-
[OPIK-6971] Treat insufficient_quota (429) as non-retryable for OpenAI providers (#7148)
frostbyte_neo 发布于
2026-06-18 09:37:44 +00:00 - [OPIK-6971] Treat insufficient_quota (429) as non-retryable for OpenAI providers
When an LLM key returns HTTP 429 insufficient_quota (out of credits), the
provider keeps failing the same way. LangChain4j maps any 429 to a retryable
RateLimitException, both in the OpenAiChatModel internal retry and in the outer
ChatCompletionService retry policy, so an exhausted key was retried repeatedly
per request (inner x outer attempts), producing an error storm.Add a QuotaAwareHttpClient decorator that inspects 429 responses and, when the
body is insufficient_quota, rethrows a NonRetriableException so both retry layers
stop immediately. The original body is carried as the exception message (so the
provider-error mapping still surfaces HTTP 402) and no HttpException is kept as
the cause (so ExceptionMapper.findRoot cannot re-map the 429 back to retryable).
rate_limit_exceeded and other statuses pass through unchanged and stay retryable.Wired into the free-model and OpenAI completions chat models and the raw OpenAI
client.Co-Authored-By: Claude Opus 4.8 (1M context) noreply@anthropic.com
- [OPIK-6971] Use assertThatThrownBy in quota test for consistency
Address review nit: keep all four QuotaAwareHttpClient tests on the same
assertion style.Co-Authored-By: Claude Opus 4.8 (1M context) noreply@anthropic.com
- [OPIK-6971] Address review: dedup builder, case-insensitive match, lombok, test cleanup
- Extract DelegatingHttpClientBuilder base (timeout forwarding) and have both
QuotaAwareHttpClient.Builder and InterceptingHttpClientBuilder extend it, instead
of duplicating the delegation logic. - Match insufficient_quota case-insensitively (StringUtils.containsIgnoreCase).
- Reuse jakarta Response.Status.TOO_MANY_REQUESTS instead of a 429 literal.
- Use @RequiredArgsConstructor for the boilerplate constructor.
- Tests: java text blocks, @InjectMocks, parameterized pass-through + casing coverage.
Co-Authored-By: Claude Opus 4.8 (1M context) noreply@anthropic.com
Co-authored-by: Claude Opus 4.8 (1M context) noreply@anthropic.com
下载附件