Updating failure rate message and accounting (facebook#2723)

Summary:
Pull Request resolved: facebook#2723

This diff adds a clarification to failure-rate-exceeded errors that "abandoned" trials are also added in the failure rate accounting, which can help users look this up on their own.

In addition the diff changes the denominator of the failure-rate computation to only consider  trials with a `terminal` status.

Differential Revision: D61914570

ax/service/scheduler.py

@@ -77,8 +77,9 @@
 """
 FAILURE_EXCEEDED_MSG = (
  "Failure rate exceeds the tolerated trial failure rate of {f_rate} (at least "
- "{n_failed} out of first {n_ran} trials failed). Checks are triggered both at "
- "the end of a optimization and if at least {min_failed} trials have failed."
+ "{n_failed} out of first {n_ran} trials failed or abandoned). Checks are triggered both at "
+ "the end of a optimization and if at least {min_failed} trials have either failed, "
+ "or have been abandoned, potentially automatically due to issues with the trial."
 )
 
 
@@ -850,13 +851,16 @@ def error_if_failure_rate_exceeded(self, force_check: bool = False) -> None:
  ):
  return
 
- num_ran_in_scheduler = (
- len(self.experiment.trials) - self._num_preexisting_trials
+ num_ran_in_scheduler = sum(
+ 1
+ for idx, t in self.experiment.trials.items()
+ if idx >= self._num_preexisting_trials and t.status.is_terminal
  )
-
  failure_rate_exceeded = (
- num_bad_in_scheduler / num_ran_in_scheduler
- ) > self.options.tolerated_trial_failure_rate
+ (num_bad_in_scheduler / num_ran_in_scheduler)
+ > self.options.tolerated_trial_failure_rate
+
+ )
 
  if failure_rate_exceeded:
  if self._num_trials_bad_due_to_err > num_bad_in_scheduler / 2: