发布

  • [NA] [SDK] test(load): add think-time to test_burst_single_loop (#6856)

    frostbyte_neo 发布于 2026-05-25 19:40:54 +00:00

    At scale 1.0 with no pacing this test took ~17 min of wall-clock under
    the load_tests workflow's xdist -n auto (50 traces/sec — pure backend-
    contention bill, vs the SDK doing thousands/sec on its own) and kept
    the shared docker-compose Opik stack saturated while three other heavy
    scenarios ran in parallel. That backend saturation is what was OOM-
    killing the ingestion-rate workers downstream.

    Add _helpers.think_time() (0.5-2 ms randomised, same as the rest of
    the suite) inside the inner loop. The SDK's in-process queue and
    batcher are still kept busy throughout — at one trace per ~1.25 ms,
    50k traces still complete in ~60 s of submission — but the runner's
    backend isn't being DoS-ed any more. The scenario is no longer the
    "worst-case zero-pacing" burst; docstring updated to match.

    The dataset-items scenario intentionally still has no think_time
    between Dataset.insert() calls — those are synchronous REST round-
    trips that already serialise on their own latency.

    下载附件