Coverage Report

Created: 2026-06-03 13:22

next uncovered line (L), next uncovered region (R), next uncovered branch (B)
be/src/storage/segment/column_writer.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/segment/column_writer.h"
19
20
#include <gen_cpp/segment_v2.pb.h>
21
22
#include <algorithm>
23
#include <cstring>
24
#include <filesystem>
25
#include <memory>
26
27
#include "common/config.h"
28
#include "common/logging.h"
29
#include "core/data_type/data_type_agg_state.h"
30
#include "core/data_type/data_type_factory.hpp"
31
#include "core/types.h"
32
#include "io/fs/file_writer.h"
33
#include "storage/index/bloom_filter/bloom_filter_index_writer.h"
34
#include "storage/index/inverted/inverted_index_writer.h"
35
#include "storage/index/ordinal_page_index.h"
36
#include "storage/index/zone_map/zone_map_index.h"
37
#include "storage/olap_common.h"
38
#include "storage/segment/encoding_info.h"
39
#include "storage/segment/options.h"
40
#include "storage/segment/page_builder.h"
41
#include "storage/segment/page_io.h"
42
#include "storage/segment/page_pointer.h"
43
#include "storage/segment/variant/variant_column_writer_impl.h"
44
#include "storage/tablet/tablet_schema.h"
45
#include "storage/types.h"
46
#include "util/block_compression.h"
47
#include "util/debug_points.h"
48
#include "util/faststring.h"
49
#include "util/rle_encoding.h"
50
#include "util/simd/bits.h"
51
52
namespace doris::segment_v2 {
53
54
class NullBitmapBuilder {
55
public:
56
464k
    NullBitmapBuilder() : _has_null(false), _bitmap_buf(512), _rle_encoder(&_bitmap_buf, 1) {}
57
58
    explicit NullBitmapBuilder(size_t reserve_bits)
59
            : _has_null(false),
60
              _bitmap_buf(BitmapSize(reserve_bits)),
61
0
              _rle_encoder(&_bitmap_buf, 1) {}
62
63
2.96M
    void reserve_for_write(size_t num_rows, size_t non_null_count) {
64
2.96M
        if (num_rows == 0) {
65
0
            return;
66
0
        }
67
2.96M
        if (non_null_count == 0 || (non_null_count == num_rows && !_has_null)) {
68
372k
            if (_bitmap_buf.capacity() < kSmallReserveBytes) {
69
0
                _bitmap_buf.reserve(kSmallReserveBytes);
70
0
            }
71
372k
            return;
72
372k
        }
73
2.58M
        size_t raw_bytes = BitmapSize(num_rows);
74
2.58M
        size_t run_est = std::min(num_rows, non_null_count * 2 + 1);
75
2.58M
        size_t run_bytes_est = run_est * kBytesPerRun + kReserveSlackBytes;
76
2.58M
        size_t raw_overhead = raw_bytes / 63 + 1;
77
2.58M
        size_t raw_est = raw_bytes + raw_overhead + kReserveSlackBytes;
78
2.58M
        size_t reserve_bytes = std::min(raw_est, run_bytes_est);
79
2.58M
        if (_bitmap_buf.capacity() < reserve_bytes) {
80
11.0k
            const size_t cap = _bitmap_buf.capacity();
81
11.0k
            const size_t grow = cap + cap / 2;
82
11.0k
            const size_t new_cap = std::max(reserve_bytes, grow);
83
11.0k
            _bitmap_buf.reserve(new_cap);
84
11.0k
        }
85
2.58M
    }
86
87
6.73M
    void add_run(bool value, size_t run) {
88
6.73M
        _has_null |= value;
89
6.73M
        _rle_encoder.Put(value, run);
90
6.73M
    }
91
92
    // Returns whether the building nullmap contains nullptr
93
463k
    bool has_null() const { return _has_null; }
94
95
132k
    Status finish(OwnedSlice* slice) {
96
132k
        _rle_encoder.Flush();
97
132k
        RETURN_IF_CATCH_EXCEPTION({ *slice = _bitmap_buf.build(); });
98
132k
        return Status::OK();
99
132k
    }
100
101
463k
    void reset() {
102
463k
        _has_null = false;
103
463k
        _rle_encoder.Clear();
104
463k
    }
105
106
29.0k
    uint64_t size() { return _bitmap_buf.size(); }
107
108
private:
109
    static constexpr size_t kSmallReserveBytes = 64;
110
    static constexpr size_t kReserveSlackBytes = 16;
111
    static constexpr size_t kBytesPerRun = 6;
112
113
    bool _has_null;
114
    faststring _bitmap_buf;
115
    RleEncoder<bool> _rle_encoder;
116
};
117
118
inline ScalarColumnWriter* get_null_writer(const ColumnWriterOptions& opts,
119
70.3k
                                           io::FileWriter* file_writer, uint32_t id) {
120
70.3k
    if (!opts.meta->is_nullable()) {
121
33.6k
        return nullptr;
122
33.6k
    }
123
124
36.6k
    FieldType null_type = FieldType::OLAP_FIELD_TYPE_TINYINT;
125
36.6k
    ColumnWriterOptions null_options;
126
36.6k
    null_options.meta = opts.meta->add_children_columns();
127
36.6k
    null_options.meta->set_column_id(id);
128
36.6k
    null_options.meta->set_unique_id(id);
129
36.6k
    null_options.meta->set_type(int(null_type));
130
36.6k
    null_options.meta->set_is_nullable(false);
131
36.6k
    null_options.meta->set_length(
132
36.6k
            cast_set<int32_t>(field_type_size(FieldType::OLAP_FIELD_TYPE_TINYINT)));
133
36.6k
    null_options.meta->set_compression(opts.meta->compression());
134
135
36.6k
    null_options.need_zone_map = false;
136
36.6k
    null_options.need_bloom_filter = false;
137
36.6k
    null_options.storage_format = opts.storage_format;
138
139
36.6k
    auto null_column_ptr = std::make_shared<TabletColumn>(
140
36.6k
            FieldAggregationMethod::OLAP_FIELD_AGGREGATION_NONE, null_type, false,
141
36.6k
            null_options.meta->unique_id(), null_options.meta->length());
142
36.6k
    null_column_ptr->set_name("nullable");
143
36.6k
    null_column_ptr->set_index_length(-1); // no short key index
144
36.6k
    null_options.meta->set_encoding(
145
36.6k
            EncodingInfo::resolve_default_encoding(opts.storage_format, *null_column_ptr));
146
36.6k
    return new ScalarColumnWriter(null_options, std::move(null_column_ptr), file_writer);
147
70.3k
}
148
149
ColumnWriter::ColumnWriter(TabletColumnPtr column, bool is_nullable, ColumnMetaPB* meta)
150
963k
        : _column(std::move(column)), _is_nullable(is_nullable), _column_meta(meta) {
151
963k
    _data_type = DataTypeFactory::instance().create_data_type(*_column_meta);
152
963k
}
153
Status ColumnWriter::create_struct_writer(const ColumnWriterOptions& opts,
154
                                          const TabletColumn* column, io::FileWriter* file_writer,
155
2.12k
                                          std::unique_ptr<ColumnWriter>* writer) {
156
    // not support empty struct
157
2.12k
    DCHECK(column->get_subtype_count() >= 1);
158
2.12k
    std::vector<std::unique_ptr<ColumnWriter>> sub_column_writers;
159
2.12k
    sub_column_writers.reserve(column->get_subtype_count());
160
11.3k
    for (uint32_t i = 0; i < column->get_subtype_count(); i++) {
161
9.27k
        const TabletColumn& sub_column = column->get_sub_column(i);
162
9.27k
        RETURN_IF_ERROR(sub_column.check_valid());
163
164
        // create sub writer
165
9.27k
        ColumnWriterOptions column_options;
166
9.27k
        column_options.meta = opts.meta->mutable_children_columns(i);
167
9.27k
        column_options.need_zone_map = false;
168
9.27k
        column_options.need_bloom_filter = sub_column.is_bf_column();
169
9.27k
        column_options.storage_format = opts.storage_format;
170
9.27k
        std::unique_ptr<ColumnWriter> sub_column_writer;
171
9.27k
        RETURN_IF_ERROR(
172
9.27k
                ColumnWriter::create(column_options, &sub_column, file_writer, &sub_column_writer));
173
9.27k
        sub_column_writers.push_back(std::move(sub_column_writer));
174
9.27k
    }
175
176
2.12k
    ScalarColumnWriter* null_writer =
177
2.12k
            get_null_writer(opts, file_writer, column->get_subtype_count() + 1);
178
179
2.12k
    *writer = std::unique_ptr<ColumnWriter>(new StructColumnWriter(
180
2.12k
            opts, std::make_shared<TabletColumn>(*column), null_writer, sub_column_writers));
181
2.12k
    return Status::OK();
182
2.12k
}
183
184
Status ColumnWriter::create_array_writer(const ColumnWriterOptions& opts,
185
                                         const TabletColumn* column, io::FileWriter* file_writer,
186
42.5k
                                         std::unique_ptr<ColumnWriter>* writer) {
187
42.5k
    DCHECK(column->get_subtype_count() == 1);
188
42.5k
    const TabletColumn& item_column = column->get_sub_column(0);
189
42.5k
    RETURN_IF_ERROR(item_column.check_valid());
190
191
    // create item writer
192
42.5k
    ColumnWriterOptions item_options;
193
42.5k
    item_options.meta = opts.meta->mutable_children_columns(0);
194
42.5k
    item_options.need_zone_map = false;
195
42.5k
    item_options.need_bloom_filter = item_column.is_bf_column();
196
42.5k
    item_options.storage_format = opts.storage_format;
197
42.5k
    std::unique_ptr<ColumnWriter> item_writer;
198
42.5k
    RETURN_IF_ERROR(ColumnWriter::create(item_options, &item_column, file_writer, &item_writer));
199
200
    // create length writer
201
42.5k
    FieldType length_type = FieldType::OLAP_FIELD_TYPE_UNSIGNED_BIGINT;
202
203
42.5k
    ColumnWriterOptions length_options;
204
42.5k
    length_options.meta = opts.meta->add_children_columns();
205
42.5k
    length_options.meta->set_column_id(2);
206
42.5k
    length_options.meta->set_unique_id(2);
207
42.5k
    length_options.meta->set_type(int(length_type));
208
42.5k
    length_options.meta->set_is_nullable(false);
209
42.5k
    length_options.meta->set_length(
210
42.5k
            cast_set<int32_t>(field_type_size(FieldType::OLAP_FIELD_TYPE_UNSIGNED_BIGINT)));
211
42.5k
    length_options.meta->set_compression(opts.meta->compression());
212
213
42.5k
    length_options.need_zone_map = false;
214
42.5k
    length_options.need_bloom_filter = false;
215
42.5k
    length_options.storage_format = opts.storage_format;
216
217
42.5k
    auto length_column_ptr = std::make_shared<TabletColumn>(
218
42.5k
            FieldAggregationMethod::OLAP_FIELD_AGGREGATION_NONE, length_type,
219
42.5k
            length_options.meta->is_nullable(), length_options.meta->unique_id(),
220
42.5k
            length_options.meta->length());
221
42.5k
    length_column_ptr->set_name("length");
222
42.5k
    length_column_ptr->set_index_length(-1); // no short key index
223
42.5k
    length_options.meta->set_encoding(
224
42.5k
            EncodingInfo::resolve_default_encoding(opts.storage_format, *length_column_ptr));
225
42.5k
    auto* length_writer =
226
42.5k
            new OffsetColumnWriter(length_options, std::move(length_column_ptr), file_writer);
227
228
42.5k
    ScalarColumnWriter* null_writer = get_null_writer(opts, file_writer, 3);
229
230
42.5k
    *writer = std::unique_ptr<ColumnWriter>(
231
42.5k
            new ArrayColumnWriter(opts, std::make_shared<TabletColumn>(*column), length_writer,
232
42.5k
                                  null_writer, std::move(item_writer)));
233
42.5k
    return Status::OK();
234
42.5k
}
235
236
Status ColumnWriter::create_map_writer(const ColumnWriterOptions& opts, const TabletColumn* column,
237
                                       io::FileWriter* file_writer,
238
25.6k
                                       std::unique_ptr<ColumnWriter>* writer) {
239
25.6k
    DCHECK(column->get_subtype_count() == 2);
240
25.6k
    if (column->get_subtype_count() < 2) {
241
0
        return Status::InternalError(
242
0
                "If you upgraded from version 1.2.*, please DROP the MAP columns and then "
243
0
                "ADD the MAP columns back.");
244
0
    }
245
    // create key & value writer
246
25.6k
    std::vector<std::unique_ptr<ColumnWriter>> inner_writer_list;
247
76.8k
    for (int i = 0; i < 2; ++i) {
248
51.2k
        const TabletColumn& item_column = column->get_sub_column(i);
249
51.2k
        RETURN_IF_ERROR(item_column.check_valid());
250
251
        // create item writer
252
51.2k
        ColumnWriterOptions item_options;
253
51.2k
        item_options.meta = opts.meta->mutable_children_columns(i);
254
51.2k
        item_options.need_zone_map = false;
255
51.2k
        item_options.need_bloom_filter = item_column.is_bf_column();
256
51.2k
        item_options.storage_format = opts.storage_format;
257
51.2k
        std::unique_ptr<ColumnWriter> item_writer;
258
51.2k
        RETURN_IF_ERROR(
259
51.2k
                ColumnWriter::create(item_options, &item_column, file_writer, &item_writer));
260
51.2k
        inner_writer_list.push_back(std::move(item_writer));
261
51.2k
    }
262
263
    // create offset writer
264
25.6k
    FieldType length_type = FieldType::OLAP_FIELD_TYPE_UNSIGNED_BIGINT;
265
266
    // Be Cautious: column unique id is used for column reader creation
267
25.6k
    ColumnWriterOptions length_options;
268
25.6k
    length_options.meta = opts.meta->add_children_columns();
269
25.6k
    length_options.meta->set_column_id(column->get_subtype_count() + 1);
270
25.6k
    length_options.meta->set_unique_id(column->get_subtype_count() + 1);
271
25.6k
    length_options.meta->set_type(int(length_type));
272
25.6k
    length_options.meta->set_is_nullable(false);
273
25.6k
    length_options.meta->set_length(
274
25.6k
            cast_set<int32_t>(field_type_size(FieldType::OLAP_FIELD_TYPE_UNSIGNED_BIGINT)));
275
25.6k
    length_options.meta->set_compression(opts.meta->compression());
276
277
25.6k
    length_options.need_zone_map = false;
278
25.6k
    length_options.need_bloom_filter = false;
279
25.6k
    length_options.storage_format = opts.storage_format;
280
281
25.6k
    auto length_column_ptr = std::make_shared<TabletColumn>(
282
25.6k
            FieldAggregationMethod::OLAP_FIELD_AGGREGATION_NONE, length_type,
283
25.6k
            length_options.meta->is_nullable(), length_options.meta->unique_id(),
284
25.6k
            length_options.meta->length());
285
25.6k
    length_column_ptr->set_name("length");
286
25.6k
    length_column_ptr->set_index_length(-1); // no short key index
287
25.6k
    length_options.meta->set_encoding(
288
25.6k
            EncodingInfo::resolve_default_encoding(opts.storage_format, *length_column_ptr));
289
25.6k
    auto* length_writer =
290
25.6k
            new OffsetColumnWriter(length_options, std::move(length_column_ptr), file_writer);
291
292
25.6k
    ScalarColumnWriter* null_writer =
293
25.6k
            get_null_writer(opts, file_writer, column->get_subtype_count() + 2);
294
295
25.6k
    *writer = std::unique_ptr<ColumnWriter>(
296
25.6k
            new MapColumnWriter(opts, std::make_shared<TabletColumn>(*column), null_writer,
297
25.6k
                                length_writer, inner_writer_list));
298
299
25.6k
    return Status::OK();
300
25.6k
}
301
302
Status ColumnWriter::create_agg_state_writer(const ColumnWriterOptions& opts,
303
                                             const TabletColumn* column,
304
                                             io::FileWriter* file_writer,
305
1.27k
                                             std::unique_ptr<ColumnWriter>* writer) {
306
1.27k
    auto data_type = DataTypeFactory::instance().create_data_type(*column);
307
1.27k
    const auto* agg_state_type = assert_cast<const DataTypeAggState*>(data_type.get());
308
1.27k
    auto type = agg_state_type->get_serialized_type()->get_primitive_type();
309
1.27k
    if (type == PrimitiveType::TYPE_STRING || type == PrimitiveType::INVALID_TYPE ||
310
1.27k
        type == PrimitiveType::TYPE_FIXED_LENGTH_OBJECT || type == PrimitiveType::TYPE_BITMAP) {
311
1.20k
        *writer = std::unique_ptr<ColumnWriter>(
312
1.20k
                new ScalarColumnWriter(opts, std::make_shared<TabletColumn>(*column), file_writer));
313
1.20k
    } else if (type == PrimitiveType::TYPE_ARRAY) {
314
22
        RETURN_IF_ERROR(create_array_writer(opts, column, file_writer, writer));
315
53
    } else if (type == PrimitiveType::TYPE_MAP) {
316
53
        RETURN_IF_ERROR(create_map_writer(opts, column, file_writer, writer));
317
18.4E
    } else {
318
18.4E
        throw Exception(ErrorCode::INTERNAL_ERROR,
319
18.4E
                        "OLAP_FIELD_TYPE_AGG_STATE meet unsupported type: {}",
320
18.4E
                        agg_state_type->get_name());
321
18.4E
    }
322
1.27k
    return Status::OK();
323
1.27k
}
324
325
Status ColumnWriter::create_variant_writer(const ColumnWriterOptions& opts,
326
                                           const TabletColumn* column, io::FileWriter* file_writer,
327
6.53k
                                           std::unique_ptr<ColumnWriter>* writer) {
328
    // Variant extracted columns have two kinds of physical writers:
329
    // - Doc-value snapshot column (`...__DORIS_VARIANT_DOC_VALUE__...`): use `VariantDocCompactWriter`
330
    //   to store the doc snapshot in a compact binary form.
331
    // - Regular extracted subcolumns: use `VariantSubcolumnWriter`.
332
    // The root VARIANT column itself uses `VariantColumnWriter`.
333
6.53k
    if (column->is_extracted_column()) {
334
612
        if (column->name().find(DOC_VALUE_COLUMN_PATH) != std::string::npos) {
335
513
            *writer = std::make_unique<VariantDocCompactWriter>(
336
513
                    opts, std::make_shared<TabletColumn>(*column));
337
513
            return Status::OK();
338
513
        }
339
99
        VLOG_DEBUG << "gen subwriter for " << column->path_info_ptr()->get_path();
340
99
        *writer = std::make_unique<VariantSubcolumnWriter>(opts,
341
99
                                                           std::make_shared<TabletColumn>(*column));
342
99
        return Status::OK();
343
612
    }
344
5.92k
    *writer = std::make_unique<VariantColumnWriter>(opts, std::make_shared<TabletColumn>(*column));
345
5.92k
    return Status::OK();
346
6.53k
}
347
348
//Todo(Amory): here should according nullable and offset and need sub to simply this function
349
Status ColumnWriter::create(const ColumnWriterOptions& opts, const TabletColumn* column,
350
836k
                            io::FileWriter* file_writer, std::unique_ptr<ColumnWriter>* writer) {
351
836k
    auto column_ptr = std::make_shared<TabletColumn>(*column);
352
836k
    if (is_scalar_type(column->type())) {
353
774k
        *writer = std::unique_ptr<ColumnWriter>(
354
774k
                new ScalarColumnWriter(opts, std::move(column_ptr), file_writer));
355
774k
        return Status::OK();
356
774k
    } else {
357
62.6k
        switch (column->type()) {
358
1.27k
        case FieldType::OLAP_FIELD_TYPE_AGG_STATE: {
359
1.27k
            RETURN_IF_ERROR(create_agg_state_writer(opts, column, file_writer, writer));
360
1.27k
            return Status::OK();
361
1.27k
        }
362
2.12k
        case FieldType::OLAP_FIELD_TYPE_STRUCT: {
363
2.12k
            RETURN_IF_ERROR(create_struct_writer(opts, column, file_writer, writer));
364
2.12k
            return Status::OK();
365
2.12k
        }
366
42.5k
        case FieldType::OLAP_FIELD_TYPE_ARRAY: {
367
42.5k
            RETURN_IF_ERROR(create_array_writer(opts, column, file_writer, writer));
368
42.5k
            return Status::OK();
369
42.5k
        }
370
10.3k
        case FieldType::OLAP_FIELD_TYPE_MAP: {
371
10.3k
            RETURN_IF_ERROR(create_map_writer(opts, column, file_writer, writer));
372
10.3k
            return Status::OK();
373
10.3k
        }
374
6.53k
        case FieldType::OLAP_FIELD_TYPE_VARIANT: {
375
            // Process columns with sparse column
376
6.53k
            RETURN_IF_ERROR(create_variant_writer(opts, column, file_writer, writer));
377
6.53k
            return Status::OK();
378
6.53k
        }
379
0
        default:
380
0
            return Status::NotSupported("unsupported type for ColumnWriter: {}",
381
0
                                        std::to_string(int(column_ptr->type())));
382
62.6k
        }
383
62.6k
    }
384
836k
}
385
386
Status ColumnWriter::append_nullable(const uint8_t* is_null_bits, const void* data,
387
651k
                                     size_t num_rows) {
388
651k
    const auto* ptr = (const uint8_t*)data;
389
651k
    BitmapIterator null_iter(is_null_bits, num_rows);
390
651k
    bool is_null = false;
391
651k
    size_t this_run = 0;
392
1.30M
    while ((this_run = null_iter.Next(&is_null)) > 0) {
393
651k
        if (is_null) {
394
0
            RETURN_IF_ERROR(append_nulls(this_run));
395
651k
        } else {
396
651k
            RETURN_IF_ERROR(append_data(&ptr, this_run));
397
651k
        }
398
651k
    }
399
651k
    return Status::OK();
400
651k
}
401
402
Status ColumnWriter::append_nullable(const uint8_t* null_map, const uint8_t** ptr,
403
0
                                     size_t num_rows) {
404
    // Fast path: use SIMD to detect all-NULL or all-non-NULL columns
405
0
    if (config::enable_rle_batch_put_optimization) {
406
0
        size_t non_null_count =
407
0
                simd::count_zero_num(reinterpret_cast<const int8_t*>(null_map), num_rows);
408
409
0
        if (non_null_count == 0) {
410
            // All NULL: skip run-length iteration, directly append all nulls
411
0
            RETURN_IF_ERROR(append_nulls(num_rows));
412
0
            *ptr += cell_size() * num_rows;
413
0
            return Status::OK();
414
0
        }
415
416
0
        if (non_null_count == num_rows) {
417
            // All non-NULL: skip run-length iteration, directly append all data
418
0
            return append_data(ptr, num_rows);
419
0
        }
420
0
    }
421
422
    // Mixed case or sparse optimization disabled: use run-length processing
423
0
    size_t offset = 0;
424
0
    auto next_run_step = [&]() {
425
0
        size_t step = 1;
426
0
        for (auto i = offset + 1; i < num_rows; ++i) {
427
0
            if (null_map[offset] == null_map[i]) {
428
0
                step++;
429
0
            } else {
430
0
                break;
431
0
            }
432
0
        }
433
0
        return step;
434
0
    };
435
436
0
    do {
437
0
        auto step = next_run_step();
438
0
        if (null_map[offset]) {
439
0
            RETURN_IF_ERROR(append_nulls(step));
440
0
            *ptr += cell_size() * step;
441
0
        } else {
442
            // TODO:
443
            //  1. `*ptr += cell_size() * step;` should do in this function, not append_data;
444
            //  2. support array vectorized load and ptr offset add
445
0
            RETURN_IF_ERROR(append_data(ptr, step));
446
0
        }
447
0
        offset += step;
448
0
    } while (offset < num_rows);
449
450
0
    return Status::OK();
451
0
}
452
453
3.35M
Status ColumnWriter::append(const uint8_t* nullmap, const void* data, size_t num_rows) {
454
3.35M
    assert(data && num_rows > 0);
455
3.35M
    const auto* ptr = (const uint8_t*)data;
456
3.35M
    if (nullmap) {
457
2.99M
        return append_nullable(nullmap, &ptr, num_rows);
458
2.99M
    } else {
459
356k
        return append_data(&ptr, num_rows);
460
356k
    }
461
3.35M
}
462
463
///////////////////////////////////////////////////////////////////////////////////
464
465
ScalarColumnWriter::ScalarColumnWriter(const ColumnWriterOptions& opts, TabletColumnPtr column,
466
                                       io::FileWriter* file_writer)
467
886k
        : ColumnWriter(std::move(column), opts.meta->is_nullable(), opts.meta),
468
886k
          _opts(opts),
469
886k
          _file_writer(file_writer),
470
886k
          _data_size(0) {
471
    // these opts.meta fields should be set by client
472
886k
    DCHECK(opts.meta->has_column_id());
473
886k
    DCHECK(opts.meta->has_unique_id());
474
886k
    DCHECK(opts.meta->has_type());
475
886k
    DCHECK(opts.meta->has_length());
476
886k
    DCHECK(opts.meta->has_encoding());
477
886k
    DCHECK(opts.meta->has_compression());
478
886k
    DCHECK(opts.meta->has_is_nullable());
479
886k
    DCHECK(file_writer != nullptr);
480
886k
    _inverted_index_builders.resize(_opts.inverted_indexes.size());
481
886k
}
482
483
886k
ScalarColumnWriter::~ScalarColumnWriter() {
484
    // delete all pages
485
886k
    _pages.clear();
486
886k
}
487
488
885k
Status ScalarColumnWriter::init() {
489
885k
    RETURN_IF_ERROR(get_block_compression_codec(_opts.meta->compression(), &_compress_codec));
490
491
885k
    PageBuilder* page_builder = nullptr;
492
493
    // Caller must set a concrete (non-DEFAULT) encoding on the meta before init.
494
885k
    if (_opts.meta->encoding() == DEFAULT_ENCODING) {
495
0
        return Status::InternalError(
496
0
                "ColumnMetaPB encoding is DEFAULT_ENCODING for column_id={}, type={}; caller must "
497
0
                "resolve to a concrete encoding before ScalarColumnWriter::init",
498
0
                _opts.meta->column_id(), get_column()->type());
499
0
    }
500
885k
    RETURN_IF_ERROR(
501
885k
            EncodingInfo::get(get_column()->type(), _opts.meta->encoding(), &_encoding_info));
502
    // create page builder
503
885k
    PageBuilderOptions opts;
504
885k
    opts.data_page_size = _opts.data_page_size;
505
885k
    opts.dict_page_size = _opts.dict_page_size;
506
885k
    opts.dict_binary_plain_encoding =
507
885k
            (_opts.storage_format == TabletStorageFormatPB::TABLET_STORAGE_FORMAT_V3)
508
885k
                    ? BinaryPlainEncodingTypePB::BINARY_PLAIN_ENCODING_V2
509
885k
                    : BinaryPlainEncodingTypePB::BINARY_PLAIN_ENCODING_V1;
510
885k
    RETURN_IF_ERROR(_encoding_info->create_page_builder(opts, &page_builder));
511
885k
    if (page_builder == nullptr) {
512
0
        return Status::NotSupported("Failed to create page builder for type {} and encoding {}",
513
0
                                    get_column()->type(), _opts.meta->encoding());
514
0
    }
515
18.4E
    VLOG_DEBUG << fmt::format(
516
18.4E
            "[verbose] scalar column writer init, column_id={}, type={}, encoding={}, "
517
18.4E
            "is_nullable={}",
518
18.4E
            _opts.meta->column_id(), get_column()->type(),
519
18.4E
            EncodingTypePB_Name(_opts.meta->encoding()), _opts.meta->is_nullable());
520
885k
    _page_builder.reset(page_builder);
521
    // create ordinal builder
522
885k
    _ordinal_index_builder = std::make_unique<OrdinalIndexWriter>();
523
    // create null bitmap builder
524
885k
    if (is_nullable()) {
525
464k
        _null_bitmap_builder = std::make_unique<NullBitmapBuilder>();
526
464k
    }
527
885k
    if (_opts.need_zone_map) {
528
623k
        RETURN_IF_ERROR(
529
623k
                ZoneMapIndexWriter::create(_data_type, get_column(), _zone_map_index_builder));
530
623k
    }
531
532
885k
    if (_opts.need_inverted_index) {
533
42.1k
        do {
534
84.4k
            for (size_t i = 0; i < _opts.inverted_indexes.size(); i++) {
535
42.3k
                DBUG_EXECUTE_IF("column_writer.init", {
536
42.3k
                    class InvertedIndexColumnWriterEmpty final : public IndexColumnWriter {
537
42.3k
                    public:
538
42.3k
                        Status init() override { return Status::OK(); }
539
42.3k
                        Status add_values(const std::string name, const void* values,
540
42.3k
                                          size_t count) override {
541
42.3k
                            return Status::OK();
542
42.3k
                        }
543
42.3k
                        Status add_array_values(size_t field_size, const void* value_ptr,
544
42.3k
                                                const uint8_t* null_map, const uint8_t* offsets_ptr,
545
42.3k
                                                size_t count) override {
546
42.3k
                            return Status::OK();
547
42.3k
                        }
548
42.3k
                        Status add_nulls(uint32_t count) override { return Status::OK(); }
549
42.3k
                        Status add_array_nulls(const uint8_t* null_map, size_t num_rows) override {
550
42.3k
                            return Status::OK();
551
42.3k
                        }
552
42.3k
                        Status finish() override { return Status::OK(); }
553
42.3k
                        int64_t size() const override { return 0; }
554
42.3k
                        void close_on_error() override {}
555
42.3k
                    };
556
557
42.3k
                    _inverted_index_builders[i] =
558
42.3k
                            std::make_unique<InvertedIndexColumnWriterEmpty>();
559
560
42.3k
                    break;
561
42.3k
                });
562
563
42.3k
                RETURN_IF_ERROR(IndexColumnWriter::create(
564
42.3k
                        get_column(), &_inverted_index_builders[i], _opts.index_file_writer,
565
42.3k
                        _opts.inverted_indexes[i]));
566
42.3k
            }
567
42.1k
        } while (false);
568
42.1k
    }
569
885k
    if (_opts.need_bloom_filter) {
570
5.40k
        if (_opts.is_ngram_bf_index) {
571
3.39k
            RETURN_IF_ERROR(NGramBloomFilterIndexWriterImpl::create(
572
3.39k
                    BloomFilterOptions(), get_column()->type(), _opts.gram_size, _opts.gram_bf_size,
573
3.39k
                    &_bloom_filter_index_builder));
574
3.39k
        } else {
575
2.01k
            RETURN_IF_ERROR(BloomFilterIndexWriter::create(_opts.bf_options, get_column()->type(),
576
2.01k
                                                           &_bloom_filter_index_builder));
577
2.01k
        }
578
5.40k
    }
579
885k
    return Status::OK();
580
885k
}
581
582
2.91M
Status ScalarColumnWriter::append_nulls(size_t num_rows) {
583
2.91M
    _null_bitmap_builder->add_run(true, num_rows);
584
2.91M
    _next_rowid += num_rows;
585
2.91M
    if (_opts.need_zone_map) {
586
2.83M
        _zone_map_index_builder->add_nulls(cast_set<uint32_t>(num_rows));
587
2.83M
    }
588
2.91M
    if (_opts.need_inverted_index) {
589
1.98M
        for (const auto& builder : _inverted_index_builders) {
590
1.98M
            RETURN_IF_ERROR(builder->add_nulls(cast_set<uint32_t>(num_rows)));
591
1.98M
        }
592
1.98M
    }
593
2.91M
    if (_opts.need_bloom_filter) {
594
576
        _bloom_filter_index_builder->add_nulls(cast_set<uint32_t>(num_rows));
595
576
    }
596
2.91M
    return Status::OK();
597
2.91M
}
598
599
// append data to page builder. this function will make sure that
600
// num_rows must be written before return. And ptr will be modified
601
// to next data should be written
602
4.17M
Status ScalarColumnWriter::append_data(const uint8_t** ptr, size_t num_rows) {
603
4.17M
    size_t remaining = num_rows;
604
8.41M
    while (remaining > 0) {
605
4.24M
        size_t num_written = remaining;
606
4.24M
        RETURN_IF_ERROR(append_data_in_current_page(ptr, &num_written));
607
608
4.24M
        remaining -= num_written;
609
610
4.24M
        if (_page_builder->is_page_full()) {
611
75.5k
            RETURN_IF_ERROR(finish_current_page());
612
75.5k
        }
613
4.24M
    }
614
4.17M
    return Status::OK();
615
4.17M
}
616
617
Status ScalarColumnWriter::_internal_append_data_in_current_page(const uint8_t* data,
618
4.31M
                                                                 size_t* num_written) {
619
4.31M
    RETURN_IF_ERROR(_page_builder->add(data, num_written));
620
4.31M
    if (_opts.need_zone_map) {
621
4.01M
        _zone_map_index_builder->add_values(data, *num_written);
622
4.01M
    }
623
4.31M
    if (_opts.need_inverted_index) {
624
2.01M
        for (const auto& builder : _inverted_index_builders) {
625
2.01M
            RETURN_IF_ERROR(builder->add_values(get_column()->name(), data, *num_written));
626
2.01M
        }
627
2.01M
    }
628
4.31M
    if (_opts.need_bloom_filter) {
629
5.68k
        RETURN_IF_ERROR(_bloom_filter_index_builder->add_values(data, *num_written));
630
5.68k
    }
631
632
4.31M
    _next_rowid += *num_written;
633
634
    // we must write null bits after write data, because we don't
635
    // know how many rows can be written into current page
636
4.31M
    if (is_nullable()) {
637
3.82M
        _null_bitmap_builder->add_run(false, *num_written);
638
3.82M
    }
639
4.31M
    return Status::OK();
640
4.31M
}
641
642
4.31M
Status ScalarColumnWriter::append_data_in_current_page(const uint8_t** data, size_t* num_written) {
643
4.31M
    RETURN_IF_ERROR(append_data_in_current_page(*data, num_written));
644
4.31M
    *data += cell_size() * (*num_written);
645
4.31M
    return Status::OK();
646
4.31M
}
647
648
Status ScalarColumnWriter::append_nullable(const uint8_t* null_map, const uint8_t** ptr,
649
2.95M
                                           size_t num_rows) {
650
    // When optimization is disabled, use base class implementation
651
2.95M
    if (!config::enable_rle_batch_put_optimization) {
652
0
        return ColumnWriter::append_nullable(null_map, ptr, num_rows);
653
0
    }
654
655
2.95M
    if (UNLIKELY(num_rows == 0)) {
656
0
        return Status::OK();
657
0
    }
658
659
    // Build run-length encoded null runs using memchr for fast boundary detection
660
2.95M
    _null_run_buffer.clear();
661
2.95M
    if (_null_run_buffer.capacity() < num_rows) {
662
474k
        _null_run_buffer.reserve(std::min(num_rows, size_t(256)));
663
474k
    }
664
665
2.95M
    size_t non_null_count = 0;
666
2.95M
    size_t offset = 0;
667
6.49M
    while (offset < num_rows) {
668
3.54M
        bool is_null = null_map[offset] != 0;
669
3.54M
        size_t remaining = num_rows - offset;
670
3.54M
        const uint8_t* run_end =
671
3.54M
                static_cast<const uint8_t*>(memchr(null_map + offset, is_null ? 0 : 1, remaining));
672
3.54M
        size_t run_length = run_end != nullptr ? (run_end - (null_map + offset)) : remaining;
673
3.54M
        _null_run_buffer.push_back(NullRun {is_null, static_cast<uint32_t>(run_length)});
674
3.54M
        if (!is_null) {
675
3.16M
            non_null_count += run_length;
676
3.16M
        }
677
3.54M
        offset += run_length;
678
3.54M
    }
679
680
    // Pre-allocate buffer based on estimated size
681
2.96M
    if (_null_bitmap_builder != nullptr) {
682
2.96M
        size_t current_rows = _next_rowid - _first_rowid;
683
2.96M
        size_t expected_rows = current_rows + num_rows;
684
2.96M
        size_t est_non_null = non_null_count;
685
2.96M
        if (num_rows > 0 && expected_rows > num_rows) {
686
2.53M
            est_non_null = (non_null_count * expected_rows) / num_rows;
687
2.53M
        }
688
2.96M
        _null_bitmap_builder->reserve_for_write(expected_rows, est_non_null);
689
2.96M
    }
690
691
2.95M
    if (non_null_count == 0) {
692
        // All NULL: skip data writing, only update null bitmap and indexes
693
48.6k
        RETURN_IF_ERROR(append_nulls(num_rows));
694
48.6k
        *ptr += cell_size() * num_rows;
695
48.6k
        return Status::OK();
696
48.6k
    }
697
698
2.90M
    if (non_null_count == num_rows) {
699
        // All non-NULL: use normal append_data which handles both data and null bitmap
700
2.84M
        return append_data(ptr, num_rows);
701
2.84M
    }
702
703
    // Process by runs
704
647k
    for (const auto& run : _null_run_buffer) {
705
647k
        size_t run_length = run.len;
706
647k
        if (run.is_null) {
707
330k
            RETURN_IF_ERROR(append_nulls(run_length));
708
330k
            *ptr += cell_size() * run_length;
709
330k
        } else {
710
            // TODO:
711
            //  1. `*ptr += cell_size() * step;` should do in this function, not append_data;
712
            //  2. support array vectorized load and ptr offset add
713
316k
            RETURN_IF_ERROR(append_data(ptr, run_length));
714
316k
        }
715
647k
    }
716
717
61.0k
    return Status::OK();
718
61.0k
}
719
720
43.3k
uint64_t ScalarColumnWriter::estimate_buffer_size() {
721
43.3k
    uint64_t size = _data_size;
722
43.3k
    size += _page_builder->size();
723
43.3k
    if (is_nullable()) {
724
29.0k
        size += _null_bitmap_builder->size();
725
29.0k
    }
726
43.3k
    size += _ordinal_index_builder->size();
727
43.3k
    if (_opts.need_zone_map) {
728
33.1k
        size += _zone_map_index_builder->size();
729
33.1k
    }
730
43.3k
    if (_opts.need_bloom_filter) {
731
267
        size += _bloom_filter_index_builder->size();
732
267
    }
733
43.3k
    return size;
734
43.3k
}
735
736
887k
Status ScalarColumnWriter::finish() {
737
887k
    RETURN_IF_ERROR(finish_current_page());
738
887k
    _opts.meta->set_num_rows(_next_rowid);
739
887k
    return Status::OK();
740
887k
}
741
742
887k
Status ScalarColumnWriter::write_data() {
743
887k
    auto offset = _file_writer->bytes_appended();
744
1.20M
    auto collect_uncompressed_bytes = [](const PageFooterPB& footer) {
745
1.20M
        return footer.uncompressed_size() + footer.ByteSizeLong() +
746
1.20M
               sizeof(uint32_t) /* footer size */ + sizeof(uint32_t) /* checksum */;
747
1.20M
    };
748
919k
    for (auto& page : _pages) {
749
919k
        _total_uncompressed_data_pages_size += collect_uncompressed_bytes(page->footer);
750
919k
        RETURN_IF_ERROR(_write_data_page(page.get()));
751
919k
    }
752
887k
    _pages.clear();
753
    // write column dict
754
887k
    if (_encoding_info->encoding() == DICT_ENCODING) {
755
286k
        OwnedSlice dict_body;
756
286k
        RETURN_IF_ERROR(_page_builder->get_dictionary_page(&dict_body));
757
286k
        EncodingTypePB dict_word_page_encoding;
758
286k
        RETURN_IF_ERROR(_page_builder->get_dictionary_page_encoding(&dict_word_page_encoding));
759
760
286k
        PageFooterPB footer;
761
286k
        footer.set_type(DICTIONARY_PAGE);
762
286k
        footer.set_uncompressed_size(cast_set<uint32_t>(dict_body.slice().get_size()));
763
286k
        footer.mutable_dict_page_footer()->set_encoding(dict_word_page_encoding);
764
286k
        _total_uncompressed_data_pages_size += collect_uncompressed_bytes(footer);
765
766
286k
        PagePointer dict_pp;
767
286k
        RETURN_IF_ERROR(PageIO::compress_and_write_page(
768
286k
                _compress_codec, _opts.compression_min_space_saving, _file_writer,
769
286k
                {dict_body.slice()}, footer, &dict_pp));
770
286k
        dict_pp.to_proto(_opts.meta->mutable_dict_page());
771
286k
    }
772
887k
    _total_compressed_data_pages_size += _file_writer->bytes_appended() - offset;
773
887k
    _page_builder.reset();
774
887k
    return Status::OK();
775
887k
}
776
777
843k
Status ScalarColumnWriter::write_ordinal_index() {
778
843k
    return _ordinal_index_builder->finish(_file_writer, _opts.meta->add_indexes());
779
843k
}
780
781
675k
Status ScalarColumnWriter::write_zone_map() {
782
675k
    if (_opts.need_zone_map) {
783
622k
        return _zone_map_index_builder->finish(_file_writer, _opts.meta->add_indexes());
784
622k
    }
785
52.4k
    return Status::OK();
786
675k
}
787
788
620k
Status ScalarColumnWriter::write_inverted_index() {
789
620k
    if (_opts.need_inverted_index) {
790
42.3k
        for (const auto& builder : _inverted_index_builders) {
791
42.3k
            RETURN_IF_ERROR(builder->finish());
792
42.3k
        }
793
42.1k
    }
794
620k
    return Status::OK();
795
620k
}
796
797
609k
Status ScalarColumnWriter::write_bloom_filter_index() {
798
609k
    if (_opts.need_bloom_filter) {
799
5.40k
        return _bloom_filter_index_builder->finish(_file_writer, _opts.meta->add_indexes());
800
5.40k
    }
801
604k
    return Status::OK();
802
609k
}
803
804
// write a data page into file and update ordinal index
805
919k
Status ScalarColumnWriter::_write_data_page(Page* page) {
806
919k
    PagePointer pp;
807
919k
    std::vector<Slice> compressed_body;
808
1.73M
    for (auto& data : page->data) {
809
1.73M
        compressed_body.push_back(data.slice());
810
1.73M
    }
811
919k
    RETURN_IF_ERROR(PageIO::write_page(_file_writer, compressed_body, page->footer, &pp));
812
919k
    _ordinal_index_builder->append_entry(page->footer.data_page_footer().first_ordinal(), pp);
813
919k
    return Status::OK();
814
919k
}
815
816
963k
Status ScalarColumnWriter::finish_current_page() {
817
963k
    if (_next_rowid == _first_rowid) {
818
43.4k
        return Status::OK();
819
43.4k
    }
820
919k
    if (_opts.need_zone_map) {
821
        // If the number of rows in the current page is less than the threshold,
822
        // we will invalidate zone map index for this page by set pass_all to true.
823
686k
        if (_next_rowid - _first_rowid < config::zone_map_row_num_threshold) {
824
484k
            _zone_map_index_builder->invalid_page_zone_map();
825
484k
        }
826
686k
        RETURN_IF_ERROR(_zone_map_index_builder->flush());
827
686k
    }
828
829
919k
    if (_opts.need_bloom_filter) {
830
5.64k
        RETURN_IF_ERROR(_bloom_filter_index_builder->flush());
831
5.64k
    }
832
833
919k
    _raw_data_bytes += _page_builder->get_raw_data_size();
834
835
    // build data page body : encoded values + [nullmap]
836
919k
    std::vector<Slice> body;
837
919k
    OwnedSlice encoded_values;
838
919k
    RETURN_IF_ERROR(_page_builder->finish(&encoded_values));
839
919k
    RETURN_IF_ERROR(_page_builder->reset());
840
919k
    body.push_back(encoded_values.slice());
841
842
919k
    OwnedSlice nullmap;
843
919k
    if (_null_bitmap_builder != nullptr) {
844
463k
        if (is_nullable() && _null_bitmap_builder->has_null()) {
845
132k
            RETURN_IF_ERROR(_null_bitmap_builder->finish(&nullmap));
846
132k
            body.push_back(nullmap.slice());
847
132k
        }
848
463k
        _null_bitmap_builder->reset();
849
463k
    }
850
851
    // prepare data page footer
852
919k
    std::unique_ptr<Page> page(new Page());
853
919k
    page->footer.set_type(DATA_PAGE);
854
919k
    page->footer.set_uncompressed_size(cast_set<uint32_t>(Slice::compute_total_size(body)));
855
919k
    auto* data_page_footer = page->footer.mutable_data_page_footer();
856
919k
    data_page_footer->set_first_ordinal(_first_rowid);
857
919k
    data_page_footer->set_num_values(_next_rowid - _first_rowid);
858
919k
    data_page_footer->set_nullmap_size(cast_set<uint32_t>(nullmap.slice().size));
859
919k
    if (_new_page_callback != nullptr) {
860
67.8k
        _new_page_callback->put_extra_info_in_page(data_page_footer);
861
67.8k
    }
862
    // trying to compress page body
863
919k
    OwnedSlice compressed_body;
864
919k
    RETURN_IF_ERROR(PageIO::compress_page_body(_compress_codec, _opts.compression_min_space_saving,
865
919k
                                               body, &compressed_body));
866
919k
    if (compressed_body.slice().empty()) {
867
        // page body is uncompressed
868
816k
        page->data.emplace_back(std::move(encoded_values));
869
816k
        page->data.emplace_back(std::move(nullmap));
870
816k
    } else {
871
        // page body is compressed
872
103k
        page->data.emplace_back(std::move(compressed_body));
873
103k
    }
874
875
919k
    _push_back_page(std::move(page));
876
919k
    _first_rowid = _next_rowid;
877
919k
    return Status::OK();
878
919k
}
879
880
////////////////////////////////////////////////////////////////////////////////
881
882
////////////////////////////////////////////////////////////////////////////////
883
// offset column writer
884
////////////////////////////////////////////////////////////////////////////////
885
886
OffsetColumnWriter::OffsetColumnWriter(const ColumnWriterOptions& opts, TabletColumnPtr column,
887
                                       io::FileWriter* file_writer)
888
68.2k
        : ScalarColumnWriter(opts, std::move(column), file_writer) {
889
    // now we only explain data in offset column as uint64
890
68.2k
    DCHECK(get_column()->type() == FieldType::OLAP_FIELD_TYPE_UNSIGNED_BIGINT);
891
68.2k
}
892
893
68.2k
OffsetColumnWriter::~OffsetColumnWriter() = default;
894
895
68.1k
Status OffsetColumnWriter::init() {
896
68.1k
    RETURN_IF_ERROR(ScalarColumnWriter::init());
897
68.1k
    register_flush_page_callback(this);
898
68.1k
    _next_offset = 0;
899
68.1k
    return Status::OK();
900
68.1k
}
901
902
68.6k
Status OffsetColumnWriter::append_data(const uint8_t** ptr, size_t num_rows) {
903
68.6k
    size_t remaining = num_rows;
904
137k
    while (remaining > 0) {
905
69.0k
        size_t num_written = remaining;
906
69.0k
        RETURN_IF_ERROR(append_data_in_current_page(ptr, &num_written));
907
        // Callers provide one extra tail offset after the written rows so the page footer can
908
        // store the next array item ordinal for the current page.
909
69.0k
        _next_offset = *(const uint64_t*)(*ptr);
910
69.0k
        remaining -= num_written;
911
912
69.0k
        if (_page_builder->is_page_full()) {
913
            // get next data for next array_item_rowid
914
409
            RETURN_IF_ERROR(finish_current_page());
915
409
        }
916
69.0k
    }
917
68.6k
    return Status::OK();
918
68.6k
}
919
920
67.8k
void OffsetColumnWriter::put_extra_info_in_page(DataPageFooterPB* footer) {
921
67.8k
    footer->set_next_array_item_ordinal(_next_offset);
922
67.8k
}
923
924
StructColumnWriter::StructColumnWriter(
925
        const ColumnWriterOptions& opts, TabletColumnPtr column, ScalarColumnWriter* null_writer,
926
        std::vector<std::unique_ptr<ColumnWriter>>& sub_column_writers)
927
2.12k
        : ColumnWriter(std::move(column), opts.meta->is_nullable(), opts.meta), _opts(opts) {
928
9.27k
    for (auto& sub_column_writer : sub_column_writers) {
929
9.27k
        _sub_column_writers.push_back(std::move(sub_column_writer));
930
9.27k
    }
931
2.12k
    _num_sub_column_writers = _sub_column_writers.size();
932
2.12k
    DCHECK(_num_sub_column_writers >= 1);
933
2.12k
    if (is_nullable()) {
934
1.85k
        _null_writer.reset(null_writer);
935
1.85k
    }
936
2.12k
}
937
938
2.12k
Status StructColumnWriter::init() {
939
9.27k
    for (auto& column_writer : _sub_column_writers) {
940
9.27k
        RETURN_IF_ERROR(column_writer->init());
941
9.27k
    }
942
2.12k
    if (is_nullable()) {
943
1.85k
        RETURN_IF_ERROR(_null_writer->init());
944
1.85k
    }
945
2.12k
    return Status::OK();
946
2.12k
}
947
948
1.50k
Status StructColumnWriter::write_inverted_index() {
949
1.50k
    if (_opts.need_inverted_index) {
950
0
        for (auto& column_writer : _sub_column_writers) {
951
0
            RETURN_IF_ERROR(column_writer->write_inverted_index());
952
0
        }
953
0
    }
954
1.50k
    return Status::OK();
955
1.50k
}
956
957
Status StructColumnWriter::append_nullable(const uint8_t* null_map, const uint8_t** ptr,
958
1.81k
                                           size_t num_rows) {
959
1.81k
    RETURN_IF_ERROR(append_data(ptr, num_rows));
960
1.81k
    RETURN_IF_ERROR(_null_writer->append_data(&null_map, num_rows));
961
1.81k
    return Status::OK();
962
1.81k
}
963
964
2.08k
Status StructColumnWriter::append_data(const uint8_t** ptr, size_t num_rows) {
965
2.08k
    const auto* results = reinterpret_cast<const uint64_t*>(*ptr);
966
11.0k
    for (size_t i = 0; i < _num_sub_column_writers; ++i) {
967
8.92k
        auto nullmap = *(results + _num_sub_column_writers + i);
968
8.92k
        auto data = *(results + i);
969
8.92k
        RETURN_IF_ERROR(_sub_column_writers[i]->append(reinterpret_cast<const uint8_t*>(nullmap),
970
8.92k
                                                       reinterpret_cast<const void*>(data),
971
8.92k
                                                       num_rows));
972
8.92k
    }
973
2.08k
    return Status::OK();
974
2.08k
}
975
976
186
uint64_t StructColumnWriter::estimate_buffer_size() {
977
186
    uint64_t size = 0;
978
780
    for (auto& column_writer : _sub_column_writers) {
979
780
        size += column_writer->estimate_buffer_size();
980
780
    }
981
186
    size += is_nullable() ? _null_writer->estimate_buffer_size() : 0;
982
186
    return size;
983
186
}
984
985
2.12k
Status StructColumnWriter::finish() {
986
9.27k
    for (auto& column_writer : _sub_column_writers) {
987
9.27k
        RETURN_IF_ERROR(column_writer->finish());
988
9.27k
    }
989
2.12k
    if (is_nullable()) {
990
1.85k
        RETURN_IF_ERROR(_null_writer->finish());
991
1.85k
    }
992
2.12k
    _opts.meta->set_num_rows(get_next_rowid());
993
2.12k
    return Status::OK();
994
2.12k
}
995
996
2.12k
Status StructColumnWriter::write_data() {
997
9.27k
    for (auto& column_writer : _sub_column_writers) {
998
9.27k
        RETURN_IF_ERROR(column_writer->write_data());
999
9.27k
    }
1000
2.12k
    if (is_nullable()) {
1001
1.85k
        RETURN_IF_ERROR(_null_writer->write_data());
1002
1.85k
    }
1003
2.12k
    return Status::OK();
1004
2.12k
}
1005
1006
2.08k
Status StructColumnWriter::write_ordinal_index() {
1007
8.92k
    for (auto& column_writer : _sub_column_writers) {
1008
8.92k
        RETURN_IF_ERROR(column_writer->write_ordinal_index());
1009
8.92k
    }
1010
2.08k
    if (is_nullable()) {
1011
1.81k
        RETURN_IF_ERROR(_null_writer->write_ordinal_index());
1012
1.81k
    }
1013
2.08k
    return Status::OK();
1014
2.08k
}
1015
1016
0
Status StructColumnWriter::append_nulls(size_t num_rows) {
1017
0
    for (auto& column_writer : _sub_column_writers) {
1018
0
        RETURN_IF_ERROR(column_writer->append_nulls(num_rows));
1019
0
    }
1020
0
    if (is_nullable()) {
1021
0
        std::vector<UInt8> null_signs(num_rows, 1);
1022
0
        const uint8_t* null_sign_ptr = null_signs.data();
1023
0
        RETURN_IF_ERROR(_null_writer->append_data(&null_sign_ptr, num_rows));
1024
0
    }
1025
0
    return Status::OK();
1026
0
}
1027
1028
0
Status StructColumnWriter::finish_current_page() {
1029
0
    return Status::NotSupported("struct writer has no data, can not finish_current_page");
1030
0
}
1031
1032
ArrayColumnWriter::ArrayColumnWriter(const ColumnWriterOptions& opts, TabletColumnPtr column,
1033
                                     OffsetColumnWriter* offset_writer,
1034
                                     ScalarColumnWriter* null_writer,
1035
                                     std::unique_ptr<ColumnWriter> item_writer)
1036
42.5k
        : ColumnWriter(std::move(column), opts.meta->is_nullable(), opts.meta),
1037
42.5k
          _item_writer(std::move(item_writer)),
1038
42.5k
          _opts(opts) {
1039
42.5k
    _offset_writer.reset(offset_writer);
1040
42.5k
    if (is_nullable()) {
1041
26.9k
        _null_writer.reset(null_writer);
1042
26.9k
    }
1043
42.5k
}
1044
1045
42.5k
Status ArrayColumnWriter::init() {
1046
42.5k
    RETURN_IF_ERROR(_offset_writer->init());
1047
42.5k
    if (is_nullable()) {
1048
26.9k
        RETURN_IF_ERROR(_null_writer->init());
1049
26.9k
    }
1050
42.5k
    RETURN_IF_ERROR(_item_writer->init());
1051
42.5k
    if (_opts.need_inverted_index) {
1052
1.84k
        auto* writer = dynamic_cast<ScalarColumnWriter*>(_item_writer.get());
1053
1.84k
        if (writer != nullptr) {
1054
1.84k
            RETURN_IF_ERROR(IndexColumnWriter::create(get_column(), &_inverted_index_writer,
1055
1.84k
                                                      _opts.index_file_writer,
1056
1.84k
                                                      _opts.inverted_indexes[0]));
1057
1.84k
        }
1058
1.84k
    }
1059
42.5k
    if (_opts.need_ann_index) {
1060
60
        auto* writer = dynamic_cast<ScalarColumnWriter*>(_item_writer.get());
1061
60
        if (writer != nullptr) {
1062
60
            _ann_index_writer = std::make_unique<AnnIndexColumnWriter>(_opts.index_file_writer,
1063
60
                                                                       _opts.ann_index);
1064
60
            RETURN_IF_ERROR(_ann_index_writer->init());
1065
60
        }
1066
60
    }
1067
42.5k
    return Status::OK();
1068
42.5k
}
1069
1070
39.5k
Status ArrayColumnWriter::write_inverted_index() {
1071
39.5k
    if (_opts.need_inverted_index) {
1072
1.84k
        return _inverted_index_writer->finish();
1073
1.84k
    }
1074
37.7k
    return Status::OK();
1075
39.5k
}
1076
1077
39.5k
Status ArrayColumnWriter::write_ann_index() {
1078
39.5k
    if (_opts.need_ann_index) {
1079
57
        return _ann_index_writer->finish();
1080
57
    }
1081
39.5k
    return Status::OK();
1082
39.5k
}
1083
1084
// batch append data for array
1085
42.3k
Status ArrayColumnWriter::append_data(const uint8_t** ptr, size_t num_rows) {
1086
    // data_ptr contains
1087
    // [size, offset_ptr, item_data_ptr, item_nullmap_ptr]
1088
42.3k
    auto data_ptr = reinterpret_cast<const uint64_t*>(*ptr);
1089
    // total number length
1090
42.3k
    size_t element_cnt = size_t((unsigned long)(*data_ptr));
1091
42.3k
    auto offset_data = *(data_ptr + 1);
1092
42.3k
    const uint8_t* offsets_ptr = (const uint8_t*)offset_data;
1093
42.3k
    auto data = *(data_ptr + 2);
1094
42.3k
    auto nested_null_map = *(data_ptr + 3);
1095
42.3k
    if (element_cnt > 0) {
1096
26.3k
        RETURN_IF_ERROR(_item_writer->append(reinterpret_cast<const uint8_t*>(nested_null_map),
1097
26.3k
                                             reinterpret_cast<const void*>(data), element_cnt));
1098
26.3k
    }
1099
42.3k
    if (_opts.need_inverted_index) {
1100
1.84k
        auto* writer = dynamic_cast<ScalarColumnWriter*>(_item_writer.get());
1101
        // now only support nested type is scala
1102
1.84k
        if (writer != nullptr) {
1103
            //NOTE: use array field name as index field, but item_writer size should be used when moving item_data_ptr
1104
1.84k
            RETURN_IF_ERROR(_inverted_index_writer->add_array_values(
1105
1.84k
                    field_type_size(_item_writer->get_column()->type()),
1106
1.84k
                    reinterpret_cast<const void*>(data),
1107
1.84k
                    reinterpret_cast<const uint8_t*>(nested_null_map), offsets_ptr, num_rows));
1108
1.84k
        }
1109
1.84k
    }
1110
1111
42.3k
    if (_opts.need_ann_index) {
1112
60
        auto* writer = dynamic_cast<ScalarColumnWriter*>(_item_writer.get());
1113
        // now only support nested type is scala
1114
60
        if (writer != nullptr) {
1115
            //NOTE: use array field name as index field, but item_writer size should be used when moving item_data_ptr
1116
60
            RETURN_IF_ERROR(_ann_index_writer->add_array_values(
1117
60
                    field_type_size(_item_writer->get_column()->type()),
1118
60
                    reinterpret_cast<const void*>(data),
1119
60
                    reinterpret_cast<const uint8_t*>(nested_null_map), offsets_ptr, num_rows));
1120
60
        } else {
1121
0
            return Status::NotSupported(
1122
0
                    "Ann index can only be build on array with scalar type. but got {} as "
1123
0
                    "nested",
1124
0
                    _item_writer->get_column()->type());
1125
0
        }
1126
60
    }
1127
1128
42.3k
    RETURN_IF_ERROR(_offset_writer->append_data(&offsets_ptr, num_rows));
1129
42.3k
    return Status::OK();
1130
42.3k
}
1131
1132
295
uint64_t ArrayColumnWriter::estimate_buffer_size() {
1133
295
    return _offset_writer->estimate_buffer_size() +
1134
295
           (is_nullable() ? _null_writer->estimate_buffer_size() : 0) +
1135
295
           _item_writer->estimate_buffer_size();
1136
295
}
1137
1138
Status ArrayColumnWriter::append_nullable(const uint8_t* null_map, const uint8_t** ptr,
1139
26.6k
                                          size_t num_rows) {
1140
26.6k
    RETURN_IF_ERROR(append_data(ptr, num_rows));
1141
26.6k
    if (is_nullable()) {
1142
26.6k
        if (_opts.need_inverted_index) {
1143
1.41k
            RETURN_IF_ERROR(_inverted_index_writer->add_array_nulls(null_map, num_rows));
1144
1.41k
        }
1145
26.6k
        RETURN_IF_ERROR(_null_writer->append_data(&null_map, num_rows));
1146
26.6k
    }
1147
26.6k
    return Status::OK();
1148
26.6k
}
1149
1150
42.6k
Status ArrayColumnWriter::finish() {
1151
42.6k
    RETURN_IF_ERROR(_offset_writer->finish());
1152
42.6k
    if (is_nullable()) {
1153
26.9k
        RETURN_IF_ERROR(_null_writer->finish());
1154
26.9k
    }
1155
42.6k
    RETURN_IF_ERROR(_item_writer->finish());
1156
42.6k
    _opts.meta->set_num_rows(get_next_rowid());
1157
42.6k
    return Status::OK();
1158
42.6k
}
1159
1160
42.6k
Status ArrayColumnWriter::write_data() {
1161
42.6k
    RETURN_IF_ERROR(_offset_writer->write_data());
1162
42.6k
    if (is_nullable()) {
1163
26.9k
        RETURN_IF_ERROR(_null_writer->write_data());
1164
26.9k
    }
1165
42.6k
    RETURN_IF_ERROR(_item_writer->write_data());
1166
42.6k
    return Status::OK();
1167
42.6k
}
1168
1169
42.1k
Status ArrayColumnWriter::write_ordinal_index() {
1170
42.1k
    RETURN_IF_ERROR(_offset_writer->write_ordinal_index());
1171
42.1k
    if (is_nullable()) {
1172
26.4k
        RETURN_IF_ERROR(_null_writer->write_ordinal_index());
1173
26.4k
    }
1174
42.1k
    if (!has_empty_items()) {
1175
26.2k
        RETURN_IF_ERROR(_item_writer->write_ordinal_index());
1176
26.2k
    }
1177
42.1k
    return Status::OK();
1178
42.1k
}
1179
1180
0
Status ArrayColumnWriter::append_nulls(size_t num_rows) {
1181
0
    const UInt64 offset = cast_set<UInt64>(_item_writer->get_next_rowid());
1182
0
    std::vector<UInt64> offsets_data(num_rows + 1, offset);
1183
0
    const uint8_t* offsets_ptr = reinterpret_cast<const uint8_t*>(offsets_data.data());
1184
0
    RETURN_IF_ERROR(_offset_writer->append_data(&offsets_ptr, num_rows));
1185
0
    return write_null_column(num_rows, true);
1186
0
}
1187
1188
0
Status ArrayColumnWriter::write_null_column(size_t num_rows, bool is_null) {
1189
0
    uint8_t null_sign = is_null ? 1 : 0;
1190
0
    while (is_nullable() && num_rows > 0) {
1191
        // TODO llj bulk write
1192
0
        const uint8_t* null_sign_ptr = &null_sign;
1193
0
        RETURN_IF_ERROR(_null_writer->append_data(&null_sign_ptr, 1));
1194
0
        --num_rows;
1195
0
    }
1196
0
    return Status::OK();
1197
0
}
1198
1199
0
Status ArrayColumnWriter::finish_current_page() {
1200
0
    return Status::NotSupported("array writer has no data, can not finish_current_page");
1201
0
}
1202
1203
/// ============================= MapColumnWriter =====================////
1204
MapColumnWriter::MapColumnWriter(const ColumnWriterOptions& opts, TabletColumnPtr column,
1205
                                 ScalarColumnWriter* null_writer, OffsetColumnWriter* offset_writer,
1206
                                 std::vector<std::unique_ptr<ColumnWriter>>& kv_writers)
1207
25.6k
        : ColumnWriter(std::move(column), opts.meta->is_nullable(), opts.meta), _opts(opts) {
1208
25.6k
    CHECK_EQ(kv_writers.size(), 2);
1209
25.6k
    _offsets_writer.reset(offset_writer);
1210
25.6k
    if (is_nullable()) {
1211
7.90k
        _null_writer.reset(null_writer);
1212
7.90k
    }
1213
51.2k
    for (auto& sub_writers : kv_writers) {
1214
51.2k
        _kv_writers.push_back(std::move(sub_writers));
1215
51.2k
    }
1216
25.6k
}
1217
1218
25.6k
Status MapColumnWriter::init() {
1219
25.6k
    RETURN_IF_ERROR(_offsets_writer->init());
1220
25.6k
    if (is_nullable()) {
1221
7.90k
        RETURN_IF_ERROR(_null_writer->init());
1222
7.90k
    }
1223
    // here register_flush_page_callback to call this.put_extra_info_in_page()
1224
    // when finish cur data page
1225
51.2k
    for (auto& sub_writer : _kv_writers) {
1226
51.2k
        RETURN_IF_ERROR(sub_writer->init());
1227
51.2k
    }
1228
25.6k
    return Status::OK();
1229
25.6k
}
1230
1231
446
uint64_t MapColumnWriter::estimate_buffer_size() {
1232
446
    size_t estimate = 0;
1233
892
    for (auto& sub_writer : _kv_writers) {
1234
892
        estimate += sub_writer->estimate_buffer_size();
1235
892
    }
1236
446
    estimate += _offsets_writer->estimate_buffer_size();
1237
446
    if (is_nullable()) {
1238
425
        estimate += _null_writer->estimate_buffer_size();
1239
425
    }
1240
446
    return estimate;
1241
446
}
1242
1243
25.6k
Status MapColumnWriter::finish() {
1244
25.6k
    RETURN_IF_ERROR(_offsets_writer->finish());
1245
25.6k
    if (is_nullable()) {
1246
7.91k
        RETURN_IF_ERROR(_null_writer->finish());
1247
7.91k
    }
1248
51.2k
    for (auto& sub_writer : _kv_writers) {
1249
51.2k
        RETURN_IF_ERROR(sub_writer->finish());
1250
51.2k
    }
1251
25.6k
    _opts.meta->set_num_rows(get_next_rowid());
1252
25.6k
    return Status::OK();
1253
25.6k
}
1254
1255
Status MapColumnWriter::append_nullable(const uint8_t* null_map, const uint8_t** ptr,
1256
7.78k
                                        size_t num_rows) {
1257
7.78k
    RETURN_IF_ERROR(append_data(ptr, num_rows));
1258
7.78k
    if (is_nullable()) {
1259
7.78k
        RETURN_IF_ERROR(_null_writer->append_data(&null_map, num_rows));
1260
7.78k
    }
1261
7.78k
    return Status::OK();
1262
7.78k
}
1263
1264
// write key value data with offsets
1265
26.3k
Status MapColumnWriter::append_data(const uint8_t** ptr, size_t num_rows) {
1266
    // data_ptr contains
1267
    // [size, offset_ptr, key_data_ptr, val_data_ptr, k_nullmap_ptr, v_nullmap_pr]
1268
    // which converted results from olap_map_convertor and later will use a structure to replace it
1269
26.3k
    auto data_ptr = reinterpret_cast<const uint64_t*>(*ptr);
1270
    // total number length
1271
26.3k
    size_t element_cnt = size_t((unsigned long)(*data_ptr));
1272
26.3k
    auto offset_data = *(data_ptr + 1);
1273
26.3k
    const uint8_t* offsets_ptr = (const uint8_t*)offset_data;
1274
1275
26.3k
    if (element_cnt > 0) {
1276
40.1k
        for (size_t i = 0; i < 2; ++i) {
1277
26.7k
            auto data = *(data_ptr + 2 + i);
1278
26.7k
            auto nested_null_map = *(data_ptr + 2 + 2 + i);
1279
26.7k
            RETURN_IF_ERROR(
1280
26.7k
                    _kv_writers[i]->append(reinterpret_cast<const uint8_t*>(nested_null_map),
1281
26.7k
                                           reinterpret_cast<const void*>(data), element_cnt));
1282
26.7k
        }
1283
13.3k
    }
1284
    // make sure the order : offset writer flush next_array_item_ordinal after kv_writers append_data
1285
    // because we use _kv_writers[0]->get_next_rowid() to set next_array_item_ordinal in offset page footer
1286
26.3k
    RETURN_IF_ERROR(_offsets_writer->append_data(&offsets_ptr, num_rows));
1287
26.3k
    return Status::OK();
1288
26.3k
}
1289
1290
25.6k
Status MapColumnWriter::write_data() {
1291
25.6k
    RETURN_IF_ERROR(_offsets_writer->write_data());
1292
25.6k
    if (is_nullable()) {
1293
7.91k
        RETURN_IF_ERROR(_null_writer->write_data());
1294
7.91k
    }
1295
51.2k
    for (auto& sub_writer : _kv_writers) {
1296
51.2k
        RETURN_IF_ERROR(sub_writer->write_data());
1297
51.2k
    }
1298
25.6k
    return Status::OK();
1299
25.6k
}
1300
1301
25.3k
Status MapColumnWriter::write_ordinal_index() {
1302
25.3k
    RETURN_IF_ERROR(_offsets_writer->write_ordinal_index());
1303
25.3k
    if (is_nullable()) {
1304
7.65k
        RETURN_IF_ERROR(_null_writer->write_ordinal_index());
1305
7.65k
    }
1306
50.7k
    for (auto& sub_writer : _kv_writers) {
1307
50.7k
        if (sub_writer->get_next_rowid() != 0) {
1308
25.6k
            RETURN_IF_ERROR(sub_writer->write_ordinal_index());
1309
25.6k
        }
1310
50.7k
    }
1311
25.3k
    return Status::OK();
1312
25.3k
}
1313
1314
0
Status MapColumnWriter::append_nulls(size_t num_rows) {
1315
0
    for (auto& sub_writer : _kv_writers) {
1316
0
        RETURN_IF_ERROR(sub_writer->append_nulls(num_rows));
1317
0
    }
1318
0
    const UInt64 offset = cast_set<UInt64>(_kv_writers[0]->get_next_rowid());
1319
0
    std::vector<UInt64> offsets_data(num_rows + 1, offset);
1320
0
    const uint8_t* offsets_ptr = reinterpret_cast<const uint8_t*>(offsets_data.data());
1321
0
    RETURN_IF_ERROR(_offsets_writer->append_data(&offsets_ptr, num_rows));
1322
1323
0
    if (is_nullable()) {
1324
0
        std::vector<UInt8> null_signs(num_rows, 1);
1325
0
        const uint8_t* null_sign_ptr = null_signs.data();
1326
0
        RETURN_IF_ERROR(_null_writer->append_data(&null_sign_ptr, num_rows));
1327
0
    }
1328
0
    return Status::OK();
1329
0
}
1330
1331
0
Status MapColumnWriter::finish_current_page() {
1332
0
    return Status::NotSupported("map writer has no data, can not finish_current_page");
1333
0
}
1334
1335
9.75k
Status MapColumnWriter::write_inverted_index() {
1336
9.75k
    if (_opts.need_inverted_index) {
1337
0
        return _index_builder->finish();
1338
0
    }
1339
9.75k
    return Status::OK();
1340
9.75k
}
1341
1342
VariantColumnWriter::VariantColumnWriter(const ColumnWriterOptions& opts, TabletColumnPtr column)
1343
5.91k
        : ColumnWriter(std::move(column), opts.meta->is_nullable(), opts.meta) {
1344
5.91k
    _impl = std::make_unique<VariantColumnWriterImpl>(opts, get_column());
1345
5.91k
}
1346
1347
5.91k
Status VariantColumnWriter::init() {
1348
5.91k
    return _impl->init();
1349
5.91k
}
1350
1351
675
Status VariantColumnWriter::append_data(const uint8_t** ptr, size_t num_rows) {
1352
675
    _next_rowid += num_rows;
1353
675
    return _impl->append_data(ptr, num_rows);
1354
675
}
1355
1356
866
uint64_t VariantColumnWriter::estimate_buffer_size() {
1357
866
    return _impl->estimate_buffer_size();
1358
866
}
1359
1360
5.92k
Status VariantColumnWriter::finish() {
1361
5.92k
    return _impl->finish();
1362
5.92k
}
1363
5.92k
Status VariantColumnWriter::write_data() {
1364
5.92k
    return _impl->write_data();
1365
5.92k
}
1366
5.92k
Status VariantColumnWriter::write_ordinal_index() {
1367
5.92k
    return _impl->write_ordinal_index();
1368
5.92k
}
1369
1370
5.92k
Status VariantColumnWriter::write_zone_map() {
1371
5.92k
    return _impl->write_zone_map();
1372
5.92k
}
1373
1374
5.90k
Status VariantColumnWriter::write_inverted_index() {
1375
5.90k
    return _impl->write_inverted_index();
1376
5.90k
}
1377
5.90k
Status VariantColumnWriter::write_bloom_filter_index() {
1378
5.90k
    return _impl->write_bloom_filter_index();
1379
5.90k
}
1380
1381
Status VariantColumnWriter::append_nullable(const uint8_t* null_map, const uint8_t** ptr,
1382
5.46k
                                            size_t num_rows) {
1383
5.46k
    return _impl->append_nullable(null_map, ptr, num_rows);
1384
5.46k
}
1385
1386
} // namespace doris::segment_v2