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
Parallelize value faceting.
Kishore Nallan0c095fbc8823
diff --git a/src/index.cpp b/src/index.cppindex 86b52a5e..6ef2fa70 100644--- a/src/index.cpp+++ b/src/index.cpp@@ -3533,7 +3533,7 @@ Option<bool> Index::search(std::vector<query_tokens_t>& field_query_tokens, consmax_candidates, facet_infos, facet_index_type);std::vector<std::vector<facet>> facet_batches(num_threads);- std::vector<facet> value_facets;+ std::vector<std::vector<facet>> value_facets;for(size_t i = 0; i < facets.size(); i++) {const auto& this_facet = facets[i];@@ -3543,7 +3543,8 @@ Option<bool> Index::search(std::vector<query_tokens_t>& field_query_tokens, consif(facet_infos[i].use_value_index) {#endif// value based faceting on a single thread- value_facets.emplace_back(this_facet.field_name, this_facet.orig_index, this_facet.facet_range_map,+ value_facets.emplace_back();+ value_facets.back().emplace_back(this_facet.field_name, this_facet.orig_index, this_facet.facet_range_map,this_facet.is_range_query, this_facet.is_sort_by_alpha,this_facet.sort_order, this_facet.sort_field);continue;@@ -3617,17 +3618,47 @@ Option<bool> Index::search(std::vector<query_tokens_t>& field_query_tokens, cons}}- // do value based faceting which must be single thread on the entire resultset (without batching)- do_facets(value_facets, facet_query, estimate_facets, facet_sample_percent,- facet_infos, group_limit, group_by_fields, group_missing_values,- all_result_ids, all_result_ids_len, max_facet_values,- is_wildcard_no_filter_query, facet_index_type);+ lock_process.unlock();+ num_processed = 0;+ num_queued = 0;- for(auto& this_facet : value_facets) {- auto& acc_facet = facets[this_facet.orig_index];- aggregate_facet(group_limit, this_facet, acc_facet);+ // do value based faceting field-wise parallel but on the entire result set+ for(size_t thread_id = 0; thread_id < value_facets.size(); thread_id++) {+ num_queued++;++ thread_pool->enqueue([this, thread_id, &facets, &value_facets, &facet_query, group_limit, group_by_fields,+ all_result_ids, all_result_ids_len, &facet_infos, max_facet_values,+ is_wildcard_no_filter_query, no_filters_provided, estimate_facets,+ facet_sample_percent, group_missing_values,+ &parent_search_begin, &parent_search_stop_ms, &parent_search_cutoff,+ &num_processed, &m_process, &cv_process, facet_index_type]() {+ search_begin_us = parent_search_begin;+ search_stop_us = parent_search_stop_ms;+ search_cutoff = false;++ auto fq = facet_query;++ do_facets({value_facets[thread_id]}, fq, estimate_facets, facet_sample_percent,+ facet_infos, group_limit, group_by_fields, group_missing_values,+ all_result_ids, all_result_ids_len, max_facet_values,+ is_wildcard_no_filter_query, facet_index_type);++ std::unique_lock<std::mutex> lock(m_process);++ auto& this_facet = value_facets[thread_id][0];+ auto& acc_facet = facets[this_facet.orig_index];+ aggregate_facet(group_limit, this_facet, acc_facet);++ num_processed++;+ parent_search_cutoff = parent_search_cutoff || search_cutoff;+ cv_process.notify_one();+ });}+ lock_process.lock();+ cv_process.wait(lock_process, [&](){ return num_processed == num_queued; });+ search_cutoff = parent_search_cutoff;+for(auto & acc_facet: facets) {for(auto& facet_kv: acc_facet.result_map) {if(group_limit) {diff --git a/test/collection_optimized_faceting_test.cpp b/test/collection_optimized_faceting_test.cppindex 023c9956..634335a1 100644--- a/test/collection_optimized_faceting_test.cpp+++ b/test/collection_optimized_faceting_test.cpp@@ -850,6 +850,46 @@ TEST_F(CollectionOptimizedFacetingTest, FacetCountOnSimilarStrings) {collectionManager.drop_collection("coll1");}+TEST_F(CollectionOptimizedFacetingTest, ConcurrentValueFacetingOnMulFields) {+ Collection *coll1;++ std::vector<field> fields = {field("c1", field_types::STRING, true),+ field("c2", field_types::STRING, true),+ field("c3", field_types::STRING, true),+ field("c4", field_types::STRING, true),+ field("points", field_types::INT32, true)};++ std::vector<sort_by> sort_fields = {sort_by("points", "DESC")};++ coll1 = collectionManager.get_collection("coll1").get();+ if (coll1 == nullptr) {+ coll1 = collectionManager.create_collection("coll1", 4, fields, "points").get();+ }++ for(size_t i = 0; i < 1000; i++) {+ nlohmann::json doc;+ doc["c1"] = "c1_" + std::to_string(i % 40);+ doc["c2"] = "c2_" + std::to_string(i % 40);+ doc["c3"] = "c3_" + std::to_string(i % 40);+ doc["c4"] = "c4_" + std::to_string(i % 40);+ doc["points"] = 25;+ coll1->add(doc.dump());+ }++ std::vector<std::string> facets = {"c1", "c2", "c3", "c4"};++ nlohmann::json results = coll1->search("*", {}, "points:[25, 50]", facets, sort_fields, {0}, 10, 1,+ token_ordering::FREQUENCY, {true}, 10, spp::sparse_hash_set<std::string>(),+ spp::sparse_hash_set<std::string>(), 10, "", 30UL, 4UL,+ "", 1UL, "", "", {}, 3UL, "<mark>", "</mark>", {},+ 4294967295UL, true, false, true, "", false, 6000000UL, 4UL,+ 7UL, fallback, 4UL, {off}, 32767UL, 32767UL, 2UL, 2UL, false,+ "", true, 0UL, max_score, 100UL, 0UL, 4294967295UL, VALUE).get();++ ASSERT_EQ(4, results["facet_counts"].size());+ collectionManager.drop_collection("coll1");+}+TEST_F(CollectionOptimizedFacetingTest, FacetByNestedIntField) {nlohmann::json schema = R"({"name": "coll1",