feat(routing): 完善平台满载避让与故障轮转

为未配置并发上限的平台增加基于运行和等待任务数的软负载,并按非满载、有效优先级、缓存亲和与负载稳定排序。\n\n平台模型 RPM、TPM 和并发额度竞争失败时只轮转候选,不触发冷却、禁用或降级;用户组额度保持不可绕过。补齐异步冷却排队恢复、满载原因、选择原因与低基数指标。\n\n验证:go test ./...;go vet ./...;PostgreSQL 原子额度、准入队列、Worker 容量回收及故障策略 HTTP acceptance。
This commit is contained in:
2026-08-03 17:42:34 +08:00
parent b125599354
commit 8362d6d27a
15 changed files with 468 additions and 152 deletions
+23 -13
View File
@@ -50,21 +50,18 @@ func isLocalRateLimitError(err error) bool {
return errors.As(err, &limitErr)
}
func platformModelRateLimitError(err error) (*store.RateLimitExceededError, bool) {
var limitErr *store.RateLimitExceededError
if !errors.As(err, &limitErr) || limitErr.ScopeType != "platform_model" {
return nil, false
}
return limitErr, true
}
func (s *Service) rateLimitReservations(ctx context.Context, user *auth.User, candidate store.RuntimeModelCandidate, body map[string]any) []store.RateLimitReservation {
out := make([]store.RateLimitReservation, 0)
out = append(out, reservationsFromPolicy(
"platform_model",
candidate.PlatformModelID,
firstNonEmptyString(candidate.DisplayName, candidate.ModelAlias, candidate.ModelName),
map[string]any{
"platformId": candidate.PlatformID,
"platformName": candidate.PlatformName,
"modelAlias": candidate.ModelAlias,
"modelName": candidate.ModelName,
},
effectiveRateLimitPolicy(candidate),
body,
)...)
// Cross-platform scopes must be evaluated first so a user-group limit can
// never be bypassed by rotating to another platform model.
if group, err := s.store.ResolveUserGroupPolicy(ctx, user); err == nil && group.ID != "" {
out = append(out, reservationsFromPolicy(
"user_group",
@@ -78,6 +75,19 @@ func (s *Service) rateLimitReservations(ctx context.Context, user *auth.User, ca
body,
)...)
}
out = append(out, reservationsFromPolicy(
"platform_model",
candidate.PlatformModelID,
firstNonEmptyString(candidate.DisplayName, candidate.ModelAlias, candidate.ModelName),
map[string]any{
"platformId": candidate.PlatformID,
"platformName": candidate.PlatformName,
"modelAlias": candidate.ModelAlias,
"modelName": candidate.ModelName,
},
effectiveRateLimitPolicy(candidate),
body,
)...)
return out
}