-
[NA] [BE] fix: classify thread online-scoring errors as processing_errors, resilient per thread id (#7299)
frostbyte_neo 发布于
2026-07-01 13:27:05 +00:00 The trace-thread scorers build their per-message chain from Flux.fromIterable(threadIds).flatMap(...),
whereas the trace/span scorers build from a Mono. Flux.flatMap is onErrorContinue-aware, so the enclosing
onErrorContinue in BaseRedisSubscriber.setupStreamListener reached into the scorer's flatMap, dropped the
failing element and recorded the failure as *_unexpected_errors_total — bypassing processMessage's
per-message onErrorResume. Client/config errors (BadRequestException, ClientErrorException) on the thread
streams were therefore counted as "unexpected", while the same errors on the trace/span streams are
correctly counted as processing_errors.Score each thread id independently and materialize per-thread errors (onErrorResume) so a single thread's
failure no longer stops scoring the sibling thread ids in the same message. After the batch, re-surface
the first failure on the Mono error path so the message is still handled by processMessage's onErrorResume
— classified as a processing error and following the normal retryable/non-retryable path — consistent
with the trace/span scorers. Applied to both OnlineScoringTraceThreadLlmAsJudgeScorer and
OnlineScoringTraceThreadUserDefinedMetricPythonScorer.Co-authored-by: Claude Opus 4.8 (1M context) noreply@anthropic.com
下载附件