// workUnitStatus extracts a summary of the status of a single work
// unit. This produces its external coordinate status and the active
// attempt (if any) on success.
func workUnitStatus(workUnit coordinate.WorkUnit) (status WorkUnitStatus, attempt coordinate.Attempt, err error) {
var attemptStatus coordinate.AttemptStatus
attempt, err = workUnit.ActiveAttempt()
if err == nil && attempt == nil {
status = Available
return
}
if err == nil {
attemptStatus, err = attempt.Status()
}
if err == nil {
switch attemptStatus {
case coordinate.Pending:
status = Pending
case coordinate.Expired:
status = Available
attempt = nil
case coordinate.Finished:
status = Finished
case coordinate.Failed:
status = Failed
case coordinate.Retryable:
status = Available
attempt = nil
default:
err = errors.New("unexpected attempt status")
}
}
return
}
// UpdateWorkUnit causes some state change in a work unit. If the
// work unit is pending, this is the principal interface to complete
// or renew it; if it is already complete this can cause it to be
// retried.
func (jobs *JobServer) UpdateWorkUnit(
workSpecName string,
workUnitKey string,
options map[string]interface{},
) (bool, string, error) {
// Note that in several corner cases, the behavior of this as
// written disagrees with Python coordinated's:
//
// * If neither "lease_time" nor "status" is specified,
// Python coordinated immediately returns False without
// checking if workUnitKey is valid
//
// * Python coordinated allows arbitrary status changes,
// including AVAILABLE -> FINISHED
//
// * This openly ignores "worker_id", as distinct from Python
// coordinated, which logs an obscure warning and changes it,
// but only on a renew
var (
attempt coordinate.Attempt
changed bool
err error
status coordinate.AttemptStatus
uwuOptions UpdateWorkUnitOptions
workSpec coordinate.WorkSpec
workUnit coordinate.WorkUnit
)
err = decode(&uwuOptions, options)
if err == nil {
workSpec, err = jobs.Namespace.WorkSpec(workSpecName)
}
if err == nil {
workUnit, err = workSpec.WorkUnit(workUnitKey)
}
if err == nil {
if workUnit == nil {
return false, fmt.Sprintf("no such work unit key=%v", workUnitKey), nil
}
}
if err == nil {
attempt, err = workUnit.ActiveAttempt()
}
if err == nil && attempt != nil {
status, err = attempt.Status()
}
if err == nil && attempt != nil {
if status == coordinate.Expired || status == coordinate.Retryable {
// The Python Coordinate API sees both of these
// statuses as "available", and we want to fall
// into the next block.
attempt = nil
}
}
if err == nil && attempt == nil {
// Caller is trying to manipulate an AVAILABLE work
// unit. Cowardly refuse to start a new attempt on
// their behalf, or to update the persistent work unit
// data this way. (In theory there's no reason we
// *couldn't* do either, though I'm not aware of any
// callers that do; add_work_unit will replace
// existing work units and is the more typical way to
// refresh data.)
err = errors.New("update_work_unit will not adjust an available work unit")
}
if err == nil {
switch status {
case coordinate.Pending:
changed = true // or there's an error
switch uwuOptions.Status {
case 0, Pending:
err = uwuRenew(attempt, uwuOptions)
case Available:
err = attempt.Expire(uwuOptions.Data)
case Finished:
err = attempt.Finish(uwuOptions.Data)
case Failed:
err = attempt.Fail(uwuOptions.Data)
default:
err = errors.New("update_work_unit invalid status")
}
case coordinate.Expired:
err = errors.New("update_work_unit logic error, trying to refresh expired unit")
case coordinate.Finished:
switch uwuOptions.Status {
case 0, Finished:
changed = false // no-op
case Available:
err = workUnit.ClearActiveAttempt()
changed = true
case Failed:
changed = false // see below
default:
err = errors.New("update_work_unit cannot change finished unit")
}
case coordinate.Failed:
switch uwuOptions.Status {
case 0, Failed:
changed = false // no-op
case Available: // "retry"
err = workUnit.ClearActiveAttempt()
changed = true
case Finished:
// The Python worker, with two separate
// processes, has a race wherein there
// could be 15 seconds to go, the parent
// kills off the child, and the child
// finishes successfully, all at the same
// time. In that case the successful
// finish should win.
err = attempt.Finish(nil)
changed = true
default:
err = errors.New("update_work_unit cannot change failed unit")
}
case coordinate.Retryable:
err = errors.New("update_work_unit logic error, trying to refresh retryable unit")
default:
err = fmt.Errorf("update_work_unit invalid attempt status %+v", status)
}
}
return changed && err == nil, "", err
}
// TestAttemptLifetime validates a basic attempt lifetime.
func (s *Suite) TestAttemptLifetime(c *check.C) {
var (
err error
data map[string]interface{}
attempt, attempt2 coordinate.Attempt
aStatus coordinate.AttemptStatus
spec coordinate.WorkSpec
unit coordinate.WorkUnit
worker coordinate.Worker
uStatus coordinate.WorkUnitStatus
)
spec, worker = s.makeWorkSpecAndWorker(c)
// Create a work unit
unit, err = spec.AddWorkUnit("a", map[string]interface{}{}, 0.0)
c.Assert(err, check.IsNil)
// The work unit should be "available"
uStatus, err = unit.Status()
c.Assert(err, check.IsNil)
c.Check(uStatus, check.Equals, coordinate.AvailableUnit)
// The work unit data should be defined but empty
data, err = unit.Data()
c.Assert(err, check.IsNil)
c.Check(data, check.HasLen, 0)
// Get an attempt for it
attempts, err := worker.RequestAttempts(coordinate.AttemptRequest{})
c.Assert(err, check.IsNil)
c.Assert(attempts, check.HasLen, 1)
attempt = attempts[0]
// The work unit should be "pending"
uStatus, err = unit.Status()
c.Assert(err, check.IsNil)
c.Check(uStatus, check.Equals, coordinate.PendingUnit)
// The attempt should be "pending" too
aStatus, err = attempt.Status()
c.Assert(err, check.IsNil)
c.Check(aStatus, check.Equals, coordinate.Pending)
// The active attempt for the unit should match this
attempt2, err = unit.ActiveAttempt()
c.Assert(err, check.IsNil)
c.Check(attempt2, AttemptMatches, attempt)
// There should be one active attempt for the worker and it should
// also match
attempts, err = worker.ActiveAttempts()
c.Assert(err, check.IsNil)
c.Check(attempts, check.HasLen, 1)
if len(attempts) > 0 {
c.Check(attempts[0], AttemptMatches, attempt)
}
// The work unit data should (still) be defined but empty
data, err = unit.Data()
c.Assert(err, check.IsNil)
c.Check(data, check.HasLen, 0)
// Now finish the attempt with some updated data
err = attempt.Finish(map[string]interface{}{
"outputs": []string{"yes"},
})
c.Assert(err, check.IsNil)
// The unit should report "finished"
uStatus, err = unit.Status()
c.Assert(err, check.IsNil)
c.Check(uStatus, check.Equals, coordinate.FinishedUnit)
// The attempt should report "finished"
aStatus, err = attempt.Status()
c.Assert(err, check.IsNil)
c.Check(aStatus, check.Equals, coordinate.Finished)
// The attempt should still be the active attempt for the unit
attempt2, err = unit.ActiveAttempt()
c.Assert(err, check.IsNil)
c.Check(attempt2, AttemptMatches, attempt)
// The attempt should not be in the active attempt list for the worker
attempts, err = worker.ActiveAttempts()
c.Assert(err, check.IsNil)
c.Check(attempts, check.HasLen, 0)
// Both the unit and the worker should have one archived attempt
attempts, err = unit.Attempts()
c.Assert(err, check.IsNil)
c.Check(attempts, check.HasLen, 1)
if len(attempts) > 0 {
c.Check(attempts[0], AttemptMatches, attempt)
}
attempts, err = worker.AllAttempts()
c.Assert(err, check.IsNil)
c.Check(attempts, check.HasLen, 1)
if len(attempts) > 0 {
c.Check(attempts[0], AttemptMatches, attempt)
}
// This should have updated the visible work unit data too
data, err = unit.Data()
c.Assert(err, check.IsNil)
c.Check(data, check.HasLen, 1)
c.Check(data["outputs"], check.HasLen, 1)
c.Check(reflect.ValueOf(data["outputs"]).Index(0).Interface(),
check.Equals, "yes")
// For bonus points, force-clear the active attempt
err = unit.ClearActiveAttempt()
c.Assert(err, check.IsNil)
// This should have pushed the unit back to available
uStatus, err = unit.Status()
c.Assert(err, check.IsNil)
c.Check(uStatus, check.Equals, coordinate.AvailableUnit)
// This also should have reset the work unit data
data, err = unit.Data()
c.Assert(err, check.IsNil)
c.Check(data, check.HasLen, 0)
// But, this should not have reset the historical attempts
attempts, err = unit.Attempts()
c.Assert(err, check.IsNil)
c.Check(attempts, check.HasLen, 1)
if len(attempts) > 0 {
c.Check(attempts[0], AttemptMatches, attempt)
}
attempts, err = worker.AllAttempts()
c.Assert(err, check.IsNil)
c.Check(attempts, check.HasLen, 1)
if len(attempts) > 0 {
c.Check(attempts[0], AttemptMatches, attempt)
}
}