From ea2a842d7d675bfd6387f438b02d09760eabb24a Mon Sep 17 00:00:00 2001 From: Kerry Lu Date: Tue, 20 Oct 2026 06:51:53 -0700 Subject: [PATCH] test(perf): set the seven new perf ceilings from 124 lane runs across v1.103.0 to v1.106.0 (#47527) Co-authored-by: devin-ai-integration[bot] <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- tests/e2e/perf/cases/mixed_size_requests_2rps.py | 14 +++++++------- tests/e2e/perf/cases/small_bodies_5rps.py | 14 +++++++------- 2 files changed, 14 insertions(+), 14 deletions(-) diff --git a/tests/e2e/perf/cases/mixed_size_requests_2rps.py b/tests/e2e/perf/cases/mixed_size_requests_2rps.py index 850658a15c5..d4f1fdddc9c 100644 --- a/tests/e2e/perf/cases/mixed_size_requests_2rps.py +++ b/tests/e2e/perf/cases/mixed_size_requests_2rps.py @@ -64,14 +64,14 @@ CASE: Final = PerformanceTestCase( Limit("unexpected_status_count", max=3), Limit("gateway_added_seconds", percentile=50, max=0.58), Limit("gateway_added_seconds", percentile=99, max=6.1), - Limit("litellm_overhead_seconds", percentile=50, max=0.05), - Limit("litellm_overhead_seconds", percentile=99, max=2.2), - Limit("litellm_response_seconds", percentile=50, max=1.2), + Limit("litellm_overhead_seconds", percentile=50, max=0.1), + Limit("litellm_overhead_seconds", percentile=99, max=3.4), + Limit("litellm_response_seconds", percentile=50, max=1.0), Limit("litellm_response_seconds", percentile=99, max=3.8), - Limit("prometheus_redis_latency_mean_seconds", max=0.003), - Limit("postgres_active_sessions", max=0.004), - Limit("postgres_slowest_statement_seconds", max=0.045), - Limit("postgres_peak_connections", max=30), + Limit("prometheus_redis_latency_mean_seconds", max=0.005), + Limit("postgres_active_sessions", max=0.016), + Limit("postgres_slowest_statement_seconds", max=0.25), + Limit("postgres_peak_connections", max=45), Limit("time_to_first_token_seconds", percentile=50, max=2.7), Limit("time_to_first_token_seconds", percentile=99, max=8.8), Limit("memory_growth_megabytes", max=600), diff --git a/tests/e2e/perf/cases/small_bodies_5rps.py b/tests/e2e/perf/cases/small_bodies_5rps.py index bc22a9a96c8..f93c64afba4 100644 --- a/tests/e2e/perf/cases/small_bodies_5rps.py +++ b/tests/e2e/perf/cases/small_bodies_5rps.py @@ -65,14 +65,14 @@ CASE: Final = PerformanceTestCase( limits=( Limit("unexpected_status_count", max=3), Limit("gateway_added_seconds", percentile=50, max=0.29), - Limit("litellm_overhead_seconds", percentile=50, max=0.02), - Limit("litellm_overhead_seconds", percentile=99, max=0.8), - Limit("litellm_response_seconds", percentile=50, max=0.8), - Limit("litellm_response_seconds", percentile=99, max=2.8), + Limit("litellm_overhead_seconds", percentile=50, max=0.012), + Limit("litellm_overhead_seconds", percentile=99, max=0.35), + Limit("litellm_response_seconds", percentile=50, max=1.0), + Limit("litellm_response_seconds", percentile=99, max=2.0), Limit("prometheus_redis_latency_mean_seconds", max=0.006), - Limit("postgres_active_sessions", max=0.006), - Limit("postgres_slowest_statement_seconds", max=0.035), - Limit("postgres_peak_connections", max=30), + Limit("postgres_active_sessions", max=0.01), + Limit("postgres_slowest_statement_seconds", max=0.025), + Limit("postgres_peak_connections", max=45), Limit("time_to_first_token_seconds", percentile=99, max=3.7), Limit("memory_growth_megabytes", max=100), Limit("postgres_calls_per_request", max=1.8),