Skip to content
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.

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

Make number of scheduler workers reloadable #11593

Merged
merged 32 commits into from
Jan 6, 2022
Merged

Conversation

angrycub
Copy link
Contributor

@angrycub angrycub commented Nov 30, 2021

This PR:

  • enables the server.num_schedulers and server.enabled_schedulers values to be hot reloadable.
  • adds an API to allow operators to make temporary changes to these values on a per server basis.
  • adds an API to get the status of the scheduler workers and their internal workloads.

This closes #11449

Copy link
Contributor

@DerekStrickland DerekStrickland left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Beautiful work.

nomad/worker.go Outdated Show resolved Hide resolved
nomad/worker.go Outdated Show resolved Hide resolved
nomad/worker.go Show resolved Hide resolved
Co-authored-by: Derek Strickland <[email protected]>
nomad/server.go Outdated Show resolved Hide resolved
nomad/server.go Show resolved Hide resolved
nomad/worker.go Outdated Show resolved Hide resolved
Copy link
Member

@tgross tgross left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I've re-reviewed the API and HTTP agent sections and I'm putting that review up for you @angrycub. I'll re-review the worker/leader/server section next.

website/content/api-docs/agent.mdx Show resolved Hide resolved
api/agent.go Outdated Show resolved Hide resolved
api/agent.go Outdated
}

// SetSchedulerWorkerConfig attempts to update the targeted agent's worker pool configuration
func (a *Agent) SetSchedulerWorkerConfig(args SchedulerWorkerPoolArgs) (*SchedulerWorkerPoolArgs, error) {
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I think we need WriteOptions here (and QueryOptions for GetSchedulerConfig above) to support ACLs and any HTTP params we might want in the future. And we always seem to want it eventually so that way we don't have to make a SetSchedulerWorkerConfigWithOptions later on.

We can probably get away without having a QueryMeta returned here because everything in QueryMeta is used for comes out of raft? None of the other agent APIs in this file have it.

Comment on lines 795 to 799
switch req.Method {
case "PUT", "POST":
return s.UpdateScheduleWorkersConfig(resp, req)
case "GET":
return s.GetScheduleWorkersConfig(resp, req)
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

We're probably not consistent about this across the code base, but good to use the constants for new code at least:

Suggested change
switch req.Method {
case "PUT", "POST":
return s.UpdateScheduleWorkersConfig(resp, req)
case "GET":
return s.GetScheduleWorkersConfig(resp, req)
switch req.Method {
case http.MethodPut, http.MethodPost:
return s.UpdateScheduleWorkersConfig(resp, req)
case http.MethodGet:
return s.GetScheduleWorkersConfig(resp, req)

}
}

func (s *HTTPServer) GetScheduleWorkersConfig(resp http.ResponseWriter, req *http.Request) (interface{}, error) {
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This is the implementation for AgentSchedulerWorkerConfigRequest and not used in other packages, right? Usually we'll want to avoid exporting it (ex. name it getScheduleWorkersConfig). Same applied for the update implementation.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I think that I'd done that to help the OpenAPI generator auto-documentation process, but will make them private is def more correct.

Comment on lines 872 to 889
type agentSchedulerWorkerConfig struct {
ServerID string `json:"server_id,omitempty"`
NumSchedulers int `json:"num_schedulers"`
EnabledSchedulers []string `json:"enabled_schedulers"`
}

type agentSchedulerWorkersInfo struct {
ServerID string `json:"server_id"`
Schedulers []agentSchedulerWorkerInfo `json:"schedulers"`
}

type agentSchedulerWorkerInfo struct {
ID string `json:"id"`
EnabledSchedulers []string `json:"enabled_schedulers"`
Started string `json:"started"`
Status string `json:"status"`
WorkloadStatus string `json:"workload_status"`
}
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Aren't these always going to be 1:1 with api structs? I think you can import the api package here and use it them directly.

t.Run(tc.name, func(t *testing.T) {

req, err := http.NewRequest(tc.request.verb, "/v1/agent/schedulers/config", bytes.NewReader([]byte(tc.request.requestBody)))
require.Nil(t, err)
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Nitpick: while require.Nil(t, err) and require.NoError(t, err) test the same thing, we tend to use NoError for clarity.

@@ -1463,3 +1464,586 @@ func TestHTTP_XSS_Monitor(t *testing.T) {
})
}
}

Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I love these giant table-driven tests!

Copy link
Member

@tgross tgross left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Ok @angrycub I've looked through the second half of the work and this is looking great. I've left a few remarks about locking in the worker.go but I think other than that there's nothing too serious here.

// this satisfies the require.Eventually test interface
checkPaused := func(count int) func() bool {
return func() bool {
workers := pausedWorkers()
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Because this is all closures inside the test function, it's probably ok to inline the body of pausedWorkers here in the checkPaused function.

Comment on lines 829 to 843
func TestWorker_WorkerInfo_String(t *testing.T) {
t.Parallel()
startTime := time.Date(2009, time.November, 10, 23, 0, 0, 0, time.UTC)
w := &Worker{
id: "uuid",
start: startTime,
status: WorkerStarted,
workloadStatus: WorkloadBackoff,
enabledSchedulers: []string{structs.JobTypeCore, structs.JobTypeBatch, structs.JobTypeSystem},
}
_, err := json.Marshal(w)
require.NoError(t, err)

require.Equal(t, `{"id":"uuid","enabled_schedulers":["_core","batch","system"],"started":"2009-11-10T23:00:00Z","status":"Started","workload_status":"Backoff"}`, fmt.Sprint(w.Info()))
}
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This test feels like it's just testing the stdlib's encoding/json package behavior... we can probably drop this one.

nomad/worker.go Outdated
}

// _newWorker creates a worker without calling its Start func. This is useful for testing.
func newWorker(ctx context.Context, srv *Server, args SchedulerWorkerPoolArgs) (*Worker, error) {
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This function doesn't ever return an error (which is pretty typical for newBlahblah functions), so I think we can drop that return value and then clean up all the cases where we're doing w, _ := newWorker(...)


// setWorkloadStatus is used internally to the worker to update the
// status of the worker based updates from the workload.
func (w *Worker) setWorkloadStatus(newStatus SchedulerWorkerStatus) {
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This is a great idea for making the scheduler worker behavior more observable. Having the trace log here is great but I can totally see writing a bpftrace script that hooks this function to read the stack args and catch all the transitions too.

(I might have done this in its own PR, but as long as it's here now we might as well enjoy it.)

nomad/worker.go Outdated
w.pauseCond.Wait()
}

w.pauseLock.Unlock()
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I think there's a race here if a Pause method call comes in after this line but before the status is set before we return. Even with the locks the Pause method's status setting calls could be interleaved such that the worker sets its status to WorkerPaused, sets the pause flag, and then its status is set to WorkerStarted.

Maybe we should move this up to the top of the function as a defer w.pauseLock.Unlock()?

nomad/worker.go Outdated
Comment on lines 349 to 354
defer func() {
w.setWorkloadStatus(WorkloadStopped)
w.markStopped()
}()
w.setStatus(WorkerStarted)
w.setWorkloadStatus(WorkloadRunning)
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

We pair up the status and workload status updates more often than not (see maybeWait above) and here we're not doing it atomically. It's probably safe here but easy to accidentally split up locked function calls later so that it's unsafe. So it might be a good idea to have a combined setStatus(workerStatus, workloadStatus) function that takes care of both and does the nice bit you've done where it only logs on change.

nomad/worker_test.go Outdated Show resolved Hide resolved
Comment on lines 863 to 883
w._start(testWorkload)
require.Eventually(t, w.IsStarted, longWait, tinyWait, "should have started")

go func() {
time.Sleep(tinyWait)
w.Pause()
}()
require.Eventually(t, w.IsPaused, longWait, tinyWait, "should have paused")

go func() {
time.Sleep(tinyWait)
w.Resume()
}()
require.Eventually(t, w.IsStarted, longWait, tinyWait, "should have restarted from pause")

go func() {
time.Sleep(tinyWait)
w.Stop()
}()
require.Eventually(t, w.IsStopped, longWait, tinyWait, "should have shutdown")
}
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

👍 This test shows exactly why I like the API you have around "pausing" vs "paused"

nomad/server.go Show resolved Hide resolved
@@ -1430,17 +1440,165 @@ func (s *Server) setupSerf(conf *serf.Config, ch chan serf.Event, path string) (
return serf.Create(conf)
}

// shouldReloadSchedulers checks the new config to determine if the scheduler worker pool
// needs to be updated. If so, returns true and a pointer to a populated SchedulerWorkerPoolArgs
func shouldReloadSchedulers(s *Server, newPoolArgs *SchedulerWorkerPoolArgs) (bool, *SchedulerWorkerPoolArgs) {
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

In retrospect we probably could have unconditionally drained and replaced the worker pool on reload; the API call is gated by ACLs so the risk of an operator DoS'ing their own scheduler seems low. And then we could come back in later to do this more clever logic in a future PR. But this is pretty nice and enables scale up/down in the future (as per your // TODO remark below).

Copy link
Member

@tgross tgross left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I added one last place where we could use setStatuses, but other than that this LGTM! Let's ship it!

nomad/worker.go Outdated
Comment on lines 387 to 388
w.setWorkloadStatus(WorkloadStopped)
w.markStopped()
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I think we missed this one:

Suggested change
w.setWorkloadStatus(WorkloadStopped)
w.markStopped()
w.setStatuses(WorkerStopped , WorkloadStopped)

@github-actions
Copy link

github-actions bot commented Nov 5, 2022

I'm going to lock this pull request because it has been closed for 120 days ⏳. This helps our maintainers find and focus on the active contributions.
If you have found a problem that seems related to this change, please open a new issue and complete the issue template so we can capture all the details necessary to investigate further.

@github-actions github-actions bot locked as resolved and limited conversation to collaborators Nov 5, 2022
Sign up for free to subscribe to this conversation on GitHub. Already have an account? Sign in.
Projects
None yet
Development

Successfully merging this pull request may close these issues.

add num_schedulers tunable to scheduler API
3 participants