plocate/database-builder.cpp

#include "database-builder.h"

#include "dprintf.h"
#include "turbopfor-encode.h"

#include <algorithm>
#include <assert.h>
#include <fcntl.h>
#include <string.h>
#include <string_view>
#include <sys/stat.h>
#include <sys/time.h>
#include <sys/types.h>
#include <unistd.h>
#include <zdict.h>
#include <zstd.h>

#define P4NENC_BOUND(n) ((n + 127) / 128 + (n + 32) * sizeof(uint32_t))

#define NUM_TRIGRAMS 16777216

using namespace std;
using namespace std::chrono;

constexpr unsigned num_overflow_slots = 16;

string zstd_compress(const string &src, ZSTD_CDict *cdict, string *tempbuf);

static inline uint32_t read_unigram(const string_view s, size_t idx)
{
	if (idx < s.size()) {
		return (unsigned char)s[idx];
	} else {
		return 0;
	}
}

static inline uint32_t read_trigram(const string_view s, size_t start)
{
	return read_unigram(s, start) |
		(read_unigram(s, start + 1) << 8) |
		(read_unigram(s, start + 2) << 16);
}

class PostingListBuilder {
public:
	inline void add_docid(uint32_t docid);
	void finish();

	string encoded;
	size_t num_docids = 0;

private:
	void write_header(uint32_t docid);
	void append_block();

	vector<uint32_t> pending_deltas;

	uint32_t last_block_end, last_docid = -1;
};

void PostingListBuilder::add_docid(uint32_t docid)
{
	// Deduplicate against the last inserted value, if any.
	if (docid == last_docid) {
		return;
	}

	if (num_docids == 0) {
		// Very first docid.
		write_header(docid);
		++num_docids;
		last_block_end = last_docid = docid;
		return;
	}

	pending_deltas.push_back(docid - last_docid - 1);
	last_docid = docid;
	if (pending_deltas.size() == 128) {
		append_block();
		pending_deltas.clear();
		last_block_end = docid;
	}
	++num_docids;
}

void PostingListBuilder::finish()
{
	if (pending_deltas.empty()) {
		return;
	}

	assert(!encoded.empty());  // write_header() should already have run.

	// No interleaving for partial blocks.
	unsigned char buf[P4NENC_BOUND(128)];
	unsigned char *end = encode_pfor_single_block<128>(pending_deltas.data(), pending_deltas.size(), /*interleaved=*/false, buf);
	encoded.append(reinterpret_cast<char *>(buf), reinterpret_cast<char *>(end));
}

void PostingListBuilder::append_block()
{
	unsigned char buf[P4NENC_BOUND(128)];
	assert(pending_deltas.size() == 128);
	unsigned char *end = encode_pfor_single_block<128>(pending_deltas.data(), 128, /*interleaved=*/true, buf);
	encoded.append(reinterpret_cast<char *>(buf), reinterpret_cast<char *>(end));
}

void PostingListBuilder::write_header(uint32_t docid)
{
	unsigned char buf[P4NENC_BOUND(1)];
	unsigned char *end = write_baseval(docid, buf);
	encoded.append(reinterpret_cast<char *>(buf), end - buf);
}

void DictionaryBuilder::add_file(string filename)
{
	if (keep_current_block) {  // Only bother saving the filenames if we're actually keeping the block.
		if (!current_block.empty()) {
			current_block.push_back('\0');
		}
		current_block += filename;
	}
	if (++num_files_in_block == block_size) {
		flush_block();
	}
}

void DictionaryBuilder::flush_block()
{
	if (keep_current_block) {
		if (slot_for_current_block == -1) {
			lengths.push_back(current_block.size());
			sampled_blocks.push_back(move(current_block));
		} else {
			lengths[slot_for_current_block] = current_block.size();
			sampled_blocks[slot_for_current_block] = move(current_block);
		}
	}
	current_block.clear();
	num_files_in_block = 0;
	++block_num;

	if (block_num < blocks_to_keep) {
		keep_current_block = true;
		slot_for_current_block = -1;
	} else {
		// Keep every block with equal probability (reservoir sampling).
		uint64_t idx = uniform_int_distribution<uint64_t>(0, block_num)(reservoir_rand);
		keep_current_block = (idx < blocks_to_keep);
		slot_for_current_block = idx;
	}
}

string DictionaryBuilder::train(size_t buf_size)
{
	string dictionary_buf;
	sort(sampled_blocks.begin(), sampled_blocks.end());  // Seemingly important for decompression speed.
	for (const string &block : sampled_blocks) {
		dictionary_buf += block;
	}

	string buf;
	buf.resize(buf_size);
	size_t ret = ZDICT_trainFromBuffer(&buf[0], buf_size, dictionary_buf.data(), lengths.data(), lengths.size());
	if (ret == size_t(-1)) {
		return "";
	}
	dprintf("Sampled %zu bytes in %zu blocks, built a dictionary of size %zu\n", dictionary_buf.size(), lengths.size(), ret);
	buf.resize(ret);

	sampled_blocks.clear();
	lengths.clear();

	return buf;
}

Corpus::Corpus(FILE *outfp, size_t block_size, ZSTD_CDict *cdict)
	: invindex(new PostingListBuilder *[NUM_TRIGRAMS]), outfp(outfp), block_size(block_size), cdict(cdict)
{
	fill(invindex.get(), invindex.get() + NUM_TRIGRAMS, nullptr);
}

Corpus::~Corpus()
{
	for (unsigned i = 0; i < NUM_TRIGRAMS; ++i) {
		delete invindex[i];
	}
}

PostingListBuilder &Corpus::get_pl_builder(uint32_t trgm)
{
	if (invindex[trgm] == nullptr) {
		invindex[trgm] = new PostingListBuilder;
	}
	return *invindex[trgm];
}

void Corpus::add_file(string filename)
{
	++num_files;
	if (!current_block.empty()) {
		current_block.push_back('\0');
	}
	current_block += filename;
	if (++num_files_in_block == block_size) {
		flush_block();
	}
}

void Corpus::flush_block()
{
	if (current_block.empty()) {
		return;
	}

	uint32_t docid = num_blocks;

	// Create trigrams.
	const char *ptr = current_block.c_str();
	while (ptr < current_block.c_str() + current_block.size()) {
		string_view s(ptr);
		if (s.size() >= 3) {
			for (size_t j = 0; j < s.size() - 2; ++j) {
				uint32_t trgm = read_trigram(s, j);
				get_pl_builder(trgm).add_docid(docid);
			}
		}
		ptr += s.size() + 1;
	}

	// Compress and add the filename block.
	filename_blocks.push_back(ftell(outfp));
	string compressed = zstd_compress(current_block, cdict, &tempbuf);
	if (fwrite(compressed.data(), compressed.size(), 1, outfp) != 1) {
		perror("fwrite()");
		exit(1);
	}

	current_block.clear();
	num_files_in_block = 0;
	++num_blocks;
}

void Corpus::finish()
{
	flush_block();
}

size_t Corpus::num_trigrams() const
{
	size_t num = 0;
	for (unsigned trgm = 0; trgm < NUM_TRIGRAMS; ++trgm) {
		if (invindex[trgm] != nullptr) {
			++num;
		}
	}
	return num;
}

string zstd_compress(const string &src, ZSTD_CDict *cdict, string *tempbuf)
{
	static ZSTD_CCtx *ctx = nullptr;
	if (ctx == nullptr) {
		ctx = ZSTD_createCCtx();
	}

	size_t max_size = ZSTD_compressBound(src.size());
	if (tempbuf->size() < max_size) {
		tempbuf->resize(max_size);
	}
	size_t size;
	if (cdict == nullptr) {
		size = ZSTD_compressCCtx(ctx, &(*tempbuf)[0], max_size, src.data(), src.size(), /*level=*/6);
	} else {
		size = ZSTD_compress_usingCDict(ctx, &(*tempbuf)[0], max_size, src.data(), src.size(), cdict);
	}
	return string(tempbuf->data(), size);
}

bool is_prime(uint32_t x)
{
	if ((x % 2) == 0 || (x % 3) == 0) {
		return false;
	}
	uint32_t limit = ceil(sqrt(x));
	for (uint32_t factor = 5; factor <= limit; ++factor) {
		if ((x % factor) == 0) {
			return false;
		}
	}
	return true;
}

uint32_t next_prime(uint32_t x)
{
	if ((x % 2) == 0) {
		++x;
	}
	while (!is_prime(x)) {
		x += 2;
	}
	return x;
}

unique_ptr<Trigram[]> create_hashtable(Corpus &corpus, const vector<uint32_t> &all_trigrams, uint32_t ht_size, uint32_t num_overflow_slots)
{
	unique_ptr<Trigram[]> ht(new Trigram[ht_size + num_overflow_slots + 1]);  // 1 for the sentinel element at the end.
	for (unsigned i = 0; i < ht_size + num_overflow_slots + 1; ++i) {
		ht[i].trgm = uint32_t(-1);
		ht[i].num_docids = 0;
		ht[i].offset = 0;
	}
	for (uint32_t trgm : all_trigrams) {
		// We don't know offset yet, so set it to zero.
		Trigram to_insert{ trgm, uint32_t(corpus.get_pl_builder(trgm).num_docids), 0 };

		uint32_t bucket = hash_trigram(trgm, ht_size);
		unsigned distance = 0;
		while (ht[bucket].num_docids != 0) {
			// Robin Hood hashing; reduces the longest distance by a lot.
			unsigned other_distance = bucket - hash_trigram(ht[bucket].trgm, ht_size);
			if (distance > other_distance) {
				swap(to_insert, ht[bucket]);
				distance = other_distance;
			}

			++bucket, ++distance;
			if (distance > num_overflow_slots) {
				return nullptr;
			}
		}
		ht[bucket] = to_insert;
	}
	return ht;
}

DatabaseBuilder::DatabaseBuilder(const char *outfile, int block_size, string dictionary)
	: outfile(outfile), block_size(block_size)
{
	umask(0027);

	string path = outfile;
	path.resize(path.find_last_of('/') + 1);
	int fd = open(path.c_str(), O_WRONLY | O_TMPFILE, 0640);
	if (fd == -1) {
		perror(path.c_str());
		exit(1);
	}

	outfp = fdopen(fd, "wb");
	if (outfp == nullptr) {
		perror(outfile);
		exit(1);
	}

	// Write the header.
	memcpy(hdr.magic, "\0plocate", 8);
	hdr.version = -1;  // Mark as broken.
	hdr.hashtable_size = 0;  // Not known yet.
	hdr.extra_ht_slots = num_overflow_slots;
	hdr.num_docids = 0;
	hdr.hash_table_offset_bytes = -1;  // We don't know these offsets yet.
	hdr.max_version = 1;
	hdr.filename_index_offset_bytes = -1;
	hdr.zstd_dictionary_length_bytes = -1;
	fwrite(&hdr, sizeof(hdr), 1, outfp);

	if (dictionary.empty()) {
		hdr.zstd_dictionary_offset_bytes = 0;
		hdr.zstd_dictionary_length_bytes = 0;
	} else {
		hdr.zstd_dictionary_offset_bytes = ftell(outfp);
		fwrite(dictionary.data(), dictionary.size(), 1, outfp);
		hdr.zstd_dictionary_length_bytes = dictionary.size();
		cdict = ZSTD_createCDict(dictionary.data(), dictionary.size(), /*level=*/6);
	}
}

Corpus *DatabaseBuilder::start_corpus()
{
	corpus_start = steady_clock::now();
	corpus = new Corpus(outfp, block_size, cdict);
	return corpus;
}

void DatabaseBuilder::finish_corpus()
{
	corpus->finish();
	hdr.num_docids = corpus->filename_blocks.size();

	// Stick an empty block at the end as sentinel.
	corpus->filename_blocks.push_back(ftell(outfp));
	const size_t bytes_for_filenames = corpus->filename_blocks.back() - corpus->filename_blocks.front();

	// Write the offsets to the filenames.
	hdr.filename_index_offset_bytes = ftell(outfp);
	const size_t bytes_for_filename_index = corpus->filename_blocks.size() * sizeof(uint64_t);
	fwrite(corpus->filename_blocks.data(), corpus->filename_blocks.size(), sizeof(uint64_t), outfp);
	corpus->filename_blocks.clear();
	corpus->filename_blocks.shrink_to_fit();

	// Finish up encoding the posting lists.
	size_t trigrams = 0, longest_posting_list = 0;
	size_t bytes_for_posting_lists = 0;
	for (unsigned trgm = 0; trgm < NUM_TRIGRAMS; ++trgm) {
		if (!corpus->seen_trigram(trgm))
			continue;
		PostingListBuilder &pl_builder = corpus->get_pl_builder(trgm);
		pl_builder.finish();
		longest_posting_list = max(longest_posting_list, pl_builder.num_docids);
		trigrams += pl_builder.num_docids;
		bytes_for_posting_lists += pl_builder.encoded.size();
	}
	size_t num_trigrams = corpus->num_trigrams();
	dprintf("%zu files, %zu different trigrams, %zu entries, avg len %.2f, longest %zu\n",
	        corpus->num_files, num_trigrams, trigrams, double(trigrams) / num_trigrams, longest_posting_list);
	dprintf("%zu bytes used for posting lists (%.2f bits/entry)\n", bytes_for_posting_lists, 8 * bytes_for_posting_lists / double(trigrams));

	dprintf("Building posting lists took %.1f ms.\n\n", 1e3 * duration<float>(steady_clock::now() - corpus_start).count());

	// Find the used trigrams.
	vector<uint32_t> all_trigrams;
	for (unsigned trgm = 0; trgm < NUM_TRIGRAMS; ++trgm) {
		if (corpus->seen_trigram(trgm)) {
			all_trigrams.push_back(trgm);
		}
	}

	// Create the hash table.
	unique_ptr<Trigram[]> hashtable;
	uint32_t ht_size = next_prime(all_trigrams.size());
	for (;;) {
		hashtable = create_hashtable(*corpus, all_trigrams, ht_size, num_overflow_slots);
		if (hashtable == nullptr) {
			dprintf("Failed creating hash table of size %u, increasing by 5%% and trying again.\n", ht_size);
			ht_size = next_prime(ht_size * 1.05);
		} else {
			dprintf("Created hash table of size %u.\n\n", ht_size);
			break;
		}
	}

	// Find the offsets for each posting list.
	size_t bytes_for_hashtable = (ht_size + num_overflow_slots + 1) * sizeof(Trigram);
	uint64_t offset = ftell(outfp) + bytes_for_hashtable;
	for (unsigned i = 0; i < ht_size + num_overflow_slots + 1; ++i) {
		hashtable[i].offset = offset;  // Needs to be there even for empty slots.
		if (hashtable[i].num_docids == 0) {
			continue;
		}

		const string &encoded = corpus->get_pl_builder(hashtable[i].trgm).encoded;
		offset += encoded.size();
	}

	// Write the hash table.
	hdr.hash_table_offset_bytes = ftell(outfp);
	hdr.hashtable_size = ht_size;
	fwrite(hashtable.get(), ht_size + num_overflow_slots + 1, sizeof(Trigram), outfp);

	// Write the actual posting lists.
	for (unsigned i = 0; i < ht_size + num_overflow_slots + 1; ++i) {
		if (hashtable[i].num_docids == 0) {
			continue;
		}
		const string &encoded = corpus->get_pl_builder(hashtable[i].trgm).encoded;
		fwrite(encoded.data(), encoded.size(), 1, outfp);
	}

	// Rewind, and write the updated header.
	hdr.version = 1;
	fseek(outfp, 0, SEEK_SET);
	fwrite(&hdr, sizeof(hdr), 1, outfp);

	// Give the file a proper name, making it visible in the file system.
	// TODO: It would be nice to be able to do this atomically, like with rename.
	unlink(outfile.c_str());
	char procpath[256];
	snprintf(procpath, sizeof(procpath), "/proc/self/fd/%d", fileno(outfp));
	if (linkat(AT_FDCWD, procpath, AT_FDCWD, outfile.c_str(), AT_SYMLINK_FOLLOW) == -1) {
		perror("linkat");
		exit(1);
	}

	fclose(outfp);

	size_t total_bytes = (bytes_for_hashtable + bytes_for_posting_lists + bytes_for_filename_index + bytes_for_filenames);

	dprintf("Block size:     %7d files\n", block_size);
	dprintf("Dictionary:     %'7.1f MB\n", hdr.zstd_dictionary_length_bytes / 1048576.0);
	dprintf("Hash table:     %'7.1f MB\n", bytes_for_hashtable / 1048576.0);
	dprintf("Posting lists:  %'7.1f MB\n", bytes_for_posting_lists / 1048576.0);
	dprintf("Filename index: %'7.1f MB\n", bytes_for_filename_index / 1048576.0);
	dprintf("Filenames:      %'7.1f MB\n", bytes_for_filenames / 1048576.0);
	dprintf("Total:          %'7.1f MB\n", total_bytes / 1048576.0);
	dprintf("\n");
}