | } |
209 | 8 | if (!is_string_type(slot->type()->get_primitive_type()) && |
210 | 8 | !is_var_len_object(slot->type()->get_primitive_type())) { |
211 | 6 | return false; |
212 | 6 | } |
213 | 2 | if (column_metadata.type != tparquet::Type::BYTE_ARRAY) { |
214 | 0 | return false; |
215 | 0 | } |
216 | | |
217 | 2 | if (!is_dictionary_encoded(column_metadata)) { |
218 | 0 | return false; |
219 | 0 | } |
220 | | |
221 | 2 | if (_slot_id_to_filter_conjuncts->find(slot_id) == _slot_id_to_filter_conjuncts->end()) { |
222 | 0 | return false; |
223 | 0 | } |
224 | | |
225 | | // TODO: The current implementation of dictionary filtering does not take into account |
226 | | // the implementation of NULL values because the dictionary itself does not contain |
227 | | // NULL value encoding. As a result, many NULL-related functions or expressions |
228 | | // cannot work properly, such as is null, is not null, coalesce, etc. |
229 | | // Here we check if the predicate expr is IN or BINARY_PRED. |
230 | | // Implementation of NULL value dictionary filtering will be carried out later. |
231 | 2 | return std::ranges::all_of(_slot_id_to_filter_conjuncts->at(slot_id), [&](const auto& ctx) { |
232 | 2 | return (ctx->root()->node_type() == TExprNodeType::IN_PRED || |
233 | 2 | ctx->root()->node_type() == TExprNodeType::BINARY_PRED) && |
234 | 2 | ctx->root()->children()[0]->node_type() == TExprNodeType::SLOT_REF; |
235 | 2 | }); |
236 | 2 | } |
237 | | |
238 | | // This function is copied from |
239 | | // https://github.com/apache/impala/blob/master/be/src/exec/parquet/hdfs-parquet-scanner.cc#L1717 |
240 | 2 | bool RowGroupReader::is_dictionary_encoded(const tparquet::ColumnMetaData& column_metadata) { |
241 | | // The Parquet spec allows for column chunks to have mixed encodings |
242 | | // where some data pages are dictionary-encoded and others are plain |
243 | | // encoded. For example, a Parquet file writer might start writing |
244 | | // a column chunk as dictionary encoded, but it will switch to plain |
245 | | // encoding if the dictionary grows too large. |
246 | | // |
247 | | // In order for dictionary filters to skip the entire row group, |
248 | | // the conjuncts must be evaluated on column chunks that are entirely |
249 | | // encoded with the dictionary encoding. There are two checks |
250 | | // available to verify this: |
251 | | // 1. The encoding_stats field on the column chunk metadata provides |
252 | | // information about the number of data pages written in each |
253 | | // format. This allows for a specific check of whether all the |
254 | | // data pages are dictionary encoded. |
255 | | // 2. The encodings field on the column chunk metadata lists the |
256 | | // encodings used. If this list contains the dictionary encoding |
257 | | // and does not include unexpected encodings (i.e. encodings not |
258 | | // associated with definition/repetition levels), then it is entirely |
259 | | // dictionary encoded. |
260 | 2 | if (column_metadata.__isset.encoding_stats) { |
261 | | // Condition #1 above |
262 | 4 | for (const tparquet::PageEncodingStats& enc_stat : column_metadata.encoding_stats) { |
263 | 4 | if (enc_stat.page_type == tparquet::PageType::DATA_PAGE && |
264 | 4 | (enc_stat.encoding != tparquet::Encoding::PLAIN_DICTIONARY && |
265 | 2 | enc_stat.encoding != tparquet::Encoding::RLE_DICTIONARY) && |
266 | 4 | enc_stat.count > 0) { |
267 | 0 | return false; |
268 | 0 | } |
269 | 4 | } |
270 | 2 | } else { |
271 | | // Condition #2 above |
272 | 0 | bool has_dict_encoding = false; |
273 | 0 | bool has_nondict_encoding = false; |
274 | 0 | for (const tparquet::Encoding::type& encoding : column_metadata.encodings) { |
275 | 0 | if (encoding == tparquet::Encoding::PLAIN_DICTIONARY || |
276 | 0 | encoding == tparquet::Encoding::RLE_DICTIONARY) { |
277 | 0 | has_dict_encoding = true; |
278 | 0 | } |
279 | | |
280 | | // RLE and BIT_PACKED are used for repetition/definition levels |
281 | 0 | if (encoding != tparquet::Encoding::PLAIN_DICTIONARY && |
282 | 0 | encoding != tparquet::Encoding::RLE_DICTIONARY && |
283 | 0 | encoding != tparquet::Encoding::RLE && encoding != tparquet::Encoding::BIT_PACKED) { |
284 | 0 | has_nondict_encoding = true; |
285 | 0 | break; |
286 | 0 | } |
287 | 0 | } |
288 | | // Not entirely dictionary encoded if: |
289 | | // 1. No dictionary encoding listed |
290 | | // OR |
291 | | // 2. Some non-dictionary encoding is listed |
292 | 0 | if (!has_dict_encoding || has_nondict_encoding) { |
293 | 0 | return false; |
294 | 0 | } |
295 | 0 | } |
296 | | |
297 | 2 | return true; |
298 | 2 | } |
299 | | |
300 | | Status RowGroupReader::next_batch(Block* block, size_t batch_size, size_t* read_rows, |
301 | 47 | bool* batch_eof) { |
302 | 47 | if (_is_row_group_filtered) { |
303 | 2 | *read_rows = 0; |
304 | 2 | *batch_eof = true; |
305 | 2 | return Status::OK(); |
306 | 2 | } |
307 | | |
308 | | // Process external table query task that select columns are all from path. |
309 | 45 | if (_read_table_columns.empty()) { |
310 | 3 | bool modify_row_ids = false; |
311 | 3 | RETURN_IF_ERROR(_read_empty_batch(batch_size, read_rows, batch_eof, &modify_row_ids)); |
312 | | |
313 | 3 | RETURN_IF_ERROR( |
314 | 3 | _fill_partition_columns(block, *read_rows, _lazy_read_ctx.partition_columns)); |
315 | 3 | RETURN_IF_ERROR(_fill_missing_columns(block, *read_rows, _lazy_read_ctx.missing_columns)); |
316 | | |
317 | 3 | RETURN_IF_ERROR(_fill_row_id_columns(block, *read_rows, modify_row_ids)); |
318 | | |
319 | 3 | Status st = VExprContext::filter_block(_lazy_read_ctx.conjuncts, block, block->columns()); |
320 | 3 | *read_rows = block->rows(); |
321 | 3 | return st; |
322 | 3 | } |
323 | 42 | if (_lazy_read_ctx.can_lazy_read) { |
324 | | // call _do_lazy_read recursively when current batch is skipped |
325 | 4 | return _do_lazy_read(block, batch_size, read_rows, batch_eof); |
326 | 38 | } else { |
327 | 38 | FilterMap filter_map; |
328 | 38 | RETURN_IF_ERROR((_read_column_data(block, _lazy_read_ctx.all_read_columns, batch_size, |
329 | 38 | read_rows, batch_eof, filter_map))); |
330 | 38 | RETURN_IF_ERROR( |
331 | 38 | _fill_partition_columns(block, *read_rows, _lazy_read_ctx.partition_columns)); |
332 | 38 | RETURN_IF_ERROR(_fill_missing_columns(block, *read_rows, _lazy_read_ctx.missing_columns)); |
333 | 38 | RETURN_IF_ERROR(_fill_row_id_columns(block, *read_rows, false)); |
334 | | |
335 | 38 | #ifndef NDEBUG |
336 | 123 | for (auto col : *block) { |
337 | 123 | col.column->sanity_check(); |
338 | 123 | DCHECK(block->rows() == col.column->size()) |
339 | 0 | << absl::Substitute("block rows = $0 , column rows = $1, col name = $2", |
340 | 0 | block->rows(), col.column->size(), col.name); |
341 | 123 | } |
342 | 38 | #endif |
343 | | |
344 | 38 | if (block->rows() == 0) { |
345 | 0 | RETURN_IF_ERROR(_convert_dict_cols_to_string_cols(block)); |
346 | 0 | *read_rows = block->rows(); |
347 | 0 | #ifndef NDEBUG |
348 | 0 | for (auto col : *block) { |
349 | 0 | col.column->sanity_check(); |
350 | 0 | DCHECK(block->rows() == col.column->size()) |
351 | 0 | << absl::Substitute("block rows = $0 , column rows = $1, col name = $2", |
352 | 0 | block->rows(), col.column->size(), col.name); |
353 | 0 | } |
354 | 0 | #endif |
355 | 0 | return Status::OK(); |
356 | 0 | } |
357 | 38 | { |
358 | 38 | SCOPED_RAW_TIMER(&_predicate_filter_time); |
359 | 38 | RETURN_IF_ERROR(_build_pos_delete_filter(*read_rows)); |
360 | | |
361 | 38 | std::vector<uint32_t> columns_to_filter; |
362 | 38 | int column_to_keep = block->columns(); |
363 | 38 | columns_to_filter.resize(column_to_keep); |
364 | 161 | for (uint32_t i = 0; i < column_to_keep; ++i) { |
365 | 123 | columns_to_filter[i] = i; |
366 | 123 | } |
367 | 38 | if (!_lazy_read_ctx.conjuncts.empty()) { |
368 | 6 | std::vector<IColumn::Filter*> filters; |
369 | 6 | if (_position_delete_ctx.has_filter) { |
370 | 0 | filters.push_back(_pos_delete_filter_ptr.get()); |
371 | 0 | } |
372 | 6 | IColumn::Filter result_filter(block->rows(), 1); |
373 | 6 | bool can_filter_all = false; |
374 | | |
375 | 6 | { |
376 | 6 | RETURN_IF_ERROR_OR_CATCH_EXCEPTION(VExprContext::execute_conjuncts( |
377 | 6 | _filter_conjuncts, &filters, block, &result_filter, &can_filter_all)); |
378 | 6 | } |
379 | | |
380 | 6 | if (can_filter_all) { |
381 | 9 | for (auto& col : columns_to_filter) { |
382 | 9 | std::move(*block->get_by_position(col).column).assume_mutable()->clear(); |
383 | 9 | } |
384 | 3 | Block::erase_useless_column(block, column_to_keep); |
385 | 3 | RETURN_IF_ERROR(_convert_dict_cols_to_string_cols(block)); |
386 | 3 | return Status::OK(); |
387 | 3 | } |
388 | | |
389 | 3 | RETURN_IF_CATCH_EXCEPTION( |
390 | 3 | Block::filter_block_internal(block, columns_to_filter, result_filter)); |
391 | 3 | Block::erase_useless_column(block, column_to_keep); |
392 | 32 | } else { |
393 | 32 | RETURN_IF_CATCH_EXCEPTION( |
394 | 32 | RETURN_IF_ERROR(_filter_block(block, column_to_keep, columns_to_filter))); |
395 | 32 | } |
396 | 35 | RETURN_IF_ERROR(_convert_dict_cols_to_string_cols(block)); |
397 | 35 | } |
398 | 35 | #ifndef NDEBUG |
399 | 114 | for (auto col : *block) { |
400 | 114 | col.column->sanity_check(); |
401 | 114 | DCHECK(block->rows() == col.column->size()) |
402 | 0 | << absl::Substitute("block rows = $0 , column rows = $1, col name = $2", |
403 | 0 | block->rows(), col.column->size(), col.name); |
404 | 114 | } |
405 | 35 | #endif |
406 | 35 | *read_rows = block->rows(); |
407 | 35 | return Status::OK(); |
408 | 35 | } |
409 | 42 | } |
410 | | |
411 | | Status RowGroupReader::_read_column_data(Block* block, |
412 | | const std::vector<std::string>& table_columns, |
413 | | size_t batch_size, size_t* read_rows, bool* batch_eof, |
414 | 47 | FilterMap& filter_map) { |
415 | 47 | size_t batch_read_rows = 0; |
416 | 47 | bool has_eof = false; |
417 | 121 | for (auto& read_col_name : table_columns) { |
418 | 121 | auto& column_with_type_and_name = |
419 | 121 | block->safe_get_by_position((*_col_name_to_block_idx)[read_col_name]); |
420 | 121 | auto& column_ptr = column_with_type_and_name.column; |
421 | 121 | auto& column_type = column_with_type_and_name.type; |
422 | 121 | bool is_dict_filter = false; |
423 | 121 | for (auto& _dict_filter_col : _dict_filter_cols) { |
424 | 0 | if (_dict_filter_col.first == read_col_name) { |
425 | 0 | MutableColumnPtr dict_column = ColumnInt32::create(); |
426 | 0 | if (!_col_name_to_block_idx->contains(read_col_name)) { |
427 | 0 | return Status::InternalError( |
428 | 0 | "Wrong read column '{}' in parquet file, block: {}", read_col_name, |
429 | 0 | block->dump_structure()); |
430 | 0 | } |
431 | 0 | if (column_type->is_nullable()) { |
432 | 0 | block->get_by_position((*_col_name_to_block_idx)[read_col_name]).type = |
433 | 0 | std::make_shared<DataTypeNullable>(std::make_shared<DataTypeInt32>()); |
434 | 0 | block->replace_by_position( |
435 | 0 | (*_col_name_to_block_idx)[read_col_name], |
436 | 0 | ColumnNullable::create(std::move(dict_column), |
437 | 0 | ColumnUInt8::create(dict_column->size(), 0))); |
438 | 0 | } else { |
439 | 0 | block->get_by_position((*_col_name_to_block_idx)[read_col_name]).type = |
440 | 0 | std::make_shared<DataTypeInt32>(); |
441 | 0 | block->replace_by_position((*_col_name_to_block_idx)[read_col_name], |
442 | 0 | std::move(dict_column)); |
443 | 0 | } |
444 | 0 | is_dict_filter = true; |
445 | 0 | break; |
446 | 0 | } |
447 | 0 | } |
448 | | |
449 | 121 | size_t col_read_rows = 0; |
450 | 121 | bool col_eof = false; |
451 | | // Should reset _filter_map_index to 0 when reading next column. |
452 | | // select_vector.reset(); |
453 | 121 | _column_readers[read_col_name]->reset_filter_map_index(); |
454 | 305 | while (!col_eof && col_read_rows < batch_size) { |
455 | 184 | size_t loop_rows = 0; |
456 | 184 | RETURN_IF_ERROR(_column_readers[read_col_name]->read_column_data( |
457 | 184 | column_ptr, column_type, _table_info_node_ptr->get_children_node(read_col_name), |
458 | 184 | filter_map, batch_size - col_read_rows, &loop_rows, &col_eof, is_dict_filter)); |
459 | 184 | VLOG_DEBUG << "[RowGroupReader] column '" << read_col_name |
460 | 0 | << "' loop_rows=" << loop_rows << " col_read_rows_so_far=" << col_read_rows |
461 | 0 | << std::endl; |
462 | 184 | col_read_rows += loop_rows; |
463 | 184 | } |
464 | 121 | VLOG_DEBUG << "[RowGroupReader] column '" << read_col_name |
465 | 0 | << "' read_rows=" << col_read_rows << std::endl; |
466 | 121 | if (batch_read_rows > 0 && batch_read_rows != col_read_rows) { |
467 | 0 | LOG(WARNING) << "[RowGroupReader] Mismatched read rows among parquet columns. " |
468 | 0 | "previous_batch_read_rows=" |
469 | 0 | << batch_read_rows << ", current_column='" << read_col_name |
470 | 0 | << "', current_col_read_rows=" << col_read_rows; |
471 | 0 | return Status::Corruption("Can't read the same number of rows among parquet columns"); |
472 | 0 | } |
473 | 121 | batch_read_rows = col_read_rows; |
474 | | |
475 | 121 | #ifndef NDEBUG |
476 | 121 | column_ptr->sanity_check(); |
477 | 121 | #endif |
478 | 121 | if (col_eof) { |
479 | 99 | has_eof = true; |
480 | 99 | } |
481 | 121 | } |
482 | | |
483 | 47 | *read_rows = batch_read_rows; |
484 | 47 | *batch_eof = has_eof; |
485 | | |
486 | 47 | return Status::OK(); |
487 | 47 | } |
488 | | |
489 | | Status RowGroupReader::_do_lazy_read(Block* block, size_t batch_size, size_t* read_rows, |
490 | 4 | bool* batch_eof) { |
491 | 4 | std::unique_ptr<FilterMap> filter_map_ptr = nullptr; |
492 | 4 | size_t pre_read_rows; |
493 | 4 | bool pre_eof; |
494 | 4 | std::vector<uint32_t> columns_to_filter; |
495 | 4 | uint32_t origin_column_num = block->columns(); |
496 | 4 | columns_to_filter.resize(origin_column_num); |
497 | 16 | for (uint32_t i = 0; i < origin_column_num; ++i) { |
498 | 12 | columns_to_filter[i] = i; |
499 | 12 | } |
500 | 4 | IColumn::Filter result_filter; |
501 | 4 | size_t pre_raw_read_rows = 0; |
502 | 6 | while (!_state->is_cancelled()) { |
503 | | // read predicate columns |
504 | 6 | pre_read_rows = 0; |
505 | 6 | pre_eof = false; |
506 | 6 | FilterMap filter_map; |
507 | 6 | RETURN_IF_ERROR(_read_column_data(block, _lazy_read_ctx.predicate_columns.first, batch_size, |
508 | 6 | &pre_read_rows, &pre_eof, filter_map)); |
509 | 6 | if (pre_read_rows == 0) { |
510 | 0 | DCHECK_EQ(pre_eof, true); |
511 | 0 | break; |
512 | 0 | } |
513 | 6 | pre_raw_read_rows += pre_read_rows; |
514 | 6 | RETURN_IF_ERROR(_fill_partition_columns(block, pre_read_rows, |
515 | 6 | _lazy_read_ctx.predicate_partition_columns)); |
516 | 6 | RETURN_IF_ERROR(_fill_missing_columns(block, pre_read_rows, |
517 | 6 | _lazy_read_ctx.predicate_missing_columns)); |
518 | 6 | RETURN_IF_ERROR(_fill_row_id_columns(block, pre_read_rows, false)); |
519 | | |
520 | 6 | RETURN_IF_ERROR(_build_pos_delete_filter(pre_read_rows)); |
521 | | |
522 | 6 | #ifndef NDEBUG |
523 | 18 | for (auto col : *block) { |
524 | 18 | if (col.column->size() == 0) { // lazy read column. |
525 | 6 | continue; |
526 | 6 | } |
527 | 12 | col.column->sanity_check(); |
528 | 12 | DCHECK(pre_read_rows == col.column->size()) |
529 | 0 | << absl::Substitute("pre_read_rows = $0 , column rows = $1, col name = $2", |
530 | 0 | pre_read_rows, col.column->size(), col.name); |
531 | 12 | } |
532 | 6 | #endif |
533 | | |
534 | 6 | bool can_filter_all = false; |
535 | 6 | { |
536 | 6 | SCOPED_RAW_TIMER(&_predicate_filter_time); |
537 | | |
538 | | // generate filter vector |
539 | 6 | if (_lazy_read_ctx.resize_first_column) { |
540 | | // VExprContext.execute has an optimization, the filtering is executed when block->rows() > 0 |
541 | | // The following process may be tricky and time-consuming, but we have no other way. |
542 | 6 | block->get_by_position(0).column->assume_mutable()->resize(pre_read_rows); |
543 | 6 | } |
544 | 6 | result_filter.assign(pre_read_rows, static_cast<unsigned char>(1)); |
545 | 6 | std::vector<IColumn::Filter*> filters; |
546 | 6 | if (_position_delete_ctx.has_filter) { |
547 | 0 | filters.push_back(_pos_delete_filter_ptr.get()); |
548 | 0 | } |
549 | | |
550 | 6 | VExprContextSPtrs filter_contexts; |
551 | 12 | for (auto& conjunct : _filter_conjuncts) { |
552 | 12 | filter_contexts.emplace_back(conjunct); |
553 | 12 | } |
554 | | |
555 | 6 | { |
556 | 6 | RETURN_IF_ERROR(VExprContext::execute_conjuncts(filter_contexts, &filters, block, |
557 | 6 | &result_filter, &can_filter_all)); |
558 | 6 | } |
559 | | |
560 | 6 | if (_lazy_read_ctx.resize_first_column) { |
561 | | // We have to clean the first column to insert right data. |
562 | 6 | block->get_by_position(0).column->assume_mutable()->clear(); |
563 | 6 | } |
564 | 6 | } |
565 | | |
566 | 0 | const uint8_t* __restrict filter_map_data = result_filter.data(); |
567 | 6 | filter_map_ptr = std::make_unique<FilterMap>(); |
568 | 6 | RETURN_IF_ERROR(filter_map_ptr->init(filter_map_data, pre_read_rows, can_filter_all)); |
569 | 6 | if (filter_map_ptr->filter_all()) { |
570 | 3 | { |
571 | 3 | SCOPED_RAW_TIMER(&_predicate_filter_time); |
572 | 3 | for (const auto& col : _lazy_read_ctx.predicate_columns.first) { |
573 | | // clean block to read predicate columns |
574 | 3 | block->get_by_position((*_col_name_to_block_idx)[col]) |
575 | 3 | .column->assume_mutable() |
576 | 3 | ->clear(); |
577 | 3 | } |
578 | 3 | for (const auto& col : _lazy_read_ctx.predicate_partition_columns) { |
579 | 3 | block->get_by_position((*_col_name_to_block_idx)[col.first]) |
580 | 3 | .column->assume_mutable() |
581 | 3 | ->clear(); |
582 | 3 | } |
583 | 3 | for (const auto& col : _lazy_read_ctx.predicate_missing_columns) { |
584 | 0 | block->get_by_position((*_col_name_to_block_idx)[col.first]) |
585 | 0 | .column->assume_mutable() |
586 | 0 | ->clear(); |
587 | 0 | } |
588 | 3 | if (_row_id_column_iterator_pair.first != nullptr) { |
589 | 0 | block->get_by_position(_row_id_column_iterator_pair.second) |
590 | 0 | .column->assume_mutable() |
591 | 0 | ->clear(); |
592 | 0 | } |
593 | 3 | Block::erase_useless_column(block, origin_column_num); |
594 | 3 | } |
595 | | |
596 | 3 | if (!pre_eof) { |
597 | | // If continuous batches are skipped, we can cache them to skip a whole page |
598 | 2 | _cached_filtered_rows += pre_read_rows; |
599 | 2 | if (pre_raw_read_rows >= config::doris_scanner_row_num) { |
600 | 0 | *read_rows = 0; |
601 | 0 | RETURN_IF_ERROR(_convert_dict_cols_to_string_cols(block)); |
602 | 0 | return Status::OK(); |
603 | 0 | } |
604 | 2 | } else { // pre_eof |
605 | | // If filter_map_ptr->filter_all() and pre_eof, we can skip whole row group. |
606 | 1 | *read_rows = 0; |
607 | 1 | *batch_eof = true; |
608 | 1 | _lazy_read_filtered_rows += (pre_read_rows + _cached_filtered_rows); |
609 | 1 | RETURN_IF_ERROR(_convert_dict_cols_to_string_cols(block)); |
610 | 1 | return Status::OK(); |
611 | 1 | } |
612 | 3 | } else { |
613 | 3 | break; |
614 | 3 | } |
615 | 6 | } |
616 | 3 | if (_state->is_cancelled()) { |
617 | 0 | return Status::Cancelled("cancelled"); |
618 | 0 | } |
619 | | |
620 | 3 | if (filter_map_ptr == nullptr) { |
621 | 0 | DCHECK_EQ(pre_read_rows + _cached_filtered_rows, 0); |
622 | 0 | *read_rows = 0; |
623 | 0 | *batch_eof = true; |
624 | 0 | return Status::OK(); |
625 | 0 | } |
626 | | |
627 | 3 | FilterMap& filter_map = *filter_map_ptr; |
628 | 3 | DorisUniqueBufferPtr<uint8_t> rebuild_filter_map = nullptr; |
629 | 3 | if (_cached_filtered_rows != 0) { |
630 | 0 | RETURN_IF_ERROR(_rebuild_filter_map(filter_map, rebuild_filter_map, pre_read_rows)); |
631 | 0 | pre_read_rows += _cached_filtered_rows; |
632 | 0 | _cached_filtered_rows = 0; |
633 | 0 | } |
634 | | |
635 | | // lazy read columns |
636 | 3 | size_t lazy_read_rows; |
637 | 3 | bool lazy_eof; |
638 | 3 | RETURN_IF_ERROR(_read_column_data(block, _lazy_read_ctx.lazy_read_columns, pre_read_rows, |
639 | 3 | &lazy_read_rows, &lazy_eof, filter_map)); |
640 | | |
641 | 3 | if (pre_read_rows != lazy_read_rows) { |
642 | 0 | return Status::Corruption("Can't read the same number of rows when doing lazy read"); |
643 | 0 | } |
644 | | // pre_eof ^ lazy_eof |
645 | | // we set pre_read_rows as batch_size for lazy read columns, so pre_eof != lazy_eof |
646 | | |
647 | | // filter data in predicate columns, and remove filter column |
648 | 3 | { |
649 | 3 | SCOPED_RAW_TIMER(&_predicate_filter_time); |
650 | 3 | if (filter_map.has_filter()) { |
651 | 0 | RETURN_IF_CATCH_EXCEPTION(Block::filter_block_internal( |
652 | 0 | block, _lazy_read_ctx.all_predicate_col_ids, result_filter)); |
653 | 0 | Block::erase_useless_column(block, origin_column_num); |
654 | |
|
655 | 3 | } else { |
656 | 3 | Block::erase_useless_column(block, origin_column_num); |
657 | 3 | } |
658 | 3 | } |
659 | | |
660 | 3 | RETURN_IF_ERROR(_convert_dict_cols_to_string_cols(block)); |
661 | | |
662 | 3 | size_t column_num = block->columns(); |
663 | 3 | size_t column_size = 0; |
664 | 12 | for (int i = 0; i < column_num; ++i) { |
665 | 9 | size_t cz = block->get_by_position(i).column->size(); |
666 | 9 | if (column_size != 0 && cz != 0) { |
667 | 6 | DCHECK_EQ(column_size, cz); |
668 | 6 | } |
669 | 9 | if (cz != 0) { |
670 | 9 | column_size = cz; |
671 | 9 | } |
672 | 9 | } |
673 | 3 | _lazy_read_filtered_rows += pre_read_rows - column_size; |
674 | 3 | *read_rows = column_size; |
675 | | |
676 | 3 | *batch_eof = pre_eof; |
677 | 3 | RETURN_IF_ERROR(_fill_partition_columns(block, column_size, _lazy_read_ctx.partition_columns)); |
678 | 3 | RETURN_IF_ERROR(_fill_missing_columns(block, column_size, _lazy_read_ctx.missing_columns)); |
679 | 3 | #ifndef NDEBUG |
680 | 9 | for (auto col : *block) { |
681 | 9 | col.column->sanity_check(); |
682 | 9 | DCHECK(block->rows() == col.column->size()) |
683 | 0 | << absl::Substitute("block rows = $0 , column rows = $1, col name = $2", |
684 | 0 | block->rows(), col.column->size(), col.name); |
685 | 9 | } |
686 | 3 | #endif |
687 | 3 | return Status::OK(); |
688 | 3 | } |
689 | | |
690 | | Status RowGroupReader::_rebuild_filter_map(FilterMap& filter_map, |
691 | | DorisUniqueBufferPtr<uint8_t>& filter_map_data, |
692 | 0 | size_t pre_read_rows) const { |
693 | 0 | if (_cached_filtered_rows == 0) { |
694 | 0 | return Status::OK(); |
695 | 0 | } |
696 | 0 | size_t total_rows = _cached_filtered_rows + pre_read_rows; |
697 | 0 | if (filter_map.filter_all()) { |
698 | 0 | RETURN_IF_ERROR(filter_map.init(nullptr, total_rows, true)); |
699 | 0 | return Status::OK(); |
700 | 0 | } |
701 | | |
702 | 0 | filter_map_data = make_unique_buffer<uint8_t>(total_rows); |
703 | 0 | auto* map = filter_map_data.get(); |
704 | 0 | for (size_t i = 0; i < _cached_filtered_rows; ++i) { |
705 | 0 | map[i] = 0; |
706 | 0 | } |
707 | 0 | const uint8_t* old_map = filter_map.filter_map_data(); |
708 | 0 | if (old_map == nullptr) { |
709 | | // select_vector.filter_all() == true is already built. |
710 | 0 | for (size_t i = _cached_filtered_rows; i < total_rows; ++i) { |
711 | 0 | map[i] = 1; |
712 | 0 | } |
713 | 0 | } else { |
714 | 0 | memcpy(map + _cached_filtered_rows, old_map, pre_read_rows); |
715 | 0 | } |
716 | 0 | RETURN_IF_ERROR(filter_map.init(map, total_rows, false)); |
717 | 0 | return Status::OK(); |
718 | 0 | } |
719 | | |
720 | | Status RowGroupReader::_fill_partition_columns( |
721 | | Block* block, size_t rows, |
722 | | const std::unordered_map<std::string, std::tuple<std::string, const SlotDescriptor*>>& |
723 | 50 | partition_columns) { |
724 | 50 | DataTypeSerDe::FormatOptions _text_formatOptions; |
725 | 50 | for (const auto& kv : partition_columns) { |
726 | 15 | auto doris_column = block->get_by_position((*_col_name_to_block_idx)[kv.first]).column; |
727 | | // obtained from block*, it is a mutable object. |
728 | 15 | auto* col_ptr = const_cast<IColumn*>(doris_column.get()); |
729 | 15 | const auto& [value, slot_desc] = kv.second; |
730 | 15 | auto _text_serde = slot_desc->get_data_type_ptr()->get_serde(); |
731 | 15 | Slice slice(value.data(), value.size()); |
732 | 15 | uint64_t num_deserialized = 0; |
733 | | // Be careful when reading empty rows from parquet row groups. |
734 | 15 | if (_text_serde->deserialize_column_from_fixed_json(*col_ptr, slice, rows, |
735 | 15 | &num_deserialized, |
736 | 15 | _text_formatOptions) != Status::OK()) { |
737 | 0 | return Status::InternalError("Failed to fill partition column: {}={}", |
738 | 0 | slot_desc->col_name(), value); |
739 | 0 | } |
740 | 15 | if (num_deserialized != rows) { |
741 | 0 | return Status::InternalError( |
742 | 0 | "Failed to fill partition column: {}={} ." |
743 | 0 | "Number of rows expected to be written : {}, number of rows actually written : " |
744 | 0 | "{}", |
745 | 0 | slot_desc->col_name(), value, num_deserialized, rows); |
746 | 0 | } |
747 | 15 | } |
748 | 50 | return Status::OK(); |
749 | 50 | } |
750 | | |
751 | | Status RowGroupReader::_fill_missing_columns( |
752 | | Block* block, size_t rows, |
753 | 50 | const std::unordered_map<std::string, VExprContextSPtr>& missing_columns) { |
754 | 50 | for (const auto& kv : missing_columns) { |
755 | 0 | if (!_col_name_to_block_idx->contains(kv.first)) { |
756 | 0 | return Status::InternalError("Missing column: {} not found in block {}", kv.first, |
757 | 0 | block->dump_structure()); |
758 | 0 | } |
759 | 0 | if (kv.second == nullptr) { |
760 | | // no default column, fill with null |
761 | 0 | auto mutable_column = block->get_by_position((*_col_name_to_block_idx)[kv.first]) |
762 | 0 | .column->assume_mutable(); |
763 | 0 | auto* nullable_column = assert_cast<ColumnNullable*>(mutable_column.get()); |
764 | 0 | nullable_column->insert_many_defaults(rows); |
765 | 0 | } else { |
766 | | // fill with default value |
767 | 0 | const auto& ctx = kv.second; |
768 | 0 | ColumnPtr result_column_ptr; |
769 | | // PT1 => dest primitive type |
770 | 0 | RETURN_IF_ERROR(ctx->execute(block, result_column_ptr)); |
771 | 0 | if (result_column_ptr->use_count() == 1) { |
772 | | // call resize because the first column of _src_block_ptr may not be filled by reader, |
773 | | // so _src_block_ptr->rows() may return wrong result, cause the column created by `ctx->execute()` |
774 | | // has only one row. |
775 | 0 | auto mutable_column = result_column_ptr->assume_mutable(); |
776 | 0 | mutable_column->resize(rows); |
777 | | // result_column_ptr maybe a ColumnConst, convert it to a normal column |
778 | 0 | result_column_ptr = result_column_ptr->convert_to_full_column_if_const(); |
779 | 0 | auto origin_column_type = |
780 | 0 | block->get_by_position((*_col_name_to_block_idx)[kv.first]).type; |
781 | 0 | bool is_nullable = origin_column_type->is_nullable(); |
782 | 0 | block->replace_by_position( |
783 | 0 | (*_col_name_to_block_idx)[kv.first], |
784 | 0 | is_nullable ? make_nullable(result_column_ptr) : result_column_ptr); |
785 | 0 | } |
786 | 0 | } |
787 | 0 | } |
788 | 50 | return Status::OK(); |
789 | 50 | } |
790 | | |
791 | | Status RowGroupReader::_read_empty_batch(size_t batch_size, size_t* read_rows, bool* batch_eof, |
792 | 3 | bool* modify_row_ids) { |
793 | 3 | *modify_row_ids = false; |
794 | 3 | if (_position_delete_ctx.has_filter) { |
795 | 0 | int64_t start_row_id = _position_delete_ctx.current_row_id; |
796 | 0 | int64_t end_row_id = std::min(_position_delete_ctx.current_row_id + (int64_t)batch_size, |
797 | 0 | _position_delete_ctx.last_row_id); |
798 | 0 | int64_t num_delete_rows = 0; |
799 | 0 | auto before_index = _position_delete_ctx.index; |
800 | 0 | while (_position_delete_ctx.index < _position_delete_ctx.end_index) { |
801 | 0 | const int64_t& delete_row_id = |
802 | 0 | _position_delete_ctx.delete_rows[_position_delete_ctx.index]; |
803 | 0 | if (delete_row_id < start_row_id) { |
804 | 0 | _position_delete_ctx.index++; |
805 | 0 | before_index = _position_delete_ctx.index; |
806 | 0 | } else if (delete_row_id < end_row_id) { |
807 | 0 | num_delete_rows++; |
808 | 0 | _position_delete_ctx.index++; |
809 | 0 | } else { // delete_row_id >= end_row_id |
810 | 0 | break; |
811 | 0 | } |
812 | 0 | } |
813 | 0 | *read_rows = end_row_id - start_row_id - num_delete_rows; |
814 | 0 | _position_delete_ctx.current_row_id = end_row_id; |
815 | 0 | *batch_eof = _position_delete_ctx.current_row_id == _position_delete_ctx.last_row_id; |
816 | |
|
817 | 0 | if (_row_id_column_iterator_pair.first != nullptr) { |
818 | 0 | *modify_row_ids = true; |
819 | 0 | _current_batch_row_ids.clear(); |
820 | 0 | _current_batch_row_ids.resize(*read_rows); |
821 | 0 | size_t idx = 0; |
822 | 0 | for (auto id = start_row_id; id < end_row_id; id++) { |
823 | 0 | if (before_index < _position_delete_ctx.index && |
824 | 0 | id == _position_delete_ctx.delete_rows[before_index]) { |
825 | 0 | before_index++; |
826 | 0 | continue; |
827 | 0 | } |
828 | 0 | _current_batch_row_ids[idx++] = (rowid_t)id; |
829 | 0 | } |
830 | 0 | } |
831 | 3 | } else { |
832 | 3 | if (batch_size < _remaining_rows) { |
833 | 2 | *read_rows = batch_size; |
834 | 2 | _remaining_rows -= batch_size; |
835 | 2 | *batch_eof = false; |
836 | 2 | } else { |
837 | 1 | *read_rows = _remaining_rows; |
838 | 1 | _remaining_rows = 0; |
839 | 1 | *batch_eof = true; |
840 | 1 | } |
841 | 3 | } |
842 | 3 | _total_read_rows += *read_rows; |
843 | 3 | return Status::OK(); |
844 | 3 | } |
845 | | |
846 | 5 | Status RowGroupReader::_get_current_batch_row_id(size_t read_rows) { |
847 | 5 | _current_batch_row_ids.clear(); |
848 | 5 | _current_batch_row_ids.resize(read_rows); |
849 | | |
850 | 5 | int64_t idx = 0; |
851 | 5 | int64_t read_range_rows = 0; |
852 | 19 | for (size_t range_idx = 0; range_idx < _read_ranges.range_size(); range_idx++) { |
853 | 14 | auto range = _read_ranges.get_range(range_idx); |
854 | 14 | if (read_rows == 0) { |
855 | 0 | break; |
856 | 0 | } |
857 | 14 | if (read_range_rows + (range.to() - range.from()) > _total_read_rows) { |
858 | 14 | int64_t fi = |
859 | 14 | std::max(_total_read_rows, read_range_rows) - read_range_rows + range.from(); |
860 | 14 | size_t len = std::min(read_rows, (size_t)(std::max(range.to(), fi) - fi)); |
861 | | |
862 | 14 | read_rows -= len; |
863 | | |
864 | 28 | for (auto i = 0; i < len; i++) { |
865 | 14 | _current_batch_row_ids[idx++] = |
866 | 14 | (rowid_t)(fi + i + _current_row_group_idx.first_row); |
867 | 14 | } |
868 | 14 | } |
869 | 14 | read_range_rows += range.to() - range.from(); |
870 | 14 | } |
871 | 5 | return Status::OK(); |
872 | 5 | } |
873 | | |
874 | | Status RowGroupReader::_fill_row_id_columns(Block* block, size_t read_rows, |
875 | 47 | bool is_current_row_ids) { |
876 | 47 | if (_row_id_column_iterator_pair.first != nullptr) { |
877 | 5 | if (!is_current_row_ids) { |
878 | 5 | RETURN_IF_ERROR(_get_current_batch_row_id(read_rows)); |
879 | 5 | } |
880 | 5 | auto col = block->get_by_position(_row_id_column_iterator_pair.second) |
881 | 5 | .column->assume_mutable(); |
882 | 5 | RETURN_IF_ERROR(_row_id_column_iterator_pair.first->read_by_rowids( |
883 | 5 | _current_batch_row_ids.data(), _current_batch_row_ids.size(), col)); |
884 | 5 | } |
885 | | |
886 | 47 | return Status::OK(); |
887 | 47 | } |
888 | | |
889 | 44 | Status RowGroupReader::_build_pos_delete_filter(size_t read_rows) { |
890 | 44 | if (!_position_delete_ctx.has_filter) { |
891 | 44 | _pos_delete_filter_ptr.reset(nullptr); |
892 | 44 | _total_read_rows += read_rows; |
893 | 44 | return Status::OK(); |
894 | 44 | } |
895 | 0 | _pos_delete_filter_ptr.reset(new IColumn::Filter(read_rows, 1)); |
896 | 0 | auto* __restrict _pos_delete_filter_data = _pos_delete_filter_ptr->data(); |
897 | 0 | while (_position_delete_ctx.index < _position_delete_ctx.end_index) { |
898 | 0 | const int64_t delete_row_index_in_row_group = |
899 | 0 | _position_delete_ctx.delete_rows[_position_delete_ctx.index] - |
900 | 0 | _position_delete_ctx.first_row_id; |
901 | 0 | int64_t read_range_rows = 0; |
902 | 0 | size_t remaining_read_rows = _total_read_rows + read_rows; |
903 | 0 | for (size_t range_idx = 0; range_idx < _read_ranges.range_size(); range_idx++) { |
904 | 0 | auto range = _read_ranges.get_range(range_idx); |
905 | 0 | if (delete_row_index_in_row_group < range.from()) { |
906 | 0 | ++_position_delete_ctx.index; |
907 | 0 | break; |
908 | 0 | } else if (delete_row_index_in_row_group < range.to()) { |
909 | 0 | int64_t index = (delete_row_index_in_row_group - range.from()) + read_range_rows - |
910 | 0 | _total_read_rows; |
911 | 0 | if (index > read_rows - 1) { |
912 | 0 | _total_read_rows += read_rows; |
913 | 0 | return Status::OK(); |
914 | 0 | } |
915 | 0 | _pos_delete_filter_data[index] = 0; |
916 | 0 | ++_position_delete_ctx.index; |
917 | 0 | break; |
918 | 0 | } else { // delete_row >= range.last_row |
919 | 0 | } |
920 | | |
921 | 0 | int64_t range_size = range.to() - range.from(); |
922 | | // Don't search next range when there is no remaining_read_rows. |
923 | 0 | if (remaining_read_rows <= range_size) { |
924 | 0 | _total_read_rows += read_rows; |
925 | 0 | return Status::OK(); |
926 | 0 | } else { |
927 | 0 | remaining_read_rows -= range_size; |
928 | 0 | read_range_rows += range_size; |
929 | 0 | } |
930 | 0 | } |
931 | 0 | } |
932 | 0 | _total_read_rows += read_rows; |
933 | 0 | return Status::OK(); |
934 | 0 | } |
935 | | |
936 | | // need exception safety |
937 | | Status RowGroupReader::_filter_block(Block* block, int column_to_keep, |
938 | 32 | const std::vector<uint32_t>& columns_to_filter) { |
939 | 32 | if (_pos_delete_filter_ptr) { |
940 | 0 | RETURN_IF_CATCH_EXCEPTION( |
941 | 0 | Block::filter_block_internal(block, columns_to_filter, (*_pos_delete_filter_ptr))); |
942 | 0 | } |
943 | 32 | Block::erase_useless_column(block, column_to_keep); |
944 | | |
945 | 32 | return Status::OK(); |
946 | 32 | } |
947 | | |
948 | 6 | Status RowGroupReader::_rewrite_dict_predicates() { |
949 | 6 | SCOPED_RAW_TIMER(&_dict_filter_rewrite_time); |
950 | 6 | for (auto it = _dict_filter_cols.begin(); it != _dict_filter_cols.end();) { |
951 | 2 | std::string& dict_filter_col_name = it->first; |
952 | 2 | int slot_id = it->second; |
953 | | // 1. Get dictionary values to a string column. |
954 | 2 | MutableColumnPtr dict_value_column = ColumnString::create(); |
955 | 2 | bool has_dict = false; |
956 | 2 | RETURN_IF_ERROR(_column_readers[dict_filter_col_name]->read_dict_values_to_column( |
957 | 2 | dict_value_column, &has_dict)); |
958 | 2 | #ifndef NDEBUG |
959 | 2 | dict_value_column->sanity_check(); |
960 | 2 | #endif |
961 | 2 | size_t dict_value_column_size = dict_value_column->size(); |
962 | 2 | DCHECK(has_dict); |
963 | | // 2. Build a temp block from the dict string column, then execute conjuncts and filter block. |
964 | | // 2.1 Build a temp block from the dict string column to match the conjuncts executing. |
965 | 2 | Block temp_block; |
966 | 2 | int dict_pos = -1; |
967 | 2 | int index = 0; |
968 | 4 | for (const auto slot_desc : _tuple_descriptor->slots()) { |
969 | 4 | if (slot_desc->id() == slot_id) { |
970 | 2 | auto data_type = slot_desc->get_data_type_ptr(); |
971 | 2 | if (data_type->is_nullable()) { |
972 | 0 | temp_block.insert( |
973 | 0 | {ColumnNullable::create( |
974 | 0 | std::move( |
975 | 0 | dict_value_column), // NOLINT(bugprone-use-after-move) |
976 | 0 | ColumnUInt8::create(dict_value_column_size, 0)), |
977 | 0 | std::make_shared<DataTypeNullable>(std::make_shared<DataTypeString>()), |
978 | 0 | ""}); |
979 | 2 | } else { |
980 | 2 | temp_block.insert( |
981 | 2 | {std::move(dict_value_column), std::make_shared<DataTypeString>(), ""}); |
982 | 2 | } |
983 | 2 | dict_pos = index; |
984 | | |
985 | 2 | } else { |
986 | 2 | temp_block.insert(ColumnWithTypeAndName(slot_desc->get_empty_mutable_column(), |
987 | 2 | slot_desc->get_data_type_ptr(), |
988 | 2 | slot_desc->col_name())); |
989 | 2 | } |
990 | 4 | ++index; |
991 | 4 | } |
992 | | |
993 | | // 2.2 Execute conjuncts. |
994 | 2 | VExprContextSPtrs ctxs; |
995 | 2 | auto iter = _slot_id_to_filter_conjuncts->find(slot_id); |
996 | 2 | if (iter != _slot_id_to_filter_conjuncts->end()) { |
997 | 2 | for (auto& ctx : iter->second) { |
998 | 2 | ctxs.push_back(ctx); |
999 | 2 | } |
1000 | 2 | } else { |
1001 | 0 | std::stringstream msg; |
1002 | 0 | msg << "_slot_id_to_filter_conjuncts: slot_id [" << slot_id << "] not found"; |
1003 | 0 | return Status::NotFound(msg.str()); |
1004 | 0 | } |
1005 | | |
1006 | 2 | if (dict_pos != 0) { |
1007 | | // VExprContext.execute has an optimization, the filtering is executed when block->rows() > 0 |
1008 | | // The following process may be tricky and time-consuming, but we have no other way. |
1009 | 0 | temp_block.get_by_position(0).column->assume_mutable()->resize(dict_value_column_size); |
1010 | 0 | } |
1011 | 2 | IColumn::Filter result_filter(temp_block.rows(), 1); |
1012 | 2 | bool can_filter_all; |
1013 | 2 | { |
1014 | 2 | RETURN_IF_ERROR(VExprContext::execute_conjuncts(ctxs, nullptr, &temp_block, |
1015 | 2 | &result_filter, &can_filter_all)); |
1016 | 2 | } |
1017 | 2 | if (dict_pos != 0) { |
1018 | | // We have to clean the first column to insert right data. |
1019 | 0 | temp_block.get_by_position(0).column->assume_mutable()->clear(); |
1020 | 0 | } |
1021 | | |
1022 | | // If can_filter_all = true, can filter this row group. |
1023 | 2 | if (can_filter_all) { |
1024 | 2 | _is_row_group_filtered = true; |
1025 | 2 | return Status::OK(); |
1026 | 2 | } |
1027 | | |
1028 | | // 3. Get dict codes. |
1029 | 0 | std::vector<int32_t> dict_codes; |
1030 | 0 | for (size_t i = 0; i < result_filter.size(); ++i) { |
1031 | 0 | if (result_filter[i]) { |
1032 | 0 | dict_codes.emplace_back(i); |
1033 | 0 | } |
1034 | 0 | } |
1035 | | |
1036 | | // About Performance: if dict_column size is too large, it will generate a large IN filter. |
1037 | 0 | if (dict_codes.size() > MAX_DICT_CODE_PREDICATE_TO_REWRITE) { |
1038 | 0 | it = _dict_filter_cols.erase(it); |
1039 | 0 | for (auto& ctx : ctxs) { |
1040 | 0 | _filter_conjuncts.push_back(ctx); |
1041 | 0 | } |
1042 | 0 | continue; |
1043 | 0 | } |
1044 | | |
1045 | | // 4. Rewrite conjuncts. |
1046 | 0 | RETURN_IF_ERROR(_rewrite_dict_conjuncts( |
1047 | 0 | dict_codes, slot_id, temp_block.get_by_position(dict_pos).column->is_nullable())); |
1048 | 0 | ++it; |
1049 | 0 | } |
1050 | 4 | return Status::OK(); |
1051 | 6 | } |
1052 | | |
1053 | | Status RowGroupReader::_rewrite_dict_conjuncts(std::vector<int32_t>& dict_codes, int slot_id, |
1054 | 0 | bool is_nullable) { |
1055 | 0 | VExprSPtr root; |
1056 | 0 | if (dict_codes.size() == 1) { |
1057 | 0 | { |
1058 | 0 | TFunction fn; |
1059 | 0 | TFunctionName fn_name; |
1060 | 0 | fn_name.__set_db_name(""); |
1061 | 0 | fn_name.__set_function_name("eq"); |
1062 | 0 | fn.__set_name(fn_name); |
1063 | 0 | fn.__set_binary_type(TFunctionBinaryType::BUILTIN); |
1064 | 0 | std::vector<TTypeDesc> arg_types; |
1065 | 0 | arg_types.push_back(create_type_desc(PrimitiveType::TYPE_INT)); |
1066 | 0 | arg_types.push_back(create_type_desc(PrimitiveType::TYPE_INT)); |
1067 | 0 | fn.__set_arg_types(arg_types); |
1068 | 0 | fn.__set_ret_type(create_type_desc(PrimitiveType::TYPE_BOOLEAN)); |
1069 | 0 | fn.__set_has_var_args(false); |
1070 | |
|
1071 | 0 | TExprNode texpr_node; |
1072 | 0 | texpr_node.__set_type(create_type_desc(PrimitiveType::TYPE_BOOLEAN)); |
1073 | 0 | texpr_node.__set_node_type(TExprNodeType::BINARY_PRED); |
1074 | 0 | texpr_node.__set_opcode(TExprOpcode::EQ); |
1075 | 0 | texpr_node.__set_fn(fn); |
1076 | 0 | texpr_node.__set_num_children(2); |
1077 | 0 | texpr_node.__set_is_nullable(is_nullable); |
1078 | 0 | root = VectorizedFnCall::create_shared(texpr_node); |
1079 | 0 | } |
1080 | 0 | { |
1081 | 0 | SlotDescriptor* slot = nullptr; |
1082 | 0 | const std::vector<SlotDescriptor*>& slots = _tuple_descriptor->slots(); |
1083 | 0 | for (auto each : slots) { |
1084 | 0 | if (each->id() == slot_id) { |
1085 | 0 | slot = each; |
1086 | 0 | break; |
1087 | 0 | } |
1088 | 0 | } |
1089 | 0 | root->add_child(VSlotRef::create_shared(slot)); |
1090 | 0 | } |
1091 | 0 | { |
1092 | 0 | TExprNode texpr_node; |
1093 | 0 | texpr_node.__set_node_type(TExprNodeType::INT_LITERAL); |
1094 | 0 | texpr_node.__set_type(create_type_desc(TYPE_INT)); |
1095 | 0 | TIntLiteral int_literal; |
1096 | 0 | int_literal.__set_value(dict_codes[0]); |
1097 | 0 | texpr_node.__set_int_literal(int_literal); |
1098 | 0 | texpr_node.__set_is_nullable(is_nullable); |
1099 | 0 | root->add_child(VLiteral::create_shared(texpr_node)); |
1100 | 0 | } |
1101 | 0 | } else { |
1102 | 0 | { |
1103 | 0 | TTypeDesc type_desc = create_type_desc(PrimitiveType::TYPE_BOOLEAN); |
1104 | 0 | TExprNode node; |
1105 | 0 | node.__set_type(type_desc); |
1106 | 0 | node.__set_node_type(TExprNodeType::IN_PRED); |
1107 | 0 | node.in_predicate.__set_is_not_in(false); |
1108 | 0 | node.__set_opcode(TExprOpcode::FILTER_IN); |
1109 | | // VdirectInPredicate assume is_nullable = false. |
1110 | 0 | node.__set_is_nullable(false); |
1111 | |
|
1112 | 0 | std::shared_ptr<HybridSetBase> hybrid_set( |
1113 | 0 | create_set(PrimitiveType::TYPE_INT, dict_codes.size(), false)); |
1114 | 0 | for (int j = 0; j < dict_codes.size(); ++j) { |
1115 | 0 | hybrid_set->insert(&dict_codes[j]); |
1116 | 0 | } |
1117 | 0 | root = VDirectInPredicate::create_shared(node, hybrid_set); |
1118 | 0 | } |
1119 | 0 | { |
1120 | 0 | SlotDescriptor* slot = nullptr; |
1121 | 0 | const std::vector<SlotDescriptor*>& slots = _tuple_descriptor->slots(); |
1122 | 0 | for (auto each : slots) { |
1123 | 0 | if (each->id() == slot_id) { |
1124 | 0 | slot = each; |
1125 | 0 | break; |
1126 | 0 | } |
1127 | 0 | } |
1128 | 0 | root->add_child(VSlotRef::create_shared(slot)); |
1129 | 0 | } |
1130 | 0 | } |
1131 | 0 | VExprContextSPtr rewritten_conjunct_ctx = VExprContext::create_shared(root); |
1132 | 0 | RETURN_IF_ERROR(rewritten_conjunct_ctx->prepare(_state, *_row_descriptor)); |
1133 | 0 | RETURN_IF_ERROR(rewritten_conjunct_ctx->open(_state)); |
1134 | 0 | _dict_filter_conjuncts.push_back(rewritten_conjunct_ctx); |
1135 | 0 | _filter_conjuncts.push_back(rewritten_conjunct_ctx); |
1136 | 0 | return Status::OK(); |
1137 | 0 | } |
1138 | | |
1139 | 42 | Status RowGroupReader::_convert_dict_cols_to_string_cols(Block* block) { |
1140 | 42 | for (auto& dict_filter_cols : _dict_filter_cols) { |
1141 | 0 | if (!_col_name_to_block_idx->contains(dict_filter_cols.first)) { |
1142 | 0 | throw Exception(ErrorCode::INTERNAL_ERROR, |
1143 | 0 | "Wrong read column '{}' in parquet file, block: {}", |
1144 | 0 | dict_filter_cols.first, block->dump_structure()); |
1145 | 0 | } |
1146 | 0 | ColumnWithTypeAndName& column_with_type_and_name = |
1147 | 0 | block->get_by_position((*_col_name_to_block_idx)[dict_filter_cols.first]); |
1148 | 0 | const ColumnPtr& column = column_with_type_and_name.column; |
1149 | 0 | if (const auto* nullable_column = check_and_get_column<ColumnNullable>(*column)) { |
1150 | 0 | const ColumnPtr& nested_column = nullable_column->get_nested_column_ptr(); |
1151 | 0 | const auto* dict_column = assert_cast<const ColumnInt32*>(nested_column.get()); |
1152 | 0 | DCHECK(dict_column); |
1153 | |
|
1154 | 0 | auto string_column = DORIS_TRY( |
1155 | 0 | _column_readers[dict_filter_cols.first]->convert_dict_column_to_string_column( |
1156 | 0 | dict_column)); |
1157 | |
|
1158 | 0 | column_with_type_and_name.type = |
1159 | 0 | std::make_shared<DataTypeNullable>(std::make_shared<DataTypeString>()); |
1160 | 0 | block->replace_by_position( |
1161 | 0 | (*_col_name_to_block_idx)[dict_filter_cols.first], |
1162 | 0 | ColumnNullable::create(std::move(string_column), |
1163 | 0 | nullable_column->get_null_map_column_ptr())); |
1164 | 0 | } else { |
1165 | 0 | const auto* dict_column = assert_cast<const ColumnInt32*>(column.get()); |
1166 | 0 | auto string_column = DORIS_TRY( |
1167 | 0 | _column_readers[dict_filter_cols.first]->convert_dict_column_to_string_column( |
1168 | 0 | dict_column)); |
1169 | |
|
1170 | 0 | column_with_type_and_name.type = std::make_shared<DataTypeString>(); |
1171 | 0 | block->replace_by_position((*_col_name_to_block_idx)[dict_filter_cols.first], |
1172 | 0 | std::move(string_column)); |
1173 | 0 | } |
1174 | 0 | } |
1175 | 42 | return Status::OK(); |
1176 | 42 | } |
1177 | | |
1178 | 35 | ParquetColumnReader::ColumnStatistics RowGroupReader::merged_column_statistics() { |
1179 | 35 | ParquetColumnReader::ColumnStatistics st; |
1180 | 104 | for (auto& reader : _column_readers) { |
1181 | 104 | auto ost = reader.second->column_statistics(); |
1182 | 104 | st.merge(ost); |
1183 | 104 | } |
1184 | 35 | return st; |
1185 | 35 | } |
1186 | | #include "common/compile_check_end.h" |
1187 | | |
1188 | | } // namespace doris |