fix(eap): Use notEmpty for exists() on the ai_conversation_id column

Author: pbhandari

snuba/protos/common.py

@@ -191,6 +191,12 @@ def _generate_subscriptable_reference(
     "attributes_array_bool",
 )
 
+# Normalized String columns whose unset value ingests as an empty string (not NULL) — e.g.
+# ai_conversation_id when a TraceItem has no conversation_id. Because the column is never NULL,
+# existence must be checked with notEmpty(...) instead of isNotNull(...), which would otherwise
+# match every row (see get_field_existence_expression).
+EMPTY_STRING_DEFAULT_COLUMNS: tuple[str, ...] = ("ai_conversation_id",)
+
 
 def type_array_typed_columns_select_expressions(attr_key: AttributeKey) -> list[FunctionCall]:
     """Native ``arrayElement`` read per typed array map column, for a deprecated untyped

snuba/web/rpc/common/common.py

@@ -17,6 +17,7 @@
 from snuba.protos.common import (
     ARRAY_TYPES,
     ATTRIBUTES_TO_COALESCE,
+    EMPTY_STRING_DEFAULT_COLUMNS,
     NORMALIZED_COLUMNS_EAP_ITEMS,
     PROTO_TYPE_TO_ATTRIBUTE_COLUMN,
     PROTO_TYPE_TO_CLICKHOUSE_TYPE,
@@ -1256,6 +1257,7 @@ def get_subscriptable_field(field: Expression) -> SubscriptableReference | None:
         # is the right existence check. Scalar map lookups still use map_key_exists.
         if isinstance(base, Column) and base.column_name in TYPED_ARRAY_MAP_COLUMNS:
             return f.notEmpty(field)
+
         return map_key_exists(field.parameters[0], field.parameters[1])
 
     if isinstance(field, FunctionCall) and field.function_name in ("arrayMap", "arrayConcat"):
@@ -1265,4 +1267,12 @@ def get_subscriptable_field(field: Expression) -> SubscriptableReference | None:
         # type_array_to_membership_array_expression_from_typed_columns).
         return f.notEmpty(field)
 
+    if isinstance(field, FunctionCall) and field.function_name == "cast":
+        # A normalized String column with an empty-string default (e.g. ai_conversation_id)
+        # is never NULL, so isNotNull would always be true. notEmpty distinguishes an unset
+        # (empty) value from a real one.
+        base = field.parameters[0]
+        if isinstance(base, Column) and base.column_name in EMPTY_STRING_DEFAULT_COLUMNS:
+            return f.notEmpty(field)
+
     return f.isNotNull(field)

tests/web/rpc/v1/test_indexed_columns.py

@@ -23,6 +23,7 @@
 )
 from sentry_protos.snuba.v1.trace_item_filter_pb2 import (
     ComparisonFilter,
+    ExistsFilter,
     TraceItemFilter,
 )
 from sentry_protos.snuba.v1.trace_item_pb2 import AnyValue, TraceItem
@@ -83,6 +84,11 @@ def setup(self, eap: None, redis_db: None) -> None:
         messages = [
             _item_message(self.base_time, SESSION_A, CONV_A, "red"),
             _item_message(self.base_time + timedelta(minutes=1), SESSION_B, CONV_B, "blue"),
+            # "green" sets neither: an unset conversation_id ingests as an empty
+            # ai_conversation_id, so exists on gen_ai.conversation.id must exclude it.
+            # (session_id is intentionally not exercised for exists: an unset session_id
+            # ingests as a random non-nil UUID, so absence is indistinguishable there.)
+            _item_message(self.base_time + timedelta(minutes=2), "", "", "green"),
         ]
         write_raw_unprocessed_events(get_writable_storage(StorageKey("eap_items")), messages)
 
@@ -178,3 +184,17 @@ def test_filter_uses_skip_index_and_matches(
         )
         assert colors == expected_colors
         assert self._explain_uses_index(sql, expected_index), sql
+
+    def test_exists_on_conversation_id_excludes_rows_without_a_value(self) -> None:
+        # gen_ai.conversation.id is optional: the "green" span never set it, so its
+        # ai_conversation_id is empty. exists() must treat empty as absent and exclude
+        # it, rather than matching every row (a non-nullable column is never NULL, so a
+        # plain isNotNull check would always be true).
+        _, colors = self._execute(
+            TraceItemFilter(
+                exists_filter=ExistsFilter(
+                    key=AttributeKey(type=AttributeKey.TYPE_STRING, name="gen_ai.conversation.id")
+                )
+            )
+        )
+        assert colors == ["blue", "red"]