Method beginInferLoad

internal/server/server.go:243–273

beginInferLoad enforces reliability.overload: "reject" returns reject=true without consuming a slot; "degrade" allows the request and sets overloadDegrade when at/above the limit. QueueLimit <= 0 disables checks. Concurrency slots are acquired with CAS so active /infer calls cannot exceed the limit

()

// "degrade" allows the request and sets overloadDegrade when at/above the limit. QueueLimit <= 0 disables checks.
// Concurrency slots are acquired with CAS so active /infer calls cannot exceed the limit (reject mode).
func (s *Server) beginInferLoad() (done func(), overloadDegrade bool, reject bool) {
	nop := func() {}
	limit := s.cfg.Reliability.Overload.QueueLimit
	if limit <= 0 {
		s.inferInflight.Add(1)
		return func() { s.inferInflight.Add(-1) }, false, false
	}
	action := strings.ToLower(strings.TrimSpace(s.cfg.Reliability.Overload.Action))
	if action == "" {
		action = "degrade"
	}
	lim := int32(limit)
	if action == "reject" {
		for {
			cur := s.inferInflight.Load()
			if cur >= lim {
				return nop, false, true
			}
			if s.inferInflight.CompareAndSwap(cur, cur+1) {
				return func() { s.inferInflight.Add(-1) }, false, false
			}
		}
	}
	for {
		cur := s.inferInflight.Load()
		degraded := cur >= lim
		if s.inferInflight.CompareAndSwap(cur, cur+1) {
			return func() { s.inferInflight.Add(-1) }, degraded, false
		}
	}
}

// cachedBackendHealth probes adapter.Health per backend with TTL cache (shared by routing and /status).
func (s *Server) cachedBackendHealth(ctx context.Context) map[string]bool {