Docs / Load & performance

Testing mode

Load & performance

Protocol-aware latency and load in the same CI suite as functional tests — not a separate load product.

Latency budget

from mcp_test_harness import assert_latency, marker

@marker(tags=["perf"])
async def test_echo_p95(mcp_server):
    await assert_latency(
        mcp_server, "echo", {"text": "hi"},
        max_ms=200, runs=20, aggregate="p95", warmup=3,
    )

Concurrent load

Fixed burst (total_calls) or closed-loop soak (duration_s). Optional weighted multi-tool calls.

from mcp_test_harness import assert_throughput, marker

@marker(tags=["perf", "load"])
async def test_echo_soak(mcp_server):
    await assert_throughput(
        mcp_server, "echo", {"text": "load"},
        concurrent=8, duration_s=5,
        min_rps=10, max_p95_ms=300, max_error_rate=0.02,
    )

Concurrency ramp

Warmup phases are measured but excluded from aggregate SLO gates.

from mcp_test_harness import assert_load_phases, marker

@marker(tags=["perf", "load"])
async def test_echo_ramp(mcp_server):
    await assert_load_phases(
        mcp_server, "echo", {"text": "load"},
        phases=[
            {"name": "warmup", "concurrent": 4, "duration_s": 1, "warmup": True},
            {"name": "c16", "concurrent": 16, "duration_s": 3},
        ],
        min_rps=10, max_p95_ms=400,
    )

CI + report

mcp-test -m perf --report-format html --report-output report.html
mcp-test export-pdf report.html

The HTML report includes a Load & performance SLOs panel (suite percentiles + perf-tagged tests). It is CI evidence, not a JMeter live GUI — pair with k6/JMeter for datacenter-scale RPS.