Skip to content

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

Option to curtail expensive queries.

Kishore Nallan6210e823907c
diff --git a/include/collection.h b/include/collection.h
index 2b3baba4..e1942d37 100644
--- a/include/collection.h
+++ b/include/collection.h
@@ -386,7 +386,8 @@ public:
bool pre_segmented_query=false,
bool enable_overrides=true,
const std::string& highlight_fields="",
- const bool exhaustive_search = false) const;
+ const bool exhaustive_search = false,
+ size_t search_stop_millis = 200) const;
Option<bool> get_filter_ids(const std::string & simple_filter_query,
std::vector<std::pair<size_t, uint32_t*>>& index_ids);
diff --git a/include/index.h b/include/index.h
index 6af2c749..8d8171b2 100644
--- a/include/index.h
+++ b/include/index.h
@@ -260,6 +260,7 @@ struct search_args {
bool exhaustive_search;
size_t concurrency;
const std::vector<const override_t*>& filter_overrides;
+ size_t search_cutoff_ms;
spp::sparse_hash_set<uint64_t> groups_processed;
std::vector<std::vector<art_leaf*>> searched_queries;
@@ -281,7 +282,8 @@ struct search_args {
bool prioritize_exact_match,
bool exhaustive_search,
size_t concurrency,
- const std::vector<const override_t*>& dynamic_overrides):
+ const std::vector<const override_t*>& dynamic_overrides,
+ size_t search_cutoff_ms):
field_query_tokens(field_query_tokens),
search_fields(search_fields), filters(filters), facets(facets),
included_ids(included_ids), excluded_ids(excluded_ids), sort_fields_std(sort_fields_std),
@@ -291,7 +293,7 @@ struct search_args {
group_by_fields(group_by_fields), group_limit(group_limit), default_sorting_field(default_sorting_field),
prioritize_exact_match(prioritize_exact_match), all_result_ids_len(0),
exhaustive_search(exhaustive_search), concurrency(concurrency),
- filter_overrides(dynamic_overrides) {
+ filter_overrides(dynamic_overrides), search_cutoff_ms(search_cutoff_ms) {
const size_t topster_size = std::max((size_t)1, max_hits); // needs to be atleast 1 since scoring is mandatory
topster = new Topster(topster_size, group_limit);
@@ -650,7 +652,8 @@ public:
const std::string& default_sorting_field,
bool prioritize_exact_match,
bool exhaustive_search,
- size_t concurrency) const;
+ size_t concurrency,
+ size_t search_cutoff_ms) const;
Option<uint32_t> remove(const uint32_t seq_id, const nlohmann::json & document, const bool is_update);
diff --git a/include/posting.h b/include/posting.h
index 682135b3..85ab60e9 100644
--- a/include/posting.h
+++ b/include/posting.h
@@ -22,7 +22,7 @@ struct compact_posting_list_t {
static compact_posting_list_t* create(uint32_t num_ids, const uint32_t* ids, const uint32_t* offset_index,
uint32_t num_offsets, uint32_t* offsets);
- posting_list_t* to_full_posting_list();
+ posting_list_t* to_full_posting_list() const;
bool contains(uint32_t id);
diff --git a/include/thread_local_vars.h b/include/thread_local_vars.h
index 37845ac5..0b5f024d 100644
--- a/include/thread_local_vars.h
+++ b/include/thread_local_vars.h
@@ -1 +1,6 @@
-extern thread_local int64_t write_log_index;
\ No newline at end of file
+#include <chrono>
+
+extern thread_local int64_t write_log_index;
+extern thread_local std::chrono::high_resolution_clock::time_point begin;
+extern thread_local int64_t search_stop_ms;
+extern thread_local bool search_cutoff;
\ No newline at end of file
diff --git a/src/collection.cpp b/src/collection.cpp
index c92da6ff..020b02d8 100644
--- a/src/collection.cpp
+++ b/src/collection.cpp
@@ -493,7 +493,8 @@ Option<nlohmann::json> Collection::search(const std::string & raw_query, const s
bool pre_segmented_query,
bool enable_overrides,
const std::string& highlight_fields,
- const bool exhaustive_search) const {
+ const bool exhaustive_search,
+ const size_t search_stop_millis) const {
std::shared_lock lock(mutex);
@@ -916,7 +917,8 @@ Option<nlohmann::json> Collection::search(const std::string & raw_query, const s
per_page, page, token_order, prefixes,
drop_tokens_threshold, typo_tokens_threshold,
group_by_fields, group_limit, default_sorting_field, prioritize_exact_match,
- exhaustive_search, 4, filter_overrides);
+ exhaustive_search, 4, filter_overrides,
+ search_stop_millis);
index->run_search(search_params);
@@ -1251,6 +1253,8 @@ Option<nlohmann::json> Collection::search(const std::string & raw_query, const s
// free search params
delete search_params;
+ result["search_cutoff"] = search_cutoff;
+
result["request_params"] = nlohmann::json::object();;
result["request_params"]["collection_name"] = name;
result["request_params"]["per_page"] = per_page;
diff --git a/src/collection_manager.cpp b/src/collection_manager.cpp
index 2935d2bf..0034c71b 100644
--- a/src/collection_manager.cpp
+++ b/src/collection_manager.cpp
@@ -535,6 +535,7 @@ Option<bool> CollectionManager::do_search(std::map<std::string, std::string>& re
const char *PRE_SEGMENTED_QUERY = "pre_segmented_query";
const char *EXHAUSTIVE_SEARCH = "exhaustive_search";
+ const char *SEARCH_CUTOFF_MS = "search_cutoff_ms";
if(req_params.count(NUM_TYPOS) == 0) {
req_params[NUM_TYPOS] = "2";
@@ -637,6 +638,10 @@ Option<bool> CollectionManager::do_search(std::map<std::string, std::string>& re
req_params[EXHAUSTIVE_SEARCH] = "false";
}
+ if(req_params.count(SEARCH_CUTOFF_MS) == 0) {
+ req_params[SEARCH_CUTOFF_MS] = "2000";
+ }
+
std::vector<std::string> query_by_weights_str;
std::vector<size_t> query_by_weights;
@@ -706,6 +711,10 @@ Option<bool> CollectionManager::do_search(std::map<std::string, std::string>& re
return Option<bool>(400,"Parameter `" + std::string(GROUP_LIMIT) + "` must be an unsigned integer.");
}
+ if(!StringUtils::is_uint32_t(req_params[SEARCH_CUTOFF_MS])) {
+ return Option<bool>(400,"Parameter `" + std::string(SEARCH_CUTOFF_MS) + "` must be an unsigned integer.");
+ }
+
bool prioritize_exact_match = (req_params[PRIORITIZE_EXACT_MATCH] == "true");
bool pre_segmented_query = (req_params[PRE_SEGMENTED_QUERY] == "true");
bool exhaustive_search = (req_params[EXHAUSTIVE_SEARCH] == "true");
@@ -813,7 +822,8 @@ Option<bool> CollectionManager::do_search(std::map<std::string, std::string>& re
pre_segmented_query,
enable_overrides,
req_params[HIGHLIGHT_FIELDS],
- exhaustive_search
+ exhaustive_search,
+ static_cast<size_t>(std::stol(req_params[SEARCH_CUTOFF_MS]))
);
uint64_t timeMillis = std::chrono::duration_cast<std::chrono::milliseconds>(
diff --git a/src/index.cpp b/src/index.cpp
index 1c61730f..12e871e6 100644
--- a/src/index.cpp
+++ b/src/index.cpp
@@ -17,8 +17,21 @@
#include <s2/s2loop.h>
#include <s2/s2builder.h>
#include <posting.h>
+#include <thread_local_vars.h>
#include "logger.h"
+#define RETURN_CIRCUIT_BREAKER if(std::chrono::duration_cast<std::chrono::milliseconds>(\
+ std::chrono::high_resolution_clock::now() - begin).count() > search_stop_ms) { \
+ search_cutoff = true; \
+ return ;\
+ }
+
+#define BREAK_CIRCUIT_BREAKER if(std::chrono::duration_cast<std::chrono::milliseconds>(\
+ std::chrono::high_resolution_clock::now() - begin).count() > search_stop_ms) { \
+ search_cutoff = true; \
+ break;\
+ }
+
spp::sparse_hash_map<uint32_t, int64_t> Index::text_match_sentinel_value;
spp::sparse_hash_map<uint32_t, int64_t> Index::seq_id_sentinel_value;
spp::sparse_hash_map<uint32_t, int64_t> Index::geo_sentinel_value;
@@ -1113,6 +1126,8 @@ void Index::search_candidates(const uint8_t & field_id, bool field_is_array,
size_t combination_limit = exhaustive_search ? Index::COMBINATION_MAX_LIMIT : Index::COMBINATION_MIN_LIMIT;
for(long long n=0; n<N && n<combination_limit; ++n) {
+ RETURN_CIRCUIT_BREAKER
+
// every element in `query_suggestion` contains a token and its associated hits
std::vector<art_leaf*> query_suggestion(token_candidates_vec.size());
@@ -1232,6 +1247,8 @@ void Index::do_filtering(uint32_t*& filter_ids, uint32_t& filter_ids_length,
for(size_t i = 0; i < filters.size(); i++) {
const filter & a_filter = filters[i];
+ RETURN_CIRCUIT_BREAKER
+
if(a_filter.field_name == "id") {
// we handle `ids` separately
std::vector<uint32> result_ids;
@@ -1616,7 +1633,8 @@ void Index::run_search(search_args* search_params) {
search_params->default_sorting_field,
search_params->prioritize_exact_match,
search_params->exhaustive_search,
- search_params->concurrency);
+ search_params->concurrency,
+ search_params->search_cutoff_ms);
}
void Index::collate_included_ids(const std::vector<std::string>& q_included_tokens,
@@ -2002,9 +2020,11 @@ void Index::search(std::vector<query_tokens_t>& field_query_tokens,
const std::string& default_sorting_field,
bool prioritize_exact_match,
const bool exhaustive_search,
- const size_t concurrency) const {
+ const size_t concurrency,
+ const size_t search_cutoff_ms) const {
- //auto begin = std::chrono::high_resolution_clock::now();
+ begin = std::chrono::high_resolution_clock::now();
+ search_stop_ms = search_cutoff_ms;
// process the filters
@@ -2520,12 +2540,19 @@ void Index::search_wildcard(const std::vector<std::string>& qtokens, const std::
uint32_t token_bits = 255;
std::vector<posting_list_t::iterator_t> plists;
+ bool check_for_circuit_break = (filter_ids_length > 1000000);
+
for(size_t i = 0; i < filter_ids_length; i++) {
const uint32_t seq_id = filter_ids[i];
score_results(sort_fields_std, (uint16_t) searched_queries.size(), field_id, false, 0, topster, {},
groups_processed, seq_id, sort_order, field_values,
geopoint_indices, group_limit, group_by_fields, token_bits,
false, false, plists);
+
+ if(check_for_circuit_break && i % (2^17) == 0) {
+ // check only once every 2^17 docs to reduce overhead
+ BREAK_CIRCUIT_BREAKER
+ }
}
collate_included_ids(qtokens, field, field_id, included_ids_map, curated_topster, searched_queries);
@@ -2637,6 +2664,8 @@ void Index::search_field(const uint8_t & field_id,
const size_t num_fuzzy_candidates = exhaustive_search ? 10000 : 4;
while(n < N && n < combination_limit) {
+ RETURN_CIRCUIT_BREAKER
+
// Outerloop generates combinations of [cost to max_cost] for each token
// For e.g. for a 3-token query: [0, 0, 0], [0, 0, 1], [0, 1, 1] etc.
std::vector<uint32_t> costs(token_to_costs.size());
diff --git a/src/posting.cpp b/src/posting.cpp
index 6b2ef148..33158b56 100644
--- a/src/posting.cpp
+++ b/src/posting.cpp
@@ -155,7 +155,7 @@ compact_posting_list_t* compact_posting_list_t::create(uint32_t num_ids, const u
return pl;
}
-posting_list_t* compact_posting_list_t::to_full_posting_list() {
+posting_list_t* compact_posting_list_t::to_full_posting_list() const {
posting_list_t* pl = new posting_list_t(256);
size_t i = 0;
diff --git a/src/thread_local_vars.cpp b/src/thread_local_vars.cpp
index 79ba502b..816a898e 100644
--- a/src/thread_local_vars.cpp
+++ b/src/thread_local_vars.cpp
@@ -2,3 +2,6 @@
#include "thread_local_vars.h"
thread_local int64_t write_log_index = 0;
+thread_local std::chrono::high_resolution_clock::time_point begin;
+thread_local int64_t search_stop_ms;
+thread_local bool search_cutoff = false;
diff --git a/test/collection_synonyms_test.cpp b/test/collection_synonyms_test.cpp
index efe69fdd..b8cb8ace 100644
--- a/test/collection_synonyms_test.cpp
+++ b/test/collection_synonyms_test.cpp
@@ -526,5 +526,47 @@ TEST_F(CollectionSynonymsTest, SynonymJsonSerialization) {
ASSERT_STREQ("ipod", obj["synonyms"][0].get<std::string>().c_str());
ASSERT_STREQ("i pod", obj["synonyms"][1].get<std::string>().c_str());
ASSERT_STREQ("pod", obj["synonyms"][2].get<std::string>().c_str());
+}
+
+TEST_F(CollectionSynonymsTest, SynonymSingleTokenExactMatch) {
+ Collection *coll1;
+ std::vector<field> fields = {field("title", field_types::STRING, false),
+ field("description", field_types::STRING, false),
+ field("points", field_types::INT32, false),};
+
+ coll1 = collectionManager.get_collection("coll1").get();
+ if(coll1 == nullptr) {
+ coll1 = collectionManager.create_collection("coll1", 1, fields, "points").get();
+ }
+
+ std::vector<std::vector<std::string>> records = {
+ {"Smashed Lemon", "Description 1", "100"},
+ {"Lulu Guinness", "Description 2", "100"},
+ {"Lululemon", "Description 3", "100"},
+ };
+
+ for(size_t i=0; i<records.size(); i++) {
+ nlohmann::json doc;
+
+ doc["id"] = std::to_string(i);
+ doc["title"] = records[i][0];
+ doc["description"] = records[i][1];
+ doc["points"] = std::stoi(records[i][2]);
+
+ ASSERT_TRUE(coll1->add(doc.dump()).ok());
+ }
+
+ synonym_t synonym1{"syn-1", {"lulu", "lemon"}, {{"lululemon"}}};
+ coll1->add_synonym(synonym1);
+
+ auto res = coll1->search("lulu lemon", {"title"}, "", {}, {}, {2}, 10, 1, FREQUENCY, {true}, 1).get();
+
+ ASSERT_EQ(2, res["hits"].size());
+ ASSERT_EQ(2, res["found"].get<uint32_t>());
+
+ ASSERT_STREQ("2", res["hits"][0]["document"]["id"].get<std::string>().c_str());
+ ASSERT_STREQ("1", res["hits"][1]["document"]["id"].get<std::string>().c_str());
+
+ collectionManager.drop_collection("coll1");
}
\ No newline at end of file