fix(signals): evict stale db connections in temporal activities

[posthog]

Problem

A new OperationalError: the connection is closed exception surfaced in the signals semantic-search Temporal activity. The traceback lands inside run_signal_semantic_search_activity at products/signals/backend/temporal/signal_queries.py, where await Team.objects.aget(pk=input.team_id) calls into psycopg's _check_connection_ok and finds the cached Django connection closed before it can create a cursor.

Long-running Temporal workers don't go through Django's request cycle, so the request_started / request_finished signals that normally invoke close_old_connections() never fire. Connections that exceed CONN_MAX_AGE or are killed by Postgres/middleboxes stay in the per-thread pool until the next ORM call fails. The same unguarded pattern exists across the rest of the signals Temporal activities, even though posthog/temporal/common/utils.py already exposes the canonical close_db_connections decorator and other products (tasks) already apply it.

Changes

Stack @close_db_connections from posthog.temporal.common.utils beneath @temporalio.activity.defn and @scoped_temporal() on every signals Temporal activity that touches the Django ORM:

• signal_queries.py — fetch_signal_type_examples_activity, run_signal_semantic_search_activity, wait_for_signal_in_clickhouse_activity, fetch_signals_for_report_activity
• summary.py — mark_report_in_progress_activity, mark_report_ready_activity, mark_report_failed_activity, mark_report_pending_input_activity, reset_report_to_potential_activity
• reingestion.py — soft_delete_report_signals_activity, delete_report_activity, reingest_signals_activity, process_team_signals_batch_activity, delete_team_reports_activity
• grouping.py — get_embedding_activity, fetch_report_contexts_activity, assign_and_emit_signal_activity
• backfill_error_tracking.py — fetch_error_tracking_issues_activity, emit_backfill_signal_activity
• emit_eval_signal.py — emit_eval_signal_activity
• report_safety_judge.py — report_safety_judge_activity
• agentic/select_repository.py — select_repository_activity
• agentic/report.py — run_agentic_report_activity

Activities that only talk to ClickHouse, Kafka, object storage, or the Temporal client (publish_report_completed_activity, pause_grouping_until_activity, get_grouping_paused_state_activity, restore_grouping_pause_activity, read_signals_from_s3_activity) are left untouched.

The decorator no-ops under settings.TEST, so existing pytest fixtures that rely on transaction=True are unaffected.

How did you test this code?

Agent-authored change.

• Verified all 9 modified Python files parse cleanly with ast.parse.
• ruff check products/signals/backend/temporal/ — all checks passed.
• ruff format --check products/signals/backend/temporal/ — all files already formatted.
• No automated tests added; this matches how the equivalent decorator is applied in products/tasks/backend/temporal/process_task/activities/*.py, where the contract is documented in posthog/temporal/common/utils.py and exercised by posthog/temporal/tests/common/test_utils.py.

Publish to changelog?

no

🤖 Agent context

PostHog Code agent worked from a signal report flagging a single new OperationalError: the connection is closed issue in run_signal_semantic_search_activity. The report identified the canonical close_db_connections decorator as the existing fix pattern (already used in the tasks product) and recommended a sweep across the rest of the signals activities. The agent verified the decorator's stacking expectations (@activity.defn on top, @close_db_connections innermost), then applied it across every signals activity that calls the Django ORM, skipping ones that only touch ClickHouse, Kafka, object storage, or the Temporal client to avoid noise.

---

Created with PostHog Code