perf(agent): exponential backoff in polling to prevent forking storm (#41222)

dify-agent-runtime/internal/server/config.go

@@ -24,6 +24,7 @@ const (
 	DefaultGCIntervalSeconds             = 60.0
 	DefaultGCFinishedJobRetentionSeconds = 300.0
 	DefaultPollInterval                  = 5 * time.Millisecond
+	DefaultMaxPollInterval               = 50 * time.Millisecond
 	DefaultPipeMonitorInterval           = 1 * time.Second
 	DefaultPipeReadyTimeout              = 10 * time.Second
 	DefaultSQLiteBusyTimeoutMs           = 5000
@@ -53,6 +54,7 @@ type Config struct {
 	MaxOutputLimitBytes          int
 	DefaultTerminateGraceSeconds float64
 	PollInterval                 time.Duration
+	MaxPollInterval              time.Duration
 	PipeMonitorInterval          time.Duration
 	PipeReadyTimeout             time.Duration
 	SQLiteBusyTimeoutMs          int
@@ -90,6 +92,7 @@ func DefaultConfig() (*Config, error) {
 		MaxOutputLimitBytes:          MaxOutputLimitBytes,
 		DefaultTerminateGraceSeconds: DefaultTerminateGraceSeconds,
 		PollInterval:                 DefaultPollInterval,
+		MaxPollInterval:              DefaultMaxPollInterval,
 		PipeMonitorInterval:          DefaultPipeMonitorInterval,
 		PipeReadyTimeout:             DefaultPipeReadyTimeout,
 		SQLiteBusyTimeoutMs:          DefaultSQLiteBusyTimeoutMs,

dify-agent-runtime/internal/server/service.go

@@ -349,6 +349,11 @@ func (s *Service) WaitJob(jobID string, req *WaitJobRequest) (*JobResult, error)
 		lastGrowthAt = &now
 	}
 
+	// Poll with exponential backoff: start tight so short jobs return with low
+	// latency, then grow the interval up to MaxPollInterval so long/idle waits
+	// stop spawning a tmux liveness probe (fork) every few milliseconds.
+	pollInterval := s.config.PollInterval
+
 	for {
 		view, err := s.GetJobStatus(jobID)
 		if err != nil {
@@ -411,7 +416,8 @@ func (s *Service) WaitJob(jobID string, req *WaitJobRequest) (*JobResult, error)
 			return s.jobResultFromView(view, row, window), nil
 		}
 
-		time.Sleep(s.config.PollInterval)
+		time.Sleep(pollInterval)
+		pollInterval = min(pollInterval*2, s.config.MaxPollInterval)
 	}
 }