Universal runner improvements.

sdks/go/pkg/beam/runners/universal/runnerlib/execute.go

@@ -41,7 +41,7 @@ func Execute(ctx context.Context, p *pipepb.Pipeline, endpoint string, opt *JobO
 	presult := &universalPipelineResult{}
 
 	bin := opt.Worker
-	if bin == "" {
+	if bin == "" && !opt.Loopback {
 		if self, ok := IsWorkerCompatibleBinary(); ok {
 			bin = self
 			log.Infof(ctx, "Using running binary as worker binary: '%v'", bin)
@@ -56,6 +56,11 @@ func Execute(ctx context.Context, p *pipepb.Pipeline, endpoint string, opt *JobO
 
 			bin = worker
 		}
+	} else if opt.Loopback {
+		// TODO(https://github.com/apache/beam/issues/27569: determine the canonical location for Beam temp files.
+		// In loopback mode, the binary is unused, so we can avoid an unnecessary compile step.
+		f, _ := os.CreateTemp(os.TempDir(), "beamloopbackworker-*")
+		bin = f.Name()
 	} else {
 		log.Infof(ctx, "Using specified worker binary: '%v'", bin)
 	}

sdks/go/pkg/beam/runners/universal/runnerlib/job.go

@@ -39,15 +39,17 @@ type JobOptions struct {
 	// Experiments are additional experiments.
 	Experiments []string
 
-	// TODO(herohde) 3/17/2018: add further parametrization as needed
-
 	// Worker is the worker binary override.
 	Worker string
 
-	// RetainDocker is an option to pass to the runner.
+	// RetainDocker is an option to pass to the runner indicating the docker containers should be cached.
 	RetainDocker bool
 
+	// Indicates a limit on parallelism the runner should impose.
 	Parallelism int
+
+	// Loopback indicates this job is running in loopback mode and will reconnect to the local process.
+	Loopback bool
 }
 
 // Prepare prepares a job to the given job service. It returns the preparation id
@@ -101,10 +103,17 @@ func WaitForCompletion(ctx context.Context, client jobpb.JobServiceClient, jobID
 		return errors.Wrap(err, "failed to get job stream")
 	}
 
+	mostRecentError := errors.New("<no error received, see runner logs>")
+	var errReceived, jobFailed bool
+
 	for {
 		msg, err := stream.Recv()
 		if err != nil {
 			if err == io.EOF {
+				if jobFailed {
+					// Connection finished with a failed status, so produce what we have.
+					return errors.Errorf("job %v failed:\n%w", jobID, mostRecentError)
+				}
 				return nil
 			}
 			return err
@@ -120,7 +129,11 @@ func WaitForCompletion(ctx context.Context, client jobpb.JobServiceClient, jobID
 			case jobpb.JobState_DONE, jobpb.JobState_CANCELLED:
 				return nil
 			case jobpb.JobState_FAILED:
-				return errors.Errorf("job %v failed", jobID)
+				jobFailed = true
+				if errReceived {
+					return errors.Errorf("job %v failed:\n%w", jobID, mostRecentError)
+				}
+				// Otherwise, wait for at least one error log from the runner, or the connection to close.
 			}
 
 		case msg.GetMessageResponse() != nil:
@@ -129,6 +142,15 @@ func WaitForCompletion(ctx context.Context, client jobpb.JobServiceClient, jobID
 			text := fmt.Sprintf("%v (%v): %v", resp.GetTime(), resp.GetMessageId(), resp.GetMessageText())
 			log.Output(ctx, messageSeverity(resp.GetImportance()), 1, text)
 
+			if resp.GetImportance() >= jobpb.JobMessage_JOB_MESSAGE_ERROR {
+				errReceived = true
+				mostRecentError = errors.New(resp.GetMessageText())
+
+				if jobFailed {
+					return errors.Errorf("job %v failed:\n%w", jobID, mostRecentError)
+				}
+			}
+
 		default:
 			return errors.Errorf("unexpected job update: %v", proto.MarshalTextString(msg))
 		}

sdks/go/pkg/beam/runners/universal/universal.go

@@ -101,6 +101,7 @@ func Execute(ctx context.Context, p *beam.Pipeline) (beam.PipelineResult, error)
 		Worker:       *jobopts.WorkerBinary,
 		RetainDocker: *jobopts.RetainDockerContainers,
 		Parallelism:  *jobopts.Parallelism,
+		Loopback:     jobopts.IsLoopback(),
 	}
 	return runnerlib.Execute(ctx, pipeline, endpoint, opt, *jobopts.Async)
 }