Merge pull request #30921 from vespa-engine/geirst/in-operator-cost-model-readjustment

Re-adjust cost model params for IN/weightedSet/dotProduct.

4 files changed, 43 insertions, 22 deletions

diff --git a/searchlib/src/tests/queryeval/iterator_benchmark/iterator_benchmark_test.cpp b/searchlib/src/tests/queryeval/iterator_benchmark/iterator_benchmark_test.cpp
index f82b9804720..590cf1a26f0 100644
--- a/searchlib/src/tests/queryeval/iterator_benchmark/iterator_benchmark_test.cpp
+++ b/searchlib/src/tests/queryeval/iterator_benchmark/iterator_benchmark_test.cpp
@@ -934,6 +934,13 @@ TEST(IteratorBenchmark, analyze_and_with_filter_vs_in)
                       num_docs);
 }
 
+TEST(IteratorBenchmark, analyze_and_with_filter_vs_in_array)
+{
+    run_and_benchmark({int32_fs, QueryOperator::Term, gen_ratios(0.1, 8.0, 15)},
+                      {int32_array_fs, QueryOperator::In, {0.1}, 100, false},
+                      num_docs);
+}
+
 TEST(IteratorBenchmark, analyze_and_with_filter_vs_or)
 {
     run_and_benchmark({int32_fs, QueryOperator::Term, gen_ratios(0.1, 8.0, 15)},
diff --git a/searchlib/src/vespa/searchlib/attribute/direct_multi_term_blueprint.h b/searchlib/src/vespa/searchlib/attribute/direct_multi_term_blueprint.h
index 413d0dd0bf4..51321a56885 100644
--- a/searchlib/src/vespa/searchlib/attribute/direct_multi_term_blueprint.h
+++ b/searchlib/src/vespa/searchlib/attribute/direct_multi_term_blueprint.h
@@ -76,28 +76,7 @@ public:
         resolve_strict(in_flow);
     }
 
-    queryeval::FlowStats calculate_flow_stats(uint32_t docid_limit) const override {
-        using OrFlow = search::queryeval::OrFlow;
-        struct MyAdapter {
-            uint32_t docid_limit;
-            MyAdapter(uint32_t docid_limit_in) noexcept : docid_limit(docid_limit_in) {}
-            double estimate(const IDirectPostingStore::LookupResult &term) const noexcept {
-                return abs_to_rel_est(term.posting_size, docid_limit);
-            }
-            double cost(const IDirectPostingStore::LookupResult &) const noexcept {
-                return search::queryeval::flow::btree_cost();
-            }
-            double strict_cost(const IDirectPostingStore::LookupResult &term) const noexcept {
-                double rel_est = abs_to_rel_est(term.posting_size, docid_limit);
-                return search::queryeval::flow::btree_strict_cost(rel_est);
-            }
-        };
-        double est = OrFlow::estimate_of(MyAdapter(docid_limit), _terms);
-        // Iterator benchmarking has shown that non-strict cost should be 1.0.
-        // Program: searchlib/src/tests/queryeval/iterator_benchmark
-        // TODO: Add more details, and consider moving constant to flow_tuning.h
-        return {est, 1.0, OrFlow::cost_of(MyAdapter(docid_limit), _terms, true) + queryeval::flow::heap_cost(est, _terms.size())};
-    }
+    queryeval::FlowStats calculate_flow_stats(uint32_t docid_limit) const override;
 
     std::unique_ptr<queryeval::SearchIterator> createLeafSearch(const fef::TermFieldMatchDataArray &tfmda) const override;
 
diff --git a/searchlib/src/vespa/searchlib/attribute/direct_multi_term_blueprint.hpp b/searchlib/src/vespa/searchlib/attribute/direct_multi_term_blueprint.hpp
index 817eab3e070..e506ec55c76 100644
--- a/searchlib/src/vespa/searchlib/attribute/direct_multi_term_blueprint.hpp
+++ b/searchlib/src/vespa/searchlib/attribute/direct_multi_term_blueprint.hpp
@@ -181,4 +181,34 @@ DirectMultiTermBlueprint<PostingStoreType, SearchType>::createFilterSearch(Filte
     return wrapper;
 }
 
+template <typename PostingStoreType, typename SearchType>
+queryeval::FlowStats
+DirectMultiTermBlueprint<PostingStoreType, SearchType>::calculate_flow_stats(uint32_t docid_limit) const
+{
+    using OrFlow = search::queryeval::OrFlow;
+    struct MyAdapter {
+        uint32_t docid_limit;
+        MyAdapter(uint32_t docid_limit_in) noexcept : docid_limit(docid_limit_in) {}
+        double estimate(const IDirectPostingStore::LookupResult &term) const noexcept {
+            return abs_to_rel_est(term.posting_size, docid_limit);
+        }
+        double cost(const IDirectPostingStore::LookupResult &) const noexcept {
+            return search::queryeval::flow::btree_cost();
+        }
+        double strict_cost(const IDirectPostingStore::LookupResult &term) const noexcept {
+            double rel_est = abs_to_rel_est(term.posting_size, docid_limit);
+            return search::queryeval::flow::btree_strict_cost(rel_est);
+        }
+    };
+    double est = OrFlow::estimate_of(MyAdapter(docid_limit), _terms);
+    // Iterator benchmarking has shown that non-strict cost is different for attributes
+    // that support using a reverse hash filter (see use_hash_filter()).
+    // Program used: searchlib/src/tests/queryeval/iterator_benchmark
+    // Tests: analyze_and_with_filter_vs_in(), analyze_and_with_filter_vs_in_array()
+    double non_strict_cost = (SearchType::supports_hash_filter && !_iattr.hasMultiValue())
+            ? queryeval::flow::reverse_hash_lookup() :
+              OrFlow::cost_of(MyAdapter(docid_limit), _terms, false);
+    return {est, non_strict_cost, OrFlow::cost_of(MyAdapter(docid_limit), _terms, true) + queryeval::flow::heap_cost(est, _terms.size())};
+}
+
 }
diff --git a/searchlib/src/vespa/searchlib/queryeval/flow_tuning.h b/searchlib/src/vespa/searchlib/queryeval/flow_tuning.h
index 51e544b2e30..c4df2d99508 100644
--- a/searchlib/src/vespa/searchlib/queryeval/flow_tuning.h
+++ b/searchlib/src/vespa/searchlib/queryeval/flow_tuning.h
@@ -30,6 +30,11 @@ inline double lookup_cost(size_t num_indirections) {
     return 1.0 + (num_indirections * 1.0);
 }
 
+// Non-strict cost of reverse lookup into a hash table (containing terms from a multi-term operator).
+inline double reverse_hash_lookup() {
+    return 1.0;
+}
+
 // Strict cost of lookup based matching in an attribute (not fast-search).
 inline double lookup_strict_cost(size_t num_indirections) {
     return lookup_cost(num_indirections);