diff --git a/litellm/types/integrations/prometheus.py b/litellm/types/integrations/prometheus.py index 0d1501664b..35b695cd05 100644 --- a/litellm/types/integrations/prometheus.py +++ b/litellm/types/integrations/prometheus.py @@ -156,6 +156,8 @@ LATENCY_BUCKETS = ( 180.0, 240.0, 300.0, + 420.0, # 7 minutes + 600.0, # 10 minutes (typical default LLM request timeout) float("inf"), ) diff --git a/tests/test_litellm/types/test_prometheus_latency_buckets.py b/tests/test_litellm/types/test_prometheus_latency_buckets.py new file mode 100644 index 0000000000..85670bb0b7 --- /dev/null +++ b/tests/test_litellm/types/test_prometheus_latency_buckets.py @@ -0,0 +1,17 @@ +"""LATENCY_BUCKETS covers long-running LLM calls (histograms are in seconds).""" + +import math + +from litellm.types.integrations.prometheus import LATENCY_BUCKETS + + +def test_latency_buckets_include_seven_and_ten_minutes(): + """Buckets beyond 5 min so histograms resolve requests up to default LLM timeouts.""" + assert 300.0 in LATENCY_BUCKETS + assert 420.0 in LATENCY_BUCKETS # 7 min + assert 600.0 in LATENCY_BUCKETS # 10 min + assert math.isinf(LATENCY_BUCKETS[-1]) + idx_300 = LATENCY_BUCKETS.index(300.0) + idx_420 = LATENCY_BUCKETS.index(420.0) + idx_600 = LATENCY_BUCKETS.index(600.0) + assert idx_300 < idx_420 < idx_600