Optimize ExistsQuery for a high number of dynamic columns (#2694)

* Optimize ExistsQuery for a high number of dynamic columns The previous algorithm checked _each_ doc in _each_ column for existence. This causes huge cost on JSON fields with e.g. 100k columns. Compute a bitset instead if we have more than one column. add `iter_docs` to the multivalued_index * add benchmark subfields=1 exists_json_union Memory: 89.3 KB (+2.01%) Avg: 0.4865ms (-26.03%) Median: 0.4865ms (-26.03%) [0.4865ms .. 0.4865ms] subfields=2 exists_json_union Memory: 68.1 KB Avg: 1.7048ms (-0.46%) Median: 1.7048ms (-0.46%) [1.7048ms .. 1.7048ms] subfields=3 exists_json_union Memory: 61.8 KB Avg: 2.0742ms (-2.22%) Median: 2.0742ms (-2.22%) [2.0742ms .. 2.0742ms] subfields=4 exists_json_union Memory: 119.8 KB (+103.44%) Avg: 3.9500ms (+42.62%) Median: 3.9500ms (+42.62%) [3.9500ms .. 3.9500ms] subfields=5 exists_json_union Memory: 120.4 KB (+107.65%) Avg: 3.9610ms (+20.65%) Median: 3.9610ms (+20.65%) [3.9610ms .. 3.9610ms] subfields=6 exists_json_union Memory: 120.6 KB (+107.49%) Avg: 3.8903ms (+3.11%) Median: 3.8903ms (+3.11%) [3.8903ms .. 3.8903ms] subfields=7 exists_json_union Memory: 120.9 KB (+106.93%) Avg: 3.6220ms (-16.22%) Median: 3.6220ms (-16.22%) [3.6220ms .. 3.6220ms] subfields=8 exists_json_union Memory: 121.3 KB (+106.23%) Avg: 4.0981ms (-15.97%) Median: 4.0981ms (-15.97%) [4.0981ms .. 4.0981ms] subfields=16 exists_json_union Memory: 123.1 KB (+103.09%) Avg: 4.3483ms (-92.26%) Median: 4.3483ms (-92.26%) [4.3483ms .. 4.3483ms] subfields=256 exists_json_union Memory: 204.6 KB (+19.85%) Avg: 3.8874ms (-99.01%) Median: 3.8874ms (-99.01%) [3.8874ms .. 3.8874ms] subfields=4096 exists_json_union Memory: 2.0 MB Avg: 3.5571ms (-99.90%) Median: 3.5571ms (-99.90%) [3.5571ms .. 3.5571ms] subfields=65536 exists_json_union Memory: 28.3 MB Avg: 14.4417ms (-99.97%) Median: 14.4417ms (-99.97%) [14.4417ms .. 14.4417ms] subfields=262144 exists_json_union Memory: 113.3 MB Avg: 66.2860ms (-99.95%) Median: 66.2860ms (-99.95%) [66.2860ms .. 66.2860ms] * rename methods
2025-12-23 02:29:57 +00:00 · 2025-09-16 18:21:03 +02:00
parent 7963b0b4aa
commit 203751f2fe
8 changed files with 195 additions and 15 deletions
--- a/benches/exists_json.rs
+++ b/benches/exists_json.rs
@@ -0,0 +1,69 @@
+use binggan::plugins::PeakMemAllocPlugin;
+use binggan::{black_box, InputGroup, PeakMemAlloc, INSTRUMENTED_SYSTEM};
+use serde_json::json;
+use tantivy::collector::Count;
+use tantivy::query::ExistsQuery;
+use tantivy::schema::{Schema, FAST, TEXT};
+use tantivy::{doc, Index};
+
+#[global_allocator]
+pub static GLOBAL: &PeakMemAlloc<std::alloc::System> = &INSTRUMENTED_SYSTEM;
+
+fn main() {
+    let doc_count: usize = 500_000;
+    let subfield_counts: &[usize] = &[1, 2, 3, 4, 5, 6, 7, 8, 16, 256, 4096, 65536, 262144];
+
+    let indices: Vec<(String, Index)> = subfield_counts
+        .iter()
+        .map(|&sub_fields| {
+            (
+                format!("subfields={sub_fields}"),
+                build_index_with_json_subfields(doc_count, sub_fields),
+            )
+        })
+        .collect();
+
+    let mut group = InputGroup::new_with_inputs(indices);
+    group.add_plugin(PeakMemAllocPlugin::new(GLOBAL));
+
+    group.config().num_iter_group = Some(1);
+    group.config().num_iter_bench = Some(1);
+    group.register("exists_json", exists_json_union);
+
+    group.run();
+}
+
+fn exists_json_union(index: &Index) {
+    let reader = index.reader().expect("reader");
+    let searcher = reader.searcher();
+    let query = ExistsQuery::new("json".to_string(), true);
+    let count = searcher.search(&query, &Count).expect("exists search");
+    // Prevents optimizer from eliding the search
+    black_box(count);
+}
+
+fn build_index_with_json_subfields(num_docs: usize, num_subfields: usize) -> Index {
+    // Schema: single JSON field stored as FAST to support ExistsQuery.
+    let mut schema_builder = Schema::builder();
+    let json_field = schema_builder.add_json_field("json", TEXT | FAST);
+    let schema = schema_builder.build();
+
+    let index = Index::create_from_tempdir(schema).expect("create index");
+    {
+        let mut index_writer = index
+            .writer_with_num_threads(1, 200_000_000)
+            .expect("writer");
+        for i in 0..num_docs {
+            let sub = i % num_subfields;
+            // Only one subpath set per document; rotate subpaths so that
+            // no single subpath is full, but the union covers all docs.
+            let v = json!({ format!("field_{sub}"): i as u64 });
+            index_writer
+                .add_document(doc!(json_field => v))
+                .expect("add_document");
+        }
+        index_writer.commit().expect("commit");
+    }
+
+    index
+}