Open Source alternative to Algolia + Pinecone and an Easier-to-Use alternative to ElasticSearch ⚡ 🔍 ✨ Fast, typo tolerant, in-memory fuzzy Search Engine for building delightful search experiences
C++ algoliadatastoreelasticsearchenterprise-searchfacetingfull-text-searchfuzzy-searchgeosearchin-memoryinstantsearchmerchandisingpineconesearchsearch-enginesemantic-searchsimilarity-searchsite-searchsynonymstypo-tolerancevector-search
Fix iterator skip performance.
Kishore Nallan4d3f6dd0de24
diff --git a/include/or_iterator.h b/include/or_iterator.hindex 566d310a..bdabfee7 100644--- a/include/or_iterator.h+++ b/include/or_iterator.h@@ -69,40 +69,43 @@ bool or_iterator_t::intersect(std::vector<or_iterator_t>& its, result_iter_stateauto id = its[0].id();if(take_id(istate, id, is_excluded)) {func(id, its);- its[0].next();- } else {- // skip iterator till next id available in filter- if(istate.filter_ids_length != 0 && !is_excluded) {- if(istate.filter_ids_index < istate.filter_ids_length) {- its[0].skip_to(istate.filter_ids[istate.filter_ids_index]);- } else {- break;- }+ }++ if(istate.filter_ids_length != 0 && !is_excluded) {+ if(istate.filter_ids_index < istate.filter_ids_length) {+ // skip iterator till next id available in filter+ its[0].skip_to(istate.filter_ids[istate.filter_ids_index]);} else {- its[0].next();+ break;}+ } else {+ its[0].next();}}break;case 2:+ if(istate.filter_ids_length != 0) {+ its[0].skip_to(istate.filter_ids[istate.filter_ids_index]);+ its[1].skip_to(istate.filter_ids[istate.filter_ids_index]);+ }+while(its.size() == it_size && !at_end2(its)) {if(equals2(its)) {auto id = its[0].id();if(take_id(istate, id, is_excluded)) {func(id, its);- advance_all2(its);- } else {- // skip iterator till next id available in filter- if(istate.filter_ids_length != 0 && !is_excluded) {- if(istate.filter_ids_index < istate.filter_ids_length) {- its[0].skip_to(istate.filter_ids[istate.filter_ids_index]);- its[1].skip_to(istate.filter_ids[istate.filter_ids_index]);- } else {- break;- }+ }++ if(istate.filter_ids_length != 0 && !is_excluded) {+ if(istate.filter_ids_index < istate.filter_ids_length) {+ // skip iterator till next id available in filter+ its[0].skip_to(istate.filter_ids[istate.filter_ids_index]);+ its[1].skip_to(istate.filter_ids[istate.filter_ids_index]);} else {- advance_all2(its);+ break;}+ } else {+ advance_all2(its);}} else {advance_non_largest2(its);@@ -110,26 +113,30 @@ bool or_iterator_t::intersect(std::vector<or_iterator_t>& its, result_iter_state}break;default:+ if(istate.filter_ids_length != 0) {+ for(auto& it: its) {+ it.skip_to(istate.filter_ids[istate.filter_ids_index]);+ }+ }+while(its.size() == it_size && !at_end(its)) {if(equals(its)) {- //LOG(INFO) << its[0].id();auto id = its[0].id();if(take_id(istate, id, is_excluded)) {func(id, its);- advance_all(its);- } else {- // skip iterator till next id available in filter- if(istate.filter_ids_length != 0 && !is_excluded) {- if(istate.filter_ids_index < istate.filter_ids_length) {- for(auto& it: its) {- it.skip_to(istate.filter_ids[istate.filter_ids_index]);- }- } else {- break;+ }++ if(istate.filter_ids_length != 0 && !is_excluded) {+ if(istate.filter_ids_index < istate.filter_ids_length) {+ // skip iterator till next id available in filter+ for(auto& it: its) {+ it.skip_to(istate.filter_ids[istate.filter_ids_index]);}} else {- advance_all2(its);+ break;}+ } else {+ advance_all2(its);}} else {advance_non_largest(its);diff --git a/include/posting.h b/include/posting.hindex 0e7595f8..9bb68c9e 100644--- a/include/posting.h+++ b/include/posting.h@@ -49,14 +49,12 @@ public:std::vector<posting_list_t*> expanded_plists;result_iter_state_t& iter_state;ThreadPool* thread_pool;- size_t parallelize_min_ids;block_intersector_t(const std::vector<void*>& raw_posting_lists,result_iter_state_t& iter_state,ThreadPool* thread_pool,size_t parallelize_min_ids = 1):- iter_state(iter_state), thread_pool(thread_pool),- parallelize_min_ids(parallelize_min_ids) {+ iter_state(iter_state), thread_pool(thread_pool) {to_expanded_plists(raw_posting_lists, plists, expanded_plists);@@ -75,8 +73,6 @@ public:template<class T>bool intersect(T func, size_t concurrency=4);-- void split_lists(size_t concurrency, std::vector<std::vector<posting_list_t::iterator_t>>& partial_its_vec);};static void to_expanded_plists(const std::vector<void*>& raw_posting_lists, std::vector<posting_list_t*>& plists,@@ -120,85 +116,17 @@ public:template<class T>bool posting_t::block_intersector_t::intersect(T func, size_t concurrency) {- // Split posting lists into N chunks and intersect them in-parallel- // 1. Sort posting lists by number of blocks- // 2. Iterate on the posting list with least number of blocks on N-block windows- // 3. On each window, pick the last block's last ID and identify blocks from other lists containing that ID- // 4. Construct N groups of iterators this way (the last block must overlap on both sides of the window)-if(plists.empty()) {return true;}- if(plists[0]->num_ids() < parallelize_min_ids) {- std::vector<posting_list_t::iterator_t> its;- its.reserve(plists.size());-- for(const auto& posting_list: plists) {- its.push_back(posting_list->new_iterator());- }-- posting_list_t::block_intersect<T>(its, iter_state, func);- return true;- }-- std::vector<std::vector<posting_list_t::iterator_t>> partial_its_vec(concurrency);- split_lists(concurrency, partial_its_vec);-- /*for(size_t i = 0; i < partial_its_vec.size(); i++) {- auto& partial_its = partial_its_vec[i];-- if (partial_its.empty()) {- continue;- }-- LOG(INFO) << "Vec " << i;+ std::vector<posting_list_t::iterator_t> its;+ its.reserve(plists.size());- for (auto& it: partial_its) {- while (it.valid()) {- LOG(INFO) << it.id();- it.next();- }-- LOG(INFO) << "---";- }- }*/-- size_t num_processed = 0;- std::mutex m_process;- std::condition_variable cv_process;- size_t num_non_empty = 0;-- for(size_t i = 0; i < partial_its_vec.size(); i++) {- auto& partial_its = partial_its_vec[i];-- if(partial_its.empty()) {- continue;- }-- num_non_empty++;-- /*for(auto& it: partial_its) {- while(it.valid()) {- LOG(INFO) << it.id();- it.next();- }-- LOG(INFO) << "---";- }*/-- thread_pool->enqueue([this, i, &func, &partial_its, &num_processed, &m_process, &cv_process]() {- auto iter_state_copy = iter_state;- iter_state_copy.index = i;- posting_list_t::block_intersect<T>(partial_its, iter_state_copy, func);- std::unique_lock<std::mutex> lock(m_process);- num_processed++;- cv_process.notify_one();- });+ for(const auto& posting_list: plists) {+ its.push_back(posting_list->new_iterator());}- std::unique_lock<std::mutex> lock_process(m_process);- cv_process.wait(lock_process, [&](){ return num_processed == num_non_empty; });-+ posting_list_t::block_intersect<T>(its, iter_state, func);return true;}diff --git a/include/posting_list.h b/include/posting_list.hindex 7500fb5c..d044333e 100644--- a/include/posting_list.h+++ b/include/posting_list.h@@ -62,6 +62,7 @@ public:class iterator_t {private:+ const std::map<last_id_t, block_t*>* id_block_map;block_t* curr_block;uint32_t curr_index;block_t* end_block;@@ -75,18 +76,20 @@ public:uint32_t* offset_index = nullptr;uint32_t* offsets = nullptr;- explicit iterator_t(block_t* start, block_t* end, bool auto_destroy = true, uint32_t field_id = 0);+ explicit iterator_t(const std::map<last_id_t, block_t*>* id_block_map,+ block_t* start, block_t* end, bool auto_destroy = true, uint32_t field_id = 0);~iterator_t();iterator_t(iterator_t&& rhs) noexcept;iterator_t& operator=(iterator_t&& rhs) noexcept;- void destroy();+ void reset_cache();[[nodiscard]] bool valid() const;void next();void skip_to(uint32_t id);void set_index(uint32_t index);[[nodiscard]] uint32_t id() const;+ [[nodiscard]] uint32_t last_block_id() const;[[nodiscard]] inline uint32_t index() const;[[nodiscard]] inline block_t* block() const;[[nodiscard]] uint32_t get_field_id() const;diff --git a/src/or_iterator.cpp b/src/or_iterator.cppindex e22ab63f..dbfe9e31 100644--- a/src/or_iterator.cpp+++ b/src/or_iterator.cpp@@ -128,7 +128,7 @@ void or_iterator_t::advance_smallest() {}if(!its[i].valid()) {- its[i].destroy();+ its[i].reset_cache();its.erase(its.cbegin() + i);i--;}@@ -152,7 +152,7 @@ bool or_iterator_t::skip_to(uint32_t id) {it.skip_to(id);if(!it.valid()) {- its[i].destroy();+ its[i].reset_cache();its.erase(its.begin() + i);i--;} else {@@ -238,7 +238,7 @@ const std::vector<posting_list_t::iterator_t>& or_iterator_t::get_its() const {or_iterator_t::~or_iterator_t() noexcept {for(auto& it: its) {- it.destroy();+ it.reset_cache();}}diff --git a/src/posting.cpp b/src/posting.cppindex e6be0a9e..8b72f078 100644--- a/src/posting.cpp+++ b/src/posting.cpp@@ -517,56 +517,3 @@ void posting_t::get_phrase_matches(const std::vector<void*>& raw_posting_lists,delete expanded_plist;}}--void posting_t::block_intersector_t::split_lists(size_t concurrency,- std::vector<std::vector<posting_list_t::iterator_t>>& partial_its_vec) {- const size_t num_blocks = this->plists[0]->num_blocks();- const size_t window_size = (num_blocks + concurrency - 1) / concurrency; // rounds up-- size_t blocks_traversed = 0;- posting_list_t::block_t* start_block = this->plists[0]->get_root();- posting_list_t::block_t* curr_block = start_block;-- size_t window_index = 0;-- while(curr_block != nullptr) {- blocks_traversed++;- if(blocks_traversed % window_size == 0 || blocks_traversed == num_blocks) {- // construct partial iterators and intersect within them-- std::vector<posting_list_t::iterator_t>& partial_its = partial_its_vec[window_index];-- for(size_t i = 0; i < this->plists.size(); i++) {- posting_list_t::block_t* p_start_block = nullptr;- posting_list_t::block_t* p_end_block = nullptr;-- // [1, 2] [3, 4] [5, 6]- // [3, 5] [6]-- if(i == 0) {- p_start_block = start_block;- p_end_block = curr_block->next;- } else {- auto start_block_first_id = start_block->ids.at(0);- auto end_block_last_id = curr_block->ids.last();-- p_start_block = this->plists[i]->block_of(start_block_first_id);- posting_list_t::block_t* last_block = this->plists[i]->block_of(end_block_last_id);-- if(p_start_block == last_block && p_start_block != nullptr) {- p_end_block = p_start_block->next;- } else {- p_end_block = last_block == nullptr ? nullptr : last_block->next;- }- }-- partial_its.emplace_back(p_start_block, p_end_block);- }-- start_block = curr_block->next;- window_index++;- }-- curr_block = curr_block->next;- }-}diff --git a/src/posting_list.cpp b/src/posting_list.cppindex f3966cb5..2de0fdde 100644--- a/src/posting_list.cpp+++ b/src/posting_list.cpp@@ -794,13 +794,6 @@ bool posting_list_t::get_offsets(const std::vector<iterator_t>& its,continue;}- /*uint32_t* offsets = curr_block->offsets.uncompress();-- uint32_t start_offset = curr_block->offset_index.at(curr_index);- uint32_t end_offset = (curr_index == curr_block->size() - 1) ?- curr_block->offsets.getLength() :- curr_block->offset_index.at(curr_index + 1);*/-uint32_t* offsets = its[j].offsets;uint32_t start_offset = its[j].offset_index[curr_index];@@ -859,8 +852,6 @@ bool posting_list_t::get_offsets(const std::vector<iterator_t>& its,// for plain string fieldsarray_token_pos[0].push_back(token_positions_t{is_last_token, positions});}-- //delete [] offsets;}return true;@@ -941,7 +932,7 @@ bool posting_list_t::equals2(std::vector<posting_list_t::iterator_t>& its) {posting_list_t::iterator_t posting_list_t::new_iterator(block_t* start_block, block_t* end_block, uint32_t field_id) {start_block = (start_block == nullptr) ? &root_block : start_block;- return posting_list_t::iterator_t(start_block, end_block, true, field_id);+ return posting_list_t::iterator_t(&id_block_map, start_block, end_block, true, field_id);}void posting_list_t::advance_all(std::vector<posting_list_t::iterator_t>& its) {@@ -1391,9 +1382,11 @@ size_t posting_list_t::get_last_offset(const posting_list_t::iterator_t& it, boo/* iterator_t operations */-posting_list_t::iterator_t::iterator_t(posting_list_t::block_t* start, posting_list_t::block_t* end,+posting_list_t::iterator_t::iterator_t(const std::map<last_id_t, block_t*>* id_block_map,+ posting_list_t::block_t* start, posting_list_t::block_t* end,bool auto_destroy, uint32_t field_id):- curr_block(start), curr_index(0), end_block(end), auto_destroy(auto_destroy), field_id(field_id) {+ id_block_map(id_block_map), curr_block(start), curr_index(0), end_block(end),+ auto_destroy(auto_destroy), field_id(field_id) {if(curr_block != end_block) {ids = curr_block->ids.uncompress();@@ -1426,6 +1419,10 @@ void posting_list_t::iterator_t::next() {}}+uint32_t posting_list_t::iterator_t::last_block_id() const {+ return ids[curr_block->size() - 1];+}+uint32_t posting_list_t::iterator_t::id() const {return ids[curr_index];}@@ -1439,52 +1436,56 @@ posting_list_t::block_t* posting_list_t::iterator_t::block() const {}void posting_list_t::iterator_t::skip_to(uint32_t id) {- bool skipped_block = false;- while(curr_block != end_block && curr_block->ids.last() < id) {- curr_block = curr_block->next;-- delete [] ids;- delete [] offset_index;- delete [] offsets;-- ids = offset_index = offsets = nullptr;-- if(curr_block != end_block) {- ids = curr_block->ids.uncompress();- offset_index = curr_block->offset_index.uncompress();- offsets = curr_block->offsets.uncompress();+ // first look to skip within current block+ if(id <= this->last_block_id()) {+ while(curr_index < curr_block->size() && this->id() < id) {+ curr_index++;}- skipped_block = true;+ return ;}- if(skipped_block) {- curr_index = 0;+ // identify the block where the id could exist and skip to that+ reset_cache();++ const auto it = id_block_map->lower_bound(id);+ if(it == id_block_map->end()) {+ return;}- while(curr_block != end_block && curr_index < curr_block->size() && this->id() < id) {+ curr_block = it->second;+ curr_index = 0;+ ids = curr_block->ids.uncompress();+ offset_index = curr_block->offset_index.uncompress();+ offsets = curr_block->offsets.uncompress();++ while(curr_index < curr_block->size() && this->id() < id) {curr_index++;}++ if(curr_index == curr_block->size()) {+ reset_cache();+ }}posting_list_t::iterator_t::~iterator_t() {if(auto_destroy) {- destroy();+ reset_cache();}}-void posting_list_t::iterator_t::destroy() {+void posting_list_t::iterator_t::reset_cache() {delete [] ids;- ids = nullptr;-delete [] offsets;- offsets = nullptr;-delete [] offset_index;- offset_index = nullptr;++ ids = offset_index = offsets = nullptr;+ curr_index = 0;+ curr_block = end_block = nullptr;}posting_list_t::iterator_t::iterator_t(iterator_t&& rhs) noexcept {+ id_block_map = rhs.id_block_map;curr_block = rhs.curr_block;curr_index = rhs.curr_index;end_block = rhs.end_block;@@ -1494,6 +1495,7 @@ posting_list_t::iterator_t::iterator_t(iterator_t&& rhs) noexcept {auto_destroy = rhs.auto_destroy;field_id = rhs.field_id;+ rhs.id_block_map = nullptr;rhs.curr_block = nullptr;rhs.end_block = nullptr;rhs.ids = nullptr;@@ -1502,6 +1504,7 @@ posting_list_t::iterator_t::iterator_t(iterator_t&& rhs) noexcept {}posting_list_t::iterator_t& posting_list_t::iterator_t::operator=(posting_list_t::iterator_t&& rhs) noexcept {+ id_block_map = rhs.id_block_map;curr_block = rhs.curr_block;curr_index = rhs.curr_index;end_block = rhs.end_block;@@ -1511,6 +1514,7 @@ posting_list_t::iterator_t& posting_list_t::iterator_t::operator=(posting_list_tauto_destroy = rhs.auto_destroy;field_id = rhs.field_id;+ rhs.id_block_map = nullptr;rhs.curr_block = nullptr;rhs.end_block = nullptr;rhs.ids = nullptr;@@ -1525,7 +1529,8 @@ void posting_list_t::iterator_t::set_index(uint32_t index) {}posting_list_t::iterator_t posting_list_t::iterator_t::clone() const {- posting_list_t::iterator_t it(nullptr, nullptr);+ posting_list_t::iterator_t it(nullptr, nullptr, nullptr);+ it.id_block_map = id_block_map;it.curr_block = curr_block;it.curr_index = curr_index;it.end_block = end_block;diff --git a/test/or_iterator_test.cpp b/test/or_iterator_test.cppindex 19c949b0..4472da61 100644--- a/test/or_iterator_test.cpp+++ b/test/or_iterator_test.cpp@@ -58,7 +58,6 @@ TEST(OrIteratorTest, IntersectTwoListsWith3SubLists) {std::vector<uint32_t> results;or_iterator_t::intersect(or_its, istate, [&results](uint32_t id, std::vector<or_iterator_t>& its) {- LOG(INFO) << "id: " << id;results.push_back(id);});@@ -69,6 +68,14 @@ TEST(OrIteratorTest, IntersectTwoListsWith3SubLists) {for(size_t i = 0; i < expected_results.size(); i++) {ASSERT_EQ(expected_results[i], results[i]);}++ for(auto p: postings1) {+ delete p;+ }++ for(auto p: postings2) {+ delete p;+ }}TEST(OrIteratorTest, IntersectTwoListsWith4SubLists) {@@ -138,4 +145,12 @@ TEST(OrIteratorTest, IntersectTwoListsWith4SubLists) {for(size_t i = 0; i < expected_results.size(); i++) {ASSERT_EQ(expected_results[i], results[i]);}++ for(auto p: postings1) {+ delete p;+ }++ for(auto p: postings2) {+ delete p;+ }}\ No newline at end of filediff --git a/test/posting_list_test.cpp b/test/posting_list_test.cppindex 7c41028b..55495152 100644--- a/test/posting_list_test.cpp+++ b/test/posting_list_test.cpp@@ -600,69 +600,6 @@ TEST_F(PostingListTest, MergeBasics) {}}-TEST_F(PostingListTest, SplittingOfListsSimple) {- std::vector<uint32_t> offsets = {0, 1, 3};-- // [0, 2] [3, 20]- // [1, 3], [5, 10], [20]- // [2, 3], [5, 7], [20]-- posting_list_t p1(2);- p1.upsert(0, offsets);- p1.upsert(2, offsets);- p1.upsert(3, offsets);- p1.upsert(20, offsets);-- posting_list_t p2(2);- p2.upsert(1, offsets);- p2.upsert(3, offsets);- p2.upsert(5, offsets);- p2.upsert(10, offsets);- p2.upsert(20, offsets);-- posting_list_t p3(2);- p3.upsert(2, offsets);- p3.upsert(3, offsets);- p3.upsert(5, offsets);- p3.upsert(7, offsets);- p3.upsert(20, offsets);-- std::vector<void*> raw_lists = {&p1, &p2, &p3};-- std::vector<posting_list_t::iterator_t> its;- result_iter_state_t iter_state;- posting_t::block_intersector_t intersector(raw_lists, iter_state, pool);-- std::vector<std::vector<posting_list_t::iterator_t>> partial_its_vec(4);- intersector.split_lists(4, partial_its_vec);-- std::vector<std::vector<std::vector<uint32_t>>> split_ids = {- {{0, 2}, {1, 3}, {2, 3}},- {{3, 20}, {1, 3, 5, 10, 20}, {2, 3, 5, 7, 20}}- };-- ASSERT_EQ(4, partial_its_vec.size());- ASSERT_EQ(3, partial_its_vec[0].size());- ASSERT_EQ(3, partial_its_vec[1].size());-- for(size_t i = 0; i < partial_its_vec.size(); i++) {- auto& partial_its = partial_its_vec[i];- for (size_t j = 0; j < partial_its.size(); j++) {- auto& it = partial_its[j];- size_t k = 0;-- while (it.valid()) {- //LOG(INFO) << it.id();- ASSERT_EQ(split_ids[i][j][k], it.id());- k++;- it.next();- }-- //LOG(INFO) << "---";- }- }-}-TEST_F(PostingListTest, IntersectionBasics) {std::vector<uint32_t> offsets = {0, 1, 3};