diff --git a/client/alloc_runner_test.go b/client/alloc_runner_test.go
index 390117926..649dbdebc 100644
--- a/client/alloc_runner_test.go
+++ b/client/alloc_runner_test.go
@@ -405,6 +405,10 @@ func TestAllocRunner_SaveRestoreState_TerminalAlloc(t *testing.T) {
 		t.Fatalf("err: %v", err)
 	}
 
+	// Ensure ar1 doesn't recreate the state file
+	ar.persistLock.Lock()
+	defer ar.persistLock.Unlock()
+
 	// Ensure both alloc runners don't destroy
 	ar.destroy = true
 
diff --git a/client/driver/docker.go b/client/driver/docker.go
index a58884812..7db042072 100644
--- a/client/driver/docker.go
+++ b/client/driver/docker.go
@@ -50,7 +50,7 @@ var (
 
 	// recoverableErrTimeouts returns a recoverable error if the error was due
 	// to timeouts
-	recoverableErrTimeouts = func(err error) *structs.RecoverableError {
+	recoverableErrTimeouts = func(err error) error {
 		r := false
 		if strings.Contains(err.Error(), "Client.Timeout exceeded while awaiting headers") ||
 			strings.Contains(err.Error(), "EOF") {
@@ -82,16 +82,24 @@ const (
 	// Docker's privileged mode.
 	dockerPrivilegedConfigOption = "docker.privileged.enabled"
 
+	// dockerCleanupImageConfigOption is the key for whether or not to
+	// cleanup images after the task exits.
+	dockerCleanupImageConfigOption  = "docker.cleanup.image"
+	dockerCleanupImageConfigDefault = true
+
 	// dockerTimeout is the length of time a request can be outstanding before
 	// it is timed out.
 	dockerTimeout = 5 * time.Minute
+
+	// dockerImageResKey is the CreatedResources key for docker images
+	dockerImageResKey = "image"
 )
 
 type DockerDriver struct {
 	DriverContext
 
-	imageID      string
 	driverConfig *DockerDriverConfig
+	imageID      string
 }
 
 type DockerDriverAuth struct {
@@ -235,8 +243,6 @@ type DockerHandle struct {
 	client            *docker.Client
 	waitClient        *docker.Client
 	logger            *log.Logger
-	cleanupImage      bool
-	imageID           string
 	containerID       string
 	version           string
 	clkSpeed          float64
@@ -353,35 +359,39 @@ func (d *DockerDriver) FSIsolation() cstructs.FSIsolation {
 	return cstructs.FSIsolationImage
 }
 
-func (d *DockerDriver) Prestart(ctx *ExecContext, task *structs.Task) error {
+func (d *DockerDriver) Prestart(ctx *ExecContext, task *structs.Task) (*CreatedResources, error) {
 	driverConfig, err := NewDockerDriverConfig(task, d.taskEnv)
 	if err != nil {
-		return err
+		return nil, err
 	}
 
+	// Set state needed by Start()
+	d.driverConfig = driverConfig
+
 	// Initialize docker API clients
 	client, _, err := d.dockerClients()
 	if err != nil {
-		return fmt.Errorf("Failed to connect to docker daemon: %s", err)
+		return nil, fmt.Errorf("Failed to connect to docker daemon: %s", err)
 	}
 
+	// Ensure the image is available
 	if err := d.createImage(driverConfig, client, ctx.TaskDir); err != nil {
-		return err
+		return nil, err
 	}
 
-	image := driverConfig.ImageName
-	// Now that we have the image we can get the image id
-	dockerImage, err := client.InspectImage(image)
+	// Regardless of whether the image was downloaded already or not, store
+	// it as a created resource. Cleanup will soft fail if the image is
+	// still in use by another contianer.
+	dockerImage, err := client.InspectImage(driverConfig.ImageName)
 	if err != nil {
-		d.logger.Printf("[ERR] driver.docker: failed getting image id for %s: %s", image, err)
-		return fmt.Errorf("Failed to determine image id for `%s`: %s", image, err)
+		d.logger.Printf("[ERR] driver.docker: failed getting image id for %q: %v", driverConfig.ImageName, err)
+		return nil, err
 	}
-	d.logger.Printf("[DEBUG] driver.docker: identified image %s as %s", image, dockerImage.ID)
 
-	// Set state needed by Start()
+	res := NewCreatedResources()
+	res.Add(dockerImageResKey, dockerImage.ID)
 	d.imageID = dockerImage.ID
-	d.driverConfig = driverConfig
-	return nil
+	return res, nil
 }
 
 func (d *DockerDriver) Start(ctx *ExecContext, task *structs.Task) (DriverHandle, error) {
@@ -426,23 +436,24 @@ func (d *DockerDriver) Start(ctx *ExecContext, task *structs.Task) (DriverHandle
 
 	config, err := d.createContainerConfig(ctx, task, d.driverConfig, syslogAddr)
 	if err != nil {
-		d.logger.Printf("[ERR] driver.docker: failed to create container configuration for image %s: %s", d.imageID, err)
+		d.logger.Printf("[ERR] driver.docker: failed to create container configuration for image %q (%q): %v", d.driverConfig.ImageName, d.imageID, err)
 		pluginClient.Kill()
-		return nil, fmt.Errorf("Failed to create container configuration for image %s: %s", d.imageID, err)
+		return nil, fmt.Errorf("Failed to create container configuration for image %q (%q): %v", d.driverConfig.ImageName, d.imageID, err)
 	}
 
-	container, rerr := d.createContainer(config)
-	if rerr != nil {
-		d.logger.Printf("[ERR] driver.docker: failed to create container: %s", rerr)
+	container, err := d.createContainer(config)
+	if err != nil {
+		d.logger.Printf("[ERR] driver.docker: failed to create container: %s", err)
 		pluginClient.Kill()
-		rerr.Err = fmt.Sprintf("Failed to create container: %s", rerr.Err)
-		return nil, rerr
+		if rerr, ok := err.(*structs.RecoverableError); ok {
+			rerr.Err = fmt.Sprintf("Failed to create container: %s", rerr.Err)
+			return nil, rerr
+		}
+		return nil, err
 	}
 
 	d.logger.Printf("[INFO] driver.docker: created container %s", container.ID)
 
-	cleanupImage := d.config.ReadBoolDefault("docker.cleanup.image", true)
-
 	// We don't need to start the container if the container is already running
 	// since we don't create containers which are already present on the host
 	// and are running
@@ -466,9 +477,7 @@ func (d *DockerDriver) Start(ctx *ExecContext, task *structs.Task) (DriverHandle
 		waitClient:     waitClient,
 		executor:       exec,
 		pluginClient:   pluginClient,
-		cleanupImage:   cleanupImage,
 		logger:         d.logger,
-		imageID:        d.imageID,
 		containerID:    container.ID,
 		version:        d.config.Version,
 		killTimeout:    GetKillTimeout(task.KillTimeout, maxKill),
@@ -484,6 +493,58 @@ func (d *DockerDriver) Start(ctx *ExecContext, task *structs.Task) (DriverHandle
 	return h, nil
 }
 
+func (d *DockerDriver) Cleanup(_ *ExecContext, res *CreatedResources) error {
+	retry := false
+	var merr multierror.Error
+	for key, resources := range res.Resources {
+		switch key {
+		case dockerImageResKey:
+			for _, value := range resources {
+				err := d.cleanupImage(value)
+				if err != nil {
+					if structs.IsRecoverable(err) {
+						retry = true
+					}
+					merr.Errors = append(merr.Errors, err)
+					continue
+				}
+
+				// Remove cleaned image from resources
+				res.Remove(dockerImageResKey, value)
+			}
+		default:
+			d.logger.Printf("[ERR] driver.docker: unknown resource to cleanup: %q", key)
+		}
+	}
+	return structs.NewRecoverableError(merr.ErrorOrNil(), retry)
+}
+
+// cleanupImage removes a Docker image. No error is returned if the image
+// doesn't exist or is still in use. Requires the global client to already be
+// initialized.
+func (d *DockerDriver) cleanupImage(id string) error {
+	if !d.config.ReadBoolDefault(dockerCleanupImageConfigOption, dockerCleanupImageConfigDefault) {
+		// Config says not to cleanup
+		return nil
+	}
+
+	if err := client.RemoveImage(id); err != nil {
+		if err == docker.ErrNoSuchImage {
+			d.logger.Printf("[DEBUG] driver.docker: unable to cleanup image %q: does not exist", id)
+			return nil
+		}
+		if derr, ok := err.(*docker.Error); ok && derr.Status == 409 {
+			d.logger.Printf("[DEBUG] driver.docker: unable to cleanup image %q: still in use", id)
+			return nil
+		}
+		// Retry on unknown errors
+		return structs.NewRecoverableError(err, true)
+	}
+
+	d.logger.Printf("[DEBUG] driver.docker: cleanup removed downloaded image: %q", id)
+	return nil
+}
+
 // dockerClients creates two *docker.Client, one for long running operations and
 // the other for shorter operations. In test / dev mode we can use ENV vars to
 // connect to the docker daemon. In production mode we will read docker.endpoint
@@ -657,7 +718,7 @@ func (d *DockerDriver) createContainerConfig(ctx *ExecContext, task *structs.Tas
 	}
 
 	config := &docker.Config{
-		Image:     driverConfig.ImageName,
+		Image:     d.imageID,
 		Hostname:  driverConfig.Hostname,
 		User:      task.User,
 		Tty:       driverConfig.TTY,
@@ -886,26 +947,28 @@ func (d *DockerDriver) createImage(driverConfig *DockerDriverConfig, client *doc
 		tag = "latest"
 	}
 
-	var dockerImage *docker.Image
-	var err error
 	// We're going to check whether the image is already downloaded. If the tag
 	// is "latest", or ForcePull is set, we have to check for a new version every time so we don't
 	// bother to check and cache the id here. We'll download first, then cache.
 	if driverConfig.ForcePull {
 		d.logger.Printf("[DEBUG] driver.docker: force pull image '%s:%s' instead of inspecting local", repo, tag)
 	} else if tag != "latest" {
-		dockerImage, err = client.InspectImage(image)
+		if dockerImage, _ := client.InspectImage(image); dockerImage != nil {
+			// Image exists, nothing to do
+			return nil
+		}
+	}
+
+	// Load the image if specified
+	if len(driverConfig.LoadImages) > 0 {
+		return d.loadImage(driverConfig, client, taskDir)
 	}
 
 	// Download the image
-	if dockerImage == nil {
-		if len(driverConfig.LoadImages) > 0 {
-			return d.loadImage(driverConfig, client, taskDir)
-		}
-
-		return d.pullImage(driverConfig, client, repo, tag)
+	if err := d.pullImage(driverConfig, client, repo, tag); err != nil {
+		return err
 	}
-	return err
+	return nil
 }
 
 // pullImage creates an image by pulling it from a docker registry
@@ -955,6 +1018,7 @@ func (d *DockerDriver) pullImage(driverConfig *DockerDriverConfig, client *docke
 		d.logger.Printf("[ERR] driver.docker: failed pulling container %s:%s: %s", repo, tag, err)
 		return d.recoverablePullError(err, driverConfig.ImageName)
 	}
+
 	d.logger.Printf("[DEBUG] driver.docker: docker pull %s:%s succeeded", repo, tag)
 	return nil
 }
@@ -980,7 +1044,7 @@ func (d *DockerDriver) loadImage(driverConfig *DockerDriverConfig, client *docke
 
 // createContainer creates the container given the passed configuration. It
 // attempts to handle any transient Docker errors.
-func (d *DockerDriver) createContainer(config docker.CreateContainerOptions) (*docker.Container, *structs.RecoverableError) {
+func (d *DockerDriver) createContainer(config docker.CreateContainerOptions) (*docker.Container, error) {
 	// Create a container
 	attempted := 0
 CREATE:
@@ -989,7 +1053,8 @@ CREATE:
 		return container, nil
 	}
 
-	d.logger.Printf("[DEBUG] driver.docker: failed to create container %q (attempt %d): %v", config.Name, attempted+1, createErr)
+	d.logger.Printf("[DEBUG] driver.docker: failed to create container %q from image %q (ID: %q) (attempt %d): %v",
+		config.Name, d.driverConfig.ImageName, d.imageID, attempted+1, createErr)
 	if strings.Contains(strings.ToLower(createErr.Error()), "container already exists") {
 		containers, err := client.ListContainers(docker.ListContainersOptions{
 			All: true,
@@ -1052,7 +1117,7 @@ CREATE:
 
 // startContainer starts the passed container. It attempts to handle any
 // transient Docker errors.
-func (d *DockerDriver) startContainer(c *docker.Container) *structs.RecoverableError {
+func (d *DockerDriver) startContainer(c *docker.Container) error {
 	// Start a container
 	attempted := 0
 START:
@@ -1076,8 +1141,6 @@ START:
 }
 
 func (d *DockerDriver) Open(ctx *ExecContext, handleID string) (DriverHandle, error) {
-	cleanupImage := d.config.ReadBoolDefault("docker.cleanup.image", true)
-
 	// Split the handle
 	pidBytes := []byte(strings.TrimPrefix(handleID, "DOCKER:"))
 	pid := &dockerPID{}
@@ -1133,9 +1196,7 @@ func (d *DockerDriver) Open(ctx *ExecContext, handleID string) (DriverHandle, er
 		waitClient:     waitClient,
 		executor:       exec,
 		pluginClient:   pluginClient,
-		cleanupImage:   cleanupImage,
 		logger:         d.logger,
-		imageID:        pid.ImageID,
 		containerID:    pid.ContainerID,
 		version:        pid.Version,
 		killTimeout:    pid.KillTimeout,
@@ -1156,7 +1217,6 @@ func (h *DockerHandle) ID() string {
 	// Return a handle to the PID
 	pid := dockerPID{
 		Version:        h.version,
-		ImageID:        h.imageID,
 		ContainerID:    h.containerID,
 		KillTimeout:    h.killTimeout,
 		MaxKillTimeout: h.maxKillTimeout,
@@ -1273,13 +1333,6 @@ func (h *DockerHandle) run() {
 		h.logger.Printf("[ERR] driver.docker: error removing container: %v", err)
 	}
 
-	// Cleanup the image
-	if h.cleanupImage {
-		if err := h.client.RemoveImage(h.imageID); err != nil {
-			h.logger.Printf("[DEBUG] driver.docker: error removing image: %v", err)
-		}
-	}
-
 	// Send the results
 	h.waitCh <- dstructs.NewWaitResult(exitCode, 0, werr)
 	close(h.waitCh)
diff --git a/client/driver/docker_test.go b/client/driver/docker_test.go
index 9b1e129d5..ee38b629a 100644
--- a/client/driver/docker_test.go
+++ b/client/driver/docker_test.go
@@ -99,7 +99,8 @@ func dockerSetupWithClient(t *testing.T, task *structs.Task, client *docker.Clie
 	driver := NewDockerDriver(tctx.DriverCtx)
 	copyImage(t, tctx.ExecCtx.TaskDir, "busybox.tar")
 
-	if err := driver.Prestart(tctx.ExecCtx, task); err != nil {
+	_, err := driver.Prestart(tctx.ExecCtx, task)
+	if err != nil {
 		tctx.AllocDir.Destroy()
 		t.Fatalf("error in prestart: %v", err)
 	}
@@ -182,9 +183,11 @@ func TestDockerDriver_StartOpen_Wait(t *testing.T) {
 	d := NewDockerDriver(ctx.DriverCtx)
 	copyImage(t, ctx.ExecCtx.TaskDir, "busybox.tar")
 
-	if err := d.Prestart(ctx.ExecCtx, task); err != nil {
+	_, err := d.Prestart(ctx.ExecCtx, task)
+	if err != nil {
 		t.Fatalf("error in prestart: %v", err)
 	}
+
 	handle, err := d.Start(ctx.ExecCtx, task)
 	if err != nil {
 		t.Fatalf("err: %v", err)
@@ -276,7 +279,8 @@ func TestDockerDriver_Start_LoadImage(t *testing.T) {
 	// Copy the image into the task's directory
 	copyImage(t, ctx.ExecCtx.TaskDir, "busybox.tar")
 
-	if err := d.Prestart(ctx.ExecCtx, task); err != nil {
+	_, err := d.Prestart(ctx.ExecCtx, task)
+	if err != nil {
 		t.Fatalf("error in prestart: %v", err)
 	}
 	handle, err := d.Start(ctx.ExecCtx, task)
@@ -340,7 +344,7 @@ func TestDockerDriver_Start_BadPull_Recoverable(t *testing.T) {
 	defer ctx.AllocDir.Destroy()
 	d := NewDockerDriver(ctx.DriverCtx)
 
-	err := d.Prestart(ctx.ExecCtx, task)
+	_, err := d.Prestart(ctx.ExecCtx, task)
 	if err == nil {
 		t.Fatalf("want error in prestart: %v", err)
 	}
@@ -391,7 +395,8 @@ func TestDockerDriver_Start_Wait_AllocDir(t *testing.T) {
 	d := NewDockerDriver(ctx.DriverCtx)
 	copyImage(t, ctx.ExecCtx.TaskDir, "busybox.tar")
 
-	if err := d.Prestart(ctx.ExecCtx, task); err != nil {
+	_, err := d.Prestart(ctx.ExecCtx, task)
+	if err != nil {
 		t.Fatalf("error in prestart: %v", err)
 	}
 	handle, err := d.Start(ctx.ExecCtx, task)
@@ -477,7 +482,6 @@ func TestDockerDriver_StartN(t *testing.T) {
 	t.Logf("Starting %d tasks", len(taskList))
 
 	// Let's spin up a bunch of things
-	var err error
 	for idx, task := range taskList {
 		ctx := testDriverContexts(t, task)
 		ctx.DriverCtx.config.Options = map[string]string{"docker.cleanup.image": "false"}
@@ -485,7 +489,8 @@ func TestDockerDriver_StartN(t *testing.T) {
 		d := NewDockerDriver(ctx.DriverCtx)
 		copyImage(t, ctx.ExecCtx.TaskDir, "busybox.tar")
 
-		if err := d.Prestart(ctx.ExecCtx, task); err != nil {
+		_, err := d.Prestart(ctx.ExecCtx, task)
+		if err != nil {
 			t.Fatalf("error in prestart #%d: %v", idx+1, err)
 		}
 		handles[idx], err = d.Start(ctx.ExecCtx, task)
@@ -535,7 +540,6 @@ func TestDockerDriver_StartNVersions(t *testing.T) {
 	t.Logf("Starting %d tasks", len(taskList))
 
 	// Let's spin up a bunch of things
-	var err error
 	for idx, task := range taskList {
 		ctx := testDriverContexts(t, task)
 		ctx.DriverCtx.config.Options = map[string]string{"docker.cleanup.image": "false"}
@@ -545,7 +549,8 @@ func TestDockerDriver_StartNVersions(t *testing.T) {
 		copyImage(t, ctx.ExecCtx.TaskDir, "busybox_musl.tar")
 		copyImage(t, ctx.ExecCtx.TaskDir, "busybox_glibc.tar")
 
-		if err := d.Prestart(ctx.ExecCtx, task); err != nil {
+		_, err := d.Prestart(ctx.ExecCtx, task)
+		if err != nil {
 			t.Fatalf("error in prestart #%d: %v", idx+1, err)
 		}
 		handles[idx], err = d.Start(ctx.ExecCtx, task)
@@ -708,7 +713,7 @@ func TestDockerDriver_ForcePull_IsInvalidConfig(t *testing.T) {
 	ctx.DriverCtx.config.Options = map[string]string{"docker.cleanup.image": "false"}
 	driver := NewDockerDriver(ctx.DriverCtx)
 
-	if err := driver.Prestart(ctx.ExecCtx, task); err == nil {
+	if _, err := driver.Prestart(ctx.ExecCtx, task); err == nil {
 		t.Fatalf("error expected in prestart")
 	}
 }
@@ -916,7 +921,8 @@ func TestDockerDriver_User(t *testing.T) {
 	defer ctx.AllocDir.Destroy()
 	copyImage(t, ctx.ExecCtx.TaskDir, "busybox.tar")
 
-	if err := driver.Prestart(ctx.ExecCtx, task); err != nil {
+	_, err := driver.Prestart(ctx.ExecCtx, task)
+	if err != nil {
 		t.Fatalf("error in prestart: %v", err)
 	}
 
@@ -1072,7 +1078,8 @@ done
 		fmt.Errorf("Failed to write data")
 	}
 
-	if err := d.Prestart(ctx.ExecCtx, task); err != nil {
+	_, err := d.Prestart(ctx.ExecCtx, task)
+	if err != nil {
 		t.Fatalf("error in prestart: %v", err)
 	}
 	handle, err := d.Start(ctx.ExecCtx, task)
@@ -1194,7 +1201,8 @@ func TestDockerDriver_VolumesDisabled(t *testing.T) {
 		task, driver, execCtx, _, cleanup := setupDockerVolumes(t, cfg, tmpvol)
 		defer cleanup()
 
-		if err := driver.Prestart(execCtx, task); err != nil {
+		_, err = driver.Prestart(execCtx, task)
+		if err != nil {
 			t.Fatalf("error in prestart: %v", err)
 		}
 		if _, err := driver.Start(execCtx, task); err == nil {
@@ -1207,7 +1215,8 @@ func TestDockerDriver_VolumesDisabled(t *testing.T) {
 		task, driver, execCtx, fn, cleanup := setupDockerVolumes(t, cfg, ".")
 		defer cleanup()
 
-		if err := driver.Prestart(execCtx, task); err != nil {
+		_, err := driver.Prestart(execCtx, task)
+		if err != nil {
 			t.Fatalf("error in prestart: %v", err)
 		}
 		handle, err := driver.Start(execCtx, task)
@@ -1243,7 +1252,8 @@ func TestDockerDriver_VolumesEnabled(t *testing.T) {
 	task, driver, execCtx, hostpath, cleanup := setupDockerVolumes(t, cfg, tmpvol)
 	defer cleanup()
 
-	if err := driver.Prestart(execCtx, task); err != nil {
+	_, err = driver.Prestart(execCtx, task)
+	if err != nil {
 		t.Fatalf("error in prestart: %v", err)
 	}
 	handle, err := driver.Start(execCtx, task)
@@ -1266,6 +1276,56 @@ func TestDockerDriver_VolumesEnabled(t *testing.T) {
 	}
 }
 
+// TestDockerDriver_Cleanup ensures Cleanup removes only downloaded images.
+func TestDockerDriver_Cleanup(t *testing.T) {
+	if !testutil.DockerIsConnected(t) {
+		t.SkipNow()
+	}
+
+	imageName := "hello-world:latest"
+	task := &structs.Task{
+		Name:   "cleanup_test",
+		Driver: "docker",
+		Config: map[string]interface{}{
+			"image": imageName,
+		},
+	}
+	tctx := testDriverContexts(t, task)
+	defer tctx.AllocDir.Destroy()
+
+	// Run Prestart
+	driver := NewDockerDriver(tctx.DriverCtx).(*DockerDriver)
+	res, err := driver.Prestart(tctx.ExecCtx, task)
+	if err != nil {
+		t.Fatalf("error in prestart: %v", err)
+	}
+	if len(res.Resources) == 0 || len(res.Resources[dockerImageResKey]) == 0 {
+		t.Fatalf("no created resources: %#v", res)
+	}
+
+	// Cleanup
+	rescopy := res.Copy()
+	if err := driver.Cleanup(tctx.ExecCtx, rescopy); err != nil {
+		t.Fatalf("Cleanup failed: %v", err)
+	}
+
+	// Make sure rescopy is updated
+	if len(rescopy.Resources) > 0 {
+		t.Errorf("Cleanup should have cleared resource map: %#v", rescopy.Resources)
+	}
+
+	// Ensure image was removed
+	if _, err := client.InspectImage(driver.driverConfig.ImageName); err == nil {
+		t.Fatalf("image exists but should have been removed. Does another %v container exist?", imageName)
+	}
+
+	// The image doesn't exist which shouldn't be an error when calling
+	// Cleanup, so call it again to make sure.
+	if err := driver.Cleanup(tctx.ExecCtx, res.Copy()); err != nil {
+		t.Fatalf("Cleanup failed: %v", err)
+	}
+}
+
 func copyImage(t *testing.T, taskDir *allocdir.TaskDir, image string) {
 	dst := filepath.Join(taskDir.LocalDir, image)
 	copyFile(filepath.Join("./test-resources/docker", image), dst, t)
diff --git a/client/driver/driver.go b/client/driver/driver.go
index c4560e978..c91d41e5d 100644
--- a/client/driver/driver.go
+++ b/client/driver/driver.go
@@ -51,6 +51,100 @@ func NewDriver(name string, ctx *DriverContext) (Driver, error) {
 // Factory is used to instantiate a new Driver
 type Factory func(*DriverContext) Driver
 
+// CreatedResources is a map of resources (eg downloaded images) created by a driver
+// that must be cleaned up.
+type CreatedResources struct {
+	Resources map[string][]string
+}
+
+func NewCreatedResources() *CreatedResources {
+	return &CreatedResources{Resources: make(map[string][]string)}
+}
+
+// Add a new resource if it doesn't already exist.
+func (r *CreatedResources) Add(k, v string) {
+	if r.Resources == nil {
+		r.Resources = map[string][]string{k: []string{v}}
+		return
+	}
+	existing, ok := r.Resources[k]
+	if !ok {
+		// Key doesn't exist, create it
+		r.Resources[k] = []string{v}
+		return
+	}
+	for _, item := range existing {
+		if item == v {
+			// resource exists, return
+			return
+		}
+	}
+
+	// Resource type exists but value did not, append it
+	r.Resources[k] = append(existing, v)
+	return
+}
+
+// Remove a resource. Return true if removed, otherwise false.
+//
+// Removes the entire key if the needle is the last value in the list.
+func (r *CreatedResources) Remove(k, needle string) bool {
+	haystack := r.Resources[k]
+	for i, item := range haystack {
+		if item == needle {
+			r.Resources[k] = append(haystack[:i], haystack[i+1:]...)
+			if len(r.Resources[k]) == 0 {
+				delete(r.Resources, k)
+			}
+			return true
+		}
+	}
+	return false
+}
+
+// Copy returns a new deep copy of CreatedResrouces.
+func (r *CreatedResources) Copy() *CreatedResources {
+	newr := CreatedResources{
+		Resources: make(map[string][]string, len(r.Resources)),
+	}
+	for k, v := range r.Resources {
+		newv := make([]string, len(v))
+		copy(newv, v)
+		newr.Resources[k] = newv
+	}
+	return &newr
+}
+
+// Merge another CreatedResources into this one. If the other CreatedResources
+// is nil this method is a noop.
+func (r *CreatedResources) Merge(o *CreatedResources) {
+	if o == nil {
+		return
+	}
+
+	for k, v := range o.Resources {
+		// New key
+		if len(r.Resources[k]) == 0 {
+			r.Resources[k] = v
+			continue
+		}
+
+		// Existing key
+	OUTER:
+		for _, item := range v {
+			for _, existing := range r.Resources[k] {
+				if item == existing {
+					// Found it, move on
+					continue OUTER
+				}
+			}
+
+			// New item, append it
+			r.Resources[k] = append(r.Resources[k], item)
+		}
+	}
+}
+
 // Driver is used for execution of tasks. This allows Nomad
 // to support many pluggable implementations of task drivers.
 // Examples could include LXC, Docker, Qemu, etc.
@@ -60,7 +154,9 @@ type Driver interface {
 
 	// Prestart prepares the task environment and performs expensive
 	// intialization steps like downloading images.
-	Prestart(*ExecContext, *structs.Task) error
+	//
+	// CreatedResources may be non-nil even when an error occurs.
+	Prestart(*ExecContext, *structs.Task) (*CreatedResources, error)
 
 	// Start is used to being task execution
 	Start(ctx *ExecContext, task *structs.Task) (DriverHandle, error)
@@ -68,6 +164,14 @@ type Driver interface {
 	// Open is used to re-open a handle to a task
 	Open(ctx *ExecContext, handleID string) (DriverHandle, error)
 
+	// Cleanup is called to remove resources which were created for a task
+	// and no longer needed.
+	//
+	// If Cleanup returns a recoverable error it may be retried. On retry
+	// it will be passed the same CreatedResources, so all successfully
+	// cleaned up resources should be removed.
+	Cleanup(*ExecContext, *CreatedResources) error
+
 	// Drivers must validate their configuration
 	Validate(map[string]interface{}) error
 
diff --git a/client/driver/driver_test.go b/client/driver/driver_test.go
index 9db1148c8..6cca2b4c7 100644
--- a/client/driver/driver_test.go
+++ b/client/driver/driver_test.go
@@ -256,3 +256,80 @@ func TestMapMergeStrStr(t *testing.T) {
 		t.Errorf("\nExpected\n%+v\nGot\n%+v\n", d, c)
 	}
 }
+
+func TestCreatedResources_AddMerge(t *testing.T) {
+	res1 := NewCreatedResources()
+	res1.Add("k1", "v1")
+	res1.Add("k1", "v2")
+	res1.Add("k1", "v1")
+	res1.Add("k2", "v1")
+
+	expected := map[string][]string{
+		"k1": {"v1", "v2"},
+		"k2": {"v1"},
+	}
+	if !reflect.DeepEqual(expected, res1.Resources) {
+		t.Fatalf("1.  %#v != expected %#v", res1.Resources, expected)
+	}
+
+	// Make sure merging nil works
+	var res2 *CreatedResources
+	res1.Merge(res2)
+	if !reflect.DeepEqual(expected, res1.Resources) {
+		t.Fatalf("2.  %#v != expected %#v", res1.Resources, expected)
+	}
+
+	// Make sure a normal merge works
+	res2 = NewCreatedResources()
+	res2.Add("k1", "v3")
+	res2.Add("k2", "v1")
+	res2.Add("k3", "v3")
+	res1.Merge(res2)
+
+	expected = map[string][]string{
+		"k1": {"v1", "v2", "v3"},
+		"k2": {"v1"},
+		"k3": {"v3"},
+	}
+	if !reflect.DeepEqual(expected, res1.Resources) {
+		t.Fatalf("3.  %#v != expected %#v", res1.Resources, expected)
+	}
+}
+
+func TestCreatedResources_CopyRemove(t *testing.T) {
+	res1 := NewCreatedResources()
+	res1.Add("k1", "v1")
+	res1.Add("k1", "v2")
+	res1.Add("k1", "v3")
+	res1.Add("k2", "v1")
+
+	// Assert Copy creates a deep copy
+	res2 := res1.Copy()
+
+	if !reflect.DeepEqual(res1, res2) {
+		t.Fatalf("%#v != %#v", res1, res2)
+	}
+
+	// Assert removing v1 from k1 returns true and updates Resources slice
+	if removed := res2.Remove("k1", "v1"); !removed {
+		t.Fatalf("expected v1 to be removed: %#v", res2)
+	}
+
+	if expected := []string{"v2", "v3"}; !reflect.DeepEqual(expected, res2.Resources["k1"]) {
+		t.Fatalf("unpexpected list for k1: %#v", res2.Resources["k1"])
+	}
+
+	// Assert removing the only value from a key removes the key
+	if removed := res2.Remove("k2", "v1"); !removed {
+		t.Fatalf("expected v1 to be removed from k2: %#v", res2.Resources)
+	}
+
+	if _, found := res2.Resources["k2"]; found {
+		t.Fatalf("k2 should have been removed from Resources: %#v", res2.Resources)
+	}
+
+	// Make sure res1 wasn't updated
+	if reflect.DeepEqual(res1, res2) {
+		t.Fatalf("res1 should not equal res2: #%v", res1)
+	}
+}
diff --git a/client/driver/exec.go b/client/driver/exec.go
index 3d1930e16..1a1b7457f 100644
--- a/client/driver/exec.go
+++ b/client/driver/exec.go
@@ -92,8 +92,8 @@ func (d *ExecDriver) Periodic() (bool, time.Duration) {
 	return true, 15 * time.Second
 }
 
-func (d *ExecDriver) Prestart(ctx *ExecContext, task *structs.Task) error {
-	return nil
+func (d *ExecDriver) Prestart(*ExecContext, *structs.Task) (*CreatedResources, error) {
+	return nil, nil
 }
 
 func (d *ExecDriver) Start(ctx *ExecContext, task *structs.Task) (DriverHandle, error) {
@@ -167,6 +167,8 @@ func (d *ExecDriver) Start(ctx *ExecContext, task *structs.Task) (DriverHandle,
 	return h, nil
 }
 
+func (d *ExecDriver) Cleanup(*ExecContext, *CreatedResources) error { return nil }
+
 type execId struct {
 	Version         string
 	KillTimeout     time.Duration
diff --git a/client/driver/exec_test.go b/client/driver/exec_test.go
index 7db6696ef..279480435 100644
--- a/client/driver/exec_test.go
+++ b/client/driver/exec_test.go
@@ -67,7 +67,7 @@ func TestExecDriver_StartOpen_Wait(t *testing.T) {
 	defer ctx.AllocDir.Destroy()
 	d := NewExecDriver(ctx.DriverCtx)
 
-	if err := d.Prestart(ctx.ExecCtx, task); err != nil {
+	if _, err := d.Prestart(ctx.ExecCtx, task); err != nil {
 		t.Fatalf("prestart err: %v", err)
 	}
 	handle, err := d.Start(ctx.ExecCtx, task)
@@ -111,7 +111,7 @@ func TestExecDriver_KillUserPid_OnPluginReconnectFailure(t *testing.T) {
 	defer ctx.AllocDir.Destroy()
 	d := NewExecDriver(ctx.DriverCtx)
 
-	if err := d.Prestart(ctx.ExecCtx, task); err != nil {
+	if _, err := d.Prestart(ctx.ExecCtx, task); err != nil {
 		t.Fatalf("prestart err: %v", err)
 	}
 	handle, err := d.Start(ctx.ExecCtx, task)
@@ -145,12 +145,20 @@ func TestExecDriver_KillUserPid_OnPluginReconnectFailure(t *testing.T) {
 		handle2.Kill()
 		t.Fatalf("expected handle2 to be nil")
 	}
+
 	// Test if the userpid is still present
-	userProc, err := os.FindProcess(id.UserPid)
+	userProc, _ := os.FindProcess(id.UserPid)
 
-	err = userProc.Signal(syscall.Signal(0))
+	for retry := 3; retry > 0; retry-- {
+		if err = userProc.Signal(syscall.Signal(0)); err != nil {
+			// Process is gone as expected; exit
+			return
+		}
 
-	if err == nil {
+		// Killing processes is async; wait and check again
+		time.Sleep(time.Second)
+	}
+	if err = userProc.Signal(syscall.Signal(0)); err == nil {
 		t.Fatalf("expected user process to die")
 	}
 }
@@ -175,7 +183,7 @@ func TestExecDriver_Start_Wait(t *testing.T) {
 	defer ctx.AllocDir.Destroy()
 	d := NewExecDriver(ctx.DriverCtx)
 
-	if err := d.Prestart(ctx.ExecCtx, task); err != nil {
+	if _, err := d.Prestart(ctx.ExecCtx, task); err != nil {
 		t.Fatalf("prestart err: %v", err)
 	}
 	handle, err := d.Start(ctx.ExecCtx, task)
@@ -229,7 +237,7 @@ func TestExecDriver_Start_Wait_AllocDir(t *testing.T) {
 	defer ctx.AllocDir.Destroy()
 	d := NewExecDriver(ctx.DriverCtx)
 
-	if err := d.Prestart(ctx.ExecCtx, task); err != nil {
+	if _, err := d.Prestart(ctx.ExecCtx, task); err != nil {
 		t.Fatalf("prestart err: %v", err)
 	}
 	handle, err := d.Start(ctx.ExecCtx, task)
@@ -283,7 +291,7 @@ func TestExecDriver_Start_Kill_Wait(t *testing.T) {
 	defer ctx.AllocDir.Destroy()
 	d := NewExecDriver(ctx.DriverCtx)
 
-	if err := d.Prestart(ctx.ExecCtx, task); err != nil {
+	if _, err := d.Prestart(ctx.ExecCtx, task); err != nil {
 		t.Fatalf("prestart err: %v", err)
 	}
 	handle, err := d.Start(ctx.ExecCtx, task)
@@ -349,7 +357,7 @@ done
 		fmt.Errorf("Failed to write data")
 	}
 
-	if err := d.Prestart(ctx.ExecCtx, task); err != nil {
+	if _, err := d.Prestart(ctx.ExecCtx, task); err != nil {
 		t.Fatalf("prestart err: %v", err)
 	}
 	handle, err := d.Start(ctx.ExecCtx, task)
@@ -414,7 +422,7 @@ func TestExecDriverUser(t *testing.T) {
 	defer ctx.AllocDir.Destroy()
 	d := NewExecDriver(ctx.DriverCtx)
 
-	if err := d.Prestart(ctx.ExecCtx, task); err != nil {
+	if _, err := d.Prestart(ctx.ExecCtx, task); err != nil {
 		t.Fatalf("prestart err: %v", err)
 	}
 	handle, err := d.Start(ctx.ExecCtx, task)
diff --git a/client/driver/java.go b/client/driver/java.go
index 366b891e8..0a5240f81 100644
--- a/client/driver/java.go
+++ b/client/driver/java.go
@@ -167,8 +167,8 @@ func (d *JavaDriver) Fingerprint(cfg *config.Config, node *structs.Node) (bool,
 	return true, nil
 }
 
-func (d *JavaDriver) Prestart(ctx *ExecContext, task *structs.Task) error {
-	return nil
+func (d *JavaDriver) Prestart(*ExecContext, *structs.Task) (*CreatedResources, error) {
+	return nil, nil
 }
 
 func NewJavaDriverConfig(task *structs.Task, env *env.TaskEnvironment) (*JavaDriverConfig, error) {
@@ -293,6 +293,8 @@ func (d *JavaDriver) Start(ctx *ExecContext, task *structs.Task) (DriverHandle,
 	return h, nil
 }
 
+func (d *JavaDriver) Cleanup(*ExecContext, *CreatedResources) error { return nil }
+
 // cgroupsMounted returns true if the cgroups are mounted on a system otherwise
 // returns false
 func (d *JavaDriver) cgroupsMounted(node *structs.Node) bool {
diff --git a/client/driver/java_test.go b/client/driver/java_test.go
index 8fa5ec26b..53285d83b 100644
--- a/client/driver/java_test.go
+++ b/client/driver/java_test.go
@@ -94,7 +94,7 @@ func TestJavaDriver_StartOpen_Wait(t *testing.T) {
 	dst := ctx.ExecCtx.TaskDir.Dir
 	copyFile("./test-resources/java/demoapp.jar", filepath.Join(dst, "demoapp.jar"), t)
 
-	if err := d.Prestart(ctx.ExecCtx, task); err != nil {
+	if _, err := d.Prestart(ctx.ExecCtx, task); err != nil {
 		t.Fatalf("prestart err: %v", err)
 	}
 	handle, err := d.Start(ctx.ExecCtx, task)
@@ -149,7 +149,7 @@ func TestJavaDriver_Start_Wait(t *testing.T) {
 	dst := ctx.ExecCtx.TaskDir.Dir
 	copyFile("./test-resources/java/demoapp.jar", filepath.Join(dst, "demoapp.jar"), t)
 
-	if err := d.Prestart(ctx.ExecCtx, task); err != nil {
+	if _, err := d.Prestart(ctx.ExecCtx, task); err != nil {
 		t.Fatalf("prestart err: %v", err)
 	}
 	handle, err := d.Start(ctx.ExecCtx, task)
@@ -215,7 +215,7 @@ func TestJavaDriver_Start_Kill_Wait(t *testing.T) {
 	dst := ctx.ExecCtx.TaskDir.Dir
 	copyFile("./test-resources/java/demoapp.jar", filepath.Join(dst, "demoapp.jar"), t)
 
-	if err := d.Prestart(ctx.ExecCtx, task); err != nil {
+	if _, err := d.Prestart(ctx.ExecCtx, task); err != nil {
 		t.Fatalf("prestart err: %v", err)
 	}
 	handle, err := d.Start(ctx.ExecCtx, task)
@@ -277,7 +277,7 @@ func TestJavaDriver_Signal(t *testing.T) {
 	dst := ctx.ExecCtx.TaskDir.Dir
 	copyFile("./test-resources/java/demoapp.jar", filepath.Join(dst, "demoapp.jar"), t)
 
-	if err := d.Prestart(ctx.ExecCtx, task); err != nil {
+	if _, err := d.Prestart(ctx.ExecCtx, task); err != nil {
 		t.Fatalf("prestart err: %v", err)
 	}
 	handle, err := d.Start(ctx.ExecCtx, task)
@@ -336,7 +336,7 @@ func TestJavaDriverUser(t *testing.T) {
 	defer ctx.AllocDir.Destroy()
 	d := NewJavaDriver(ctx.DriverCtx)
 
-	if err := d.Prestart(ctx.ExecCtx, task); err != nil {
+	if _, err := d.Prestart(ctx.ExecCtx, task); err != nil {
 		t.Fatalf("prestart err: %v", err)
 	}
 	handle, err := d.Start(ctx.ExecCtx, task)
diff --git a/client/driver/lxc.go b/client/driver/lxc.go
index 8e9f7aa56..5b541a3f8 100644
--- a/client/driver/lxc.go
+++ b/client/driver/lxc.go
@@ -171,8 +171,8 @@ func (d *LxcDriver) Fingerprint(cfg *config.Config, node *structs.Node) (bool, e
 	return true, nil
 }
 
-func (d *LxcDriver) Prestart(ctx *ExecContext, task *structs.Task) error {
-	return nil
+func (d *LxcDriver) Prestart(*ExecContext, *structs.Task) (*CreatedResources, error) {
+	return nil, nil
 }
 
 // Start starts the LXC Driver
@@ -286,6 +286,8 @@ func (d *LxcDriver) Start(ctx *ExecContext, task *structs.Task) (DriverHandle, e
 	return &handle, nil
 }
 
+func (d *LxcDriver) Cleanup(*ExecContext, *CreatedResources) error { return nil }
+
 // Open creates the driver to monitor an existing LXC container
 func (d *LxcDriver) Open(ctx *ExecContext, handleID string) (DriverHandle, error) {
 	pid := &lxcPID{}
diff --git a/client/driver/lxc_test.go b/client/driver/lxc_test.go
index 5f9b8852a..b6188bc6b 100644
--- a/client/driver/lxc_test.go
+++ b/client/driver/lxc_test.go
@@ -72,7 +72,7 @@ func TestLxcDriver_Start_Wait(t *testing.T) {
 	defer ctx.AllocDir.Destroy()
 	d := NewLxcDriver(ctx.DriverCtx)
 
-	if err := d.Prestart(ctx.ExecCtx, task); err != nil {
+	if _, err := d.Prestart(ctx.ExecCtx, task); err != nil {
 		t.Fatalf("prestart err: %v", err)
 	}
 	handle, err := d.Start(ctx.ExecCtx, task)
@@ -148,7 +148,7 @@ func TestLxcDriver_Open_Wait(t *testing.T) {
 	defer ctx.AllocDir.Destroy()
 	d := NewLxcDriver(ctx.DriverCtx)
 
-	if err := d.Prestart(ctx.ExecCtx, task); err != nil {
+	if _, err := d.Prestart(ctx.ExecCtx, task); err != nil {
 		t.Fatalf("prestart err: %v", err)
 	}
 	handle, err := d.Start(ctx.ExecCtx, task)
diff --git a/client/driver/mock_driver.go b/client/driver/mock_driver.go
index b6d5737b5..b81eca64f 100644
--- a/client/driver/mock_driver.go
+++ b/client/driver/mock_driver.go
@@ -8,6 +8,7 @@ import (
 	"fmt"
 	"log"
 	"os"
+	"strconv"
 	"time"
 
 	"github.com/mitchellh/mapstructure"
@@ -62,6 +63,8 @@ type MockDriverConfig struct {
 type MockDriver struct {
 	DriverContext
 	fingerprint.StaticFingerprinter
+
+	cleanupFailNum int
 }
 
 // NewMockDriver is a factory method which returns a new Mock Driver
@@ -79,8 +82,8 @@ func (d *MockDriver) FSIsolation() cstructs.FSIsolation {
 	return cstructs.FSIsolationNone
 }
 
-func (d *MockDriver) Prestart(ctx *ExecContext, task *structs.Task) error {
-	return nil
+func (d *MockDriver) Prestart(*ExecContext, *structs.Task) (*CreatedResources, error) {
+	return nil, nil
 }
 
 // Start starts the mock driver
@@ -124,6 +127,24 @@ func (m *MockDriver) Start(ctx *ExecContext, task *structs.Task) (DriverHandle,
 	return &h, nil
 }
 
+// Cleanup deletes all keys except for Config.Options["cleanup_fail_on"] for
+// Config.Options["cleanup_fail_num"] times. For failures it will return a
+// recoverable error.
+func (m *MockDriver) Cleanup(ctx *ExecContext, res *CreatedResources) error {
+	var err error
+	failn, _ := strconv.Atoi(m.config.Options["cleanup_fail_num"])
+	failk := m.config.Options["cleanup_fail_on"]
+	for k := range res.Resources {
+		if k == failk && m.cleanupFailNum < failn {
+			m.cleanupFailNum++
+			err = structs.NewRecoverableError(fmt.Errorf("mock_driver failure on %q call %d/%d", k, m.cleanupFailNum, failn), true)
+		} else {
+			delete(res.Resources, k)
+		}
+	}
+	return err
+}
+
 // Validate validates the mock driver configuration
 func (m *MockDriver) Validate(map[string]interface{}) error {
 	return nil
diff --git a/client/driver/qemu.go b/client/driver/qemu.go
index 98575c5e6..746740829 100644
--- a/client/driver/qemu.go
+++ b/client/driver/qemu.go
@@ -136,8 +136,8 @@ func (d *QemuDriver) Fingerprint(cfg *config.Config, node *structs.Node) (bool,
 	return true, nil
 }
 
-func (d *QemuDriver) Prestart(ctx *ExecContext, task *structs.Task) error {
-	return nil
+func (d *QemuDriver) Prestart(*ExecContext, *structs.Task) (*CreatedResources, error) {
+	return nil, nil
 }
 
 // Run an existing Qemu image. Start() will pull down an existing, valid Qemu
@@ -336,6 +336,8 @@ func (d *QemuDriver) Open(ctx *ExecContext, handleID string) (DriverHandle, erro
 	return h, nil
 }
 
+func (d *QemuDriver) Cleanup(*ExecContext, *CreatedResources) error { return nil }
+
 func (h *qemuHandle) ID() string {
 	id := qemuId{
 		Version:        h.version,
diff --git a/client/driver/qemu_test.go b/client/driver/qemu_test.go
index 16ba2ba31..8e7393060 100644
--- a/client/driver/qemu_test.go
+++ b/client/driver/qemu_test.go
@@ -80,7 +80,7 @@ func TestQemuDriver_StartOpen_Wait(t *testing.T) {
 	dst := ctx.ExecCtx.TaskDir.Dir
 	copyFile("./test-resources/qemu/linux-0.2.img", filepath.Join(dst, "linux-0.2.img"), t)
 
-	if err := d.Prestart(ctx.ExecCtx, task); err != nil {
+	if _, err := d.Prestart(ctx.ExecCtx, task); err != nil {
 		t.Fatalf("Prestart faild: %v", err)
 	}
 
@@ -146,7 +146,7 @@ func TestQemuDriverUser(t *testing.T) {
 	defer ctx.AllocDir.Destroy()
 	d := NewQemuDriver(ctx.DriverCtx)
 
-	if err := d.Prestart(ctx.ExecCtx, task); err != nil {
+	if _, err := d.Prestart(ctx.ExecCtx, task); err != nil {
 		t.Fatalf("Prestart faild: %v", err)
 	}
 
diff --git a/client/driver/raw_exec.go b/client/driver/raw_exec.go
index eb4e1b790..e7dbe2bb2 100644
--- a/client/driver/raw_exec.go
+++ b/client/driver/raw_exec.go
@@ -106,8 +106,8 @@ func (d *RawExecDriver) Fingerprint(cfg *config.Config, node *structs.Node) (boo
 	return false, nil
 }
 
-func (d *RawExecDriver) Prestart(ctx *ExecContext, task *structs.Task) error {
-	return nil
+func (d *RawExecDriver) Prestart(*ExecContext, *structs.Task) (*CreatedResources, error) {
+	return nil, nil
 }
 
 func (d *RawExecDriver) Start(ctx *ExecContext, task *structs.Task) (DriverHandle, error) {
@@ -177,6 +177,8 @@ func (d *RawExecDriver) Start(ctx *ExecContext, task *structs.Task) (DriverHandl
 	return h, nil
 }
 
+func (d *RawExecDriver) Cleanup(*ExecContext, *CreatedResources) error { return nil }
+
 type rawExecId struct {
 	Version        string
 	KillTimeout    time.Duration
diff --git a/client/driver/raw_exec_test.go b/client/driver/raw_exec_test.go
index c91d1c52f..c3efcfd71 100644
--- a/client/driver/raw_exec_test.go
+++ b/client/driver/raw_exec_test.go
@@ -77,7 +77,7 @@ func TestRawExecDriver_StartOpen_Wait(t *testing.T) {
 	defer ctx.AllocDir.Destroy()
 	d := NewRawExecDriver(ctx.DriverCtx)
 
-	if err := d.Prestart(ctx.ExecCtx, task); err != nil {
+	if _, err := d.Prestart(ctx.ExecCtx, task); err != nil {
 		t.Fatalf("prestart err: %v", err)
 	}
 	handle, err := d.Start(ctx.ExecCtx, task)
@@ -126,7 +126,7 @@ func TestRawExecDriver_Start_Wait(t *testing.T) {
 	defer ctx.AllocDir.Destroy()
 	d := NewRawExecDriver(ctx.DriverCtx)
 
-	if err := d.Prestart(ctx.ExecCtx, task); err != nil {
+	if _, err := d.Prestart(ctx.ExecCtx, task); err != nil {
 		t.Fatalf("prestart err: %v", err)
 	}
 	handle, err := d.Start(ctx.ExecCtx, task)
@@ -180,7 +180,7 @@ func TestRawExecDriver_Start_Wait_AllocDir(t *testing.T) {
 	defer ctx.AllocDir.Destroy()
 	d := NewRawExecDriver(ctx.DriverCtx)
 
-	if err := d.Prestart(ctx.ExecCtx, task); err != nil {
+	if _, err := d.Prestart(ctx.ExecCtx, task); err != nil {
 		t.Fatalf("prestart err: %v", err)
 	}
 	handle, err := d.Start(ctx.ExecCtx, task)
@@ -233,7 +233,7 @@ func TestRawExecDriver_Start_Kill_Wait(t *testing.T) {
 	defer ctx.AllocDir.Destroy()
 	d := NewRawExecDriver(ctx.DriverCtx)
 
-	if err := d.Prestart(ctx.ExecCtx, task); err != nil {
+	if _, err := d.Prestart(ctx.ExecCtx, task); err != nil {
 		t.Fatalf("prestart err: %v", err)
 	}
 	handle, err := d.Start(ctx.ExecCtx, task)
@@ -286,7 +286,7 @@ func TestRawExecDriverUser(t *testing.T) {
 	defer ctx.AllocDir.Destroy()
 	d := NewRawExecDriver(ctx.DriverCtx)
 
-	if err := d.Prestart(ctx.ExecCtx, task); err != nil {
+	if _, err := d.Prestart(ctx.ExecCtx, task); err != nil {
 		t.Fatalf("prestart err: %v", err)
 	}
 	handle, err := d.Start(ctx.ExecCtx, task)
@@ -335,7 +335,7 @@ done
 		fmt.Errorf("Failed to write data")
 	}
 
-	if err := d.Prestart(ctx.ExecCtx, task); err != nil {
+	if _, err := d.Prestart(ctx.ExecCtx, task); err != nil {
 		t.Fatalf("prestart err: %v", err)
 	}
 	handle, err := d.Start(ctx.ExecCtx, task)
diff --git a/client/driver/rkt.go b/client/driver/rkt.go
index d99621368..116659729 100644
--- a/client/driver/rkt.go
+++ b/client/driver/rkt.go
@@ -207,8 +207,8 @@ func (d *RktDriver) Periodic() (bool, time.Duration) {
 	return true, 15 * time.Second
 }
 
-func (d *RktDriver) Prestart(ctx *ExecContext, task *structs.Task) error {
-	return nil
+func (d *RktDriver) Prestart(ctx *ExecContext, task *structs.Task) (*CreatedResources, error) {
+	return nil, nil
 }
 
 // Run an existing Rkt image.
@@ -451,6 +451,8 @@ func (d *RktDriver) Start(ctx *ExecContext, task *structs.Task) (DriverHandle, e
 	return h, nil
 }
 
+func (d *RktDriver) Cleanup(*ExecContext, *CreatedResources) error { return nil }
+
 func (d *RktDriver) Open(ctx *ExecContext, handleID string) (DriverHandle, error) {
 	// Parse the handle
 	pidBytes := []byte(strings.TrimPrefix(handleID, "Rkt:"))
diff --git a/client/driver/rkt_test.go b/client/driver/rkt_test.go
index 25dd5c151..72699986c 100644
--- a/client/driver/rkt_test.go
+++ b/client/driver/rkt_test.go
@@ -98,7 +98,7 @@ func TestRktDriver_Start_DNS(t *testing.T) {
 	defer ctx.AllocDir.Destroy()
 	d := NewRktDriver(ctx.DriverCtx)
 
-	if err := d.Prestart(ctx.ExecCtx, task); err != nil {
+	if _, err := d.Prestart(ctx.ExecCtx, task); err != nil {
 		t.Fatalf("error in prestart: %v", err)
 	}
 	handle, err := d.Start(ctx.ExecCtx, task)
@@ -149,7 +149,7 @@ func TestRktDriver_Start_Wait(t *testing.T) {
 	defer ctx.AllocDir.Destroy()
 	d := NewRktDriver(ctx.DriverCtx)
 
-	if err := d.Prestart(ctx.ExecCtx, task); err != nil {
+	if _, err := d.Prestart(ctx.ExecCtx, task); err != nil {
 		t.Fatalf("error in prestart: %v", err)
 	}
 	handle, err := d.Start(ctx.ExecCtx, task)
@@ -210,7 +210,7 @@ func TestRktDriver_Start_Wait_Skip_Trust(t *testing.T) {
 	defer ctx.AllocDir.Destroy()
 	d := NewRktDriver(ctx.DriverCtx)
 
-	if err := d.Prestart(ctx.ExecCtx, task); err != nil {
+	if _, err := d.Prestart(ctx.ExecCtx, task); err != nil {
 		t.Fatalf("error in prestart: %v", err)
 	}
 	handle, err := d.Start(ctx.ExecCtx, task)
@@ -281,7 +281,7 @@ func TestRktDriver_Start_Wait_AllocDir(t *testing.T) {
 	defer ctx.AllocDir.Destroy()
 	d := NewRktDriver(ctx.DriverCtx)
 
-	if err := d.Prestart(ctx.ExecCtx, task); err != nil {
+	if _, err := d.Prestart(ctx.ExecCtx, task); err != nil {
 		t.Fatalf("error in prestart: %v", err)
 	}
 	handle, err := d.Start(ctx.ExecCtx, task)
@@ -343,7 +343,7 @@ func TestRktDriverUser(t *testing.T) {
 	defer ctx.AllocDir.Destroy()
 	d := NewRktDriver(ctx.DriverCtx)
 
-	if err := d.Prestart(ctx.ExecCtx, task); err != nil {
+	if _, err := d.Prestart(ctx.ExecCtx, task); err != nil {
 		t.Fatalf("error in prestart: %v", err)
 	}
 	handle, err := d.Start(ctx.ExecCtx, task)
@@ -384,7 +384,7 @@ func TestRktTrustPrefix(t *testing.T) {
 	defer ctx.AllocDir.Destroy()
 	d := NewRktDriver(ctx.DriverCtx)
 
-	if err := d.Prestart(ctx.ExecCtx, task); err != nil {
+	if _, err := d.Prestart(ctx.ExecCtx, task); err != nil {
 		t.Fatalf("error in prestart: %v", err)
 	}
 	handle, err := d.Start(ctx.ExecCtx, task)
@@ -462,7 +462,7 @@ func TestRktDriver_PortsMapping(t *testing.T) {
 	defer ctx.AllocDir.Destroy()
 	d := NewRktDriver(ctx.DriverCtx)
 
-	if err := d.Prestart(ctx.ExecCtx, task); err != nil {
+	if _, err := d.Prestart(ctx.ExecCtx, task); err != nil {
 		t.Fatalf("error in prestart: %v", err)
 	}
 	handle, err := d.Start(ctx.ExecCtx, task)
diff --git a/client/task_runner.go b/client/task_runner.go
index f8331f803..443846e58 100644
--- a/client/task_runner.go
+++ b/client/task_runner.go
@@ -93,6 +93,11 @@ type TaskRunner struct {
 	// Must acquire persistLock when accessing
 	taskDirBuilt bool
 
+	// createdResources are all the resources created by the task driver
+	// across all attempts to start the task.
+	createdResources     *driver.CreatedResources
+	createdResourcesLock sync.Mutex
+
 	// payloadRendered tracks whether the payload has been rendered to disk
 	payloadRendered bool
 
@@ -130,7 +135,10 @@ type TaskRunner struct {
 	// waitCh closing marks the run loop as having exited
 	waitCh chan struct{}
 
-	// serialize SaveState calls
+	// persistLock must be acquired when accessing fields stored by
+	// SaveState. SaveState is called asynchronously to TaskRunner.Run by
+	// AllocRunner, so all state fields must be synchronized using this
+	// lock.
 	persistLock sync.Mutex
 }
 
@@ -141,6 +149,7 @@ type taskRunnerState struct {
 	HandleID           string
 	ArtifactDownloaded bool
 	TaskDirBuilt       bool
+	CreatedResources   *driver.CreatedResources
 	PayloadRendered    bool
 }
 
@@ -177,22 +186,23 @@ func NewTaskRunner(logger *log.Logger, config *config.Config,
 	restartTracker := newRestartTracker(tg.RestartPolicy, alloc.Job.Type)
 
 	tc := &TaskRunner{
-		config:         config,
-		updater:        updater,
-		logger:         logger,
-		restartTracker: restartTracker,
-		alloc:          alloc,
-		task:           task,
-		taskDir:        taskDir,
-		vaultClient:    vaultClient,
-		vaultFuture:    NewTokenFuture().Set(""),
-		updateCh:       make(chan *structs.Allocation, 64),
-		destroyCh:      make(chan struct{}),
-		waitCh:         make(chan struct{}),
-		startCh:        make(chan struct{}, 1),
-		unblockCh:      make(chan struct{}),
-		restartCh:      make(chan *structs.TaskEvent),
-		signalCh:       make(chan SignalEvent),
+		config:           config,
+		updater:          updater,
+		logger:           logger,
+		restartTracker:   restartTracker,
+		alloc:            alloc,
+		task:             task,
+		taskDir:          taskDir,
+		createdResources: driver.NewCreatedResources(),
+		vaultClient:      vaultClient,
+		vaultFuture:      NewTokenFuture().Set(""),
+		updateCh:         make(chan *structs.Allocation, 64),
+		destroyCh:        make(chan struct{}),
+		waitCh:           make(chan struct{}),
+		startCh:          make(chan struct{}, 1),
+		unblockCh:        make(chan struct{}),
+		restartCh:        make(chan *structs.TaskEvent),
+		signalCh:         make(chan SignalEvent),
 	}
 
 	return tc
@@ -237,6 +247,7 @@ func (r *TaskRunner) RestoreState() error {
 	}
 	r.artifactsDownloaded = snap.ArtifactDownloaded
 	r.taskDirBuilt = snap.TaskDirBuilt
+	r.createdResources = snap.CreatedResources
 	r.payloadRendered = snap.PayloadRendered
 
 	if err := r.setTaskEnv(); err != nil {
@@ -292,12 +303,17 @@ func (r *TaskRunner) SaveState() error {
 	r.persistLock.Lock()
 	defer r.persistLock.Unlock()
 
+	r.createdResourcesLock.Lock()
+	res := r.createdResources.Copy()
+	r.createdResourcesLock.Unlock()
+
 	snap := taskRunnerState{
 		Task:               r.task,
 		Version:            r.config.Version,
 		ArtifactDownloaded: r.artifactsDownloaded,
 		TaskDirBuilt:       r.taskDirBuilt,
 		PayloadRendered:    r.payloadRendered,
+		CreatedResources:   res,
 	}
 
 	r.handleLock.Lock()
@@ -870,6 +886,7 @@ func (r *TaskRunner) run() {
 			select {
 			case success := <-prestartResultCh:
 				if !success {
+					r.cleanup()
 					r.setState(structs.TaskStateDead, nil)
 					return
 				}
@@ -958,6 +975,7 @@ func (r *TaskRunner) run() {
 				running := r.running
 				r.runningLock.Unlock()
 				if !running {
+					r.cleanup()
 					r.setState(structs.TaskStateDead, r.destroyEvent)
 					return
 				}
@@ -976,6 +994,11 @@ func (r *TaskRunner) run() {
 
 				r.killTask(killEvent)
 				close(stopCollection)
+
+				// Wait for handler to exit before calling cleanup
+				<-handleWaitCh
+				r.cleanup()
+
 				r.setState(structs.TaskStateDead, nil)
 				return
 			}
@@ -984,6 +1007,7 @@ func (r *TaskRunner) run() {
 	RESTART:
 		restart := r.shouldRestart()
 		if !restart {
+			r.cleanup()
 			r.setState(structs.TaskStateDead, nil)
 			return
 		}
@@ -997,6 +1021,44 @@ func (r *TaskRunner) run() {
 	}
 }
 
+// cleanup calls Driver.Cleanup when a task is stopping. Errors are logged.
+func (r *TaskRunner) cleanup() {
+	drv, err := r.createDriver()
+	if err != nil {
+		r.logger.Printf("[ERR] client: error creating driver to cleanup resources: %v", err)
+		return
+	}
+
+	r.createdResourcesLock.Lock()
+	res := r.createdResources.Copy()
+	r.createdResourcesLock.Unlock()
+
+	ctx := driver.NewExecContext(r.taskDir, r.alloc.ID)
+	attempts := 1
+	var cleanupErr error
+	for retry := true; retry; attempts++ {
+		cleanupErr = drv.Cleanup(ctx, res)
+		retry = structs.IsRecoverable(cleanupErr)
+
+		// Copy current createdResources state in case SaveState is
+		// called between retries
+		r.createdResourcesLock.Lock()
+		r.createdResources = res.Copy()
+		r.createdResourcesLock.Unlock()
+
+		// Retry 3 times with sleeps between
+		if !retry || attempts > 3 {
+			break
+		}
+		time.Sleep(time.Duration(attempts) * time.Second)
+	}
+
+	if cleanupErr != nil {
+		r.logger.Printf("[ERR] client: error cleaning up resources for task %q after %d attempts: %v", r.task.Name, attempts, cleanupErr)
+	}
+	return
+}
+
 // shouldRestart returns if the task should restart. If the return value is
 // true, the task's restart policy has already been considered and any wait time
 // between restarts has been applied.
@@ -1095,7 +1157,14 @@ func (r *TaskRunner) startTask() error {
 
 	// Run prestart
 	ctx := driver.NewExecContext(r.taskDir, r.alloc.ID)
-	if err := drv.Prestart(ctx, r.task); err != nil {
+	res, err := drv.Prestart(ctx, r.task)
+
+	// Merge newly created resources into previously created resources
+	r.createdResourcesLock.Lock()
+	r.createdResources.Merge(res)
+	r.createdResourcesLock.Unlock()
+
+	if err != nil {
 		wrapped := fmt.Errorf("failed to initialize task %q for alloc %q: %v",
 			r.task.Name, r.alloc.ID, err)
 
diff --git a/client/task_runner_test.go b/client/task_runner_test.go
index e76426294..7a5b9cb8c 100644
--- a/client/task_runner_test.go
+++ b/client/task_runner_test.go
@@ -1305,3 +1305,53 @@ func TestTaskRunner_SimpleRun_Dispatch(t *testing.T) {
 		t.Fatalf("Bad; got %v; want %v", string(data), string(expected))
 	}
 }
+
+func TestTaskRunner_CleanupOK(t *testing.T) {
+	alloc := mock.Alloc()
+	task := alloc.Job.TaskGroups[0].Tasks[0]
+	task.Driver = "mock_driver"
+	key := "ERR"
+
+	ctx := testTaskRunnerFromAlloc(t, false, alloc)
+	ctx.tr.config.Options = map[string]string{
+		"cleanup_fail_on":  key,
+		"cleanup_fail_num": "1",
+	}
+	ctx.tr.MarkReceived()
+
+	ctx.tr.createdResources.Resources[key] = []string{"x", "y"}
+	ctx.tr.createdResources.Resources["foo"] = []string{"z"}
+
+	defer ctx.Cleanup()
+	ctx.tr.Run()
+
+	// Since we only failed once, createdResources should be empty
+	if len(ctx.tr.createdResources.Resources) > 0 {
+		t.Fatalf("expected all created resources to be removed: %#v", ctx.tr.createdResources.Resources)
+	}
+}
+
+func TestTaskRunner_CleanupFail(t *testing.T) {
+	alloc := mock.Alloc()
+	task := alloc.Job.TaskGroups[0].Tasks[0]
+	task.Driver = "mock_driver"
+	key := "ERR"
+	ctx := testTaskRunnerFromAlloc(t, false, alloc)
+	ctx.tr.config.Options = map[string]string{
+		"cleanup_fail_on":  key,
+		"cleanup_fail_num": "5",
+	}
+	ctx.tr.MarkReceived()
+
+	ctx.tr.createdResources.Resources[key] = []string{"x"}
+	ctx.tr.createdResources.Resources["foo"] = []string{"y", "z"}
+
+	defer ctx.Cleanup()
+	ctx.tr.Run()
+
+	// Since we failed > 3 times, the failed key should remain
+	expected := map[string][]string{key: {"x"}}
+	if !reflect.DeepEqual(expected, ctx.tr.createdResources.Resources) {
+		t.Fatalf("expected %#v but found: %#v", expected, ctx.tr.createdResources.Resources)
+	}
+}
diff --git a/nomad/node_endpoint.go b/nomad/node_endpoint.go
index 5328d9671..d2787535a 100644
--- a/nomad/node_endpoint.go
+++ b/nomad/node_endpoint.go
@@ -944,7 +944,7 @@ func (n *Node) DeriveVaultToken(args *structs.DeriveVaultTokenRequest,
 	// setErr is a helper for setting the recoverable error on the reply and
 	// logging it
 	setErr := func(e error, recoverable bool) {
-		reply.Error = structs.NewRecoverableError(e, recoverable)
+		reply.Error = structs.NewRecoverableError(e, recoverable).(*structs.RecoverableError)
 		n.srv.logger.Printf("[ERR] nomad.client: DeriveVaultToken failed (recoverable %v): %v", recoverable, e)
 	}
 
@@ -1134,7 +1134,7 @@ func (n *Node) DeriveVaultToken(args *structs.DeriveVaultTokenRequest,
 		if rerr, ok := createErr.(*structs.RecoverableError); ok {
 			reply.Error = rerr
 		} else {
-			reply.Error = structs.NewRecoverableError(createErr, false)
+			reply.Error = structs.NewRecoverableError(createErr, false).(*structs.RecoverableError)
 		}
 
 		return nil
diff --git a/nomad/structs/structs.go b/nomad/structs/structs.go
index fa63def23..ceb7df38a 100644
--- a/nomad/structs/structs.go
+++ b/nomad/structs/structs.go
@@ -4054,7 +4054,7 @@ type RecoverableError struct {
 
 // NewRecoverableError is used to wrap an error and mark it as recoverable or
 // not.
-func NewRecoverableError(e error, recoverable bool) *RecoverableError {
+func NewRecoverableError(e error, recoverable bool) error {
 	if e == nil {
 		return nil
 	}
@@ -4068,3 +4068,12 @@ func NewRecoverableError(e error, recoverable bool) *RecoverableError {
 func (r *RecoverableError) Error() string {
 	return r.Err
 }
+
+// IsRecoverable returns true if error is a RecoverableError with
+// Recoverable=true. Otherwise false is returned.
+func IsRecoverable(e error) bool {
+	if re, ok := e.(*RecoverableError); ok {
+		return re.Recoverable
+	}
+	return false
+}
diff --git a/nomad/structs/structs_test.go b/nomad/structs/structs_test.go
index b0ddb7aaf..72c097224 100644
--- a/nomad/structs/structs_test.go
+++ b/nomad/structs/structs_test.go
@@ -1,6 +1,7 @@
 package structs
 
 import (
+	"fmt"
 	"reflect"
 	"strings"
 	"testing"
@@ -1539,3 +1540,21 @@ func TestDispatchInputConfig_Validate(t *testing.T) {
 		t.Fatalf("bad: %v", err)
 	}
 }
+
+func TestIsRecoverable(t *testing.T) {
+	if IsRecoverable(nil) {
+		t.Errorf("nil should not be recoverable")
+	}
+	if IsRecoverable(NewRecoverableError(nil, true)) {
+		t.Errorf("NewRecoverableError(nil, true) should not be recoverable")
+	}
+	if IsRecoverable(fmt.Errorf("i promise im recoverable")) {
+		t.Errorf("Custom errors should not be recoverable")
+	}
+	if IsRecoverable(NewRecoverableError(fmt.Errorf(""), false)) {
+		t.Errorf("Explicitly unrecoverable errors should not be recoverable")
+	}
+	if !IsRecoverable(NewRecoverableError(fmt.Errorf(""), true)) {
+		t.Errorf("Explicitly recoverable errors *should* be recoverable")
+	}
+}