Skip to content

[shim] Persist task state to clean up tasks after a restart #4220

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
un-def merged 2 commits into from
Aug 27, 2026
Merged

[shim] Persist task state to clean up tasks after a restart #4220

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
e6d0dde
[shim] Persist task state to clean up tasks after a restart
un-def Aug 26, 2026
83c7ec3
Dummy commit to trigger Actions
un-def Aug 27, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
118 changes: 90 additions & 28 deletions runner/internal/shim/docker.go
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -17,6 +17,7 @@ import (
"sync"
"time"

dockertypes "github.com/docker/docker/api/types"
"github.com/docker/docker/api/types/container"
"github.com/docker/docker/api/types/filters"
"github.com/docker/docker/api/types/image"
Expand Down Expand Up @@ -156,6 +157,8 @@ type DockerRunner struct {
gpuVendor gpu.GpuVendor
gpuLock *GpuLock
tasks TaskStorage
// stateMu serializes task state file updates, see saveTaskState()
stateMu sync.Mutex
}

func NewDockerRunner(ctx context.Context, dockerParams DockerParameters) (*DockerRunner, error) {
Expand Down Expand Up @@ -203,6 +206,9 @@ func NewDockerRunner(ctx context.Context, dockerParams DockerParameters) (*Docke
if err := runner.restoreStateFromContainers(ctx); err != nil {
return nil, fmt.Errorf("failed to restore state from containers: %w", err)
}
// Must be called after the tasks are restored, as it uses them to tell the dirs of
// the live tasks from the orphaned ones
runner.sweepOrphanedTaskDirs(ctx)

return runner, nil
}
Expand Down Expand Up @@ -283,29 +289,70 @@ func (d *DockerRunner) restoreStateFromContainers(ctx context.Context) error {
break
}
}
if len(gpuIDs) > 0 {
state, restored := readRestoredTaskState(ctx, taskID, runnerDir)
config := state.Config
if !restored {
// Containers created by shim versions that did not write the state file have
// no config to restore. The volumes can still be recovered from the container
// mounts, unlike the host SSH keys, which are only known to the state file
config.Volumes = volumesFromMounts(containerShort.Mounts)
}
if state.CleanedUp {
// The resources of this task, its GPUs included, have already been released,
// therefore the task owns nothing
gpuIDs = nil
} else if len(gpuIDs) > 0 {
// A GPU already locked by another restored task is not locked again and,
// therefore, is not owned by this task -- otherwise, cleaning up this task
// would release a GPU that the other one is still using
gpuIDs = d.gpuLock.Lock(ctx, gpuIDs)
log.Debug(ctx, "locked GPU(s) due to running task", "task", taskID, "gpus", gpuIDs)
}
// Tasks are restored as running regardless of the container state, letting
// ProcessTasks() decide whether the container is still running and, if it is
// not, why it finished. This way, the termination reason of a task that
// finished while the shim was not running is not lost
task := NewTask(taskID, TaskStatusRunning, containerName, containerID, gpuIDs, ports, runnerDir)
// A task with a recorded termination reason has been terminated before the shim
// restarted, and its reason must not be overridden by the container exit code.
// Otherwise the task is restored as running regardless of the container state,
// letting ProcessTasks() decide whether the container is still running and, if
// it is not, why it finished
status := TaskStatusRunning
if state.TerminationReason != "" {
status = TaskStatusTerminated
}
task := NewTask(taskID, status)
task.TerminationReason = state.TerminationReason
task.TerminationMessage = state.TerminationMessage
task.config = config
task.containerName = containerName
task.containerID = containerID
task.gpuIDs = gpuIDs
task.ports = ports
task.runnerDir = runnerDir
task.cleanedUp = state.CleanedUp
if !d.tasks.Add(task) {
log.Error(ctx, "duplicate restored task", "task", taskID)
// Nothing will release the GPUs of a task that is not stored
d.gpuLock.Release(ctx, gpuIDs)
continue
}
log.Debug(ctx, "restored task", "task", taskID, "state", containerShort.State, "gpus", gpuIDs)
log.Debug(
ctx, "restored task",
"task", taskID, "status", status, "state", containerShort.State, "gpus", gpuIDs,
)
}
return nil
}

// volumesFromMounts recovers the volumes attached to a task from its container mounts.
// Only the volume names are recovered, which is all that is needed to unmount them
func volumesFromMounts(mounts []dockertypes.MountPoint) []VolumeInfo {
var volumes []VolumeInfo
for _, mount := range mounts {
if name, found := strings.CutPrefix(mount.Source, volumeMountPointDir+"/"); found {
volumes = append(volumes, VolumeInfo{Name: name})
}
}
return volumes
}

func (d *DockerRunner) Resources(ctx context.Context) Resources {
cpuCount := host.GetCpuCount(ctx)
totalMemory, err := host.GetTotalMemory(ctx)
Expand Down Expand Up @@ -375,7 +422,7 @@ func (d *DockerRunner) Submit(ctx context.Context, cfg TaskConfig) error {
// of the task accordingly. mutate is called before the local copy is updated, so it
// can read the local copy to publish fields set by the caller, e.g., containerID.
// The local copy is left intact if the update is rejected.
func (d *DockerRunner) commit(task *Task, mutate func(*Task)) error {
func (d *DockerRunner) commit(ctx context.Context, task *Task, mutate func(*Task)) error {
updatedTask, err := d.tasks.Modify(task.ID, func(t *Task) error {
mutate(t)
return nil
Expand All @@ -384,6 +431,7 @@ func (d *DockerRunner) commit(task *Task, mutate func(*Task)) error {
return err
}
*task = updatedTask
d.saveTaskState(ctx, task.ID)
return nil
}

Expand All @@ -395,7 +443,7 @@ func (d *DockerRunner) commitTerminated(ctx context.Context, task *Task) {
// the last successful commit is the actual state, nothing to commit
return
}
if err := d.commit(task, func(t *Task) {
if err := d.commit(ctx, task, func(t *Task) {
t.SetStatusTerminated(task.TerminationReason, task.TerminationMessage)
}); err != nil && !errors.Is(err, ErrNotFound) {
log.Error(ctx, "failed to commit terminated status", "task", task.ID, "err", err)
Expand Down Expand Up @@ -439,7 +487,7 @@ func (d *DockerRunner) Start(ctx context.Context, taskID string) (err error) {
}
}
// Hand the task over to ProcessTasks()
if commitErr := d.commit(&task, func(t *Task) {
if commitErr := d.commit(ctx, &task, func(t *Task) {
t.startInFlight = false
if task.containerID != "" {
t.containerID = task.containerID
Expand All @@ -455,7 +503,7 @@ func (d *DockerRunner) Start(ctx context.Context, taskID string) (err error) {
}
}()

if err := d.commit(&task, func(t *Task) {
if err := d.commit(ctx, &task, func(t *Task) {
t.startInFlight = true
t.SetStatusPreparing()
}); err != nil {
Expand All @@ -471,7 +519,7 @@ func (d *DockerRunner) Start(ctx context.Context, taskID string) (err error) {
log.Trace(ctx, "runner dir", "task", task.ID, "path", runnerDir)
// Resources are committed as soon as they are acquired, so that they are not
// lost if the task is updated by another goroutine, e.g., terminated by the server
if err := d.commit(&task, func(t *Task) { t.runnerDir = runnerDir }); err != nil {
if err := d.commit(ctx, &task, func(t *Task) { t.runnerDir = runnerDir }); err != nil {
return fmt.Errorf("%w: failed to update task %s: %w", ErrInternal, task.ID, err)
}

Expand All @@ -487,7 +535,7 @@ func (d *DockerRunner) Start(ctx context.Context, taskID string) (err error) {
} else {
gpuIDs = []string{}
}
if err := d.commit(&task, func(t *Task) { t.gpuIDs = gpuIDs }); err != nil {
if err := d.commit(ctx, &task, func(t *Task) { t.gpuIDs = gpuIDs }); err != nil {
return fmt.Errorf("%w: failed to update task %s: %w", ErrInternal, task.ID, err)
}

Expand Down Expand Up @@ -520,7 +568,7 @@ func (d *DockerRunner) Start(ctx context.Context, taskID string) (err error) {
log.Debug(ctx, "Pulling image")
pullCtx, cancelPull := context.WithTimeout(ctx, ImagePullTimeout)
defer cancelPull()
if err := d.commit(&task, func(t *Task) { t.SetStatusPulling(cancelPull) }); err != nil {
if err := d.commit(ctx, &task, func(t *Task) { t.SetStatusPulling(cancelPull) }); err != nil {
return fmt.Errorf("%w: failed to update task %s: %w", ErrInternal, task.ID, err)
}
// Although it's called "runner dir", we also use it for shim task-related data.
Expand All @@ -534,7 +582,7 @@ func (d *DockerRunner) Start(ctx context.Context, taskID string) (err error) {
}

log.Debug(ctx, "Creating container", "task", task.ID, "name", task.containerName)
if err := d.commit(&task, func(t *Task) { t.SetStatusCreating() }); err != nil {
if err := d.commit(ctx, &task, func(t *Task) { t.SetStatusCreating() }); err != nil {
return fmt.Errorf("%w: failed to update task %s: %w", ErrInternal, task.ID, err)
}
if err := d.createContainer(ctx, &task, createContainerOptions{}); err != nil {
Expand Down Expand Up @@ -574,7 +622,7 @@ func (d *DockerRunner) Start(ctx context.Context, taskID string) (err error) {
}

// The container is running, the task is now processed in the background
if err := d.commit(&task, func(t *Task) {
if err := d.commit(ctx, &task, func(t *Task) {
// startContainer sets the ports field, the retry above may have
// replaced the container
t.containerID = task.containerID
Expand Down Expand Up @@ -608,16 +656,7 @@ func (d *DockerRunner) cleanupLocked(ctx context.Context, task *Task) {
return
}
log.Debug(ctx, "releasing task resources", "task", task.ID)
cfg := task.config
if err := unmountVolumes(ctx, cfg); err != nil {
log.Error(ctx, "failed to unmount volumes", "task", task.ID, "err", err)
}
if len(cfg.HostSshKeys) > 0 {
ak := AuthorizedKeys{user: cfg.HostSshUser, lookup: user.Lookup}
if err := ak.RemovePublicKeys(cfg.HostSshKeys); err != nil {
log.Error(ctx, "failed to remove public keys", "task", task.ID, "err", err)
}
}
releaseTaskResources(ctx, task.config)
if len(task.gpuIDs) > 0 {
releasedGpuIDs := d.gpuLock.Release(ctx, task.gpuIDs)
log.Debug(ctx, "released GPU(s)", "task", task.ID, "gpus", releasedGpuIDs)
Expand All @@ -631,6 +670,22 @@ func (d *DockerRunner) cleanupLocked(ctx context.Context, task *Task) {
}); err != nil && !errors.Is(err, ErrNotFound) {
log.Error(ctx, "failed to commit cleaned up state", "task", task.ID, "err", err)
}
d.saveTaskState(ctx, task.ID)
}

// releaseTaskResources releases the host resources acquired for a task: volumes and
// host SSH keys. Unlike GPU locks, which are only kept in memory, these outlive the
// shim process, therefore they are released by task config and not by task
func releaseTaskResources(ctx context.Context, cfg TaskConfig) {
if err := unmountVolumes(ctx, cfg); err != nil {
log.Error(ctx, "failed to unmount volumes", "err", err)
}
if len(cfg.HostSshKeys) > 0 {
ak := AuthorizedKeys{user: cfg.HostSshUser, lookup: user.Lookup}
if err := ak.RemovePublicKeys(cfg.HostSshKeys); err != nil {
log.Error(ctx, "failed to remove public keys", "err", err)
}
}
}

// Terminate aborts running operations (pulling an image, running a container) and sets task status to terminated
Expand Down Expand Up @@ -731,7 +786,10 @@ func (d *DockerRunner) remove(ctx context.Context, task *Task) (err error) {
// Failed attempts to remove or rename runner dir are considered non-fatal
if err := os.RemoveAll(task.runnerDir); err != nil {
log.Error(ctx, "failed to remove runner directory", "dir", task.runnerDir, "err", err)
trashName := fmt.Sprintf(".trash-%s-%d", task.runnerDir, time.Now().UnixMicro())
trashName := filepath.Join(
filepath.Dir(task.runnerDir),
fmt.Sprintf(".trash-%s-%d", filepath.Base(task.runnerDir), time.Now().UnixMicro()),
)
if err := os.Rename(task.runnerDir, trashName); err != nil {
log.Error(ctx, "failed to rename runner directory", "dir", task.runnerDir, "err", err)
}
Expand Down Expand Up @@ -1317,8 +1375,12 @@ func (c *CLIArgs) DockerPorts() []int {
return []int{c.Runner.HTTPPort, c.Runner.SSHPort}
}

func (c *CLIArgs) RunnersDir() string {
return filepath.Join(c.Shim.HomeDir, "runners")
}

func (c *CLIArgs) MakeRunnerDir(name string) (string, error) {
runnerTemp := filepath.Join(c.Shim.HomeDir, "runners", name)
runnerTemp := filepath.Join(c.RunnersDir(), name)
if err := os.MkdirAll(runnerTemp, 0o755); err != nil {
return "", fmt.Errorf("create runner directory: %w", err)
}
Expand Down
Loading
Loading