studio/flamenco
studio/flamenco
12
30
Fork 38
Code Issues 64 Pull Requests 5 Releases Wiki Activity

Support pausing jobs #104313

Manually merged
Sybren A. Stüvel merged 28 commits from David-Zhang-10/flamenco:paused-job-status into main 2024-07-01 17:53:44 +02:00
Conversation 1 Commits 28 Files Changed 12 +14 -10
2 changed files with 14 additions and 10 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files
Showing only changes of commit 4b4511f9b6 - Show all commits

8
internal/manager/task_state_machine/task_state_machine.go
View File

@ -180,16 +180,22 @@ func (sm *StateMachine) updateJobOnTaskStatusCanceled(ctx context.Context, logge
return sm.JobStatusChange(ctx, job, api.JobStatusCanceled, "canceled task was last runnable task of job, canceling job") return sm.JobStatusChange(ctx, job, api.JobStatusCanceled, "canceled task was last runnable task of job, canceling job")
} }
David-Zhang-10 marked this conversation as resolved Outdated
Sybren A. Stüvel commented 2024-06-17 15:35:21 +02:00
Outdated
Review
Copy Link

This code block feels a bit out of place. If I understand correctly, this is the flow that this code would handle:

  1. User cancels a task.
  2. There are other tasks that are still runnable but not active (i.e. status in {queued, soft-failed, paused}).
  3. Job is in status pause-requested.

To me this doesn't look like a status in which the pausing was complete.

This logic is also repeated quite a bit below. I think it's better to make a new function responsible for answering the question "is the job pausing, and is that process complete now?", and call that function from the places that need this.

Finally, some reordering can be done to optimise the code. Currently every check will count the tasks in the job, even when job.Status != api.JobStatusPauseRequested. It's good practice to do the cheapest check first, and the most expensive check last.

This code block feels a bit out of place. If I understand correctly, this is the flow that this code would handle: 1. User cancels a task. 2. There are other tasks that are still runnable but not active (i.e. status in `{queued, soft-failed, paused}`). 3. Job is in status `pause-requested`. To me this doesn't look like a status in which the pausing was complete. This logic is also repeated quite a bit below. I think it's better to make a new function responsible for answering the question "_is the job pausing, and is that process complete now?_", and call that function from the places that need this. Finally, some reordering can be done to optimise the code. Currently every check will count the tasks in the job, even when `job.Status != api.JobStatusPauseRequested`. It's good practice to do the cheapest check first, and the most expensive check last.
David Zhang commented 2024-06-18 22:03:23 +02:00
Outdated
Review
Copy Link

The comment makes sense to me! The only thing I am confused about is what you meant by "the pausing was complete".

The comment makes sense to me! The only thing I am confused about is what you meant by "the pausing was complete".
Sybren A. Stüvel commented 2024-06-19 10:32:31 +02:00
Outdated
Review
Copy Link

The only thing I am confused about is what you meant by "the pausing was complete".

"Pausing is complete" for me means that the work that should be done when the job goes to pause-requested is done, and it can be moved to state paused.

> The only thing I am confused about is what you meant by "the pausing was complete". "Pausing is complete" for me means that the work that should be done when the job goes to `pause-requested` is done, and it can be moved to state `paused`.
David Zhang commented 2024-06-19 17:45:08 +02:00
Outdated
Review
Copy Link

So for the pausing to be complete, we need all tasks to be in paused state, including tasks that are runnable but not in active state? In other words, instead of checking the number of active tasks here, we should check if the job has any tasks other than paused?

So for the pausing to be complete, we need all tasks to be in `paused` state, including tasks that are runnable but not in `active` state? In other words, instead of checking the number of active tasks here, we should check if the job has any tasks other than `paused`?
// Deal with the special case when the job is in pause-requested status.
if job.Status != api.JobStatusPauseRequested {
return nil
} else {
numActive, _, err := sm.persist.CountTasksOfJobInStatus(ctx, job, api.TaskStatusActive) numActive, _, err := sm.persist.CountTasksOfJobInStatus(ctx, job, api.TaskStatusActive)
if err != nil { if err != nil {
David-Zhang-10 marked this conversation as resolved Outdated
Sybren A. Stüvel commented 2024-06-17 15:42:03 +02:00
Outdated
Review
Copy Link

This log entry is giving the wrong information. If updateJobOnTaskStatusCanceled() is called, a task was cancelled and thus the log message cannot be correct. Same for the job status change reason "all tasks completed".

The same is true for other copies of this code. Don't make your code lie; in a task-failure-handling part of the code, don't make it log "all tasks completed".

This log entry is giving the wrong information. If `updateJobOnTaskStatusCanceled()` is called, a task was cancelled and thus the log message cannot be correct. Same for the job status change reason "all tasks completed". The same is true for other copies of this code. Don't make your code lie; in a task-failure-handling part of the code, don't make it log "all tasks completed".
return err return err
} }
if numActive == 0 && job.Status == api.JobStatusPauseRequested { if numActive == 0 {
// there is no active task, and the job is in pause-requested status, so we can pause the job // there is no active task, and the job is in pause-requested status, so we can pause the job
logger.Info().Msg("all tasks of job are completed, job is paused") logger.Info().Msg("all tasks of job are completed, job is paused")
return sm.JobStatusChange(ctx, job, api.JobStatusPaused, "all tasks completed") return sm.JobStatusChange(ctx, job, api.JobStatusPaused, "all tasks completed")
} }
}
// Execution should not reach here.
return nil return nil
} }

2
internal/manager/task_state_machine/task_state_machine_test.go
View File

@ -191,8 +191,6 @@ func TestTaskStatusChangeCancelSingleTask(t *testing.T) {
mocks.persist.EXPECT().CountTasksOfJobInStatus(ctx, job, mocks.persist.EXPECT().CountTasksOfJobInStatus(ctx, job,
api.TaskStatusActive, api.TaskStatusQueued, api.TaskStatusSoftFailed, api.TaskStatusPaused). api.TaskStatusActive, api.TaskStatusQueued, api.TaskStatusSoftFailed, api.TaskStatusPaused).
Return(1, 2, nil) Return(1, 2, nil)
mocks.persist.EXPECT().CountTasksOfJobInStatus(ctx, job,
api.TaskStatusActive).Return(0, 2, nil)
require.NoError(t, sm.TaskStatusChange(ctx, task, api.TaskStatusCanceled)) require.NoError(t, sm.TaskStatusChange(ctx, task, api.TaskStatusCanceled))
// T2: queued > cancelled --> J: cancel-requested > canceled // T2: queued > cancelled --> J: cancel-requested > canceled