Coverage Report

Created: 2026-06-11 05:27

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
548k
    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
1.33M
    void reserve_for_write(size_t num_rows, size_t non_null_count) {
64
1.33M
        if (num_rows == 0) {
65
0
            return;
66
0
        }
67
1.33M
        if (non_null_count == 0 || (non_null_count == num_rows && !_has_null)) {
68
418k
            if (_bitmap_buf.capacity() < kSmallReserveBytes) {
69
40
                _bitmap_buf.reserve(kSmallReserveBytes);
70
40
            }
71
418k
            return;
72
418k
        }
73
912k
        size_t raw_bytes = BitmapSize(num_rows);
74
912k
        size_t run_est = std::min(num_rows, non_null_count * 2 + 1);
75
912k
        size_t run_bytes_est = run_est * kBytesPerRun + kReserveSlackBytes;
76
912k
        size_t raw_overhead = raw_bytes / 63 + 1;
77
912k
        size_t raw_est = raw_bytes + raw_overhead + kReserveSlackBytes;
78
912k
        size_t reserve_bytes = std::min(raw_est, run_bytes_est);
79
912k
        if (_bitmap_buf.capacity() < reserve_bytes) {
80
2.34k
            const size_t cap = _bitmap_buf.capacity();
81
2.34k
            const size_t grow = cap + cap / 2;
82
2.34k
            const size_t new_cap = std::max(reserve_bytes, grow);
83
2.34k
            _bitmap_buf.reserve(new_cap);
84
2.34k
        }
85
912k
    }
86
87
3.76M
    void add_run(bool value, size_t run) {
88
3.76M
        _has_null |= value;
89
3.76M
        _rle_encoder.Put(value, run);
90
3.76M
    }
91
92
    // Returns whether the building nullmap contains nullptr
93
546k
    bool has_null() const { return _has_null; }
94
95
175k
    Status finish(OwnedSlice* slice) {
96
175k
        _rle_encoder.Flush();
97
175k
        RETURN_IF_CATCH_EXCEPTION({ *slice = _bitmap_buf.build(); });
98
175k
        return Status::OK();
99
175k
    }
100
101
546k
    void reset() {
102
546k
        _has_null = false;
103
546k
        _rle_encoder.Clear();
104
546k
    }
105
106
47.6k
    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
71.2k
                                           io::FileWriter* file_writer, uint32_t id) {
120
71.2k
    if (!opts.meta->is_nullable()) {
121
32.8k
        return nullptr;
122
32.8k
    }
123
124
38.4k
    FieldType null_type = FieldType::OLAP_FIELD_TYPE_TINYINT;
125
38.4k
    ColumnWriterOptions null_options;
126
38.4k
    null_options.meta = opts.meta->add_children_columns();
127
38.4k
    null_options.meta->set_column_id(id);
128
38.4k
    null_options.meta->set_unique_id(id);
129
38.4k
    null_options.meta->set_type(int(null_type));
130
38.4k
    null_options.meta->set_is_nullable(false);
131
38.4k
    null_options.meta->set_length(
132
38.4k
            cast_set<int32_t>(field_type_size(FieldType::OLAP_FIELD_TYPE_TINYINT)));
133
38.4k
    null_options.meta->set_compression(opts.meta->compression());
134
135
38.4k
    null_options.need_zone_map = false;
136
38.4k
    null_options.need_bloom_filter = false;
137
38.4k
    null_options.storage_format = opts.storage_format;
138
139
38.4k
    auto null_column_ptr = std::make_shared<TabletColumn>(
140
38.4k
            FieldAggregationMethod::OLAP_FIELD_AGGREGATION_NONE, null_type, false,
141
38.4k
            null_options.meta->unique_id(), null_options.meta->length());
142
38.4k
    null_column_ptr->set_name("nullable");
143
38.4k
    null_column_ptr->set_index_length(-1); // no short key index
144
38.4k
    null_options.meta->set_encoding(
145
38.4k
            EncodingInfo::resolve_default_encoding(opts.storage_format, *null_column_ptr));
146
38.4k
    return new ScalarColumnWriter(null_options, std::move(null_column_ptr), file_writer);
147
71.2k
}
148
149
ColumnWriter::ColumnWriter(TabletColumnPtr column, bool is_nullable, ColumnMetaPB* meta)
150
1.08M
        : _column(std::move(column)), _is_nullable(is_nullable), _column_meta(meta) {
151
1.08M
    _data_type = DataTypeFactory::instance().create_data_type(*_column_meta);
152
1.08M
}
153
Status ColumnWriter::create_struct_writer(const ColumnWriterOptions& opts,
154
                                          const TabletColumn* column, io::FileWriter* file_writer,
155
2.27k
                                          std::unique_ptr<ColumnWriter>* writer) {
156
    // not support empty struct
157
2.27k
    DCHECK(column->get_subtype_count() >= 1);
158
2.27k
    std::vector<std::unique_ptr<ColumnWriter>> sub_column_writers;
159
2.27k
    sub_column_writers.reserve(column->get_subtype_count());
160
12.3k
    for (uint32_t i = 0; i < column->get_subtype_count(); i++) {
161
10.0k
        const TabletColumn& sub_column = column->get_sub_column(i);
162
10.0k
        RETURN_IF_ERROR(sub_column.check_valid());
163
164
        // create sub writer
165
10.0k
        ColumnWriterOptions column_options;
166
10.0k
        column_options.meta = opts.meta->mutable_children_columns(i);
167
10.0k
        column_options.need_zone_map = false;
168
10.0k
        column_options.need_bloom_filter = sub_column.is_bf_column();
169
10.0k
        column_options.storage_format = opts.storage_format;
170
10.0k
        std::unique_ptr<ColumnWriter> sub_column_writer;
171
10.0k
        RETURN_IF_ERROR(
172
10.0k
                ColumnWriter::create(column_options, &sub_column, file_writer, &sub_column_writer));
173
10.0k
        sub_column_writers.push_back(std::move(sub_column_writer));
174
10.0k
    }
175
176
2.27k
    ScalarColumnWriter* null_writer =
177
2.27k
            get_null_writer(opts, file_writer, column->get_subtype_count() + 1);
178
179
2.27k
    *writer = std::unique_ptr<ColumnWriter>(new StructColumnWriter(
180
2.27k
            opts, std::make_shared<TabletColumn>(*column), null_writer, sub_column_writers));
181
2.27k
    return Status::OK();
182
2.27k
}
183
184
Status ColumnWriter::create_array_writer(const ColumnWriterOptions& opts,
185
                                         const TabletColumn* column, io::FileWriter* file_writer,
186
43.7k
                                         std::unique_ptr<ColumnWriter>* writer) {
187
43.7k
    DCHECK(column->get_subtype_count() == 1);
188
43.7k
    const TabletColumn& item_column = column->get_sub_column(0);
189
43.7k
    RETURN_IF_ERROR(item_column.check_valid());
190
191
    // create item writer
192
43.7k
    ColumnWriterOptions item_options;
193
43.7k
    item_options.meta = opts.meta->mutable_children_columns(0);
194
43.7k
    item_options.need_zone_map = false;
195
43.7k
    item_options.need_bloom_filter = item_column.is_bf_column();
196
43.7k
    item_options.storage_format = opts.storage_format;
197
43.7k
    std::unique_ptr<ColumnWriter> item_writer;
198
43.7k
    RETURN_IF_ERROR(ColumnWriter::create(item_options, &item_column, file_writer, &item_writer));
199
200
    // create length writer
201
43.7k
    FieldType length_type = FieldType::OLAP_FIELD_TYPE_UNSIGNED_BIGINT;
202
203
43.7k
    ColumnWriterOptions length_options;
204
43.7k
    length_options.meta = opts.meta->add_children_columns();
205
43.7k
    length_options.meta->set_column_id(2);
206
43.7k
    length_options.meta->set_unique_id(2);
207
43.7k
    length_options.meta->set_type(int(length_type));
208
43.7k
    length_options.meta->set_is_nullable(false);
209
43.7k
    length_options.meta->set_length(
210
43.7k
            cast_set<int32_t>(field_type_size(FieldType::OLAP_FIELD_TYPE_UNSIGNED_BIGINT)));
211
43.7k
    length_options.meta->set_compression(opts.meta->compression());
212
213
43.7k
    length_options.need_zone_map = false;
214
43.7k
    length_options.need_bloom_filter = false;
215
43.7k
    length_options.storage_format = opts.storage_format;
216
217
43.7k
    auto length_column_ptr = std::make_shared<TabletColumn>(
218
43.7k
            FieldAggregationMethod::OLAP_FIELD_AGGREGATION_NONE, length_type,
219
43.7k
            length_options.meta->is_nullable(), length_options.meta->unique_id(),
220
43.7k
            length_options.meta->length());
221
43.7k
    length_column_ptr->set_name("length");
222
43.7k
    length_column_ptr->set_index_length(-1); // no short key index
223
43.7k
    length_options.meta->set_encoding(
224
43.7k
            EncodingInfo::resolve_default_encoding(opts.storage_format, *length_column_ptr));
225
43.7k
    auto* length_writer =
226
43.7k
            new OffsetColumnWriter(length_options, std::move(length_column_ptr), file_writer);
227
228
43.7k
    ScalarColumnWriter* null_writer = get_null_writer(opts, file_writer, 3);
229
230
43.7k
    *writer = std::unique_ptr<ColumnWriter>(
231
43.7k
            new ArrayColumnWriter(opts, std::make_shared<TabletColumn>(*column), length_writer,
232
43.7k
                                  null_writer, std::move(item_writer)));
233
43.7k
    return Status::OK();
234
43.7k
}
235
236
Status ColumnWriter::create_map_writer(const ColumnWriterOptions& opts, const TabletColumn* column,
237
                                       io::FileWriter* file_writer,
238
25.2k
                                       std::unique_ptr<ColumnWriter>* writer) {
239
25.2k
    DCHECK(column->get_subtype_count() == 2);
240
25.2k
    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
25.2k
    std::vector<std::unique_ptr<ColumnWriter>> inner_writer_list;
247
75.8k
    for (int i = 0; i < 2; ++i) {
248
50.5k
        const TabletColumn& item_column = column->get_sub_column(i);
249
50.5k
        RETURN_IF_ERROR(item_column.check_valid());
250
251
        // create item writer
252
50.5k
        ColumnWriterOptions item_options;
253
50.5k
        item_options.meta = opts.meta->mutable_children_columns(i);
254
50.5k
        item_options.need_zone_map = false;
255
50.5k
        item_options.need_bloom_filter = item_column.is_bf_column();
256
50.5k
        item_options.storage_format = opts.storage_format;
257
50.5k
        std::unique_ptr<ColumnWriter> item_writer;
258
50.5k
        RETURN_IF_ERROR(
259
50.5k
                ColumnWriter::create(item_options, &item_column, file_writer, &item_writer));
260
50.5k
        inner_writer_list.push_back(std::move(item_writer));
261
50.5k
    }
262
263
    // create offset writer
264
25.2k
    FieldType length_type = FieldType::OLAP_FIELD_TYPE_UNSIGNED_BIGINT;
265
266
    // Be Cautious: column unique id is used for column reader creation
267
25.2k
    ColumnWriterOptions length_options;
268
25.2k
    length_options.meta = opts.meta->add_children_columns();
269
25.2k
    length_options.meta->set_column_id(column->get_subtype_count() + 1);
270
25.2k
    length_options.meta->set_unique_id(column->get_subtype_count() + 1);
271
25.2k
    length_options.meta->set_type(int(length_type));
272
25.2k
    length_options.meta->set_is_nullable(false);
273
25.2k
    length_options.meta->set_length(
274
25.2k
            cast_set<int32_t>(field_type_size(FieldType::OLAP_FIELD_TYPE_UNSIGNED_BIGINT)));
275
25.2k
    length_options.meta->set_compression(opts.meta->compression());
276
277
25.2k
    length_options.need_zone_map = false;
278
25.2k
    length_options.need_bloom_filter = false;
279
25.2k
    length_options.storage_format = opts.storage_format;
280
281
25.2k
    auto length_column_ptr = std::make_shared<TabletColumn>(
282
25.2k
            FieldAggregationMethod::OLAP_FIELD_AGGREGATION_NONE, length_type,
283
25.2k
            length_options.meta->is_nullable(), length_options.meta->unique_id(),
284
25.2k
            length_options.meta->length());
285
25.2k
    length_column_ptr->set_name("length");
286
25.2k
    length_column_ptr->set_index_length(-1); // no short key index
287
25.2k
    length_options.meta->set_encoding(
288
25.2k
            EncodingInfo::resolve_default_encoding(opts.storage_format, *length_column_ptr));
289
25.2k
    auto* length_writer =
290
25.2k
            new OffsetColumnWriter(length_options, std::move(length_column_ptr), file_writer);
291
292
25.2k
    ScalarColumnWriter* null_writer =
293
25.2k
            get_null_writer(opts, file_writer, column->get_subtype_count() + 2);
294
295
25.2k
    *writer = std::unique_ptr<ColumnWriter>(
296
25.2k
            new MapColumnWriter(opts, std::make_shared<TabletColumn>(*column), null_writer,
297
25.2k
                                length_writer, inner_writer_list));
298
299
25.2k
    return Status::OK();
300
25.2k
}
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
1.19k
        *writer = std::unique_ptr<ColumnWriter>(
312
1.19k
                new ScalarColumnWriter(opts, std::make_shared<TabletColumn>(*column), file_writer));
313
1.19k
    } else if (type == PrimitiveType::TYPE_ARRAY) {
314
22
        RETURN_IF_ERROR(create_array_writer(opts, column, file_writer, writer));
315
63
    } else if (type == PrimitiveType::TYPE_MAP) {
316
52
        RETURN_IF_ERROR(create_map_writer(opts, column, file_writer, writer));
317
52
    } else {
318
11
        throw Exception(ErrorCode::INTERNAL_ERROR,
319
11
                        "OLAP_FIELD_TYPE_AGG_STATE meet unsupported type: {}",
320
11
                        agg_state_type->get_name());
321
11
    }
322
1.26k
    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
6.33k
                                           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
6.33k
    if (column->is_extracted_column()) {
334
326
        if (column->name().find(DOC_VALUE_COLUMN_PATH) != std::string::npos) {
335
244
            *writer = std::make_unique<VariantDocCompactWriter>(
336
244
                    opts, std::make_shared<TabletColumn>(*column));
337
244
            return Status::OK();
338
244
        }
339
82
        VLOG_DEBUG << "gen subwriter for " << column->path_info_ptr()->get_path();
340
82
        *writer = std::make_unique<VariantSubcolumnWriter>(opts,
341
82
                                                           std::make_shared<TabletColumn>(*column));
342
82
        return Status::OK();
343
326
    }
344
6.00k
    *writer = std::make_unique<VariantColumnWriter>(opts, std::make_shared<TabletColumn>(*column));
345
6.00k
    return Status::OK();
346
6.33k
}
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
954k
                            io::FileWriter* file_writer, std::unique_ptr<ColumnWriter>* writer) {
351
954k
    auto column_ptr = std::make_shared<TabletColumn>(*column);
352
954k
    if (is_scalar_type(column->type())) {
353
889k
        *writer = std::unique_ptr<ColumnWriter>(
354
889k
                new ScalarColumnWriter(opts, std::move(column_ptr), file_writer));
355
889k
        return Status::OK();
356
889k
    } else {
357
64.0k
        switch (column->type()) {
358
1.26k
        case FieldType::OLAP_FIELD_TYPE_AGG_STATE: {
359
1.26k
            RETURN_IF_ERROR(create_agg_state_writer(opts, column, file_writer, writer));
360
1.26k
            return Status::OK();
361
1.26k
        }
362
2.27k
        case FieldType::OLAP_FIELD_TYPE_STRUCT: {
363
2.27k
            RETURN_IF_ERROR(create_struct_writer(opts, column, file_writer, writer));
364
2.27k
            return Status::OK();
365
2.27k
        }
366
43.7k
        case FieldType::OLAP_FIELD_TYPE_ARRAY: {
367
43.7k
            RETURN_IF_ERROR(create_array_writer(opts, column, file_writer, writer));
368
43.7k
            return Status::OK();
369
43.7k
        }
370
11.1k
        case FieldType::OLAP_FIELD_TYPE_MAP: {
371
11.1k
            RETURN_IF_ERROR(create_map_writer(opts, column, file_writer, writer));
372
11.1k
            return Status::OK();
373
11.1k
        }
374
6.33k
        case FieldType::OLAP_FIELD_TYPE_VARIANT: {
375
            // Process columns with sparse column
376
6.33k
            RETURN_IF_ERROR(create_variant_writer(opts, column, file_writer, writer));
377
6.33k
            return Status::OK();
378
6.33k
        }
379
0
        default:
380
0
            return Status::NotSupported("unsupported type for ColumnWriter: {}",
381
0
                                        std::to_string(int(column_ptr->type())));
382
64.0k
        }
383
64.0k
    }
384
954k
}
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
1.76M
Status ColumnWriter::append(const uint8_t* nullmap, const void* data, size_t num_rows) {
454
1.76M
    assert(data && num_rows > 0);
455
1.76M
    const auto* ptr = (const uint8_t*)data;
456
1.76M
    if (nullmap) {
457
1.36M
        return append_nullable(nullmap, &ptr, num_rows);
458
1.36M
    } else {
459
391k
        return append_data(&ptr, num_rows);
460
391k
    }
461
1.76M
}
462
463
///////////////////////////////////////////////////////////////////////////////////
464
465
ScalarColumnWriter::ScalarColumnWriter(const ColumnWriterOptions& opts, TabletColumnPtr column,
466
                                       io::FileWriter* file_writer)
467
1.00M
        : ColumnWriter(std::move(column), opts.meta->is_nullable(), opts.meta),
468
1.00M
          _opts(opts),
469
1.00M
          _file_writer(file_writer),
470
1.00M
          _data_size(0) {
471
    // these opts.meta fields should be set by client
472
1.00M
    DCHECK(opts.meta->has_column_id());
473
1.00M
    DCHECK(opts.meta->has_unique_id());
474
1.00M
    DCHECK(opts.meta->has_type());
475
1.00M
    DCHECK(opts.meta->has_length());
476
1.00M
    DCHECK(opts.meta->has_encoding());
477
1.00M
    DCHECK(opts.meta->has_compression());
478
1.00M
    DCHECK(opts.meta->has_is_nullable());
479
1.00M
    DCHECK(file_writer != nullptr);
480
1.00M
    _inverted_index_builders.resize(_opts.inverted_indexes.size());
481
1.00M
}
482
483
1.00M
ScalarColumnWriter::~ScalarColumnWriter() {
484
    // delete all pages
485
1.00M
    _pages.clear();
486
1.00M
}
487
488
1.00M
Status ScalarColumnWriter::init() {
489
1.00M
    RETURN_IF_ERROR(get_block_compression_codec(_opts.meta->compression(), &_compress_codec));
490
491
1.00M
    PageBuilder* page_builder = nullptr;
492
493
    // Caller must set a concrete (non-DEFAULT) encoding on the meta before init.
494
1.00M
    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
1.00M
    RETURN_IF_ERROR(
501
1.00M
            EncodingInfo::get(get_column()->type(), _opts.meta->encoding(), &_encoding_info));
502
    // create page builder
503
1.00M
    PageBuilderOptions opts;
504
1.00M
    opts.data_page_size = _opts.data_page_size;
505
1.00M
    opts.dict_page_size = _opts.dict_page_size;
506
    // V3 segments store the dictionary word page (and the dict-overflow fallback plain page)
507
    // with the V3 binary plain layout; pre-V3 segments keep V1.
508
1.00M
    opts.dict_binary_plain_encoding =
509
1.00M
            (_opts.storage_format == TabletStorageFormatPB::TABLET_STORAGE_FORMAT_V3)
510
1.00M
                    ? PLAIN_ENCODING_V3
511
1.00M
                    : PLAIN_ENCODING;
512
1.00M
    RETURN_IF_ERROR(_encoding_info->create_page_builder(opts, &page_builder));
513
1.00M
    if (page_builder == nullptr) {
514
0
        return Status::NotSupported("Failed to create page builder for type {} and encoding {}",
515
0
                                    get_column()->type(), _opts.meta->encoding());
516
0
    }
517
18.4E
    VLOG_DEBUG << fmt::format(
518
18.4E
            "[verbose] scalar column writer init, column_id={}, type={}, encoding={}, "
519
18.4E
            "is_nullable={}",
520
18.4E
            _opts.meta->column_id(), get_column()->type(),
521
18.4E
            EncodingTypePB_Name(_opts.meta->encoding()), _opts.meta->is_nullable());
522
1.00M
    _page_builder.reset(page_builder);
523
    // create ordinal builder
524
1.00M
    _ordinal_index_builder = std::make_unique<OrdinalIndexWriter>();
525
    // create null bitmap builder
526
1.00M
    if (is_nullable()) {
527
548k
        _null_bitmap_builder = std::make_unique<NullBitmapBuilder>();
528
548k
    }
529
1.00M
    if (_opts.need_zone_map) {
530
737k
        RETURN_IF_ERROR(
531
737k
                ZoneMapIndexWriter::create(_data_type, get_column(), _zone_map_index_builder));
532
737k
    }
533
534
1.00M
    if (_opts.need_inverted_index) {
535
33.4k
        do {
536
67.0k
            for (size_t i = 0; i < _opts.inverted_indexes.size(); i++) {
537
33.6k
                DBUG_EXECUTE_IF("column_writer.init", {
538
33.6k
                    class InvertedIndexColumnWriterEmpty final : public IndexColumnWriter {
539
33.6k
                    public:
540
33.6k
                        Status init() override { return Status::OK(); }
541
33.6k
                        Status add_values(const std::string name, const void* values,
542
33.6k
                                          size_t count) override {
543
33.6k
                            return Status::OK();
544
33.6k
                        }
545
33.6k
                        Status add_array_values(size_t field_size, const void* value_ptr,
546
33.6k
                                                const uint8_t* null_map, const uint8_t* offsets_ptr,
547
33.6k
                                                size_t count) override {
548
33.6k
                            return Status::OK();
549
33.6k
                        }
550
33.6k
                        Status add_nulls(uint32_t count) override { return Status::OK(); }
551
33.6k
                        Status add_array_nulls(const uint8_t* null_map, size_t num_rows) override {
552
33.6k
                            return Status::OK();
553
33.6k
                        }
554
33.6k
                        Status finish() override { return Status::OK(); }
555
33.6k
                        int64_t size() const override { return 0; }
556
33.6k
                        void close_on_error() override {}
557
33.6k
                    };
558
559
33.6k
                    _inverted_index_builders[i] =
560
33.6k
                            std::make_unique<InvertedIndexColumnWriterEmpty>();
561
562
33.6k
                    break;
563
33.6k
                });
564
565
33.6k
                RETURN_IF_ERROR(IndexColumnWriter::create(
566
33.6k
                        get_column(), &_inverted_index_builders[i], _opts.index_file_writer,
567
33.6k
                        _opts.inverted_indexes[i]));
568
33.6k
            }
569
33.4k
        } while (false);
570
33.4k
    }
571
1.00M
    if (_opts.need_bloom_filter) {
572
5.34k
        if (_opts.is_ngram_bf_index) {
573
3.38k
            RETURN_IF_ERROR(NGramBloomFilterIndexWriterImpl::create(
574
3.38k
                    BloomFilterOptions(), get_column()->type(), _opts.gram_size, _opts.gram_bf_size,
575
3.38k
                    &_bloom_filter_index_builder));
576
3.38k
        } else {
577
1.95k
            RETURN_IF_ERROR(BloomFilterIndexWriter::create(_opts.bf_options, get_column()->type(),
578
1.95k
                                                           &_bloom_filter_index_builder));
579
1.95k
        }
580
5.34k
    }
581
1.00M
    return Status::OK();
582
1.00M
}
583
584
1.43M
Status ScalarColumnWriter::append_nulls(size_t num_rows) {
585
1.43M
    _null_bitmap_builder->add_run(true, num_rows);
586
1.43M
    _next_rowid += num_rows;
587
1.43M
    if (_opts.need_zone_map) {
588
1.37M
        _zone_map_index_builder->add_nulls(cast_set<uint32_t>(num_rows));
589
1.37M
    }
590
1.43M
    if (_opts.need_inverted_index) {
591
427k
        for (const auto& builder : _inverted_index_builders) {
592
427k
            RETURN_IF_ERROR(builder->add_nulls(cast_set<uint32_t>(num_rows)));
593
427k
        }
594
427k
    }
595
1.43M
    if (_opts.need_bloom_filter) {
596
579
        _bloom_filter_index_builder->add_nulls(cast_set<uint32_t>(num_rows));
597
579
    }
598
1.43M
    return Status::OK();
599
1.43M
}
600
601
// append data to page builder. this function will make sure that
602
// num_rows must be written before return. And ptr will be modified
603
// to next data should be written
604
2.72M
Status ScalarColumnWriter::append_data(const uint8_t** ptr, size_t num_rows) {
605
2.72M
    size_t remaining = num_rows;
606
5.51M
    while (remaining > 0) {
607
2.79M
        size_t num_written = remaining;
608
2.79M
        RETURN_IF_ERROR(append_data_in_current_page(ptr, &num_written));
609
610
2.79M
        remaining -= num_written;
611
612
2.79M
        if (_page_builder->is_page_full()) {
613
75.5k
            RETURN_IF_ERROR(finish_current_page());
614
75.5k
        }
615
2.79M
    }
616
2.72M
    return Status::OK();
617
2.72M
}
618
619
Status ScalarColumnWriter::_internal_append_data_in_current_page(const uint8_t* data,
620
2.86M
                                                                 size_t* num_written) {
621
2.86M
    RETURN_IF_ERROR(_page_builder->add(data, num_written));
622
2.86M
    if (_opts.need_zone_map) {
623
2.58M
        _zone_map_index_builder->add_values(data, *num_written);
624
2.58M
    }
625
2.86M
    if (_opts.need_inverted_index) {
626
458k
        for (const auto& builder : _inverted_index_builders) {
627
458k
            RETURN_IF_ERROR(builder->add_values(get_column()->name(), data, *num_written));
628
458k
        }
629
458k
    }
630
2.86M
    if (_opts.need_bloom_filter) {
631
5.64k
        RETURN_IF_ERROR(_bloom_filter_index_builder->add_values(data, *num_written));
632
5.64k
    }
633
634
2.86M
    _next_rowid += *num_written;
635
636
    // we must write null bits after write data, because we don't
637
    // know how many rows can be written into current page
638
2.86M
    if (is_nullable()) {
639
2.33M
        _null_bitmap_builder->add_run(false, *num_written);
640
2.33M
    }
641
2.86M
    return Status::OK();
642
2.86M
}
643
644
2.86M
Status ScalarColumnWriter::append_data_in_current_page(const uint8_t** data, size_t* num_written) {
645
2.86M
    RETURN_IF_ERROR(append_data_in_current_page(*data, num_written));
646
2.86M
    *data += cell_size() * (*num_written);
647
2.86M
    return Status::OK();
648
2.86M
}
649
650
Status ScalarColumnWriter::append_nullable(const uint8_t* null_map, const uint8_t** ptr,
651
1.32M
                                           size_t num_rows) {
652
    // When optimization is disabled, use base class implementation
653
1.32M
    if (!config::enable_rle_batch_put_optimization) {
654
0
        return ColumnWriter::append_nullable(null_map, ptr, num_rows);
655
0
    }
656
657
1.32M
    if (UNLIKELY(num_rows == 0)) {
658
0
        return Status::OK();
659
0
    }
660
661
    // Build run-length encoded null runs using memchr for fast boundary detection
662
1.32M
    _null_run_buffer.clear();
663
1.32M
    if (_null_run_buffer.capacity() < num_rows) {
664
549k
        _null_run_buffer.reserve(std::min(num_rows, size_t(256)));
665
549k
    }
666
667
1.32M
    size_t non_null_count = 0;
668
1.32M
    size_t offset = 0;
669
3.46M
    while (offset < num_rows) {
670
2.14M
        bool is_null = null_map[offset] != 0;
671
2.14M
        size_t remaining = num_rows - offset;
672
2.14M
        const uint8_t* run_end =
673
2.14M
                static_cast<const uint8_t*>(memchr(null_map + offset, is_null ? 0 : 1, remaining));
674
2.14M
        size_t run_length = run_end != nullptr ? (run_end - (null_map + offset)) : remaining;
675
2.14M
        _null_run_buffer.push_back(NullRun {is_null, static_cast<uint32_t>(run_length)});
676
2.14M
        if (!is_null) {
677
1.67M
            non_null_count += run_length;
678
1.67M
        }
679
2.14M
        offset += run_length;
680
2.14M
    }
681
682
    // Pre-allocate buffer based on estimated size
683
1.33M
    if (_null_bitmap_builder != nullptr) {
684
1.33M
        size_t current_rows = _next_rowid - _first_rowid;
685
1.33M
        size_t expected_rows = current_rows + num_rows;
686
1.33M
        size_t est_non_null = non_null_count;
687
1.33M
        if (num_rows > 0 && expected_rows > num_rows) {
688
894k
            est_non_null = (non_null_count * expected_rows) / num_rows;
689
894k
        }
690
1.33M
        _null_bitmap_builder->reserve_for_write(expected_rows, est_non_null);
691
1.33M
    }
692
693
1.32M
    if (non_null_count == 0) {
694
        // All NULL: skip data writing, only update null bitmap and indexes
695
51.9k
        RETURN_IF_ERROR(append_nulls(num_rows));
696
51.9k
        *ptr += cell_size() * num_rows;
697
51.9k
        return Status::OK();
698
51.9k
    }
699
700
1.27M
    if (non_null_count == num_rows) {
701
        // All non-NULL: use normal append_data which handles both data and null bitmap
702
1.24M
        return append_data(ptr, num_rows);
703
1.24M
    }
704
705
    // Process by runs
706
844k
    for (const auto& run : _null_run_buffer) {
707
844k
        size_t run_length = run.len;
708
844k
        if (run.is_null) {
709
416k
            RETURN_IF_ERROR(append_nulls(run_length));
710
416k
            *ptr += cell_size() * run_length;
711
427k
        } else {
712
            // TODO:
713
            //  1. `*ptr += cell_size() * step;` should do in this function, not append_data;
714
            //  2. support array vectorized load and ptr offset add
715
427k
            RETURN_IF_ERROR(append_data(ptr, run_length));
716
427k
        }
717
844k
    }
718
719
27.2k
    return Status::OK();
720
27.2k
}
721
722
82.4k
uint64_t ScalarColumnWriter::estimate_buffer_size() {
723
82.4k
    uint64_t size = _data_size;
724
82.4k
    size += _page_builder->size();
725
82.4k
    if (is_nullable()) {
726
47.6k
        size += _null_bitmap_builder->size();
727
47.6k
    }
728
82.4k
    size += _ordinal_index_builder->size();
729
82.4k
    if (_opts.need_zone_map) {
730
68.8k
        size += _zone_map_index_builder->size();
731
68.8k
    }
732
82.4k
    if (_opts.need_bloom_filter) {
733
267
        size += _bloom_filter_index_builder->size();
734
267
    }
735
82.4k
    return size;
736
82.4k
}
737
738
1.00M
Status ScalarColumnWriter::finish() {
739
1.00M
    RETURN_IF_ERROR(finish_current_page());
740
1.00M
    _opts.meta->set_num_rows(_next_rowid);
741
1.00M
    return Status::OK();
742
1.00M
}
743
744
1.00M
Status ScalarColumnWriter::write_data() {
745
1.00M
    auto offset = _file_writer->bytes_appended();
746
1.33M
    auto collect_uncompressed_bytes = [](const PageFooterPB& footer) {
747
1.33M
        return footer.uncompressed_size() + footer.ByteSizeLong() +
748
1.33M
               sizeof(uint32_t) /* footer size */ + sizeof(uint32_t) /* checksum */;
749
1.33M
    };
750
1.03M
    for (auto& page : _pages) {
751
1.03M
        _total_uncompressed_data_pages_size += collect_uncompressed_bytes(page->footer);
752
1.03M
        RETURN_IF_ERROR(_write_data_page(page.get()));
753
1.03M
    }
754
1.00M
    _pages.clear();
755
    // write column dict
756
1.00M
    if (_encoding_info->encoding() == DICT_ENCODING) {
757
297k
        OwnedSlice dict_body;
758
297k
        RETURN_IF_ERROR(_page_builder->get_dictionary_page(&dict_body));
759
297k
        EncodingTypePB dict_word_page_encoding;
760
297k
        RETURN_IF_ERROR(_page_builder->get_dictionary_page_encoding(&dict_word_page_encoding));
761
762
297k
        PageFooterPB footer;
763
297k
        footer.set_type(DICTIONARY_PAGE);
764
297k
        footer.set_uncompressed_size(cast_set<uint32_t>(dict_body.slice().get_size()));
765
297k
        footer.mutable_dict_page_footer()->set_encoding(dict_word_page_encoding);
766
297k
        _total_uncompressed_data_pages_size += collect_uncompressed_bytes(footer);
767
768
297k
        PagePointer dict_pp;
769
297k
        RETURN_IF_ERROR(PageIO::compress_and_write_page(
770
297k
                _compress_codec, _opts.compression_min_space_saving, _file_writer,
771
297k
                {dict_body.slice()}, footer, &dict_pp));
772
297k
        dict_pp.to_proto(_opts.meta->mutable_dict_page());
773
297k
    }
774
1.00M
    _total_compressed_data_pages_size += _file_writer->bytes_appended() - offset;
775
1.00M
    _page_builder.reset();
776
1.00M
    return Status::OK();
777
1.00M
}
778
779
962k
Status ScalarColumnWriter::write_ordinal_index() {
780
962k
    return _ordinal_index_builder->finish(_file_writer, _opts.meta->add_indexes());
781
962k
}
782
783
790k
Status ScalarColumnWriter::write_zone_map() {
784
790k
    if (_opts.need_zone_map) {
785
737k
        return _zone_map_index_builder->finish(_file_writer, _opts.meta->add_indexes());
786
737k
    }
787
53.2k
    return Status::OK();
788
790k
}
789
790
686k
Status ScalarColumnWriter::write_inverted_index() {
791
686k
    if (_opts.need_inverted_index) {
792
33.6k
        for (const auto& builder : _inverted_index_builders) {
793
33.6k
            RETURN_IF_ERROR(builder->finish());
794
33.6k
        }
795
33.4k
    }
796
686k
    return Status::OK();
797
686k
}
798
799
684k
Status ScalarColumnWriter::write_bloom_filter_index() {
800
684k
    if (_opts.need_bloom_filter) {
801
5.34k
        return _bloom_filter_index_builder->finish(_file_writer, _opts.meta->add_indexes());
802
5.34k
    }
803
678k
    return Status::OK();
804
684k
}
805
806
// write a data page into file and update ordinal index
807
1.03M
Status ScalarColumnWriter::_write_data_page(Page* page) {
808
1.03M
    PagePointer pp;
809
1.03M
    std::vector<Slice> compressed_body;
810
1.98M
    for (auto& data : page->data) {
811
1.98M
        compressed_body.push_back(data.slice());
812
1.98M
    }
813
1.03M
    RETURN_IF_ERROR(PageIO::write_page(_file_writer, compressed_body, page->footer, &pp));
814
1.03M
    _ordinal_index_builder->append_entry(page->footer.data_page_footer().first_ordinal(), pp);
815
1.03M
    return Status::OK();
816
1.03M
}
817
818
1.08M
Status ScalarColumnWriter::finish_current_page() {
819
1.08M
    if (_next_rowid == _first_rowid) {
820
43.2k
        return Status::OK();
821
43.2k
    }
822
1.03M
    if (_opts.need_zone_map) {
823
        // If the number of rows in the current page is less than the threshold,
824
        // we will invalidate zone map index for this page by set pass_all to true.
825
803k
        if (_next_rowid - _first_rowid < config::zone_map_row_num_threshold) {
826
542k
            _zone_map_index_builder->invalid_page_zone_map();
827
542k
        }
828
803k
        RETURN_IF_ERROR(_zone_map_index_builder->flush());
829
803k
    }
830
831
1.03M
    if (_opts.need_bloom_filter) {
832
5.58k
        RETURN_IF_ERROR(_bloom_filter_index_builder->flush());
833
5.58k
    }
834
835
1.03M
    _raw_data_bytes += _page_builder->get_raw_data_size();
836
837
    // build data page body : encoded values + [nullmap]
838
1.03M
    std::vector<Slice> body;
839
1.03M
    OwnedSlice encoded_values;
840
1.03M
    RETURN_IF_ERROR(_page_builder->finish(&encoded_values));
841
1.03M
    RETURN_IF_ERROR(_page_builder->reset());
842
1.03M
    body.push_back(encoded_values.slice());
843
844
1.03M
    OwnedSlice nullmap;
845
1.03M
    if (_null_bitmap_builder != nullptr) {
846
546k
        if (is_nullable() && _null_bitmap_builder->has_null()) {
847
175k
            RETURN_IF_ERROR(_null_bitmap_builder->finish(&nullmap));
848
175k
            body.push_back(nullmap.slice());
849
175k
        }
850
546k
        _null_bitmap_builder->reset();
851
546k
    }
852
853
    // prepare data page footer
854
1.03M
    std::unique_ptr<Page> page(new Page());
855
1.03M
    page->footer.set_type(DATA_PAGE);
856
1.03M
    page->footer.set_uncompressed_size(cast_set<uint32_t>(Slice::compute_total_size(body)));
857
1.03M
    auto* data_page_footer = page->footer.mutable_data_page_footer();
858
1.03M
    data_page_footer->set_first_ordinal(_first_rowid);
859
1.03M
    data_page_footer->set_num_values(_next_rowid - _first_rowid);
860
1.03M
    data_page_footer->set_nullmap_size(cast_set<uint32_t>(nullmap.slice().size));
861
1.03M
    if (_new_page_callback != nullptr) {
862
68.9k
        _new_page_callback->put_extra_info_in_page(data_page_footer);
863
68.9k
    }
864
    // trying to compress page body
865
1.03M
    OwnedSlice compressed_body;
866
1.03M
    RETURN_IF_ERROR(PageIO::compress_page_body(_compress_codec, _opts.compression_min_space_saving,
867
1.03M
                                               body, &compressed_body));
868
1.03M
    if (compressed_body.slice().empty()) {
869
        // page body is uncompressed
870
949k
        page->data.emplace_back(std::move(encoded_values));
871
949k
        page->data.emplace_back(std::move(nullmap));
872
949k
    } else {
873
        // page body is compressed
874
89.2k
        page->data.emplace_back(std::move(compressed_body));
875
89.2k
    }
876
877
1.03M
    _push_back_page(std::move(page));
878
1.03M
    _first_rowid = _next_rowid;
879
1.03M
    return Status::OK();
880
1.03M
}
881
882
////////////////////////////////////////////////////////////////////////////////
883
884
////////////////////////////////////////////////////////////////////////////////
885
// offset column writer
886
////////////////////////////////////////////////////////////////////////////////
887
888
OffsetColumnWriter::OffsetColumnWriter(const ColumnWriterOptions& opts, TabletColumnPtr column,
889
                                       io::FileWriter* file_writer)
890
69.0k
        : ScalarColumnWriter(opts, std::move(column), file_writer) {
891
    // now we only explain data in offset column as uint64
892
69.0k
    DCHECK(get_column()->type() == FieldType::OLAP_FIELD_TYPE_UNSIGNED_BIGINT);
893
69.0k
}
894
895
69.0k
OffsetColumnWriter::~OffsetColumnWriter() = default;
896
897
68.8k
Status OffsetColumnWriter::init() {
898
68.8k
    RETURN_IF_ERROR(ScalarColumnWriter::init());
899
68.8k
    register_flush_page_callback(this);
900
68.8k
    _next_offset = 0;
901
68.8k
    return Status::OK();
902
68.8k
}
903
904
68.9k
Status OffsetColumnWriter::append_data(const uint8_t** ptr, size_t num_rows) {
905
68.9k
    size_t remaining = num_rows;
906
138k
    while (remaining > 0) {
907
69.5k
        size_t num_written = remaining;
908
69.5k
        RETURN_IF_ERROR(append_data_in_current_page(ptr, &num_written));
909
        // Callers provide one extra tail offset after the written rows so the page footer can
910
        // store the next array item ordinal for the current page.
911
69.5k
        _next_offset = *(const uint64_t*)(*ptr);
912
69.5k
        remaining -= num_written;
913
914
69.5k
        if (_page_builder->is_page_full()) {
915
            // get next data for next array_item_rowid
916
594
            RETURN_IF_ERROR(finish_current_page());
917
594
        }
918
69.5k
    }
919
68.9k
    return Status::OK();
920
68.9k
}
921
922
68.9k
void OffsetColumnWriter::put_extra_info_in_page(DataPageFooterPB* footer) {
923
68.9k
    footer->set_next_array_item_ordinal(_next_offset);
924
68.9k
}
925
926
StructColumnWriter::StructColumnWriter(
927
        const ColumnWriterOptions& opts, TabletColumnPtr column, ScalarColumnWriter* null_writer,
928
        std::vector<std::unique_ptr<ColumnWriter>>& sub_column_writers)
929
2.27k
        : ColumnWriter(std::move(column), opts.meta->is_nullable(), opts.meta), _opts(opts) {
930
10.0k
    for (auto& sub_column_writer : sub_column_writers) {
931
10.0k
        _sub_column_writers.push_back(std::move(sub_column_writer));
932
10.0k
    }
933
2.27k
    _num_sub_column_writers = _sub_column_writers.size();
934
2.27k
    DCHECK(_num_sub_column_writers >= 1);
935
2.27k
    if (is_nullable()) {
936
2.00k
        _null_writer.reset(null_writer);
937
2.00k
    }
938
2.27k
}
939
940
2.27k
Status StructColumnWriter::init() {
941
10.0k
    for (auto& column_writer : _sub_column_writers) {
942
10.0k
        RETURN_IF_ERROR(column_writer->init());
943
10.0k
    }
944
2.27k
    if (is_nullable()) {
945
2.00k
        RETURN_IF_ERROR(_null_writer->init());
946
2.00k
    }
947
2.27k
    return Status::OK();
948
2.27k
}
949
950
1.64k
Status StructColumnWriter::write_inverted_index() {
951
1.64k
    if (_opts.need_inverted_index) {
952
0
        for (auto& column_writer : _sub_column_writers) {
953
0
            RETURN_IF_ERROR(column_writer->write_inverted_index());
954
0
        }
955
0
    }
956
1.64k
    return Status::OK();
957
1.64k
}
958
959
Status StructColumnWriter::append_nullable(const uint8_t* null_map, const uint8_t** ptr,
960
1.96k
                                           size_t num_rows) {
961
1.96k
    RETURN_IF_ERROR(append_data(ptr, num_rows));
962
1.96k
    RETURN_IF_ERROR(_null_writer->append_data(&null_map, num_rows));
963
1.96k
    return Status::OK();
964
1.96k
}
965
966
2.23k
Status StructColumnWriter::append_data(const uint8_t** ptr, size_t num_rows) {
967
2.23k
    const auto* results = reinterpret_cast<const uint64_t*>(*ptr);
968
11.9k
    for (size_t i = 0; i < _num_sub_column_writers; ++i) {
969
9.68k
        auto nullmap = *(results + _num_sub_column_writers + i);
970
9.68k
        auto data = *(results + i);
971
9.68k
        RETURN_IF_ERROR(_sub_column_writers[i]->append(reinterpret_cast<const uint8_t*>(nullmap),
972
9.68k
                                                       reinterpret_cast<const void*>(data),
973
9.68k
                                                       num_rows));
974
9.68k
    }
975
2.23k
    return Status::OK();
976
2.23k
}
977
978
186
uint64_t StructColumnWriter::estimate_buffer_size() {
979
186
    uint64_t size = 0;
980
780
    for (auto& column_writer : _sub_column_writers) {
981
780
        size += column_writer->estimate_buffer_size();
982
780
    }
983
186
    size += is_nullable() ? _null_writer->estimate_buffer_size() : 0;
984
186
    return size;
985
186
}
986
987
2.27k
Status StructColumnWriter::finish() {
988
10.0k
    for (auto& column_writer : _sub_column_writers) {
989
10.0k
        RETURN_IF_ERROR(column_writer->finish());
990
10.0k
    }
991
2.27k
    if (is_nullable()) {
992
2.00k
        RETURN_IF_ERROR(_null_writer->finish());
993
2.00k
    }
994
2.27k
    _opts.meta->set_num_rows(get_next_rowid());
995
2.27k
    return Status::OK();
996
2.27k
}
997
998
2.27k
Status StructColumnWriter::write_data() {
999
10.0k
    for (auto& column_writer : _sub_column_writers) {
1000
10.0k
        RETURN_IF_ERROR(column_writer->write_data());
1001
10.0k
    }
1002
2.27k
    if (is_nullable()) {
1003
2.00k
        RETURN_IF_ERROR(_null_writer->write_data());
1004
2.00k
    }
1005
2.27k
    return Status::OK();
1006
2.27k
}
1007
1008
2.23k
Status StructColumnWriter::write_ordinal_index() {
1009
9.68k
    for (auto& column_writer : _sub_column_writers) {
1010
9.68k
        RETURN_IF_ERROR(column_writer->write_ordinal_index());
1011
9.68k
    }
1012
2.23k
    if (is_nullable()) {
1013
1.96k
        RETURN_IF_ERROR(_null_writer->write_ordinal_index());
1014
1.96k
    }
1015
2.23k
    return Status::OK();
1016
2.23k
}
1017
1018
0
Status StructColumnWriter::append_nulls(size_t num_rows) {
1019
0
    for (auto& column_writer : _sub_column_writers) {
1020
0
        RETURN_IF_ERROR(column_writer->append_nulls(num_rows));
1021
0
    }
1022
0
    if (is_nullable()) {
1023
0
        std::vector<UInt8> null_signs(num_rows, 1);
1024
0
        const uint8_t* null_sign_ptr = null_signs.data();
1025
0
        RETURN_IF_ERROR(_null_writer->append_data(&null_sign_ptr, num_rows));
1026
0
    }
1027
0
    return Status::OK();
1028
0
}
1029
1030
0
Status StructColumnWriter::finish_current_page() {
1031
0
    return Status::NotSupported("struct writer has no data, can not finish_current_page");
1032
0
}
1033
1034
ArrayColumnWriter::ArrayColumnWriter(const ColumnWriterOptions& opts, TabletColumnPtr column,
1035
                                     OffsetColumnWriter* offset_writer,
1036
                                     ScalarColumnWriter* null_writer,
1037
                                     std::unique_ptr<ColumnWriter> item_writer)
1038
43.7k
        : ColumnWriter(std::move(column), opts.meta->is_nullable(), opts.meta),
1039
43.7k
          _item_writer(std::move(item_writer)),
1040
43.7k
          _opts(opts) {
1041
43.7k
    _offset_writer.reset(offset_writer);
1042
43.7k
    if (is_nullable()) {
1043
27.8k
        _null_writer.reset(null_writer);
1044
27.8k
    }
1045
43.7k
}
1046
1047
43.6k
Status ArrayColumnWriter::init() {
1048
43.6k
    RETURN_IF_ERROR(_offset_writer->init());
1049
43.6k
    if (is_nullable()) {
1050
27.9k
        RETURN_IF_ERROR(_null_writer->init());
1051
27.9k
    }
1052
43.6k
    RETURN_IF_ERROR(_item_writer->init());
1053
43.6k
    if (_opts.need_inverted_index) {
1054
1.95k
        auto* writer = dynamic_cast<ScalarColumnWriter*>(_item_writer.get());
1055
1.95k
        if (writer != nullptr) {
1056
1.95k
            RETURN_IF_ERROR(IndexColumnWriter::create(get_column(), &_inverted_index_writer,
1057
1.95k
                                                      _opts.index_file_writer,
1058
1.95k
                                                      _opts.inverted_indexes[0]));
1059
1.95k
        }
1060
1.95k
    }
1061
43.6k
    if (_opts.need_ann_index) {
1062
60
        auto* writer = dynamic_cast<ScalarColumnWriter*>(_item_writer.get());
1063
60
        if (writer != nullptr) {
1064
60
            _ann_index_writer = std::make_unique<AnnIndexColumnWriter>(_opts.index_file_writer,
1065
60
                                                                       _opts.ann_index);
1066
60
            RETURN_IF_ERROR(_ann_index_writer->init());
1067
60
        }
1068
60
    }
1069
43.6k
    return Status::OK();
1070
43.6k
}
1071
1072
40.7k
Status ArrayColumnWriter::write_inverted_index() {
1073
40.7k
    if (_opts.need_inverted_index) {
1074
1.95k
        return _inverted_index_writer->finish();
1075
1.95k
    }
1076
38.8k
    return Status::OK();
1077
40.7k
}
1078
1079
40.7k
Status ArrayColumnWriter::write_ann_index() {
1080
40.7k
    if (_opts.need_ann_index) {
1081
57
        return _ann_index_writer->finish();
1082
57
    }
1083
40.6k
    return Status::OK();
1084
40.7k
}
1085
1086
// batch append data for array
1087
43.5k
Status ArrayColumnWriter::append_data(const uint8_t** ptr, size_t num_rows) {
1088
    // data_ptr contains
1089
    // [size, offset_ptr, item_data_ptr, item_nullmap_ptr]
1090
43.5k
    auto data_ptr = reinterpret_cast<const uint64_t*>(*ptr);
1091
    // total number length
1092
43.5k
    size_t element_cnt = size_t((unsigned long)(*data_ptr));
1093
43.5k
    auto offset_data = *(data_ptr + 1);
1094
43.5k
    const uint8_t* offsets_ptr = (const uint8_t*)offset_data;
1095
43.5k
    auto data = *(data_ptr + 2);
1096
43.5k
    auto nested_null_map = *(data_ptr + 3);
1097
43.5k
    if (element_cnt > 0) {
1098
27.2k
        RETURN_IF_ERROR(_item_writer->append(reinterpret_cast<const uint8_t*>(nested_null_map),
1099
27.2k
                                             reinterpret_cast<const void*>(data), element_cnt));
1100
27.2k
    }
1101
43.5k
    if (_opts.need_inverted_index) {
1102
2.04k
        auto* writer = dynamic_cast<ScalarColumnWriter*>(_item_writer.get());
1103
        // now only support nested type is scala
1104
2.04k
        if (writer != nullptr) {
1105
            //NOTE: use array field name as index field, but item_writer size should be used when moving item_data_ptr
1106
2.04k
            RETURN_IF_ERROR(_inverted_index_writer->add_array_values(
1107
2.04k
                    field_type_size(_item_writer->get_column()->type()),
1108
2.04k
                    reinterpret_cast<const void*>(data),
1109
2.04k
                    reinterpret_cast<const uint8_t*>(nested_null_map), offsets_ptr, num_rows));
1110
2.04k
        }
1111
2.04k
    }
1112
1113
43.5k
    if (_opts.need_ann_index) {
1114
60
        auto* writer = dynamic_cast<ScalarColumnWriter*>(_item_writer.get());
1115
        // now only support nested type is scala
1116
60
        if (writer != nullptr) {
1117
            //NOTE: use array field name as index field, but item_writer size should be used when moving item_data_ptr
1118
60
            RETURN_IF_ERROR(_ann_index_writer->add_array_values(
1119
60
                    field_type_size(_item_writer->get_column()->type()),
1120
60
                    reinterpret_cast<const void*>(data),
1121
60
                    reinterpret_cast<const uint8_t*>(nested_null_map), offsets_ptr, num_rows));
1122
60
        } else {
1123
0
            return Status::NotSupported(
1124
0
                    "Ann index can only be build on array with scalar type. but got {} as "
1125
0
                    "nested",
1126
0
                    _item_writer->get_column()->type());
1127
0
        }
1128
60
    }
1129
1130
43.5k
    RETURN_IF_ERROR(_offset_writer->append_data(&offsets_ptr, num_rows));
1131
43.5k
    return Status::OK();
1132
43.5k
}
1133
1134
651
uint64_t ArrayColumnWriter::estimate_buffer_size() {
1135
651
    return _offset_writer->estimate_buffer_size() +
1136
651
           (is_nullable() ? _null_writer->estimate_buffer_size() : 0) +
1137
651
           _item_writer->estimate_buffer_size();
1138
651
}
1139
1140
Status ArrayColumnWriter::append_nullable(const uint8_t* null_map, const uint8_t** ptr,
1141
27.6k
                                          size_t num_rows) {
1142
27.6k
    RETURN_IF_ERROR(append_data(ptr, num_rows));
1143
27.6k
    if (is_nullable()) {
1144
27.6k
        if (_opts.need_inverted_index) {
1145
1.57k
            RETURN_IF_ERROR(_inverted_index_writer->add_array_nulls(null_map, num_rows));
1146
1.57k
        }
1147
27.6k
        RETURN_IF_ERROR(_null_writer->append_data(&null_map, num_rows));
1148
27.6k
    }
1149
27.6k
    return Status::OK();
1150
27.6k
}
1151
1152
43.8k
Status ArrayColumnWriter::finish() {
1153
43.8k
    RETURN_IF_ERROR(_offset_writer->finish());
1154
43.8k
    if (is_nullable()) {
1155
27.9k
        RETURN_IF_ERROR(_null_writer->finish());
1156
27.9k
    }
1157
43.8k
    RETURN_IF_ERROR(_item_writer->finish());
1158
43.8k
    _opts.meta->set_num_rows(get_next_rowid());
1159
43.8k
    return Status::OK();
1160
43.8k
}
1161
1162
43.8k
Status ArrayColumnWriter::write_data() {
1163
43.8k
    RETURN_IF_ERROR(_offset_writer->write_data());
1164
43.8k
    if (is_nullable()) {
1165
27.9k
        RETURN_IF_ERROR(_null_writer->write_data());
1166
27.9k
    }
1167
43.8k
    RETURN_IF_ERROR(_item_writer->write_data());
1168
43.8k
    return Status::OK();
1169
43.8k
}
1170
1171
43.3k
Status ArrayColumnWriter::write_ordinal_index() {
1172
43.3k
    RETURN_IF_ERROR(_offset_writer->write_ordinal_index());
1173
43.3k
    if (is_nullable()) {
1174
27.4k
        RETURN_IF_ERROR(_null_writer->write_ordinal_index());
1175
27.4k
    }
1176
43.3k
    if (!has_empty_items()) {
1177
27.1k
        RETURN_IF_ERROR(_item_writer->write_ordinal_index());
1178
27.1k
    }
1179
43.3k
    return Status::OK();
1180
43.3k
}
1181
1182
0
Status ArrayColumnWriter::append_nulls(size_t num_rows) {
1183
0
    const UInt64 offset = cast_set<UInt64>(_item_writer->get_next_rowid());
1184
0
    std::vector<UInt64> offsets_data(num_rows + 1, offset);
1185
0
    const uint8_t* offsets_ptr = reinterpret_cast<const uint8_t*>(offsets_data.data());
1186
0
    RETURN_IF_ERROR(_offset_writer->append_data(&offsets_ptr, num_rows));
1187
0
    return write_null_column(num_rows, true);
1188
0
}
1189
1190
0
Status ArrayColumnWriter::write_null_column(size_t num_rows, bool is_null) {
1191
0
    uint8_t null_sign = is_null ? 1 : 0;
1192
0
    while (is_nullable() && num_rows > 0) {
1193
        // TODO llj bulk write
1194
0
        const uint8_t* null_sign_ptr = &null_sign;
1195
0
        RETURN_IF_ERROR(_null_writer->append_data(&null_sign_ptr, 1));
1196
0
        --num_rows;
1197
0
    }
1198
0
    return Status::OK();
1199
0
}
1200
1201
0
Status ArrayColumnWriter::finish_current_page() {
1202
0
    return Status::NotSupported("array writer has no data, can not finish_current_page");
1203
0
}
1204
1205
/// ============================= MapColumnWriter =====================////
1206
MapColumnWriter::MapColumnWriter(const ColumnWriterOptions& opts, TabletColumnPtr column,
1207
                                 ScalarColumnWriter* null_writer, OffsetColumnWriter* offset_writer,
1208
                                 std::vector<std::unique_ptr<ColumnWriter>>& kv_writers)
1209
25.2k
        : ColumnWriter(std::move(column), opts.meta->is_nullable(), opts.meta), _opts(opts) {
1210
25.2k
    CHECK_EQ(kv_writers.size(), 2);
1211
25.2k
    _offsets_writer.reset(offset_writer);
1212
25.2k
    if (is_nullable()) {
1213
8.60k
        _null_writer.reset(null_writer);
1214
8.60k
    }
1215
50.5k
    for (auto& sub_writers : kv_writers) {
1216
50.5k
        _kv_writers.push_back(std::move(sub_writers));
1217
50.5k
    }
1218
25.2k
}
1219
1220
25.2k
Status MapColumnWriter::init() {
1221
25.2k
    RETURN_IF_ERROR(_offsets_writer->init());
1222
25.2k
    if (is_nullable()) {
1223
8.61k
        RETURN_IF_ERROR(_null_writer->init());
1224
8.61k
    }
1225
    // here register_flush_page_callback to call this.put_extra_info_in_page()
1226
    // when finish cur data page
1227
50.5k
    for (auto& sub_writer : _kv_writers) {
1228
50.5k
        RETURN_IF_ERROR(sub_writer->init());
1229
50.5k
    }
1230
25.2k
    return Status::OK();
1231
25.2k
}
1232
1233
994
uint64_t MapColumnWriter::estimate_buffer_size() {
1234
994
    size_t estimate = 0;
1235
1.98k
    for (auto& sub_writer : _kv_writers) {
1236
1.98k
        estimate += sub_writer->estimate_buffer_size();
1237
1.98k
    }
1238
994
    estimate += _offsets_writer->estimate_buffer_size();
1239
994
    if (is_nullable()) {
1240
973
        estimate += _null_writer->estimate_buffer_size();
1241
973
    }
1242
994
    return estimate;
1243
994
}
1244
1245
25.2k
Status MapColumnWriter::finish() {
1246
25.2k
    RETURN_IF_ERROR(_offsets_writer->finish());
1247
25.2k
    if (is_nullable()) {
1248
8.61k
        RETURN_IF_ERROR(_null_writer->finish());
1249
8.61k
    }
1250
50.5k
    for (auto& sub_writer : _kv_writers) {
1251
50.5k
        RETURN_IF_ERROR(sub_writer->finish());
1252
50.5k
    }
1253
25.2k
    _opts.meta->set_num_rows(get_next_rowid());
1254
25.2k
    return Status::OK();
1255
25.2k
}
1256
1257
Status MapColumnWriter::append_nullable(const uint8_t* null_map, const uint8_t** ptr,
1258
8.56k
                                        size_t num_rows) {
1259
8.56k
    RETURN_IF_ERROR(append_data(ptr, num_rows));
1260
8.56k
    if (is_nullable()) {
1261
8.56k
        RETURN_IF_ERROR(_null_writer->append_data(&null_map, num_rows));
1262
8.56k
    }
1263
8.56k
    return Status::OK();
1264
8.56k
}
1265
1266
// write key value data with offsets
1267
25.4k
Status MapColumnWriter::append_data(const uint8_t** ptr, size_t num_rows) {
1268
    // data_ptr contains
1269
    // [size, offset_ptr, key_data_ptr, val_data_ptr, k_nullmap_ptr, v_nullmap_pr]
1270
    // which converted results from olap_map_convertor and later will use a structure to replace it
1271
25.4k
    auto data_ptr = reinterpret_cast<const uint64_t*>(*ptr);
1272
    // total number length
1273
25.4k
    size_t element_cnt = size_t((unsigned long)(*data_ptr));
1274
25.4k
    auto offset_data = *(data_ptr + 1);
1275
25.4k
    const uint8_t* offsets_ptr = (const uint8_t*)offset_data;
1276
1277
25.4k
    if (element_cnt > 0) {
1278
39.2k
        for (size_t i = 0; i < 2; ++i) {
1279
26.1k
            auto data = *(data_ptr + 2 + i);
1280
26.1k
            auto nested_null_map = *(data_ptr + 2 + 2 + i);
1281
26.1k
            RETURN_IF_ERROR(
1282
26.1k
                    _kv_writers[i]->append(reinterpret_cast<const uint8_t*>(nested_null_map),
1283
26.1k
                                           reinterpret_cast<const void*>(data), element_cnt));
1284
26.1k
        }
1285
13.0k
    }
1286
    // make sure the order : offset writer flush next_array_item_ordinal after kv_writers append_data
1287
    // because we use _kv_writers[0]->get_next_rowid() to set next_array_item_ordinal in offset page footer
1288
25.4k
    RETURN_IF_ERROR(_offsets_writer->append_data(&offsets_ptr, num_rows));
1289
25.4k
    return Status::OK();
1290
25.4k
}
1291
1292
25.2k
Status MapColumnWriter::write_data() {
1293
25.2k
    RETURN_IF_ERROR(_offsets_writer->write_data());
1294
25.2k
    if (is_nullable()) {
1295
8.61k
        RETURN_IF_ERROR(_null_writer->write_data());
1296
8.61k
    }
1297
50.5k
    for (auto& sub_writer : _kv_writers) {
1298
50.5k
        RETURN_IF_ERROR(sub_writer->write_data());
1299
50.5k
    }
1300
25.2k
    return Status::OK();
1301
25.2k
}
1302
1303
25.0k
Status MapColumnWriter::write_ordinal_index() {
1304
25.0k
    RETURN_IF_ERROR(_offsets_writer->write_ordinal_index());
1305
25.0k
    if (is_nullable()) {
1306
8.35k
        RETURN_IF_ERROR(_null_writer->write_ordinal_index());
1307
8.35k
    }
1308
50.0k
    for (auto& sub_writer : _kv_writers) {
1309
50.0k
        if (sub_writer->get_next_rowid() != 0) {
1310
25.3k
            RETURN_IF_ERROR(sub_writer->write_ordinal_index());
1311
25.3k
        }
1312
50.0k
    }
1313
25.0k
    return Status::OK();
1314
25.0k
}
1315
1316
0
Status MapColumnWriter::append_nulls(size_t num_rows) {
1317
0
    for (auto& sub_writer : _kv_writers) {
1318
0
        RETURN_IF_ERROR(sub_writer->append_nulls(num_rows));
1319
0
    }
1320
0
    const UInt64 offset = cast_set<UInt64>(_kv_writers[0]->get_next_rowid());
1321
0
    std::vector<UInt64> offsets_data(num_rows + 1, offset);
1322
0
    const uint8_t* offsets_ptr = reinterpret_cast<const uint8_t*>(offsets_data.data());
1323
0
    RETURN_IF_ERROR(_offsets_writer->append_data(&offsets_ptr, num_rows));
1324
1325
0
    if (is_nullable()) {
1326
0
        std::vector<UInt8> null_signs(num_rows, 1);
1327
0
        const uint8_t* null_sign_ptr = null_signs.data();
1328
0
        RETURN_IF_ERROR(_null_writer->append_data(&null_sign_ptr, num_rows));
1329
0
    }
1330
0
    return Status::OK();
1331
0
}
1332
1333
0
Status MapColumnWriter::finish_current_page() {
1334
0
    return Status::NotSupported("map writer has no data, can not finish_current_page");
1335
0
}
1336
1337
10.2k
Status MapColumnWriter::write_inverted_index() {
1338
10.2k
    if (_opts.need_inverted_index) {
1339
0
        return _index_builder->finish();
1340
0
    }
1341
10.2k
    return Status::OK();
1342
10.2k
}
1343
1344
VariantColumnWriter::VariantColumnWriter(const ColumnWriterOptions& opts, TabletColumnPtr column)
1345
5.99k
        : ColumnWriter(std::move(column), opts.meta->is_nullable(), opts.meta) {
1346
5.99k
    _impl = std::make_unique<VariantColumnWriterImpl>(opts, get_column());
1347
5.99k
}
1348
1349
6.00k
Status VariantColumnWriter::init() {
1350
6.00k
    return _impl->init();
1351
6.00k
}
1352
1353
685
Status VariantColumnWriter::append_data(const uint8_t** ptr, size_t num_rows) {
1354
685
    _next_rowid += num_rows;
1355
685
    return _impl->append_data(ptr, num_rows);
1356
685
}
1357
1358
868
uint64_t VariantColumnWriter::estimate_buffer_size() {
1359
868
    return _impl->estimate_buffer_size();
1360
868
}
1361
1362
6.03k
Status VariantColumnWriter::finish() {
1363
6.03k
    return _impl->finish();
1364
6.03k
}
1365
6.03k
Status VariantColumnWriter::write_data() {
1366
6.03k
    return _impl->write_data();
1367
6.03k
}
1368
6.03k
Status VariantColumnWriter::write_ordinal_index() {
1369
6.03k
    return _impl->write_ordinal_index();
1370
6.03k
}
1371
1372
6.03k
Status VariantColumnWriter::write_zone_map() {
1373
6.03k
    return _impl->write_zone_map();
1374
6.03k
}
1375
1376
6.00k
Status VariantColumnWriter::write_inverted_index() {
1377
6.00k
    return _impl->write_inverted_index();
1378
6.00k
}
1379
6.00k
Status VariantColumnWriter::write_bloom_filter_index() {
1380
6.00k
    return _impl->write_bloom_filter_index();
1381
6.00k
}
1382
1383
Status VariantColumnWriter::append_nullable(const uint8_t* null_map, const uint8_t** ptr,
1384
5.60k
                                            size_t num_rows) {
1385
5.60k
    return _impl->append_nullable(null_map, ptr, num_rows);
1386
5.60k
}
1387
1388
} // namespace doris::segment_v2