diff --git a/cmd/svcinit/main.go b/cmd/svcinit/main.go
index cc40ec1..faf40de 100644
--- a/cmd/svcinit/main.go
+++ b/cmd/svcinit/main.go
@@ -285,17 +285,16 @@ func main() {
 
 			testCancel()
 
-			// TODO(zbarsky): what is the right behavior here when services are crashing in ibazel mode?
-
-			// This is a brittle way of draining a channel in a nonblocking way,
-			// consider instead signalling cancellation of the services with a
-			// context, letting them close the channel, and using a waitgroup to
-			// wait for them to exit.
+		// This is a brittle way of draining a channel in a nonblocking way,
+		// consider instead signalling cancellation of the services with a
+		// context, letting them close the channel, and using a waitgroup to
+		// wait for them to exit.
+		// See: https://github.com/hermeticbuild/rules_itest/issues/72
 		Drain:
 			for {
 				select {
-				case <-servicesErrCh:
-					// nothing
+				case crashErr := <-servicesErrCh:
+					log.Printf("Discarding pending service error before reload: %v", crashErr)
 				default:
 					break Drain
 				}
diff --git a/runner/runner.go b/runner/runner.go
index 5fd3cf7..d9f57c7 100644
--- a/runner/runner.go
+++ b/runner/runner.go
@@ -9,6 +9,7 @@ import (
 	"reflect"
 	"runtime"
 	"sync"
+	"syscall"
 	"time"
 
 	"rules_itest/logger"
@@ -198,9 +199,20 @@ func (r *Runner) UpdateSpecs(serviceSpecs ServiceSpecs, ibazelCmd []byte) error
 	}
 
 	for _, label := range updateActions.toReloadLabels {
-		_, err := r.serviceInstances[label].stdin.Write(ibazelCmd)
+		old := r.serviceInstances[label]
+		_, err := old.stdin.Write(ibazelCmd)
 		if err != nil {
-			return err
+			// Service likely crashed — fall back to a full restart.
+			log.Printf("%s hot-reload stdin write failed, falling back to restart: %v", colorize(old.VersionedServiceSpec), err)
+			old.stdin.Close()
+			if stopErr := old.StopWithSignal(syscall.SIGKILL); stopErr != nil {
+				log.Printf("%s stop during crash fallback failed: %v", colorize(old.VersionedServiceSpec), stopErr)
+			}
+			delete(r.serviceInstances, label)
+			r.serviceInstances[label], err = prepareServiceInstance(r.ctx, serviceSpecs[label])
+			if err != nil {
+				return err
+			}
 		}
 	}
 
diff --git a/runner/service_instance.go b/runner/service_instance.go
index 99b8cf6..e645d96 100644
--- a/runner/service_instance.go
+++ b/runner/service_instance.go
@@ -249,8 +249,10 @@ func (s *ServiceInstance) StopWithSignal(signal syscall.Signal) error {
 		s.killed = true
 	}()
 
-	if err != nil {
-		return err
+	// If the process has already exited (raced between killGroup above and here),
+	// return nil — the killGroup call above already attempted group cleanup.
+	if s.isDone() {
+		return nil
 	}
 
 	if signal == syscall.SIGKILL {