git.sesse.net Git - plocate/blob - database-builder.cpp

   1 #include "database-builder.h"
   2
   3 #include "dprintf.h"
   4 #include "turbopfor-encode.h"
   5
   6 #include <algorithm>
   7 #include <assert.h>
   8 #include <fcntl.h>
   9 #include <string.h>
  10 #include <string_view>
  11 #include <sys/stat.h>
  12 #include <sys/time.h>
  13 #include <sys/types.h>
  14 #include <unistd.h>
  15 #include <zdict.h>
  16 #include <zstd.h>
  17
  18 #define P4NENC_BOUND(n) ((n + 127) / 128 + (n + 32) * sizeof(uint32_t))
  19
  20 #define NUM_TRIGRAMS 16777216
  21
  22 using namespace std;
  23 using namespace std::chrono;
  24
  25 constexpr unsigned num_overflow_slots = 16;
  26
  27 string zstd_compress(const string &src, ZSTD_CDict *cdict, string *tempbuf);
  28
  29 static inline uint32_t read_unigram(const string_view s, size_t idx)
  30 {
  31         if (idx < s.size()) {
  32                 return (unsigned char)s[idx];
  33         } else {
  34                 return 0;
  35         }
  36 }
  37
  38 static inline uint32_t read_trigram(const string_view s, size_t start)
  39 {
  40         return read_unigram(s, start) |
  41                 (read_unigram(s, start + 1) << 8) |
  42                 (read_unigram(s, start + 2) << 16);
  43 }
  44
  45 class PostingListBuilder {
  46 public:
  47         inline void add_docid(uint32_t docid);
  48         void finish();
  49
  50         string encoded;
  51         size_t num_docids = 0;
  52
  53 private:
  54         void write_header(uint32_t docid);
  55         void append_block();
  56
  57         vector<uint32_t> pending_deltas;
  58
  59         uint32_t last_block_end, last_docid = -1;
  60 };
  61
  62 void PostingListBuilder::add_docid(uint32_t docid)
  63 {
  64         // Deduplicate against the last inserted value, if any.
  65         if (docid == last_docid) {
  66                 return;
  67         }
  68
  69         if (num_docids == 0) {
  70                 // Very first docid.
  71                 write_header(docid);
  72                 ++num_docids;
  73                 last_block_end = last_docid = docid;
  74                 return;
  75         }
  76
  77         pending_deltas.push_back(docid - last_docid - 1);
  78         last_docid = docid;
  79         if (pending_deltas.size() == 128) {
  80                 append_block();
  81                 pending_deltas.clear();
  82                 last_block_end = docid;
  83         }
  84         ++num_docids;
  85 }
  86
  87 void PostingListBuilder::finish()
  88 {
  89         if (pending_deltas.empty()) {
  90                 return;
  91         }
  92
  93         assert(!encoded.empty());  // write_header() should already have run.
  94
  95         // No interleaving for partial blocks.
  96         unsigned char buf[P4NENC_BOUND(128)];
  97         unsigned char *end = encode_pfor_single_block<128>(pending_deltas.data(), pending_deltas.size(), /*interleaved=*/false, buf);
  98         encoded.append(reinterpret_cast<char *>(buf), reinterpret_cast<char *>(end));
  99 }
 100
 101 void PostingListBuilder::append_block()
 102 {
 103         unsigned char buf[P4NENC_BOUND(128)];
 104         assert(pending_deltas.size() == 128);
 105         unsigned char *end = encode_pfor_single_block<128>(pending_deltas.data(), 128, /*interleaved=*/true, buf);
 106         encoded.append(reinterpret_cast<char *>(buf), reinterpret_cast<char *>(end));
 107 }
 108
 109 void PostingListBuilder::write_header(uint32_t docid)
 110 {
 111         unsigned char buf[P4NENC_BOUND(1)];
 112         unsigned char *end = write_baseval(docid, buf);
 113         encoded.append(reinterpret_cast<char *>(buf), end - buf);
 114 }
 115
 116 void DictionaryBuilder::add_file(string filename)
 117 {
 118         if (keep_current_block) {  // Only bother saving the filenames if we're actually keeping the block.
 119                 if (!current_block.empty()) {
 120                         current_block.push_back('\0');
 121                 }
 122                 current_block += filename;
 123         }
 124         if (++num_files_in_block == block_size) {
 125                 flush_block();
 126         }
 127 }
 128
 129 void DictionaryBuilder::flush_block()
 130 {
 131         if (keep_current_block) {
 132                 if (slot_for_current_block == -1) {
 133                         lengths.push_back(current_block.size());
 134                         sampled_blocks.push_back(move(current_block));
 135                 } else {
 136                         lengths[slot_for_current_block] = current_block.size();
 137                         sampled_blocks[slot_for_current_block] = move(current_block);
 138                 }
 139         }
 140         current_block.clear();
 141         num_files_in_block = 0;
 142         ++block_num;
 143
 144         if (block_num < blocks_to_keep) {
 145                 keep_current_block = true;
 146                 slot_for_current_block = -1;
 147         } else {
 148                 // Keep every block with equal probability (reservoir sampling).
 149                 uint64_t idx = uniform_int_distribution<uint64_t>(0, block_num)(reservoir_rand);
 150                 keep_current_block = (idx < blocks_to_keep);
 151                 slot_for_current_block = idx;
 152         }
 153 }
 154
 155 string DictionaryBuilder::train(size_t buf_size)
 156 {
 157         string dictionary_buf;
 158         sort(sampled_blocks.begin(), sampled_blocks.end());  // Seemingly important for decompression speed.
 159         for (const string &block : sampled_blocks) {
 160                 dictionary_buf += block;
 161         }
 162
 163         string buf;
 164         buf.resize(buf_size);
 165         size_t ret = ZDICT_trainFromBuffer(&buf[0], buf_size, dictionary_buf.data(), lengths.data(), lengths.size());
 166         if (ret == size_t(-1)) {
 167                 return "";
 168         }
 169         dprintf("Sampled %zu bytes in %zu blocks, built a dictionary of size %zu\n", dictionary_buf.size(), lengths.size(), ret);
 170         buf.resize(ret);
 171
 172         sampled_blocks.clear();
 173         lengths.clear();
 174
 175         return buf;
 176 }
 177
 178 Corpus::Corpus(FILE *outfp, size_t block_size, ZSTD_CDict *cdict)
 179         : invindex(new PostingListBuilder *[NUM_TRIGRAMS]), outfp(outfp), block_size(block_size), cdict(cdict)
 180 {
 181         fill(invindex.get(), invindex.get() + NUM_TRIGRAMS, nullptr);
 182 }
 183
 184 Corpus::~Corpus()
 185 {
 186         for (unsigned i = 0; i < NUM_TRIGRAMS; ++i) {
 187                 delete invindex[i];
 188         }
 189 }
 190
 191 PostingListBuilder &Corpus::get_pl_builder(uint32_t trgm)
 192 {
 193         if (invindex[trgm] == nullptr) {
 194                 invindex[trgm] = new PostingListBuilder;
 195         }
 196         return *invindex[trgm];
 197 }
 198
 199 void Corpus::add_file(string filename)
 200 {
 201         ++num_files;
 202         if (!current_block.empty()) {
 203                 current_block.push_back('\0');
 204         }
 205         current_block += filename;
 206         if (++num_files_in_block == block_size) {
 207                 flush_block();
 208         }
 209 }
 210
 211 void Corpus::flush_block()
 212 {
 213         if (current_block.empty()) {
 214                 return;
 215         }
 216
 217         uint32_t docid = num_blocks;
 218
 219         // Create trigrams.
 220         const char *ptr = current_block.c_str();
 221         while (ptr < current_block.c_str() + current_block.size()) {
 222                 string_view s(ptr);
 223                 if (s.size() >= 3) {
 224                         for (size_t j = 0; j < s.size() - 2; ++j) {
 225                                 uint32_t trgm = read_trigram(s, j);
 226                                 get_pl_builder(trgm).add_docid(docid);
 227                         }
 228                 }
 229                 ptr += s.size() + 1;
 230         }
 231
 232         // Compress and add the filename block.
 233         filename_blocks.push_back(ftell(outfp));
 234         string compressed = zstd_compress(current_block, cdict, &tempbuf);
 235         if (fwrite(compressed.data(), compressed.size(), 1, outfp) != 1) {
 236                 perror("fwrite()");
 237                 exit(1);
 238         }
 239
 240         current_block.clear();
 241         num_files_in_block = 0;
 242         ++num_blocks;
 243 }
 244
 245 void Corpus::finish()
 246 {
 247         flush_block();
 248 }
 249
 250 size_t Corpus::num_trigrams() const
 251 {
 252         size_t num = 0;
 253         for (unsigned trgm = 0; trgm < NUM_TRIGRAMS; ++trgm) {
 254                 if (invindex[trgm] != nullptr) {
 255                         ++num;
 256                 }
 257         }
 258         return num;
 259 }
 260
 261 string zstd_compress(const string &src, ZSTD_CDict *cdict, string *tempbuf)
 262 {
 263         static ZSTD_CCtx *ctx = nullptr;
 264         if (ctx == nullptr) {
 265                 ctx = ZSTD_createCCtx();
 266         }
 267
 268         size_t max_size = ZSTD_compressBound(src.size());
 269         if (tempbuf->size() < max_size) {
 270                 tempbuf->resize(max_size);
 271         }
 272         size_t size;
 273         if (cdict == nullptr) {
 274                 size = ZSTD_compressCCtx(ctx, &(*tempbuf)[0], max_size, src.data(), src.size(), /*level=*/6);
 275         } else {
 276                 size = ZSTD_compress_usingCDict(ctx, &(*tempbuf)[0], max_size, src.data(), src.size(), cdict);
 277         }
 278         return string(tempbuf->data(), size);
 279 }
 280
 281 bool is_prime(uint32_t x)
 282 {
 283         if ((x % 2) == 0 || (x % 3) == 0) {
 284                 return false;
 285         }
 286         uint32_t limit = ceil(sqrt(x));
 287         for (uint32_t factor = 5; factor <= limit; ++factor) {
 288                 if ((x % factor) == 0) {
 289                         return false;
 290                 }
 291         }
 292         return true;
 293 }
 294
 295 uint32_t next_prime(uint32_t x)
 296 {
 297         if ((x % 2) == 0) {
 298                 ++x;
 299         }
 300         while (!is_prime(x)) {
 301                 x += 2;
 302         }
 303         return x;
 304 }
 305
 306 unique_ptr<Trigram[]> create_hashtable(Corpus &corpus, const vector<uint32_t> &all_trigrams, uint32_t ht_size, uint32_t num_overflow_slots)
 307 {
 308         unique_ptr<Trigram[]> ht(new Trigram[ht_size + num_overflow_slots + 1]);  // 1 for the sentinel element at the end.
 309         for (unsigned i = 0; i < ht_size + num_overflow_slots + 1; ++i) {
 310                 ht[i].trgm = uint32_t(-1);
 311                 ht[i].num_docids = 0;
 312                 ht[i].offset = 0;
 313         }
 314         for (uint32_t trgm : all_trigrams) {
 315                 // We don't know offset yet, so set it to zero.
 316                 Trigram to_insert{ trgm, uint32_t(corpus.get_pl_builder(trgm).num_docids), 0 };
 317
 318                 uint32_t bucket = hash_trigram(trgm, ht_size);
 319                 unsigned distance = 0;
 320                 while (ht[bucket].num_docids != 0) {
 321                         // Robin Hood hashing; reduces the longest distance by a lot.
 322                         unsigned other_distance = bucket - hash_trigram(ht[bucket].trgm, ht_size);
 323                         if (distance > other_distance) {
 324                                 swap(to_insert, ht[bucket]);
 325                                 distance = other_distance;
 326                         }
 327
 328                         ++bucket, ++distance;
 329                         if (distance > num_overflow_slots) {
 330                                 return nullptr;
 331                         }
 332                 }
 333                 ht[bucket] = to_insert;
 334         }
 335         return ht;
 336 }
 337
 338 DatabaseBuilder::DatabaseBuilder(const char *outfile, int block_size, string dictionary)
 339         : outfile(outfile), block_size(block_size)
 340 {
 341         umask(0027);
 342
 343         string path = outfile;
 344         path.resize(path.find_last_of('/') + 1);
 345         int fd = open(path.c_str(), O_WRONLY | O_TMPFILE, 0640);
 346         if (fd == -1) {
 347                 perror(path.c_str());
 348                 exit(1);
 349         }
 350
 351         outfp = fdopen(fd, "wb");
 352         if (outfp == nullptr) {
 353                 perror(outfile);
 354                 exit(1);
 355         }
 356
 357         // Write the header.
 358         memcpy(hdr.magic, "\0plocate", 8);
 359         hdr.version = -1;  // Mark as broken.
 360         hdr.hashtable_size = 0;  // Not known yet.
 361         hdr.extra_ht_slots = num_overflow_slots;
 362         hdr.num_docids = 0;
 363         hdr.hash_table_offset_bytes = -1;  // We don't know these offsets yet.
 364         hdr.max_version = 1;
 365         hdr.filename_index_offset_bytes = -1;
 366         hdr.zstd_dictionary_length_bytes = -1;
 367         fwrite(&hdr, sizeof(hdr), 1, outfp);
 368
 369         if (dictionary.empty()) {
 370                 hdr.zstd_dictionary_offset_bytes = 0;
 371                 hdr.zstd_dictionary_length_bytes = 0;
 372         } else {
 373                 hdr.zstd_dictionary_offset_bytes = ftell(outfp);
 374                 fwrite(dictionary.data(), dictionary.size(), 1, outfp);
 375                 hdr.zstd_dictionary_length_bytes = dictionary.size();
 376                 cdict = ZSTD_createCDict(dictionary.data(), dictionary.size(), /*level=*/6);
 377         }
 378 }
 379
 380 Corpus *DatabaseBuilder::start_corpus()
 381 {
 382         corpus_start = steady_clock::now();
 383         corpus = new Corpus(outfp, block_size, cdict);
 384         return corpus;
 385 }
 386
 387 void DatabaseBuilder::finish_corpus()
 388 {
 389         corpus->finish();
 390         hdr.num_docids = corpus->filename_blocks.size();
 391
 392         // Stick an empty block at the end as sentinel.
 393         corpus->filename_blocks.push_back(ftell(outfp));
 394         const size_t bytes_for_filenames = corpus->filename_blocks.back() - corpus->filename_blocks.front();
 395
 396         // Write the offsets to the filenames.
 397         hdr.filename_index_offset_bytes = ftell(outfp);
 398         const size_t bytes_for_filename_index = corpus->filename_blocks.size() * sizeof(uint64_t);
 399         fwrite(corpus->filename_blocks.data(), corpus->filename_blocks.size(), sizeof(uint64_t), outfp);
 400         corpus->filename_blocks.clear();
 401         corpus->filename_blocks.shrink_to_fit();
 402
 403         // Finish up encoding the posting lists.
 404         size_t trigrams = 0, longest_posting_list = 0;
 405         size_t bytes_for_posting_lists = 0;
 406         for (unsigned trgm = 0; trgm < NUM_TRIGRAMS; ++trgm) {
 407                 if (!corpus->seen_trigram(trgm))
 408                         continue;
 409                 PostingListBuilder &pl_builder = corpus->get_pl_builder(trgm);
 410                 pl_builder.finish();
 411                 longest_posting_list = max(longest_posting_list, pl_builder.num_docids);
 412                 trigrams += pl_builder.num_docids;
 413                 bytes_for_posting_lists += pl_builder.encoded.size();
 414         }
 415         size_t num_trigrams = corpus->num_trigrams();
 416         dprintf("%zu files, %zu different trigrams, %zu entries, avg len %.2f, longest %zu\n",
 417                 corpus->num_files, num_trigrams, trigrams, double(trigrams) / num_trigrams, longest_posting_list);
 418         dprintf("%zu bytes used for posting lists (%.2f bits/entry)\n", bytes_for_posting_lists, 8 * bytes_for_posting_lists / double(trigrams));
 419
 420         dprintf("Building posting lists took %.1f ms.\n\n", 1e3 * duration<float>(steady_clock::now() - corpus_start).count());
 421
 422         // Find the used trigrams.
 423         vector<uint32_t> all_trigrams;
 424         for (unsigned trgm = 0; trgm < NUM_TRIGRAMS; ++trgm) {
 425                 if (corpus->seen_trigram(trgm)) {
 426                         all_trigrams.push_back(trgm);
 427                 }
 428         }
 429
 430         // Create the hash table.
 431         unique_ptr<Trigram[]> hashtable;
 432         uint32_t ht_size = next_prime(all_trigrams.size());
 433         for (;;) {
 434                 hashtable = create_hashtable(*corpus, all_trigrams, ht_size, num_overflow_slots);
 435                 if (hashtable == nullptr) {
 436                         dprintf("Failed creating hash table of size %u, increasing by 5%% and trying again.\n", ht_size);
 437                         ht_size = next_prime(ht_size * 1.05);
 438                 } else {
 439                         dprintf("Created hash table of size %u.\n\n", ht_size);
 440                         break;
 441                 }
 442         }
 443
 444         // Find the offsets for each posting list.
 445         size_t bytes_for_hashtable = (ht_size + num_overflow_slots + 1) * sizeof(Trigram);
 446         uint64_t offset = ftell(outfp) + bytes_for_hashtable;
 447         for (unsigned i = 0; i < ht_size + num_overflow_slots + 1; ++i) {
 448                 hashtable[i].offset = offset;  // Needs to be there even for empty slots.
 449                 if (hashtable[i].num_docids == 0) {
 450                         continue;
 451                 }
 452
 453                 const string &encoded = corpus->get_pl_builder(hashtable[i].trgm).encoded;
 454                 offset += encoded.size();
 455         }
 456
 457         // Write the hash table.
 458         hdr.hash_table_offset_bytes = ftell(outfp);
 459         hdr.hashtable_size = ht_size;
 460         fwrite(hashtable.get(), ht_size + num_overflow_slots + 1, sizeof(Trigram), outfp);
 461
 462         // Write the actual posting lists.
 463         for (unsigned i = 0; i < ht_size + num_overflow_slots + 1; ++i) {
 464                 if (hashtable[i].num_docids == 0) {
 465                         continue;
 466                 }
 467                 const string &encoded = corpus->get_pl_builder(hashtable[i].trgm).encoded;
 468                 fwrite(encoded.data(), encoded.size(), 1, outfp);
 469         }
 470
 471         // Rewind, and write the updated header.
 472         hdr.version = 1;
 473         fseek(outfp, 0, SEEK_SET);
 474         fwrite(&hdr, sizeof(hdr), 1, outfp);
 475
 476         // Give the file a proper name, making it visible in the file system.
 477         // TODO: It would be nice to be able to do this atomically, like with rename.
 478         unlink(outfile.c_str());
 479         char procpath[256];
 480         snprintf(procpath, sizeof(procpath), "/proc/self/fd/%d", fileno(outfp));
 481         if (linkat(AT_FDCWD, procpath, AT_FDCWD, outfile.c_str(), AT_SYMLINK_FOLLOW) == -1) {
 482                 perror("linkat");
 483                 exit(1);
 484         }
 485
 486         fclose(outfp);
 487
 488         size_t total_bytes = (bytes_for_hashtable + bytes_for_posting_lists + bytes_for_filename_index + bytes_for_filenames);
 489
 490         dprintf("Block size:     %7d files\n", block_size);
 491         dprintf("Dictionary:     %'7.1f MB\n", hdr.zstd_dictionary_length_bytes / 1048576.0);
 492         dprintf("Hash table:     %'7.1f MB\n", bytes_for_hashtable / 1048576.0);
 493         dprintf("Posting lists:  %'7.1f MB\n", bytes_for_posting_lists / 1048576.0);
 494         dprintf("Filename index: %'7.1f MB\n", bytes_for_filename_index / 1048576.0);
 495         dprintf("Filenames:      %'7.1f MB\n", bytes_for_filenames / 1048576.0);
 496         dprintf("Total:          %'7.1f MB\n", total_bytes / 1048576.0);
 497         dprintf("\n");
 498 }