The floor was an absolute fleet number, so it asserted replicas x per-replica rate
and went red on how many gateway pods happened to be warm rather than on the request
path. The test now measures one replica first, with a short serial pass that only ever
occupies a single pod, and requires the concurrent phase to reach at least that rate.
A serial latency budget carries the request-path assertion the floor used to imply,
and both hold at one replica or seven.
Zero-error runs that "sustained 16.7 RPS" were queueing, not slow requests: the load
model is a mock_response deployment with no upstream, a single-worker replica serves
it in about 57ms, and 100 closed-loop users against 1/0.057 RPS of capacity sit at
6s each by Little's law.
The runner also kept locust's --json summary and threw away everything else, so a run
where 93% of requests failed said nothing about what they got. It now passes --csv,
reads the failure breakdown back, and reports locust's own generator-saturation
warnings, both folded into the assertion messages.
Resolves LIT-5054
CodSpeed benchmarks the SDK with no IO, so it can't catch regressions that
only appear under real concurrent load through the full proxy stack (auth,
routing, logging, spend, Postgres, Redis). This adds a Locust load test under
tests/e2e/load that drives concurrent POST /chat/completions traffic against a
mock deployment (litellm_params.mock_response), so the measured throughput
reflects proxy overhead rather than a provider's latency, and asserts an
aggregate RPS SLO with a failure-ratio guard. The test is marked load and the
parent conftest sorts load-marked items last so it never perturbs
latency-sensitive suites. Covers reliability.perf.throughput.under_slo.
