diff --git a/litellm/proxy/db/db_transaction_queue/spend_log_cleanup.py b/litellm/proxy/db/db_transaction_queue/spend_log_cleanup.py index db6a2943aa3..2fd34000466 100644 --- a/litellm/proxy/db/db_transaction_queue/spend_log_cleanup.py +++ b/litellm/proxy/db/db_transaction_queue/spend_log_cleanup.py @@ -1261,13 +1261,14 @@ class SpendLogCleanup: key_columns=("request_id", "guardrail_id"), time_column="start_time", deadline=deadline, - batch_cap=step_cap, + batch_cap=state.batches_left, ) + guardrail_batches: Final = math.ceil(guardrail_index_result.rows_deleted / self.batch_size) advanced: Final = _SizeStepState( logs_deleted=state.logs_deleted + logs_result.rows_deleted, index_deleted=state.index_deleted + index_result.rows_deleted, guardrail_index_deleted=state.guardrail_index_deleted + guardrail_index_result.rows_deleted, - batches_left=state.batches_left - step_cap, + batches_left=state.batches_left - max(step_cap, guardrail_batches), stop_reason=None, ) if rows_still_needed(advanced.rows_deleted) <= 0: diff --git a/tests/unit/proxy/test_spend_log_cleanup.py b/tests/unit/proxy/test_spend_log_cleanup.py index c9973c7cd7d..dfb2d1817cf 100644 --- a/tests/unit/proxy/test_spend_log_cleanup.py +++ b/tests/unit/proxy/test_spend_log_cleanup.py @@ -1606,7 +1606,7 @@ async def test_a_size_pass_that_fires_while_an_age_pass_is_running_in_the_same_p age_cleaner = SpendLogCleanup(general_settings={"maximum_spend_logs_retention_period": "7d"}) age_cleaner.pod_lock_manager = None - size_cleaner = _size_cleaner({}) + size_cleaner = _size_cleaner({"maximum_spend_logs_cleanup_max_batches": 3}) slow_run = asyncio.ensure_future(age_cleaner.cleanup_old_spend_logs(slow_client)) await asyncio.wait_for(first_batch_done.wait(), timeout=5) @@ -1885,6 +1885,25 @@ async def test_size_pass_deletes_oldest_first_in_steps_until_the_live_estimate_i assert "VACUUM FULL or pg_repack" in summary +@pytest.mark.asyncio +async def test_size_pass_deletes_every_guardrail_row_under_the_step_cutoff_not_one_per_spend_log(caplog): + db = MagicMock() + _wire_tx(db) + # one step of 3 SpendLogs batches, but two guardrails per request leave twice as many guardrail rows under it + db.execute_raw = AsyncMock(side_effect=[1000, 1000, 1000, 0, 1000, 1000, 1000, 1000, 1000, 1000, 0]) + _wire_size_probes(db, [_storage(3 * 1024**3, live=3000, dead=0)], cutoff=datetime(2026, 9, 1, tzinfo=timezone.utc)) + client = MagicMock() + client.db = db + cleaner = _size_cleaner({"maximum_spend_logs_cleanup_target_size": "512MB"}) + + with caplog.at_level(logging.INFO, logger="LiteLLM Proxy"): + await cleaner.cleanup_spend_logs_by_size(client) + + summary = next(r.getMessage() for r in caplog.records if "Spend log cleanup run finished" in r.getMessage()) + assert "LiteLLM_SpendLogs (size rule): deleted=3000" in summary, summary + assert "LiteLLM_SpendLogGuardrailIndex (size rule): deleted=6000" in summary, summary + + @pytest.mark.asyncio async def test_size_pass_stops_at_the_run_batch_cap_and_reports_how_far_above_the_target_it_is(caplog): db = MagicMock()