fix(worker): 记住轮询间隔内的容量需求

异步视频任务在 River 轮询空档会释放本地 lease，采样瞬间可能观察不到已经触顶的领取需求，导致自适应容量长期停在初始值。本提交记录采样窗口内出现过的容量饱和，同时保持空闲 Worker 不盲目升容。\n\n移植时保留主线现有的单节点容量阶梯、请求下限和严格失败门禁，仅合入失败路径先记录并排空任务与回调、再删除模拟器资源的清理顺序，并补充脚本守卫。\n\n验证：go test ./... -count=1；go vet ./...；bash -n；ShellCheck；production-acceptance-script-test.sh。
2026-08-05 00:53:06 +08:00
parent ebdb96e7d7
commit 9ce9053d7b
4 changed files with 85 additions and 15 deletions
@@ -77,6 +77,7 @@ type Controller struct {
 	claimLimit          int
 	healthySamples      int
 	healthyCount        int
+	demandedSinceSample bool

 	preparing  int
 	waiting    int
@@ -179,10 +180,20 @@ func (c *Controller) TryStart() (*Lease, bool) {
 	c.mu.Lock()
 	defer c.mu.Unlock()
 	activeLimit := min(c.activeLimit, c.claimLimit)
-	if activeLimit <= 0 || c.activeLocked() >= activeLimit || c.preparing+c.finalizing >= c.heavyLimit {
+	if activeLimit <= 0 {
+		return nil, false
+	}
+	if c.activeLocked() >= activeLimit {
+		c.demandedSinceSample = true
+		return nil, false
+	}
+	if c.preparing+c.finalizing >= c.heavyLimit {
 		return nil, false
 	}
 	c.preparing++
+	if c.activeLocked() >= activeLimit {
+		c.demandedSinceSample = true
+	}
 	c.last = c.snapshotLocked(ResourceSample{SampledAt: c.last.SampledAt})
 	return &Lease{controller: c, phase: PhasePreparing}, true
 }
@@ -276,6 +287,9 @@ func (l *Lease) Phase() Phase {
 }

 func (c *Controller) adjustLocked(state PressureState, memory, cpu, database float64) {
+	activeLimit := min(c.activeLimit, c.claimLimit)
+	demanded := c.demandedSinceSample || (activeLimit > 0 && c.activeLocked() >= activeLimit)
+	c.demandedSinceSample = false
 	switch state {
 	case PressureCritical:
 		c.healthyCount = 0
@@ -287,7 +301,7 @@ func (c *Controller) adjustLocked(state PressureState, memory, cpu, database flo
 		c.activeLimit = max(1, c.activeLimit-step)
 		c.heavyLimit = min(c.heavyLimit, max(1, (c.activeLimit+3)/4))
 	default:
-		if memory > .60 || cpu > .65 || database > .65 || c.activeLocked() < min(c.activeLimit, c.claimLimit) {
+		if memory > .60 || cpu > .65 || database > .65 || !demanded {
 			c.healthyCount = 0
 			return
 		}
@@ -29,6 +29,43 @@ func TestControllerStartsConservativelyAndGrowsUnderSustainedDemand(t *testing.T
 	}
 }

+func TestControllerRemembersDemandAcrossShortPollingRuns(t *testing.T) {
+	controller := New(Config{Mode: ModeAdaptive, HardLimit: 8, InitialActive: 4, InitialHeavy: 4, HealthySamples: 2})
+	for range 2 {
+		leases := make([]*Lease, 0, 4)
+		for range 4 {
+			lease, ok := controller.TryStart()
+			if !ok {
+				t.Fatal("short polling task was not admitted")
+			}
+			leases = append(leases, lease)
+		}
+		for _, lease := range leases {
+			lease.Release()
+		}
+		controller.Observe(ResourceSample{
+			MemoryCurrentBytes: 30,
+			MemoryLimitBytes:   100,
+			CPUUtilization:     .3,
+			DBConnections:      3,
+			DBMaxConnections:   20,
+		})
+	}
+	if snapshot := controller.Snapshot(); snapshot.ActiveLimit != 5 || snapshot.SafeCapacity != 5 {
+		t.Fatalf("grown snapshot=%+v, want active=5 safe=5", snapshot)
+	}
+}
+
+func TestControllerDoesNotGrowWhileIdleWithoutDemand(t *testing.T) {
+	controller := New(Config{Mode: ModeAdaptive, HardLimit: 8, InitialActive: 4, InitialHeavy: 1, HealthySamples: 2})
+	for range 4 {
+		controller.Observe(ResourceSample{MemoryCurrentBytes: 20, MemoryLimitBytes: 100, CPUUtilization: .2})
+	}
+	if snapshot := controller.Snapshot(); snapshot.ActiveLimit != 4 {
+		t.Fatalf("idle snapshot=%+v, want active=4", snapshot)
+	}
+}
+
 func TestControllerBusyAndCriticalPressureReduceNewClaims(t *testing.T) {
 	controller := New(Config{Mode: ModeAdaptive, HardLimit: 16, InitialActive: 8, InitialHeavy: 2})
 	busy := controller.Observe(ResourceSample{MemoryCurrentBytes: 80, MemoryLimitBytes: 100})
@@ -381,12 +381,6 @@ cleanup() {
  if [[ $remote_load_drivers_installed == true ]]; then
    cleanup_remote_load_drivers >/dev/null 2>&1 || true
  fi
-  if [[ $single_node_emulator_installed == true ]]; then
-    cleanup_single_node_emulator >/dev/null 2>&1 || true
-  fi
-  if [[ $single_node_acceptance == true ]]; then
-    cleanup_single_node_kubernetes_resources >/dev/null 2>&1 || true
-  fi
  if (( status != 0 )) && [[ -n $run_id && $failure_recorded != true ]]; then
    set +e
    if [[ $single_node_acceptance == true ]]; then
@@ -398,6 +392,12 @@ cleanup() {
    mark_run_failed "${failure_reason:-acceptance workflow exited unexpectedly}"
    set -e
  fi
+  if [[ $single_node_emulator_installed == true ]]; then
+    cleanup_single_node_emulator >/dev/null 2>&1 || true
+  fi
+  if [[ $single_node_acceptance == true ]]; then
+    cleanup_single_node_kubernetes_resources >/dev/null 2>&1 || true
+  fi
  rm -rf -- "$temporary_root"
  exit "$status"
 }
@@ -247,6 +247,25 @@ for invalid_capacity_state in \
  fi
 done

+# shellcheck disable=SC2016 # Match the literal arithmetic expression in the target script.
+grep -Fq 'requests=$((slots * 6))' "$script"
+if grep -Fq '(( requests < 64 ))' "$script"; then
+  echo 'single-node video ladder still overdrives low-capacity steps with a fixed request floor' >&2
+  exit 1
+fi
+
+awk '
+  /^cleanup\(\)/ { capture = 1 }
+  capture { print }
+  capture && /^}/ { exit }
+' "$script" >"$tmp/cleanup-function.sh"
+mark_failed_line=$(grep -n 'mark_run_failed' "$tmp/cleanup-function.sh" | cut -d: -f1)
+cleanup_resources_line=$(grep -n 'cleanup_single_node_kubernetes_resources' "$tmp/cleanup-function.sh" | cut -d: -f1)
+[[ -n $mark_failed_line && -n $cleanup_resources_line && $mark_failed_line -lt $cleanup_resources_line ]] || {
+  echo 'single-node cleanup removes the callback collector before failed outbox work drains' >&2
+  exit 1
+}
+
 passing_pressure_row='2026-08-01T00:00:00Z,0,0,899,149,200,84,1535,0,32,32,0,0,0,0,2,48,0,8,8,0,0,0,16,16,0,0,1535,100'
 pressure_row_passes_live_gates "$passing_pressure_row"
 [[ $(pressure_row_live_gate_reason '2026-08-01T00:00:00Z,0,0,900,149,200,79,1535,0,32,32,0,0,0,0,2,48,0,8,8,0,0,0,16,16,0,0,1535,100') == \