Optimise queries search for a chain of OR strings (#3250)

This is a performance enhancement motivated by users who are generating queries with many string comparisons on a single column, for example from cocoa's "IN" queries. The idea is to combine string equality conditions from a single "OR" query node and store them in an unordered_set. With N elements to search, and C conditions, the runtime changes from O(N*C) to O(N). The added benchmark goes from 30 seconds to 2 seconds. This change does not try to optimise indexed columns which should be running O(log(N)*C). The benchmark with indexes turned on runs in 3.5 seconds. Since N is likely the dominant term, using indexes should still be fastest in practice when compared to this optimisation.
realm · Mar 5, 2019 · 993363c · 993363c
1 parent e119c5d
commit 993363c
Show file tree

Hide file tree

Showing 6 changed files with 261 additions and 17 deletions.
diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -13,7 +13,9 @@
 -----------
 
 ### Internals
-* None.
+* Optimised queries for unindexed string columns when the query has a chain of OR conditions.
+  This will improve performance of "IN" queries generated by bindings.
+  ([#22](https://github.com/realm/engineering/issues/22).
 
 ----------------------------------------------
 

diff --git a/src/realm/query_engine.cpp b/src/realm/query_engine.cpp
@@ -19,6 +19,7 @@
 #include <realm/query_engine.hpp>
 
 #include <realm/query_expression.hpp>
+#include <realm/utilities.hpp>
 
 using namespace realm;
 
@@ -325,8 +326,75 @@ void StringNode<Equal>::_search_index_init()
     }
 }
 
+void StringNode<Equal>::consume_condition(StringNode<Equal>* other)
+{
+    // If a search index is present, don't try to combine conditions since index search is most likely faster.
+    // Assuming N elements to search and M conditions to check:
+    // 1) search index present:                     O(log(N)*M)
+    // 2) no search index, combine conditions:      O(N)
+    // 3) no search index, conditions not combined: O(N*M)
+    // In practice N is much larger than M, so if we have a search index, choose 1, otherwise if possible choose 2.
+    REALM_ASSERT(m_condition_column == other->m_condition_column);
+    REALM_ASSERT(other->m_needles.empty());
+    if (m_needles.empty()) {
+        m_needles.insert(bool(m_value) ? StringData(*m_value) : StringData());
+    }
+    if (bool(other->m_value)) {
+        m_needle_storage.push_back(StringBuffer());
+        m_needle_storage.back().append(*other->m_value);
+        m_needles.insert(StringData(m_needle_storage.back().data(), m_needle_storage.back().size()));
+    }
+    else {
+        m_needles.insert(StringData());
+    }
+}
+
+// Requirements of template types:
+// ArrayType must support: size() -> size_t, and get(size_t) -> ElementType
+// ElementType must be convertable to a StringData via data() and size()
+template<class ArrayType, class ElementType>
+size_t StringNode<Equal>::find_first_in(ArrayType& array, size_t begin, size_t end)
+{
+    if (m_needles.empty())
+        return not_found;
+
+    size_t n = array.size();
+    if (end == npos)
+        end = n;
+    REALM_ASSERT_7(begin, <=, n, &&, end, <=, n);
+    REALM_ASSERT_3(begin, <=, end);
+
+    const auto not_in_set = m_needles.end();
+    // For a small number of conditions it is faster to cycle through
+    // and check them individually. The threshold depends on how fast
+    // our hashing of StringData is (see `StringData.hash()`). The
+    // number 20 was found empirically when testing small strings
+    // with N==100k
+    if (m_needles.size() < 20) {
+        for (size_t i = begin; i < end; ++i) {
+            ElementType element = array.get(i);
+            StringData value_2{element.data(), element.size()};
+            for (auto it = m_needles.begin(); it != not_in_set; ++it) {
+                if (*it == value_2)
+                return i;
+            }
+        }
+    }
+    else {
+        for (size_t i = begin; i < end; ++i) {
+            ElementType element = array.get(i);
+            StringData value_2{element.data(), element.size()};
+            if (m_needles.find(value_2) != not_in_set)
+                return i;
+        }
+    }
+
+    return not_found;
+}
+
 size_t StringNode<Equal>::_find_first_local(size_t start, size_t end)
 {
+    const bool multi_target_search = !m_needles.empty();
     // Normal string column, with long or short leaf
     for (size_t s = start; s < end; ++s) {
         const StringColumn* asc = static_cast<const StringColumn*>(m_condition_column);
@@ -345,14 +413,23 @@ size_t StringNode<Equal>::_find_first_local(size_t start, size_t end)
         }
         size_t end2 = (end > m_leaf_end ? m_leaf_end - m_leaf_start : end - m_leaf_start);
 
-        if (m_leaf_type == StringColumn::leaf_type_Small)
-            s = static_cast<const ArrayString&>(*m_leaf).find_first(m_value, s - m_leaf_start, end2);
-        else if (m_leaf_type == StringColumn::leaf_type_Medium)
-            s = static_cast<const ArrayStringLong&>(*m_leaf).find_first(m_value, s - m_leaf_start, end2);
-        else
-            s = static_cast<const ArrayBigBlobs&>(*m_leaf).find_first(str_to_bin(m_value), true, s - m_leaf_start,
-                                                                      end2);
-
+        if (multi_target_search) {
+            if (m_leaf_type == StringColumn::leaf_type_Small)
+                s = find_first_in<const ArrayString&, StringData>(static_cast<const ArrayString&>(*m_leaf), s - m_leaf_start, end2);
+            else if (m_leaf_type == StringColumn::leaf_type_Medium)
+                s = find_first_in<const ArrayStringLong&, StringData>(static_cast<const ArrayStringLong&>(*m_leaf), s - m_leaf_start, end2);
+            else
+                s = find_first_in<const ArrayBigBlobs&, BinaryData>(static_cast<const ArrayBigBlobs&>(*m_leaf), s - m_leaf_start, end2);
+        }
+        else {
+            if (m_leaf_type == StringColumn::leaf_type_Small)
+                s = static_cast<const ArrayString&>(*m_leaf).find_first(m_value, s - m_leaf_start, end2);
+            else if (m_leaf_type == StringColumn::leaf_type_Medium)
+                s = static_cast<const ArrayStringLong&>(*m_leaf).find_first(m_value, s - m_leaf_start, end2);
+            else
+                s = static_cast<const ArrayBigBlobs&>(*m_leaf).find_first(str_to_bin(m_value), true, s - m_leaf_start,
+                                                                          end2);
+        }
         if (s == not_found)
             s = m_leaf_end - 1;
         else
@@ -362,6 +439,30 @@ size_t StringNode<Equal>::_find_first_local(size_t start, size_t end)
     return not_found;
 }
 
+std::string StringNode<Equal>::describe(util::serializer::SerialisationState& state) const
+{
+    if (m_needles.empty()) {
+        return StringNodeEqualBase::describe(state);
+    }
+
+    // FIXME: once the parser supports it, print something like "column IN {s1, s2, s3}"
+    REALM_ASSERT(m_condition_column != nullptr);
+    std::string desc;
+    bool is_first = true;
+    for (auto it : m_needles) {
+        StringData sd(it.data(), it.size());
+        desc += (is_first ? "" : " or ")
+            + state.describe_column(ParentNode::m_table, m_condition_column->get_column_index())
+            + " " + Equal::description() + " " + util::serializer::print_value(sd);
+        is_first = false;
+    }
+    if (!is_first) {
+        desc = "(" + desc + ")";
+    }
+    return desc;
+}
+
+
 void StringNode<EqualIns>::_search_index_init()
 {
     if (m_column_type == col_type_StringEnum) {

diff --git a/src/realm/query_engine.hpp b/src/realm/query_engine.hpp
@@ -113,9 +113,11 @@ AggregateState      State of the aggregate - contains a state variable that stor
 #include <realm/util/miscellaneous.hpp>
 #include <realm/util/serializer.hpp>
 #include <realm/util/shared_ptr.hpp>
+#include <realm/util/string_buffer.hpp>
 #include <realm/utilities.hpp>
 
 #include <map>
+#include <unordered_set>
 
 #if REALM_X86_OR_X64_TRUE && defined(_MSC_FULL_VER) && _MSC_FULL_VER >= 160040219
 #include <immintrin.h>
@@ -1184,6 +1186,11 @@ class StringNodeBase : public ParentNode {
         do_verify_column(m_condition_column);
     }
 
+    bool has_search_index() const
+    {
+        return m_condition_column->has_search_index();
+    }
+
     void init() override
     {
         ParentNode::init();
@@ -1301,7 +1308,6 @@ class StringNode : public StringNodeBase {
         StringNodeBase::init();
     }
 
-
     size_t find_first_local(size_t start, size_t end) override
     {
         TConditionFunction cond;
@@ -1537,9 +1543,9 @@ class StringNodeEqualBase : public StringNodeBase {
     size_t m_last_start;
 };
 
-
 // Specialization for Equal condition on Strings - we specialize because we can utilize indexes (if they exist) for
-// Equal.
+// Equal. This specialisation also supports combining other StringNode<Equal> conditions into itself in order to
+// optimise the non-indexed linear search that can be happen when many conditions are OR'd together in an "IN" query.
 // Future optimization: make specialization for greater, notequal, etc
 template <>
 class StringNode<Equal> : public StringNodeEqualBase {
@@ -1548,13 +1554,22 @@ class StringNode<Equal> : public StringNodeEqualBase {
 
     void _search_index_init() override;
 
+    void consume_condition(StringNode<Equal>* other);
+
     std::unique_ptr<ParentNode> clone(QueryNodeHandoverPatches* patches) const override
     {
         return std::unique_ptr<ParentNode>(new StringNode<Equal>(*this, patches));
     }
 
+    virtual std::string describe(util::serializer::SerialisationState& state) const override;
+
 private:
+    template<class ArrayType, class ElementType>
+    size_t find_first_in(ArrayType& array, size_t begin, size_t end);
+
     size_t _find_first_local(size_t start, size_t end) override;
+    std::unordered_set<StringData> m_needles;
+    std::vector<StringBuffer> m_needle_storage;
 };
 
 
@@ -1603,7 +1618,6 @@ class StringNode<EqualIns> : public StringNodeEqualBase {
     size_t _find_first_local(size_t start, size_t end) override;
 };
 
-
 // OR node contains at least two node pointers: Two or more conditions to OR
 // together in m_conditions, and the next AND condition (if any) in m_child.
 //
@@ -1641,11 +1655,9 @@ class OrNode : public ParentNode {
             condition->verify_column();
         }
     }
+
     std::string describe(util::serializer::SerialisationState& state) const override
     {
-        if (m_conditions.size() >= 2) {
-
-        }
         std::string s;
         for (size_t i = 0; i < m_conditions.size(); ++i) {
             if (m_conditions[i]) {
@@ -1661,13 +1673,37 @@ class OrNode : public ParentNode {
         return s;
     }
 
-
     void init() override
     {
         ParentNode::init();
 
         m_dD = 10.0;
 
+        StringNode<Equal>* first = nullptr;
+        std::sort(m_conditions.begin(), m_conditions.end(),
+                  [](auto& a, auto& b) { return a->m_condition_column_idx < b->m_condition_column_idx; });
+        auto it = m_conditions.begin();
+        while (it != m_conditions.end()) {
+            // Only try to optimize on StringNode<Equal> conditions without search index
+            if ((first = dynamic_cast<StringNode<Equal>*>(it->get())) && !first->has_search_index()) {
+                auto col_ndx = first->m_condition_column_idx;
+                auto next = it + 1;
+                while (next != m_conditions.end() && (*next)->m_condition_column_idx == col_ndx) {
+                    if (auto advance = dynamic_cast<StringNode<Equal>*>(next->get())) {
+                        first->consume_condition(advance);
+                        next = m_conditions.erase(next);
+                    }
+                    else {
+                        ++next;
+                    }
+                }
+                it = next;
+            }
+            else {
+                ++it;
+            }
+        }
+
         m_start.clear();
         m_start.resize(m_conditions.size(), 0);
 

diff --git a/test/benchmark-common-tasks/main.cpp b/test/benchmark-common-tasks/main.cpp
@@ -366,6 +366,51 @@ struct BenchmarkQuery : BenchmarkWithStrings {
     }
 };
 
+struct BenchmarkQueryChainedOrStrings : BenchmarkWithStringsTable {
+    const size_t num_queried_matches = 1000;
+    const size_t num_rows = 100000;
+    std::vector<std::string> values_to_query;
+    const char* name() const
+    {
+        return "QueryChainedOrStrings";
+    }
+
+    void before_all(SharedGroup& group)
+    {
+        BenchmarkWithStringsTable::before_all(group);
+        WriteTransaction tr(group);
+        TableRef t = tr.get_table("StringOnly");
+        t->add_empty_row(num_rows);
+        REALM_ASSERT(num_rows > num_queried_matches);
+        Random r;
+        for (size_t i = 0; i < num_rows; ++i) {
+            std::stringstream ss;
+            ss << i;
+            auto s = ss.str();
+            t->set_string(0, i, s);
+        }
+        //t->add_search_index(0);
+        for (size_t i = 0; i < num_queried_matches; ++i) {
+            size_t ndx_to_match = (num_rows / num_queried_matches) * i;
+            values_to_query.push_back(t->get_string(0, ndx_to_match));
+        }
+        tr.commit();
+    }
+
+    void operator()(SharedGroup& group)
+    {
+        ReadTransaction tr(group);
+        ConstTableRef table = tr.get_table("StringOnly");
+        Query query = table->where();
+        for (size_t i = 0; i < values_to_query.size(); ++i) {
+            query.Or().equal(0, values_to_query[i]);
+        }
+        TableView results = query.find_all();
+        REALM_ASSERT_EX(results.size() == num_queried_matches, results.size(), num_queried_matches, values_to_query.size());
+        static_cast<void>(results);
+    }
+};
+
 struct BenchmarkSize : BenchmarkWithStrings {
     const char* name() const
     {
@@ -1003,6 +1048,7 @@ int benchmark_common_tasks_main()
     BENCH(BenchmarkQueryInsensitiveString);
     BENCH(BenchmarkQueryInsensitiveStringIndexed);
     BENCH(BenchmarkNonInitatorOpen);
+    BENCH(BenchmarkQueryChainedOrStrings);
 
 #undef BENCH
     return 0;

diff --git a/test/test_parser.cpp b/test/test_parser.cpp
@@ -2730,4 +2730,54 @@ TEST(Parser_Between)
     CHECK(message.find("Invalid Predicate. The 'between' operator is not supported yet, please rewrite the expression using '>' and '<'.") != std::string::npos);
 }
 
+TEST(Parser_ChainedStringEqualQueries)
+{
+    Group g;
+    TableRef table = g.add_table("table");
+    size_t a_col_ndx = table->add_column(type_String, "a", false);
+    size_t b_col_ndx = table->add_column(type_String, "b", true);
+    size_t c_col_ndx = table->add_column(type_String, "c", false);
+    size_t d_col_ndx = table->add_column(type_String, "d", true);
+
+    table->add_search_index(c_col_ndx);
+    table->add_search_index(d_col_ndx);
+
+    table->add_empty_row(100);
+    std::vector<std::string> populated_data;
+    std::stringstream ss;
+    for (size_t i = 0; i < table->size(); ++i) {
+        ss.str({});
+        ss << i;
+        std::string sd (ss.str());
+        populated_data.push_back(sd);
+        table->set_string(a_col_ndx, i, sd);
+        table->set_string(b_col_ndx, i, sd);
+        table->set_string(c_col_ndx, i, sd);
+        table->set_string(d_col_ndx, i, sd);
+    }
+    table->add_empty_row(); // one null/empty string
+
+    verify_query(test_context, table, "a == '0' or a == '1' or a == '2'", 3);
+    verify_query(test_context, table, "a == '0' or b == '2' or a == '3' or b == '4'", 4);
+    verify_query(test_context, table, "(a == '0' or b == '2' or a == '3' or b == '4') and (c == '0' or d == '2' or c == '3' or d == '4')", 4);
+    verify_query(test_context, table, "a == '' or a == null", 1);
+    verify_query(test_context, table, "b == '' or b == null", 1);
+    verify_query(test_context, table, "c == '' or c == null", 1);
+    verify_query(test_context, table, "d == '' or d == null", 1);
+    verify_query(test_context, table, "(a == null or a == '') and (b == null or b == '') and (c == null or c == '') and (d == null or d == '')", 1);
+
+    Random rd;
+    rd.shuffle(populated_data.begin(), populated_data.end());
+    std::string query;
+    bool first = true;
+    char column_to_query = 0;
+    for (auto s : populated_data) {
+        std::string column_name(1, 'a' + column_to_query);
+        query += (first ? "" : " or " ) + column_name + " == '" + s + "'";
+        first = false;
+        column_to_query = (column_to_query + 1) % 4;
+    }
+    verify_query(test_context, table, query, populated_data.size());
+}
+
 #endif // TEST_PARSER