increase limit, introduce randomness

Author: nora-shap

src/sentry/seer/sentry_data_models.py

@@ -30,7 +30,7 @@ class TraceData(BaseModel):
     spans: list[Span]
 
 
-class EvidenceTraceData(BaseModel):  # hate this name
+class EvidenceTraceData(BaseModel):
     trace_id: str
     transaction_name: str
 

src/sentry/tasks/llm_issue_detection/detection.py

@@ -31,6 +31,7 @@
 SEER_TIMEOUT_S = 120
 SEER_RETRIES = 1
 START_TIME_DELTA_MINUTES = 30
+TRANSACTION_BATCH_SIZE = 100
 
 
 seer_issue_detection_connection_pool = connection_from_url(
@@ -208,8 +209,8 @@ def detect_llm_issues_for_project(project_id: int) -> None:
     """
     Process a single project for LLM issue detection.
 
-    Gets the project's top 50 transaction spans from the last START_TIME_DELTA_MINUTES, sorted by -sum(span.duration).
-    From the 50 longest transactions, dedupes on normalized transaction_name.
+    Gets the project's top TRANSACTION_BATCH_SIZE transaction spans from the last START_TIME_DELTA_MINUTES, sorted by -sum(span.duration).
+    From those transactions, dedupes on normalized transaction_name.
     For each deduped transaction, gets first trace_id from the start of time window, which has small random variation.
     Sends these trace_ids to seer, which uses get_trace_waterfall to construct an EAPTrace to analyze.
     """
@@ -224,7 +225,7 @@ def detect_llm_issues_for_project(project_id: int) -> None:
         return
 
     evidence_traces = get_project_top_transaction_traces_for_llm_detection(
-        project_id, limit=100, start_time_delta_minutes=START_TIME_DELTA_MINUTES
+        project_id, limit=TRANSACTION_BATCH_SIZE, start_time_delta_minutes=START_TIME_DELTA_MINUTES
     )
     if not evidence_traces:
         return

src/sentry/tasks/llm_issue_detection/trace_data.py

@@ -33,22 +33,27 @@ def get_project_top_transaction_traces_for_llm_detection(
         logger.exception("Project does not exist", extra={"project_id": project_id})
         return []
 
-    random_offset = random.randint(1, 8)
     end_time = datetime.now(UTC)
     start_time = end_time - timedelta(minutes=start_time_delta_minutes)
-
-    # use for both queries to ensure they are searching the same time window
-    snuba_params = SnubaParams(
-        start=start_time,
-        end=end_time,
-        projects=[project],
-        organization=project.organization,
-    )
     config = SearchResolverConfig(auto_fields=True)
 
-    # Step 1: Get top transactions by total time in time window
+    def _build_snuba_params(start: datetime) -> SnubaParams:
+        """
+        Both queries have different start times and the same end time.
+        """
+        return SnubaParams(
+            start=start,
+            end=end_time,
+            projects=[project],
+            organization=project.organization,
+        )
+
+    transaction_snuba_params = _build_snuba_params(start_time)
+    random_offset = random.randint(1, 8)
+    trace_snuba_params = _build_snuba_params(start_time + timedelta(minutes=random_offset))
+
     transactions_result = Spans.run_table_query(
-        params=snuba_params,
+        params=transaction_snuba_params,
         query_string="is_transaction:true",
         selected_columns=[
             "transaction",
@@ -74,10 +79,9 @@ def get_project_top_transaction_traces_for_llm_detection(
         if normalized_name in seen_names:
             continue
 
-        # Step 2: Get ONE trace for this transaction from THE SAME time window
         escaped_transaction_name = UNESCAPED_QUOTE_RE.sub('\\"', transaction_name)
         trace_result = Spans.run_table_query(
-            params=snuba_params,
+            params=trace_snuba_params,
             query_string=f'is_transaction:true transaction:"{escaped_transaction_name}"',
             selected_columns=["trace", "precise.start_ts"],
             orderby=["precise.start_ts"],  # First trace in the window

tests/sentry/tasks/test_llm_issue_detection.py

@@ -9,7 +9,10 @@
     detect_llm_issues_for_project,
     run_llm_issue_detection,
 )
-from sentry.tasks.llm_issue_detection.detection import START_TIME_DELTA_MINUTES
+from sentry.tasks.llm_issue_detection.detection import (
+    START_TIME_DELTA_MINUTES,
+    TRANSACTION_BATCH_SIZE,
+)
 from sentry.tasks.llm_issue_detection.trace_data import (
     get_project_top_transaction_traces_for_llm_detection,
 )
@@ -48,7 +51,9 @@ def test_detect_llm_issues_no_transactions(self, mock_get_transactions, mock_see
         detect_llm_issues_for_project(self.project.id)
 
         mock_get_transactions.assert_called_once_with(
-            self.project.id, limit=100, start_time_delta_minutes=START_TIME_DELTA_MINUTES
+            self.project.id,
+            limit=TRANSACTION_BATCH_SIZE,
+            start_time_delta_minutes=START_TIME_DELTA_MINUTES,
         )
         mock_seer_request.assert_not_called()
 
@@ -308,7 +313,7 @@ def test_returns_deduped_transaction_traces(self) -> None:
         self.store_spans([span1, span2, span3], is_eap=True)
 
         evidence_traces = get_project_top_transaction_traces_for_llm_detection(
-            self.project.id, limit=50, start_time_delta_minutes=30
+            self.project.id, limit=TRANSACTION_BATCH_SIZE, start_time_delta_minutes=30
         )
 
         assert len(evidence_traces) == 2