perf: GIN index for trace attribute filtering

server/migrations/0011_spans_attrs_gin.sql

@@ -0,0 +1,10 @@
+-- GIN index on span attributes so the traces list can filter by attribute
+-- (attributes @> '{"key":"value"}') without seq-scanning the spans table.
+-- Mirrors logs_attrs_gin; jsonb_path_ops is compact and serves the @> operator.
+--
+-- SET LOCAL statement_timeout = 0 for this migration only: the pool sets a 60s
+-- statement_timeout, but building a GIN over the existing spans can take longer,
+-- and we don't want the build aborted (which would crash-loop startup). The build
+-- holds a write lock on spans, so ingest pauses briefly once while it runs.
+SET LOCAL statement_timeout = 0;
+CREATE INDEX IF NOT EXISTS spans_attrs_gin ON spans USING gin (attributes jsonb_path_ops);

server/src/api.rs

@@ -60,8 +60,10 @@ pub async fn list_traces(
         .and_then(|s| s.split_once('='))
         .filter(|(k, _)| !k.is_empty())
         .map(|(k, v)| serde_json::json!({ k: v }));
-    // service + time bound the spans scanned (index-friendly); the trace-level
-    // filters (name / attribute / errors / duration) are applied with HAVING so
+    // service + time bound the spans scanned (index-friendly). The attribute
+    // filter selects whole traces that contain a matching span via a subquery the
+    // spans_attrs_gin index can serve (a HAVING bool_or couldn't use the index).
+    // The remaining trace-level filters (name / errors / duration) are HAVING, so
     // the per-trace aggregates stay computed over the whole trace.
     let rows = sqlx::query_as::<_, TraceSummary>(
         "SELECT trace_id,
@@ -77,10 +79,11 @@ pub async fn list_traces(
          WHERE ($1::text IS NULL OR service = $1)
            AND ($2::timestamptz IS NULL OR start_time >= $2)
            AND ($3::timestamptz IS NULL OR start_time <= $3)
+           AND ($6::jsonb IS NULL
+                OR trace_id IN (SELECT trace_id FROM spans WHERE attributes @> $6))
          GROUP BY trace_id
          HAVING ($5::text IS NULL
                  OR (array_agg(name ORDER BY start_time))[1] ILIKE '%' || $5 || '%')
-            AND ($6::jsonb IS NULL OR bool_or(attributes @> $6))
             AND (NOT $7::bool OR count(*) FILTER (WHERE status_code = 2) > 0)
             AND ($8::float8 IS NULL
                  OR extract(epoch FROM (max(end_time) - min(start_time))) * 1000.0 >= $8)