feat(parquet): RFC 0005 §3.6 — bloom filters on trace_id and span_id

crates/ourios-parquet/src/writer.rs

@@ -30,7 +30,9 @@
 //!   columns (`tenant_id`, `attributes`, `resource_attributes`,
 //!   `body`, both `params` list-element leaves,
 //!   `separators.list.element`).
-//! - Bloom filter on `template_id` (B2 predicate-pushdown).
+//! - Bloom filters on `template_id` (B2 predicate-pushdown),
+//!   `trace_id` / `span_id` (RFC 0031 L3 exact-id lookup — random
+//!   ids defeat min/max statistics), and every promoted column.
 //!
 //! [`CLAUDE.md`]: ../../../../CLAUDE.md
 
@@ -806,6 +808,17 @@ fn writer_properties(zstd: ZstdLevel, promoted: &PromotedAttributes) -> WriterPr
     let template_id = ColumnPath::new(vec![crate::columns::TEMPLATE_ID.to_string()]);
     builder = builder.set_column_bloom_filter_enabled(template_id, true);
 
+    // Bloom filters on the trace-context ids: random 16/8-byte values
+    // defeat min/max statistics entirely, so an exact-id lookup (the
+    // RFC 0031 L3 class) degenerates to a whole-column scan without
+    // them — measured at 72.4 MB for a 9-row trace on the 4.9M-record
+    // otel-demo-v8 corpus (comparative run #12). Same §3.6 pattern as
+    // `template_id` and the promoted columns.
+    for column in [crate::columns::TRACE_ID, crate::columns::SPAN_ID] {
+        let path = ColumnPath::new(vec![column.to_string()]);
+        builder = builder.set_column_bloom_filter_enabled(path, true);
+    }
+
     // RFC 0022 §3.1: promoted attribute columns are the attribute
     // predicate-pushdown surface — bloom filter each (dictionary and
     // page-level statistics are already the global defaults). A

crates/ourios-parquet/tests/it/main.rs

@@ -24,4 +24,5 @@ mod round_trip;
 mod row_vs_path_validation;
 mod schema_pin;
 mod sizing;
+mod trace_bloom;
 mod zstd_level;

crates/ourios-parquet/tests/it/trace_bloom.rs

@@ -0,0 +1,72 @@
+//! RFC 0005 §3.6 — bloom filters on the trace-context id columns.
+//!
+//! Random 16/8-byte ids defeat min/max statistics entirely, so an
+//! exact-id lookup (the RFC 0031 L3 class) degenerates to a
+//! whole-column scan without blooms — measured at 72.4 MB for a 9-row
+//! trace on the 4.9M-record otel-demo-v8 corpus (comparative run #12).
+//! This pins the writer emitting blooms for `trace_id` and `span_id`,
+//! alongside the pre-existing `template_id` one.
+
+use ourios_core::audit::ParamType;
+use ourios_core::record::{BodyKind, MinedRecord, Param};
+use ourios_core::tenant::TenantId;
+use ourios_parquet::{DEFAULT_ZSTD_LEVEL, columns, encode_records_to_parquet};
+use parquet::file::reader::{FileReader, SerializedFileReader};
+
+const TS0: u64 = 1_775_127_480_000_000_000;
+
+fn rec(trace_id: Option<[u8; 16]>, span_id: Option<[u8; 8]>) -> MinedRecord {
+    MinedRecord {
+        tenant_id: TenantId::new("a"),
+        template_id: 1,
+        template_version: 1,
+        severity_number: 9,
+        severity_text: Some("INFO".to_string()),
+        scope_name: None,
+        scope_version: None,
+        scope_attributes: Vec::new(),
+        resource_schema_url: None,
+        scope_schema_url: None,
+        time_unix_nano: TS0,
+        observed_time_unix_nano: Some(TS0 + 1_000),
+        attributes: Vec::new(),
+        dropped_attributes_count: 0,
+        resource_attributes: Vec::new(),
+        trace_id,
+        span_id,
+        flags: 0x01,
+        event_name: None,
+        body_kind: BodyKind::String,
+        params: vec![Param {
+            type_tag: ParamType::Num,
+            value: "42".to_string(),
+        }],
+        separators: vec![String::new(), " ".to_string()],
+        body: None,
+        confidence: 1.0,
+        lossy_flag: false,
+    }
+}
+
+#[test]
+fn trace_context_columns_carry_bloom_filters() {
+    let records = [
+        rec(Some([0xAB; 16]), Some([0xCD; 8])),
+        rec(Some([0x11; 16]), Some([0x22; 8])),
+        rec(None, None),
+    ];
+    let bytes = encode_records_to_parquet(&records, DEFAULT_ZSTD_LEVEL).expect("encode");
+    let reader = SerializedFileReader::new(bytes::Bytes::from(bytes)).expect("footer");
+    let rg = reader.metadata().row_group(0);
+
+    for name in [columns::TRACE_ID, columns::SPAN_ID, columns::TEMPLATE_ID] {
+        let col = (0..rg.num_columns())
+            .map(|i| rg.column(i))
+            .find(|c| c.column_path().string() == name)
+            .expect("column chunk present");
+        assert!(
+            col.bloom_filter_offset().is_some(),
+            "{name}: bloom filter written",
+        );
+    }
+}