fix(flags): sort evaluation contexts in the Rust cache builder

products/feature_flags/backend/test/test_flags_cache.py

@@ -1178,7 +1178,7 @@
 _REPO_ROOT = Path(__file__).resolve().parents[4]
 
 
 def _extract_schema(obj: Any) -> Any:
     """Extract a type skeleton from a JSON-like object for structural comparison.
 
     Returns a nested structure of dicts, lists, and leaf type names.
@@ -1225,7 +1225,7 @@
     return type(obj).__name__
 
 
 def _compare_schemas(fixture_schema: Any, result_schema: Any, path: str = "") -> list[str]:
     """Compare two type skeletons and return human-readable diffs."""
     diffs: list[str] = []
     # If the fixture declares a field as null, the serializer must also return null.
@@ -1317,7 +1317,7 @@
         super().setUp()
         clear_flags_cache(self.team, kinds=["redis", "s3"])
 
     def test_serializer_output_matches_fixture_schema(self):
         """Golden fixture validation: Python serializer output must match the fixture schema.
 
         Recursively compares the type skeleton (keys + value types at every nesting
@@ -2437,14 +2437,19 @@
             created_by=self.user,
             filters={"groups": [{"properties": [], "rollout_percentage": 100}]},
         )
-        ctx = EvaluationContext.objects.create(team=self.team, name="original-context-name")
+        # Out of alphabetical order on purpose. The assertions below also cover the sort that
+        # ArrayAgg(distinct=True) applies, which the Rust builder matches with
+        # ARRAY_AGG(DISTINCT ...). Remove distinct=True and the two write different arrays.
+        ctx = EvaluationContext.objects.create(team=self.team, name="zulu-context")
+        other_ctx = EvaluationContext.objects.create(team=self.team, name="alpha-context")
         FeatureFlagEvaluationContext.objects.create(feature_flag=flag, evaluation_context=ctx)
+        FeatureFlagEvaluationContext.objects.create(feature_flag=flag, evaluation_context=other_ctx)
 
         # Warm the cache
         update_flags_cache(self.team)
 
         # Rename the context directly in DB (bypassing signals to simulate stale cache)
-        EvaluationContext.objects.filter(id=ctx.id).update(name="renamed-context-name")
+        EvaluationContext.objects.filter(id=ctx.id).update(name="mike-context")
 
         # Verify should detect the mismatch
         result = verify_team_flags(self.team, verbose=True)
@@ -2456,8 +2461,8 @@
 
         field_diffs = result["diffs"][0]["field_diffs"]
         eval_tag_diff = next(d for d in field_diffs if d["field"] == "evaluation_contexts")
-        self.assertEqual(eval_tag_diff["cached_value"], ["original-context-name"])
-        self.assertEqual(eval_tag_diff["db_value"], ["renamed-context-name"])
+        self.assertEqual(eval_tag_diff["cached_value"], ["alpha-context", "zulu-context"])
+        self.assertEqual(eval_tag_diff["db_value"], ["alpha-context", "mike-context"])
 
         # Mismatch result should include db_data for cache fix optimization
         self.assertIn("db_data", result)

rust/feature-flags/src/flags/feature_flag_list.rs

@@ -146,8 +146,15 @@ impl FeatureFlagList {
                   f.ensure_experience_continuity,
                   f.version,
                   f.evaluation_runtime,
+                  -- DISTINCT makes Postgres sort the names before it aggregates them, which
+                  -- matches ArrayAgg(..., distinct=True) on the two Python paths that write
+                  -- this cache entry: products/feature_flags/backend/flags_cache.py (batch) and
+                  -- products/feature_flags/backend/models/feature_flag.py (single team, which is
+                  -- also what the cache verifier compares against). Without DISTINCT the array
+                  -- order follows the join, so the sides disagree for a flag with two or more
+                  -- contexts. Keep all three in lockstep.
                   COALESCE(
-                      ARRAY_AGG(ctx.name) FILTER (WHERE ctx.name IS NOT NULL),
+                      ARRAY_AGG(DISTINCT ctx.name) FILTER (WHERE ctx.name IS NOT NULL),
                       '{}'::text[]
                   ) AS evaluation_tags,
                   bucketing_identifier,
@@ -504,15 +511,13 @@ mod tests {
         assert_eq!(flag.key, "flag1");
         assert_eq!(flag.team_id, team.id);
 
-        // Check that evaluation tags were properly fetched
+        // ARRAY_AGG(DISTINCT ...) sorts, so this is not the insertion order.
         let tags = flag
             .evaluation_tags
             .as_ref()
             .expect("Should have evaluation tags");
-        assert_eq!(tags.len(), 3);
-        assert!(tags.contains(&"docs-page".to_string()));
-        assert!(tags.contains(&"marketing-site".to_string()));
-        assert!(tags.contains(&"app".to_string()));
+        let tag_names: Vec<&str> = tags.iter().map(String::as_str).collect();
+        assert_eq!(tag_names, ["app", "docs-page", "marketing-site"]);
     }
 
     #[tokio::test]