Coverage Report

Created: 2026-06-05 04:09

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
465k
    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
3.14M
    void reserve_for_write(size_t num_rows, size_t non_null_count) {
64
3.14M
        if (num_rows == 0) {
65
0
            return;
66
0
        }
67
3.14M
        if (non_null_count == 0 || (non_null_count == num_rows && !_has_null)) {
68
374k
            if (_bitmap_buf.capacity() < kSmallReserveBytes) {
69
1
                _bitmap_buf.reserve(kSmallReserveBytes);
70
1
            }
71
374k
            return;
72
374k
        }
73
2.77M
        size_t raw_bytes = BitmapSize(num_rows);
74
2.77M
        size_t run_est = std::min(num_rows, non_null_count * 2 + 1);
75
2.77M
        size_t run_bytes_est = run_est * kBytesPerRun + kReserveSlackBytes;
76
2.77M
        size_t raw_overhead = raw_bytes / 63 + 1;
77
2.77M
        size_t raw_est = raw_bytes + raw_overhead + kReserveSlackBytes;
78
2.77M
        size_t reserve_bytes = std::min(raw_est, run_bytes_est);
79
2.77M
        if (_bitmap_buf.capacity() < reserve_bytes) {
80
10.8k
            const size_t cap = _bitmap_buf.capacity();
81
10.8k
            const size_t grow = cap + cap / 2;
82
10.8k
            const size_t new_cap = std::max(reserve_bytes, grow);
83
10.8k
            _bitmap_buf.reserve(new_cap);
84
10.8k
        }
85
2.77M
    }
86
87
7.27M
    void add_run(bool value, size_t run) {
88
7.27M
        _has_null |= value;
89
7.27M
        _rle_encoder.Put(value, run);
90
7.27M
    }
91
92
    // Returns whether the building nullmap contains nullptr
93
464k
    bool has_null() const { return _has_null; }
94
95
131k
    Status finish(OwnedSlice* slice) {
96
131k
        _rle_encoder.Flush();
97
131k
        RETURN_IF_CATCH_EXCEPTION({ *slice = _bitmap_buf.build(); });
98
131k
        return Status::OK();
99
131k
    }
100
101
464k
    void reset() {
102
464k
        _has_null = false;
103
464k
        _rle_encoder.Clear();
104
464k
    }
105
106
29.1k
    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
68.5k
                                           io::FileWriter* file_writer, uint32_t id) {
120
68.5k
    if (!opts.meta->is_nullable()) {
121
32.0k
        return nullptr;
122
32.0k
    }
123
124
36.5k
    FieldType null_type = FieldType::OLAP_FIELD_TYPE_TINYINT;
125
36.5k
    ColumnWriterOptions null_options;
126
36.5k
    null_options.meta = opts.meta->add_children_columns();
127
36.5k
    null_options.meta->set_column_id(id);
128
36.5k
    null_options.meta->set_unique_id(id);
129
36.5k
    null_options.meta->set_type(int(null_type));
130
36.5k
    null_options.meta->set_is_nullable(false);
131
36.5k
    null_options.meta->set_length(
132
36.5k
            cast_set<int32_t>(field_type_size(FieldType::OLAP_FIELD_TYPE_TINYINT)));
133
36.5k
    null_options.meta->set_compression(opts.meta->compression());
134
135
36.5k
    null_options.need_zone_map = false;
136
36.5k
    null_options.need_bloom_filter = false;
137
36.5k
    null_options.storage_format = opts.storage_format;
138
139
36.5k
    auto null_column_ptr = std::make_shared<TabletColumn>(
140
36.5k
            FieldAggregationMethod::OLAP_FIELD_AGGREGATION_NONE, null_type, false,
141
36.5k
            null_options.meta->unique_id(), null_options.meta->length());
142
36.5k
    null_column_ptr->set_name("nullable");
143
36.5k
    null_column_ptr->set_index_length(-1); // no short key index
144
36.5k
    null_options.meta->set_encoding(
145
36.5k
            EncodingInfo::resolve_default_encoding(opts.storage_format, *null_column_ptr));
146
36.5k
    return new ScalarColumnWriter(null_options, std::move(null_column_ptr), file_writer);
147
68.5k
}
148
149
ColumnWriter::ColumnWriter(TabletColumnPtr column, bool is_nullable, ColumnMetaPB* meta)
150
953k
        : _column(std::move(column)), _is_nullable(is_nullable), _column_meta(meta) {
151
953k
    _data_type = DataTypeFactory::instance().create_data_type(*_column_meta);
152
953k
}
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
9.25k
        const TabletColumn& sub_column = column->get_sub_column(i);
162
9.25k
        RETURN_IF_ERROR(sub_column.check_valid());
163
164
        // create sub writer
165
9.25k
        ColumnWriterOptions column_options;
166
9.25k
        column_options.meta = opts.meta->mutable_children_columns(i);
167
9.25k
        column_options.need_zone_map = false;
168
9.25k
        column_options.need_bloom_filter = sub_column.is_bf_column();
169
9.25k
        column_options.storage_format = opts.storage_format;
170
9.25k
        std::unique_ptr<ColumnWriter> sub_column_writer;
171
9.25k
        RETURN_IF_ERROR(
172
9.25k
                ColumnWriter::create(column_options, &sub_column, file_writer, &sub_column_writer));
173
9.25k
        sub_column_writers.push_back(std::move(sub_column_writer));
174
9.25k
    }
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
42.6k
                                         std::unique_ptr<ColumnWriter>* writer) {
187
42.6k
    DCHECK(column->get_subtype_count() == 1);
188
42.6k
    const TabletColumn& item_column = column->get_sub_column(0);
189
42.6k
    RETURN_IF_ERROR(item_column.check_valid());
190
191
    // create item writer
192
42.6k
    ColumnWriterOptions item_options;
193
42.6k
    item_options.meta = opts.meta->mutable_children_columns(0);
194
42.6k
    item_options.need_zone_map = false;
195
42.6k
    item_options.need_bloom_filter = item_column.is_bf_column();
196
42.6k
    item_options.storage_format = opts.storage_format;
197
42.6k
    std::unique_ptr<ColumnWriter> item_writer;
198
42.6k
    RETURN_IF_ERROR(ColumnWriter::create(item_options, &item_column, file_writer, &item_writer));
199
200
    // create length writer
201
42.6k
    FieldType length_type = FieldType::OLAP_FIELD_TYPE_UNSIGNED_BIGINT;
202
203
42.6k
    ColumnWriterOptions length_options;
204
42.6k
    length_options.meta = opts.meta->add_children_columns();
205
42.6k
    length_options.meta->set_column_id(2);
206
42.6k
    length_options.meta->set_unique_id(2);
207
42.6k
    length_options.meta->set_type(int(length_type));
208
42.6k
    length_options.meta->set_is_nullable(false);
209
42.6k
    length_options.meta->set_length(
210
42.6k
            cast_set<int32_t>(field_type_size(FieldType::OLAP_FIELD_TYPE_UNSIGNED_BIGINT)));
211
42.6k
    length_options.meta->set_compression(opts.meta->compression());
212
213
42.6k
    length_options.need_zone_map = false;
214
42.6k
    length_options.need_bloom_filter = false;
215
42.6k
    length_options.storage_format = opts.storage_format;
216
217
42.6k
    auto length_column_ptr = std::make_shared<TabletColumn>(
218
42.6k
            FieldAggregationMethod::OLAP_FIELD_AGGREGATION_NONE, length_type,
219
42.6k
            length_options.meta->is_nullable(), length_options.meta->unique_id(),
220
42.6k
            length_options.meta->length());
221
42.6k
    length_column_ptr->set_name("length");
222
42.6k
    length_column_ptr->set_index_length(-1); // no short key index
223
42.6k
    length_options.meta->set_encoding(
224
42.6k
            EncodingInfo::resolve_default_encoding(opts.storage_format, *length_column_ptr));
225
42.6k
    auto* length_writer =
226
42.6k
            new OffsetColumnWriter(length_options, std::move(length_column_ptr), file_writer);
227
228
42.6k
    ScalarColumnWriter* null_writer = get_null_writer(opts, file_writer, 3);
229
230
42.6k
    *writer = std::unique_ptr<ColumnWriter>(
231
42.6k
            new ArrayColumnWriter(opts, std::make_shared<TabletColumn>(*column), length_writer,
232
42.6k
                                  null_writer, std::move(item_writer)));
233
42.6k
    return Status::OK();
234
42.6k
}
235
236
Status ColumnWriter::create_map_writer(const ColumnWriterOptions& opts, const TabletColumn* column,
237
                                       io::FileWriter* file_writer,
238
23.8k
                                       std::unique_ptr<ColumnWriter>* writer) {
239
23.8k
    DCHECK(column->get_subtype_count() == 2);
240
23.8k
    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
23.8k
    std::vector<std::unique_ptr<ColumnWriter>> inner_writer_list;
247
71.6k
    for (int i = 0; i < 2; ++i) {
248
47.7k
        const TabletColumn& item_column = column->get_sub_column(i);
249
47.7k
        RETURN_IF_ERROR(item_column.check_valid());
250
251
        // create item writer
252
47.7k
        ColumnWriterOptions item_options;
253
47.7k
        item_options.meta = opts.meta->mutable_children_columns(i);
254
47.7k
        item_options.need_zone_map = false;
255
47.7k
        item_options.need_bloom_filter = item_column.is_bf_column();
256
47.7k
        item_options.storage_format = opts.storage_format;
257
47.7k
        std::unique_ptr<ColumnWriter> item_writer;
258
47.7k
        RETURN_IF_ERROR(
259
47.7k
                ColumnWriter::create(item_options, &item_column, file_writer, &item_writer));
260
47.7k
        inner_writer_list.push_back(std::move(item_writer));
261
47.7k
    }
262
263
    // create offset writer
264
23.8k
    FieldType length_type = FieldType::OLAP_FIELD_TYPE_UNSIGNED_BIGINT;
265
266
    // Be Cautious: column unique id is used for column reader creation
267
23.8k
    ColumnWriterOptions length_options;
268
23.8k
    length_options.meta = opts.meta->add_children_columns();
269
23.8k
    length_options.meta->set_column_id(column->get_subtype_count() + 1);
270
23.8k
    length_options.meta->set_unique_id(column->get_subtype_count() + 1);
271
23.8k
    length_options.meta->set_type(int(length_type));
272
23.8k
    length_options.meta->set_is_nullable(false);
273
23.8k
    length_options.meta->set_length(
274
23.8k
            cast_set<int32_t>(field_type_size(FieldType::OLAP_FIELD_TYPE_UNSIGNED_BIGINT)));
275
23.8k
    length_options.meta->set_compression(opts.meta->compression());
276
277
23.8k
    length_options.need_zone_map = false;
278
23.8k
    length_options.need_bloom_filter = false;
279
23.8k
    length_options.storage_format = opts.storage_format;
280
281
23.8k
    auto length_column_ptr = std::make_shared<TabletColumn>(
282
23.8k
            FieldAggregationMethod::OLAP_FIELD_AGGREGATION_NONE, length_type,
283
23.8k
            length_options.meta->is_nullable(), length_options.meta->unique_id(),
284
23.8k
            length_options.meta->length());
285
23.8k
    length_column_ptr->set_name("length");
286
23.8k
    length_column_ptr->set_index_length(-1); // no short key index
287
23.8k
    length_options.meta->set_encoding(
288
23.8k
            EncodingInfo::resolve_default_encoding(opts.storage_format, *length_column_ptr));
289
23.8k
    auto* length_writer =
290
23.8k
            new OffsetColumnWriter(length_options, std::move(length_column_ptr), file_writer);
291
292
23.8k
    ScalarColumnWriter* null_writer =
293
23.8k
            get_null_writer(opts, file_writer, column->get_subtype_count() + 2);
294
295
23.8k
    *writer = std::unique_ptr<ColumnWriter>(
296
23.8k
            new MapColumnWriter(opts, std::make_shared<TabletColumn>(*column), null_writer,
297
23.8k
                                length_writer, inner_writer_list));
298
299
23.8k
    return Status::OK();
300
23.8k
}
301
302
Status ColumnWriter::create_agg_state_writer(const ColumnWriterOptions& opts,
303
                                             const TabletColumn* column,
304
                                             io::FileWriter* file_writer,
305
1.26k
                                             std::unique_ptr<ColumnWriter>* writer) {
306
1.26k
    auto data_type = DataTypeFactory::instance().create_data_type(*column);
307
1.26k
    const auto* agg_state_type = assert_cast<const DataTypeAggState*>(data_type.get());
308
1.26k
    auto type = agg_state_type->get_serialized_type()->get_primitive_type();
309
1.26k
    if (type == PrimitiveType::TYPE_STRING || type == PrimitiveType::INVALID_TYPE ||
310
1.26k
        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
21
        RETURN_IF_ERROR(create_array_writer(opts, column, file_writer, writer));
315
52
    } else if (type == PrimitiveType::TYPE_MAP) {
316
52
        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
1.26k
    return Status::OK();
323
1.26k
}
324
325
Status ColumnWriter::create_variant_writer(const ColumnWriterOptions& opts,
326
                                           const TabletColumn* column, io::FileWriter* file_writer,
327
6.34k
                                           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
6.34k
    if (column->is_extracted_column()) {
334
422
        if (column->name().find(DOC_VALUE_COLUMN_PATH) != std::string::npos) {
335
339
            *writer = std::make_unique<VariantDocCompactWriter>(
336
339
                    opts, std::make_shared<TabletColumn>(*column));
337
339
            return Status::OK();
338
339
        }
339
83
        VLOG_DEBUG << "gen subwriter for " << column->path_info_ptr()->get_path();
340
83
        *writer = std::make_unique<VariantSubcolumnWriter>(opts,
341
83
                                                           std::make_shared<TabletColumn>(*column));
342
83
        return Status::OK();
343
422
    }
344
5.91k
    *writer = std::make_unique<VariantColumnWriter>(opts, std::make_shared<TabletColumn>(*column));
345
5.91k
    return Status::OK();
346
6.34k
}
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
829k
                            io::FileWriter* file_writer, std::unique_ptr<ColumnWriter>* writer) {
351
829k
    auto column_ptr = std::make_shared<TabletColumn>(*column);
352
829k
    if (is_scalar_type(column->type())) {
353
768k
        *writer = std::unique_ptr<ColumnWriter>(
354
768k
                new ScalarColumnWriter(opts, std::move(column_ptr), file_writer));
355
768k
        return Status::OK();
356
768k
    } else {
357
61.9k
        switch (column->type()) {
358
1.28k
        case FieldType::OLAP_FIELD_TYPE_AGG_STATE: {
359
1.28k
            RETURN_IF_ERROR(create_agg_state_writer(opts, column, file_writer, writer));
360
1.28k
            return Status::OK();
361
1.28k
        }
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
10.1k
        case FieldType::OLAP_FIELD_TYPE_MAP: {
371
10.1k
            RETURN_IF_ERROR(create_map_writer(opts, column, file_writer, writer));
372
10.1k
            return Status::OK();
373
10.1k
        }
374
6.34k
        case FieldType::OLAP_FIELD_TYPE_VARIANT: {
375
            // Process columns with sparse column
376
6.34k
            RETURN_IF_ERROR(create_variant_writer(opts, column, file_writer, writer));
377
6.34k
            return Status::OK();
378
6.34k
        }
379
0
        default:
380
0
            return Status::NotSupported("unsupported type for ColumnWriter: {}",
381
0
                                        std::to_string(int(column_ptr->type())));
382
61.9k
        }
383
61.9k
    }
384
829k
}
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
3.53M
Status ColumnWriter::append(const uint8_t* nullmap, const void* data, size_t num_rows) {
454
3.53M
    assert(data && num_rows > 0);
455
3.53M
    const auto* ptr = (const uint8_t*)data;
456
3.53M
    if (nullmap) {
457
3.18M
        return append_nullable(nullmap, &ptr, num_rows);
458
3.18M
    } else {
459
348k
        return append_data(&ptr, num_rows);
460
348k
    }
461
3.53M
}
462
463
///////////////////////////////////////////////////////////////////////////////////
464
465
ScalarColumnWriter::ScalarColumnWriter(const ColumnWriterOptions& opts, TabletColumnPtr column,
466
                                       io::FileWriter* file_writer)
467
879k
        : ColumnWriter(std::move(column), opts.meta->is_nullable(), opts.meta),
468
879k
          _opts(opts),
469
879k
          _file_writer(file_writer),
470
879k
          _data_size(0) {
471
    // these opts.meta fields should be set by client
472
879k
    DCHECK(opts.meta->has_column_id());
473
879k
    DCHECK(opts.meta->has_unique_id());
474
879k
    DCHECK(opts.meta->has_type());
475
879k
    DCHECK(opts.meta->has_length());
476
879k
    DCHECK(opts.meta->has_encoding());
477
879k
    DCHECK(opts.meta->has_compression());
478
879k
    DCHECK(opts.meta->has_is_nullable());
479
879k
    DCHECK(file_writer != nullptr);
480
879k
    _inverted_index_builders.resize(_opts.inverted_indexes.size());
481
879k
}
482
483
879k
ScalarColumnWriter::~ScalarColumnWriter() {
484
    // delete all pages
485
879k
    _pages.clear();
486
879k
}
487
488
877k
Status ScalarColumnWriter::init() {
489
877k
    RETURN_IF_ERROR(get_block_compression_codec(_opts.meta->compression(), &_compress_codec));
490
491
877k
    PageBuilder* page_builder = nullptr;
492
493
    // Caller must set a concrete (non-DEFAULT) encoding on the meta before init.
494
877k
    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
877k
    RETURN_IF_ERROR(
501
877k
            EncodingInfo::get(get_column()->type(), _opts.meta->encoding(), &_encoding_info));
502
    // create page builder
503
877k
    PageBuilderOptions opts;
504
877k
    opts.data_page_size = _opts.data_page_size;
505
877k
    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
877k
    opts.dict_binary_plain_encoding =
509
877k
            (_opts.storage_format == TabletStorageFormatPB::TABLET_STORAGE_FORMAT_V3)
510
877k
                    ? PLAIN_ENCODING_V3
511
877k
                    : PLAIN_ENCODING;
512
877k
    RETURN_IF_ERROR(_encoding_info->create_page_builder(opts, &page_builder));
513
877k
    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
877k
    _page_builder.reset(page_builder);
523
    // create ordinal builder
524
877k
    _ordinal_index_builder = std::make_unique<OrdinalIndexWriter>();
525
    // create null bitmap builder
526
877k
    if (is_nullable()) {
527
465k
        _null_bitmap_builder = std::make_unique<NullBitmapBuilder>();
528
465k
    }
529
877k
    if (_opts.need_zone_map) {
530
622k
        RETURN_IF_ERROR(
531
622k
                ZoneMapIndexWriter::create(_data_type, get_column(), _zone_map_index_builder));
532
622k
    }
533
534
877k
    if (_opts.need_inverted_index) {
535
42.1k
        do {
536
84.4k
            for (size_t i = 0; i < _opts.inverted_indexes.size(); i++) {
537
42.2k
                DBUG_EXECUTE_IF("column_writer.init", {
538
42.2k
                    class InvertedIndexColumnWriterEmpty final : public IndexColumnWriter {
539
42.2k
                    public:
540
42.2k
                        Status init() override { return Status::OK(); }
541
42.2k
                        Status add_values(const std::string name, const void* values,
542
42.2k
                                          size_t count) override {
543
42.2k
                            return Status::OK();
544
42.2k
                        }
545
42.2k
                        Status add_array_values(size_t field_size, const void* value_ptr,
546
42.2k
                                                const uint8_t* null_map, const uint8_t* offsets_ptr,
547
42.2k
                                                size_t count) override {
548
42.2k
                            return Status::OK();
549
42.2k
                        }
550
42.2k
                        Status add_nulls(uint32_t count) override { return Status::OK(); }
551
42.2k
                        Status add_array_nulls(const uint8_t* null_map, size_t num_rows) override {
552
42.2k
                            return Status::OK();
553
42.2k
                        }
554
42.2k
                        Status finish() override { return Status::OK(); }
555
42.2k
                        int64_t size() const override { return 0; }
556
42.2k
                        void close_on_error() override {}
557
42.2k
                    };
558
559
42.2k
                    _inverted_index_builders[i] =
560
42.2k
                            std::make_unique<InvertedIndexColumnWriterEmpty>();
561
562
42.2k
                    break;
563
42.2k
                });
564
565
42.2k
                RETURN_IF_ERROR(IndexColumnWriter::create(
566
42.2k
                        get_column(), &_inverted_index_builders[i], _opts.index_file_writer,
567
42.2k
                        _opts.inverted_indexes[i]));
568
42.2k
            }
569
42.1k
        } while (false);
570
42.1k
    }
571
877k
    if (_opts.need_bloom_filter) {
572
5.43k
        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
2.04k
            RETURN_IF_ERROR(BloomFilterIndexWriter::create(_opts.bf_options, get_column()->type(),
578
2.04k
                                                           &_bloom_filter_index_builder));
579
2.04k
        }
580
5.43k
    }
581
877k
    return Status::OK();
582
877k
}
583
584
3.18M
Status ScalarColumnWriter::append_nulls(size_t num_rows) {
585
3.18M
    _null_bitmap_builder->add_run(true, num_rows);
586
3.18M
    _next_rowid += num_rows;
587
3.18M
    if (_opts.need_zone_map) {
588
3.10M
        _zone_map_index_builder->add_nulls(cast_set<uint32_t>(num_rows));
589
3.10M
    }
590
3.18M
    if (_opts.need_inverted_index) {
591
2.07M
        for (const auto& builder : _inverted_index_builders) {
592
2.07M
            RETURN_IF_ERROR(builder->add_nulls(cast_set<uint32_t>(num_rows)));
593
2.07M
        }
594
2.07M
    }
595
3.18M
    if (_opts.need_bloom_filter) {
596
564
        _bloom_filter_index_builder->add_nulls(cast_set<uint32_t>(num_rows));
597
564
    }
598
3.18M
    return Status::OK();
599
3.18M
}
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
4.43M
Status ScalarColumnWriter::append_data(const uint8_t** ptr, size_t num_rows) {
605
4.43M
    size_t remaining = num_rows;
606
8.93M
    while (remaining > 0) {
607
4.50M
        size_t num_written = remaining;
608
4.50M
        RETURN_IF_ERROR(append_data_in_current_page(ptr, &num_written));
609
610
4.50M
        remaining -= num_written;
611
612
4.50M
        if (_page_builder->is_page_full()) {
613
75.8k
            RETURN_IF_ERROR(finish_current_page());
614
75.8k
        }
615
4.50M
    }
616
4.43M
    return Status::OK();
617
4.43M
}
618
619
Status ScalarColumnWriter::_internal_append_data_in_current_page(const uint8_t* data,
620
4.56M
                                                                 size_t* num_written) {
621
4.56M
    RETURN_IF_ERROR(_page_builder->add(data, num_written));
622
4.56M
    if (_opts.need_zone_map) {
623
4.27M
        _zone_map_index_builder->add_values(data, *num_written);
624
4.27M
    }
625
4.56M
    if (_opts.need_inverted_index) {
626
2.10M
        for (const auto& builder : _inverted_index_builders) {
627
2.10M
            RETURN_IF_ERROR(builder->add_values(get_column()->name(), data, *num_written));
628
2.10M
        }
629
2.10M
    }
630
4.56M
    if (_opts.need_bloom_filter) {
631
5.72k
        RETURN_IF_ERROR(_bloom_filter_index_builder->add_values(data, *num_written));
632
5.72k
    }
633
634
4.56M
    _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
4.56M
    if (is_nullable()) {
639
4.09M
        _null_bitmap_builder->add_run(false, *num_written);
640
4.09M
    }
641
4.56M
    return Status::OK();
642
4.56M
}
643
644
4.57M
Status ScalarColumnWriter::append_data_in_current_page(const uint8_t** data, size_t* num_written) {
645
4.57M
    RETURN_IF_ERROR(append_data_in_current_page(*data, num_written));
646
4.57M
    *data += cell_size() * (*num_written);
647
4.57M
    return Status::OK();
648
4.57M
}
649
650
Status ScalarColumnWriter::append_nullable(const uint8_t* null_map, const uint8_t** ptr,
651
3.14M
                                           size_t num_rows) {
652
    // When optimization is disabled, use base class implementation
653
3.14M
    if (!config::enable_rle_batch_put_optimization) {
654
0
        return ColumnWriter::append_nullable(null_map, ptr, num_rows);
655
0
    }
656
657
3.14M
    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
3.14M
    _null_run_buffer.clear();
663
3.14M
    if (_null_run_buffer.capacity() < num_rows) {
664
479k
        _null_run_buffer.reserve(std::min(num_rows, size_t(256)));
665
479k
    }
666
667
3.14M
    size_t non_null_count = 0;
668
3.14M
    size_t offset = 0;
669
7.03M
    while (offset < num_rows) {
670
3.89M
        bool is_null = null_map[offset] != 0;
671
3.89M
        size_t remaining = num_rows - offset;
672
3.89M
        const uint8_t* run_end =
673
3.89M
                static_cast<const uint8_t*>(memchr(null_map + offset, is_null ? 0 : 1, remaining));
674
3.89M
        size_t run_length = run_end != nullptr ? (run_end - (null_map + offset)) : remaining;
675
3.89M
        _null_run_buffer.push_back(NullRun {is_null, static_cast<uint32_t>(run_length)});
676
3.89M
        if (!is_null) {
677
3.43M
            non_null_count += run_length;
678
3.43M
        }
679
3.89M
        offset += run_length;
680
3.89M
    }
681
682
    // Pre-allocate buffer based on estimated size
683
3.14M
    if (_null_bitmap_builder != nullptr) {
684
3.14M
        size_t current_rows = _next_rowid - _first_rowid;
685
3.14M
        size_t expected_rows = current_rows + num_rows;
686
3.14M
        size_t est_non_null = non_null_count;
687
3.14M
        if (num_rows > 0 && expected_rows > num_rows) {
688
2.72M
            est_non_null = (non_null_count * expected_rows) / num_rows;
689
2.72M
        }
690
3.14M
        _null_bitmap_builder->reserve_for_write(expected_rows, est_non_null);
691
3.14M
    }
692
693
3.14M
    if (non_null_count == 0) {
694
        // All NULL: skip data writing, only update null bitmap and indexes
695
50.1k
        RETURN_IF_ERROR(append_nulls(num_rows));
696
50.1k
        *ptr += cell_size() * num_rows;
697
50.1k
        return Status::OK();
698
50.1k
    }
699
700
3.09M
    if (non_null_count == num_rows) {
701
        // All non-NULL: use normal append_data which handles both data and null bitmap
702
3.03M
        return append_data(ptr, num_rows);
703
3.03M
    }
704
705
    // Process by runs
706
814k
    for (const auto& run : _null_run_buffer) {
707
814k
        size_t run_length = run.len;
708
814k
        if (run.is_null) {
709
415k
            RETURN_IF_ERROR(append_nulls(run_length));
710
415k
            *ptr += cell_size() * run_length;
711
415k
        } 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
398k
            RETURN_IF_ERROR(append_data(ptr, run_length));
716
398k
        }
717
814k
    }
718
719
62.1k
    return Status::OK();
720
62.1k
}
721
722
43.5k
uint64_t ScalarColumnWriter::estimate_buffer_size() {
723
43.5k
    uint64_t size = _data_size;
724
43.5k
    size += _page_builder->size();
725
43.5k
    if (is_nullable()) {
726
29.1k
        size += _null_bitmap_builder->size();
727
29.1k
    }
728
43.5k
    size += _ordinal_index_builder->size();
729
43.5k
    if (_opts.need_zone_map) {
730
33.3k
        size += _zone_map_index_builder->size();
731
33.3k
    }
732
43.5k
    if (_opts.need_bloom_filter) {
733
268
        size += _bloom_filter_index_builder->size();
734
268
    }
735
43.5k
    return size;
736
43.5k
}
737
738
880k
Status ScalarColumnWriter::finish() {
739
880k
    RETURN_IF_ERROR(finish_current_page());
740
880k
    _opts.meta->set_num_rows(_next_rowid);
741
880k
    return Status::OK();
742
880k
}
743
744
880k
Status ScalarColumnWriter::write_data() {
745
880k
    auto offset = _file_writer->bytes_appended();
746
1.19M
    auto collect_uncompressed_bytes = [](const PageFooterPB& footer) {
747
1.19M
        return footer.uncompressed_size() + footer.ByteSizeLong() +
748
1.19M
               sizeof(uint32_t) /* footer size */ + sizeof(uint32_t) /* checksum */;
749
1.19M
    };
750
913k
    for (auto& page : _pages) {
751
913k
        _total_uncompressed_data_pages_size += collect_uncompressed_bytes(page->footer);
752
913k
        RETURN_IF_ERROR(_write_data_page(page.get()));
753
913k
    }
754
880k
    _pages.clear();
755
    // write column dict
756
880k
    if (_encoding_info->encoding() == DICT_ENCODING) {
757
283k
        OwnedSlice dict_body;
758
283k
        RETURN_IF_ERROR(_page_builder->get_dictionary_page(&dict_body));
759
283k
        EncodingTypePB dict_word_page_encoding;
760
283k
        RETURN_IF_ERROR(_page_builder->get_dictionary_page_encoding(&dict_word_page_encoding));
761
762
283k
        PageFooterPB footer;
763
283k
        footer.set_type(DICTIONARY_PAGE);
764
283k
        footer.set_uncompressed_size(cast_set<uint32_t>(dict_body.slice().get_size()));
765
283k
        footer.mutable_dict_page_footer()->set_encoding(dict_word_page_encoding);
766
283k
        _total_uncompressed_data_pages_size += collect_uncompressed_bytes(footer);
767
768
283k
        PagePointer dict_pp;
769
283k
        RETURN_IF_ERROR(PageIO::compress_and_write_page(
770
283k
                _compress_codec, _opts.compression_min_space_saving, _file_writer,
771
283k
                {dict_body.slice()}, footer, &dict_pp));
772
283k
        dict_pp.to_proto(_opts.meta->mutable_dict_page());
773
283k
    }
774
880k
    _total_compressed_data_pages_size += _file_writer->bytes_appended() - offset;
775
880k
    _page_builder.reset();
776
880k
    return Status::OK();
777
880k
}
778
779
837k
Status ScalarColumnWriter::write_ordinal_index() {
780
837k
    return _ordinal_index_builder->finish(_file_writer, _opts.meta->add_indexes());
781
837k
}
782
783
674k
Status ScalarColumnWriter::write_zone_map() {
784
674k
    if (_opts.need_zone_map) {
785
621k
        return _zone_map_index_builder->finish(_file_writer, _opts.meta->add_indexes());
786
621k
    }
787
52.1k
    return Status::OK();
788
674k
}
789
790
615k
Status ScalarColumnWriter::write_inverted_index() {
791
615k
    if (_opts.need_inverted_index) {
792
42.2k
        for (const auto& builder : _inverted_index_builders) {
793
42.2k
            RETURN_IF_ERROR(builder->finish());
794
42.2k
        }
795
42.0k
    }
796
615k
    return Status::OK();
797
615k
}
798
799
604k
Status ScalarColumnWriter::write_bloom_filter_index() {
800
604k
    if (_opts.need_bloom_filter) {
801
5.43k
        return _bloom_filter_index_builder->finish(_file_writer, _opts.meta->add_indexes());
802
5.43k
    }
803
599k
    return Status::OK();
804
604k
}
805
806
// write a data page into file and update ordinal index
807
913k
Status ScalarColumnWriter::_write_data_page(Page* page) {
808
913k
    PagePointer pp;
809
913k
    std::vector<Slice> compressed_body;
810
1.72M
    for (auto& data : page->data) {
811
1.72M
        compressed_body.push_back(data.slice());
812
1.72M
    }
813
913k
    RETURN_IF_ERROR(PageIO::write_page(_file_writer, compressed_body, page->footer, &pp));
814
913k
    _ordinal_index_builder->append_entry(page->footer.data_page_footer().first_ordinal(), pp);
815
913k
    return Status::OK();
816
913k
}
817
818
956k
Status ScalarColumnWriter::finish_current_page() {
819
956k
    if (_next_rowid == _first_rowid) {
820
42.9k
        return Status::OK();
821
42.9k
    }
822
913k
    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
685k
        if (_next_rowid - _first_rowid < config::zone_map_row_num_threshold) {
826
485k
            _zone_map_index_builder->invalid_page_zone_map();
827
485k
        }
828
685k
        RETURN_IF_ERROR(_zone_map_index_builder->flush());
829
685k
    }
830
831
913k
    if (_opts.need_bloom_filter) {
832
5.68k
        RETURN_IF_ERROR(_bloom_filter_index_builder->flush());
833
5.68k
    }
834
835
913k
    _raw_data_bytes += _page_builder->get_raw_data_size();
836
837
    // build data page body : encoded values + [nullmap]
838
913k
    std::vector<Slice> body;
839
913k
    OwnedSlice encoded_values;
840
913k
    RETURN_IF_ERROR(_page_builder->finish(&encoded_values));
841
913k
    RETURN_IF_ERROR(_page_builder->reset());
842
913k
    body.push_back(encoded_values.slice());
843
844
913k
    OwnedSlice nullmap;
845
913k
    if (_null_bitmap_builder != nullptr) {
846
464k
        if (is_nullable() && _null_bitmap_builder->has_null()) {
847
131k
            RETURN_IF_ERROR(_null_bitmap_builder->finish(&nullmap));
848
131k
            body.push_back(nullmap.slice());
849
131k
        }
850
464k
        _null_bitmap_builder->reset();
851
464k
    }
852
853
    // prepare data page footer
854
913k
    std::unique_ptr<Page> page(new Page());
855
913k
    page->footer.set_type(DATA_PAGE);
856
913k
    page->footer.set_uncompressed_size(cast_set<uint32_t>(Slice::compute_total_size(body)));
857
913k
    auto* data_page_footer = page->footer.mutable_data_page_footer();
858
913k
    data_page_footer->set_first_ordinal(_first_rowid);
859
913k
    data_page_footer->set_num_values(_next_rowid - _first_rowid);
860
913k
    data_page_footer->set_nullmap_size(cast_set<uint32_t>(nullmap.slice().size));
861
913k
    if (_new_page_callback != nullptr) {
862
66.1k
        _new_page_callback->put_extra_info_in_page(data_page_footer);
863
66.1k
    }
864
    // trying to compress page body
865
913k
    OwnedSlice compressed_body;
866
913k
    RETURN_IF_ERROR(PageIO::compress_page_body(_compress_codec, _opts.compression_min_space_saving,
867
913k
                                               body, &compressed_body));
868
913k
    if (compressed_body.slice().empty()) {
869
        // page body is uncompressed
870
810k
        page->data.emplace_back(std::move(encoded_values));
871
810k
        page->data.emplace_back(std::move(nullmap));
872
810k
    } else {
873
        // page body is compressed
874
103k
        page->data.emplace_back(std::move(compressed_body));
875
103k
    }
876
877
913k
    _push_back_page(std::move(page));
878
913k
    _first_rowid = _next_rowid;
879
913k
    return Status::OK();
880
913k
}
881
882
////////////////////////////////////////////////////////////////////////////////
883
884
////////////////////////////////////////////////////////////////////////////////
885
// offset column writer
886
////////////////////////////////////////////////////////////////////////////////
887
888
OffsetColumnWriter::OffsetColumnWriter(const ColumnWriterOptions& opts, TabletColumnPtr column,
889
                                       io::FileWriter* file_writer)
890
66.4k
        : ScalarColumnWriter(opts, std::move(column), file_writer) {
891
    // now we only explain data in offset column as uint64
892
66.4k
    DCHECK(get_column()->type() == FieldType::OLAP_FIELD_TYPE_UNSIGNED_BIGINT);
893
66.4k
}
894
895
66.5k
OffsetColumnWriter::~OffsetColumnWriter() = default;
896
897
66.3k
Status OffsetColumnWriter::init() {
898
66.3k
    RETURN_IF_ERROR(ScalarColumnWriter::init());
899
66.3k
    register_flush_page_callback(this);
900
66.3k
    _next_offset = 0;
901
66.3k
    return Status::OK();
902
66.3k
}
903
904
66.2k
Status OffsetColumnWriter::append_data(const uint8_t** ptr, size_t num_rows) {
905
66.2k
    size_t remaining = num_rows;
906
132k
    while (remaining > 0) {
907
66.5k
        size_t num_written = remaining;
908
66.5k
        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
66.5k
        _next_offset = *(const uint64_t*)(*ptr);
912
66.5k
        remaining -= num_written;
913
914
66.5k
        if (_page_builder->is_page_full()) {
915
            // get next data for next array_item_rowid
916
405
            RETURN_IF_ERROR(finish_current_page());
917
405
        }
918
66.5k
    }
919
66.2k
    return Status::OK();
920
66.2k
}
921
922
66.1k
void OffsetColumnWriter::put_extra_info_in_page(DataPageFooterPB* footer) {
923
66.1k
    footer->set_next_array_item_ordinal(_next_offset);
924
66.1k
}
925
926
StructColumnWriter::StructColumnWriter(
927
        const ColumnWriterOptions& opts, TabletColumnPtr column, ScalarColumnWriter* null_writer,
928
        std::vector<std::unique_ptr<ColumnWriter>>& sub_column_writers)
929
2.11k
        : ColumnWriter(std::move(column), opts.meta->is_nullable(), opts.meta), _opts(opts) {
930
9.24k
    for (auto& sub_column_writer : sub_column_writers) {
931
9.24k
        _sub_column_writers.push_back(std::move(sub_column_writer));
932
9.24k
    }
933
2.11k
    _num_sub_column_writers = _sub_column_writers.size();
934
2.11k
    DCHECK(_num_sub_column_writers >= 1);
935
2.11k
    if (is_nullable()) {
936
1.84k
        _null_writer.reset(null_writer);
937
1.84k
    }
938
2.11k
}
939
940
2.11k
Status StructColumnWriter::init() {
941
9.26k
    for (auto& column_writer : _sub_column_writers) {
942
9.26k
        RETURN_IF_ERROR(column_writer->init());
943
9.26k
    }
944
2.11k
    if (is_nullable()) {
945
1.85k
        RETURN_IF_ERROR(_null_writer->init());
946
1.85k
    }
947
2.11k
    return Status::OK();
948
2.11k
}
949
950
1.50k
Status StructColumnWriter::write_inverted_index() {
951
1.50k
    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
1.50k
    return Status::OK();
957
1.50k
}
958
959
Status StructColumnWriter::append_nullable(const uint8_t* null_map, const uint8_t** ptr,
960
1.80k
                                           size_t num_rows) {
961
1.80k
    RETURN_IF_ERROR(append_data(ptr, num_rows));
962
1.80k
    RETURN_IF_ERROR(_null_writer->append_data(&null_map, num_rows));
963
1.80k
    return Status::OK();
964
1.80k
}
965
966
2.07k
Status StructColumnWriter::append_data(const uint8_t** ptr, size_t num_rows) {
967
2.07k
    const auto* results = reinterpret_cast<const uint64_t*>(*ptr);
968
10.9k
    for (size_t i = 0; i < _num_sub_column_writers; ++i) {
969
8.91k
        auto nullmap = *(results + _num_sub_column_writers + i);
970
8.91k
        auto data = *(results + i);
971
8.91k
        RETURN_IF_ERROR(_sub_column_writers[i]->append(reinterpret_cast<const uint8_t*>(nullmap),
972
8.91k
                                                       reinterpret_cast<const void*>(data),
973
8.91k
                                                       num_rows));
974
8.91k
    }
975
2.07k
    return Status::OK();
976
2.07k
}
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
2.12k
Status StructColumnWriter::finish() {
988
9.26k
    for (auto& column_writer : _sub_column_writers) {
989
9.26k
        RETURN_IF_ERROR(column_writer->finish());
990
9.26k
    }
991
2.12k
    if (is_nullable()) {
992
1.85k
        RETURN_IF_ERROR(_null_writer->finish());
993
1.85k
    }
994
2.12k
    _opts.meta->set_num_rows(get_next_rowid());
995
2.12k
    return Status::OK();
996
2.12k
}
997
998
2.12k
Status StructColumnWriter::write_data() {
999
9.26k
    for (auto& column_writer : _sub_column_writers) {
1000
9.26k
        RETURN_IF_ERROR(column_writer->write_data());
1001
9.26k
    }
1002
2.12k
    if (is_nullable()) {
1003
1.85k
        RETURN_IF_ERROR(_null_writer->write_data());
1004
1.85k
    }
1005
2.12k
    return Status::OK();
1006
2.12k
}
1007
1008
2.07k
Status StructColumnWriter::write_ordinal_index() {
1009
8.90k
    for (auto& column_writer : _sub_column_writers) {
1010
8.90k
        RETURN_IF_ERROR(column_writer->write_ordinal_index());
1011
8.90k
    }
1012
2.07k
    if (is_nullable()) {
1013
1.80k
        RETURN_IF_ERROR(_null_writer->write_ordinal_index());
1014
1.80k
    }
1015
2.07k
    return Status::OK();
1016
2.07k
}
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
42.5k
        : ColumnWriter(std::move(column), opts.meta->is_nullable(), opts.meta),
1039
42.5k
          _item_writer(std::move(item_writer)),
1040
42.5k
          _opts(opts) {
1041
42.5k
    _offset_writer.reset(offset_writer);
1042
42.5k
    if (is_nullable()) {
1043
26.9k
        _null_writer.reset(null_writer);
1044
26.9k
    }
1045
42.5k
}
1046
1047
42.5k
Status ArrayColumnWriter::init() {
1048
42.5k
    RETURN_IF_ERROR(_offset_writer->init());
1049
42.5k
    if (is_nullable()) {
1050
26.9k
        RETURN_IF_ERROR(_null_writer->init());
1051
26.9k
    }
1052
42.5k
    RETURN_IF_ERROR(_item_writer->init());
1053
42.5k
    if (_opts.need_inverted_index) {
1054
1.80k
        auto* writer = dynamic_cast<ScalarColumnWriter*>(_item_writer.get());
1055
1.80k
        if (writer != nullptr) {
1056
1.80k
            RETURN_IF_ERROR(IndexColumnWriter::create(get_column(), &_inverted_index_writer,
1057
1.80k
                                                      _opts.index_file_writer,
1058
1.80k
                                                      _opts.inverted_indexes[0]));
1059
1.80k
        }
1060
1.80k
    }
1061
42.5k
    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
42.5k
    return Status::OK();
1070
42.5k
}
1071
1072
39.4k
Status ArrayColumnWriter::write_inverted_index() {
1073
39.4k
    if (_opts.need_inverted_index) {
1074
1.79k
        return _inverted_index_writer->finish();
1075
1.79k
    }
1076
37.6k
    return Status::OK();
1077
39.4k
}
1078
1079
39.4k
Status ArrayColumnWriter::write_ann_index() {
1080
39.4k
    if (_opts.need_ann_index) {
1081
57
        return _ann_index_writer->finish();
1082
57
    }
1083
39.3k
    return Status::OK();
1084
39.4k
}
1085
1086
// batch append data for array
1087
42.2k
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
42.2k
    auto data_ptr = reinterpret_cast<const uint64_t*>(*ptr);
1091
    // total number length
1092
42.2k
    size_t element_cnt = size_t((unsigned long)(*data_ptr));
1093
42.2k
    auto offset_data = *(data_ptr + 1);
1094
42.2k
    const uint8_t* offsets_ptr = (const uint8_t*)offset_data;
1095
42.2k
    auto data = *(data_ptr + 2);
1096
42.2k
    auto nested_null_map = *(data_ptr + 3);
1097
42.2k
    if (element_cnt > 0) {
1098
26.2k
        RETURN_IF_ERROR(_item_writer->append(reinterpret_cast<const uint8_t*>(nested_null_map),
1099
26.2k
                                             reinterpret_cast<const void*>(data), element_cnt));
1100
26.2k
    }
1101
42.2k
    if (_opts.need_inverted_index) {
1102
1.80k
        auto* writer = dynamic_cast<ScalarColumnWriter*>(_item_writer.get());
1103
        // now only support nested type is scala
1104
1.80k
        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
1.80k
            RETURN_IF_ERROR(_inverted_index_writer->add_array_values(
1107
1.80k
                    field_type_size(_item_writer->get_column()->type()),
1108
1.80k
                    reinterpret_cast<const void*>(data),
1109
1.80k
                    reinterpret_cast<const uint8_t*>(nested_null_map), offsets_ptr, num_rows));
1110
1.80k
        }
1111
1.80k
    }
1112
1113
42.2k
    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
42.2k
    RETURN_IF_ERROR(_offset_writer->append_data(&offsets_ptr, num_rows));
1131
42.2k
    return Status::OK();
1132
42.2k
}
1133
1134
287
uint64_t ArrayColumnWriter::estimate_buffer_size() {
1135
287
    return _offset_writer->estimate_buffer_size() +
1136
287
           (is_nullable() ? _null_writer->estimate_buffer_size() : 0) +
1137
287
           _item_writer->estimate_buffer_size();
1138
287
}
1139
1140
Status ArrayColumnWriter::append_nullable(const uint8_t* null_map, const uint8_t** ptr,
1141
26.5k
                                          size_t num_rows) {
1142
26.5k
    RETURN_IF_ERROR(append_data(ptr, num_rows));
1143
26.5k
    if (is_nullable()) {
1144
26.5k
        if (_opts.need_inverted_index) {
1145
1.40k
            RETURN_IF_ERROR(_inverted_index_writer->add_array_nulls(null_map, num_rows));
1146
1.40k
        }
1147
26.5k
        RETURN_IF_ERROR(_null_writer->append_data(&null_map, num_rows));
1148
26.5k
    }
1149
26.5k
    return Status::OK();
1150
26.5k
}
1151
1152
42.6k
Status ArrayColumnWriter::finish() {
1153
42.6k
    RETURN_IF_ERROR(_offset_writer->finish());
1154
42.6k
    if (is_nullable()) {
1155
26.9k
        RETURN_IF_ERROR(_null_writer->finish());
1156
26.9k
    }
1157
42.6k
    RETURN_IF_ERROR(_item_writer->finish());
1158
42.6k
    _opts.meta->set_num_rows(get_next_rowid());
1159
42.6k
    return Status::OK();
1160
42.6k
}
1161
1162
42.6k
Status ArrayColumnWriter::write_data() {
1163
42.6k
    RETURN_IF_ERROR(_offset_writer->write_data());
1164
42.6k
    if (is_nullable()) {
1165
26.9k
        RETURN_IF_ERROR(_null_writer->write_data());
1166
26.9k
    }
1167
42.6k
    RETURN_IF_ERROR(_item_writer->write_data());
1168
42.6k
    return Status::OK();
1169
42.6k
}
1170
1171
42.1k
Status ArrayColumnWriter::write_ordinal_index() {
1172
42.1k
    RETURN_IF_ERROR(_offset_writer->write_ordinal_index());
1173
42.1k
    if (is_nullable()) {
1174
26.5k
        RETURN_IF_ERROR(_null_writer->write_ordinal_index());
1175
26.5k
    }
1176
42.1k
    if (!has_empty_items()) {
1177
26.1k
        RETURN_IF_ERROR(_item_writer->write_ordinal_index());
1178
26.1k
    }
1179
42.1k
    return Status::OK();
1180
42.1k
}
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
23.8k
        : ColumnWriter(std::move(column), opts.meta->is_nullable(), opts.meta), _opts(opts) {
1210
23.8k
    CHECK_EQ(kv_writers.size(), 2);
1211
23.8k
    _offsets_writer.reset(offset_writer);
1212
23.8k
    if (is_nullable()) {
1213
7.77k
        _null_writer.reset(null_writer);
1214
7.77k
    }
1215
47.7k
    for (auto& sub_writers : kv_writers) {
1216
47.7k
        _kv_writers.push_back(std::move(sub_writers));
1217
47.7k
    }
1218
23.8k
}
1219
1220
23.8k
Status MapColumnWriter::init() {
1221
23.8k
    RETURN_IF_ERROR(_offsets_writer->init());
1222
23.8k
    if (is_nullable()) {
1223
7.78k
        RETURN_IF_ERROR(_null_writer->init());
1224
7.78k
    }
1225
    // here register_flush_page_callback to call this.put_extra_info_in_page()
1226
    // when finish cur data page
1227
47.7k
    for (auto& sub_writer : _kv_writers) {
1228
47.7k
        RETURN_IF_ERROR(sub_writer->init());
1229
47.7k
    }
1230
23.8k
    return Status::OK();
1231
23.8k
}
1232
1233
436
uint64_t MapColumnWriter::estimate_buffer_size() {
1234
436
    size_t estimate = 0;
1235
872
    for (auto& sub_writer : _kv_writers) {
1236
872
        estimate += sub_writer->estimate_buffer_size();
1237
872
    }
1238
436
    estimate += _offsets_writer->estimate_buffer_size();
1239
436
    if (is_nullable()) {
1240
415
        estimate += _null_writer->estimate_buffer_size();
1241
415
    }
1242
436
    return estimate;
1243
436
}
1244
1245
23.8k
Status MapColumnWriter::finish() {
1246
23.8k
    RETURN_IF_ERROR(_offsets_writer->finish());
1247
23.8k
    if (is_nullable()) {
1248
7.78k
        RETURN_IF_ERROR(_null_writer->finish());
1249
7.78k
    }
1250
47.7k
    for (auto& sub_writer : _kv_writers) {
1251
47.7k
        RETURN_IF_ERROR(sub_writer->finish());
1252
47.7k
    }
1253
23.8k
    _opts.meta->set_num_rows(get_next_rowid());
1254
23.8k
    return Status::OK();
1255
23.8k
}
1256
1257
Status MapColumnWriter::append_nullable(const uint8_t* null_map, const uint8_t** ptr,
1258
7.62k
                                        size_t num_rows) {
1259
7.62k
    RETURN_IF_ERROR(append_data(ptr, num_rows));
1260
7.62k
    if (is_nullable()) {
1261
7.61k
        RETURN_IF_ERROR(_null_writer->append_data(&null_map, num_rows));
1262
7.61k
    }
1263
7.62k
    return Status::OK();
1264
7.62k
}
1265
1266
// write key value data with offsets
1267
23.9k
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
23.9k
    auto data_ptr = reinterpret_cast<const uint64_t*>(*ptr);
1272
    // total number length
1273
23.9k
    size_t element_cnt = size_t((unsigned long)(*data_ptr));
1274
23.9k
    auto offset_data = *(data_ptr + 1);
1275
23.9k
    const uint8_t* offsets_ptr = (const uint8_t*)offset_data;
1276
1277
23.9k
    if (element_cnt > 0) {
1278
35.1k
        for (size_t i = 0; i < 2; ++i) {
1279
23.4k
            auto data = *(data_ptr + 2 + i);
1280
23.4k
            auto nested_null_map = *(data_ptr + 2 + 2 + i);
1281
23.4k
            RETURN_IF_ERROR(
1282
23.4k
                    _kv_writers[i]->append(reinterpret_cast<const uint8_t*>(nested_null_map),
1283
23.4k
                                           reinterpret_cast<const void*>(data), element_cnt));
1284
23.4k
        }
1285
11.7k
    }
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
23.9k
    RETURN_IF_ERROR(_offsets_writer->append_data(&offsets_ptr, num_rows));
1289
23.9k
    return Status::OK();
1290
23.9k
}
1291
1292
23.8k
Status MapColumnWriter::write_data() {
1293
23.8k
    RETURN_IF_ERROR(_offsets_writer->write_data());
1294
23.8k
    if (is_nullable()) {
1295
7.78k
        RETURN_IF_ERROR(_null_writer->write_data());
1296
7.78k
    }
1297
47.7k
    for (auto& sub_writer : _kv_writers) {
1298
47.7k
        RETURN_IF_ERROR(sub_writer->write_data());
1299
47.7k
    }
1300
23.8k
    return Status::OK();
1301
23.8k
}
1302
1303
23.6k
Status MapColumnWriter::write_ordinal_index() {
1304
23.6k
    RETURN_IF_ERROR(_offsets_writer->write_ordinal_index());
1305
23.6k
    if (is_nullable()) {
1306
7.52k
        RETURN_IF_ERROR(_null_writer->write_ordinal_index());
1307
7.52k
    }
1308
47.2k
    for (auto& sub_writer : _kv_writers) {
1309
47.2k
        if (sub_writer->get_next_rowid() != 0) {
1310
22.7k
            RETURN_IF_ERROR(sub_writer->write_ordinal_index());
1311
22.7k
        }
1312
47.2k
    }
1313
23.6k
    return Status::OK();
1314
23.6k
}
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
9.36k
Status MapColumnWriter::write_inverted_index() {
1338
9.36k
    if (_opts.need_inverted_index) {
1339
0
        return _index_builder->finish();
1340
0
    }
1341
9.36k
    return Status::OK();
1342
9.36k
}
1343
1344
VariantColumnWriter::VariantColumnWriter(const ColumnWriterOptions& opts, TabletColumnPtr column)
1345
5.91k
        : ColumnWriter(std::move(column), opts.meta->is_nullable(), opts.meta) {
1346
5.91k
    _impl = std::make_unique<VariantColumnWriterImpl>(opts, get_column());
1347
5.91k
}
1348
1349
5.90k
Status VariantColumnWriter::init() {
1350
5.90k
    return _impl->init();
1351
5.90k
}
1352
1353
675
Status VariantColumnWriter::append_data(const uint8_t** ptr, size_t num_rows) {
1354
675
    _next_rowid += num_rows;
1355
675
    return _impl->append_data(ptr, num_rows);
1356
675
}
1357
1358
864
uint64_t VariantColumnWriter::estimate_buffer_size() {
1359
864
    return _impl->estimate_buffer_size();
1360
864
}
1361
1362
5.95k
Status VariantColumnWriter::finish() {
1363
5.95k
    return _impl->finish();
1364
5.95k
}
1365
5.95k
Status VariantColumnWriter::write_data() {
1366
5.95k
    return _impl->write_data();
1367
5.95k
}
1368
5.95k
Status VariantColumnWriter::write_ordinal_index() {
1369
5.95k
    return _impl->write_ordinal_index();
1370
5.95k
}
1371
1372
5.94k
Status VariantColumnWriter::write_zone_map() {
1373
5.94k
    return _impl->write_zone_map();
1374
5.94k
}
1375
1376
5.93k
Status VariantColumnWriter::write_inverted_index() {
1377
5.93k
    return _impl->write_inverted_index();
1378
5.93k
}
1379
5.93k
Status VariantColumnWriter::write_bloom_filter_index() {
1380
5.93k
    return _impl->write_bloom_filter_index();
1381
5.93k
}
1382
1383
Status VariantColumnWriter::append_nullable(const uint8_t* null_map, const uint8_t** ptr,
1384
5.46k
                                            size_t num_rows) {
1385
5.46k
    return _impl->append_nullable(null_map, ptr, num_rows);
1386
5.46k
}
1387
1388
} // namespace doris::segment_v2