发布

  • [OPIK-5689] [BE] fix: stop retrying rate-limited LLM calls and reduce log verbosity (#6132)

    frostbyte_neo 发布于 2026-04-08 14:46:00 +00:00

    • [OPIK-5689] [BE] fix: stop retrying rate-limited LLM calls and reduce log verbosity

    Map OpenAI rate_limit_exceeded error to HTTP 429 so it's treated as a
    non-retryable client error, preventing unnecessary Redis stream retries.
    Remove full stack traces from error/retry logs across the online scoring
    pipeline to drastically cut log volume.

    • [OPIK-5689] [BE] fix: address PR review comments
    • Keep full Throwable in log calls (restore stack traces) but keep
      warn level for handled errors in BaseRedisSubscriber and
      ChatCompletionService
    • Consolidate duplicate log in OnlineScoringTraceThreadLlmAsJudgeScorer
      into single log.error with Throwable
    • Map insufficient_quota to 402 instead of 429
    • Fix test naming convention, use whole-object assertions, merge
      standalone tests into parameterized
    下载附件