fix(acceptance): 为高并发验收启用有界排队

生产同构 Gemini 千请求在真实候选并发饱和后因 queueing disabled 返回 429,同时单台压测机建立千条 WAN TCP 产生连接层 reset。仅对 acceptance 与 acceptance_canary 覆盖为 10000 条、最长 15 分钟的有界等待队列,保留候选原始并发和 RPM/TPM 上限;压测传输强制 HTTP/2 复用连接,避免单源连接风暴。正式 production 策略不变。\n\n验证:acceptance-load 与 runner 定向测试通过,新增生产策略不变和验收队列边界单测;gofmt、bash -n、ShellCheck、git diff --check 通过。
This commit is contained in:
2026-07-31 01:34:16 +08:00
parent 0c9960d2e6
commit 2bfeb3e179
4 changed files with 64 additions and 1 deletions
+1
View File
@@ -525,6 +525,7 @@ func (s *Service) executeWithToken(ctx context.Context, task store.GatewayTask,
admittedPlatformModelID := ""
if distributedAdmission && len(candidates) > 0 {
admissionScopes, groupID := s.admissionScopes(ctx, user, candidates[0])
admissionScopes = acceptanceAdmissionScopes(task, admissionScopes)
hasConcurrentLimit := false
for _, scope := range admissionScopes {
if scope.ConcurrentLimit > 0 {