Uploaded image for project: 'Flink'
  1. Flink
  2. FLINK-21513

Rethink up-/down-/restartingTime metrics




      While thinking about FLINK-21510 I stumbled upon some issues in the the semantics of these metrics, both from a user perspective and from our own, and I think we need to clarify some things.


      This metric describes the time since the job transitioned RUNNING state.
      It is meant as a measure for how stably a deployment is.

      In the default scheduler this transitions happens before we do any actual scheduling work, and as a result this also includes the time it takes for the JM to request slots and deploy tasks. In practive this means we start the timer once the job has been submitted and the JobMaster/Scheduler/EG have been initialized.

      For the adaptive scheduler this now puts us a bit into an odd situation because it first acquires slots before actually transitioning the EG into a RUNNING state, so as is we'd end up measuring 2 slightly different things.

      The question now is whether this is a problem.

      While we could certainly stick with the definition of "time since EG switched to RUNNING", it raises the question what the semantics of this metric are should a scheduler use a different data-structure than the EG.
      In other words, what I'm looking for is a definition that is independent from existing data-structures; a crude example could be "The time since the job is in a state where the deployment of a task is possible.".

      An alternative for the adaptive scheduler would be to measure the time since we transitioned to WaitingForResources, with which we would also include the slot acquisition, but it would be inconsistent with the logs and UI (because they only display an INITIALIZING job).


      This metric describes the time since the job transitioned into a RESTARTING state.
      It is meant as a measure for how long the recovery in case of a job failure takes.

      In the default scheduler this in practice is the time between a failure arriving at the JM and the cancellation of tasks being completed / restart backoff (whichever is higher).

      This is consistent with the semantics of the upTime metric, because upTime also includes the time required for acquiring slots and deploying tasks.

      For the adaptive scheduler we can follow similar semantics, by measuring the time we spend in the Restarting state.
      However, if we stick to the definition of upTime as time spent in RUNNING, then we will end up with a gap for the time spent in WaitingForResources.


      This metric describes the time between the job transitioning from FAILING to RUNNING.
      It is meant as a measure for how long the recovery in case of a job failure takes.

      You may be wondering what the difference between downTime and restartingTime is meant to be. Unfortunately I do not have the answer to that.
      Presumably, at the time they were added, they were covering different parts of the recovery process, but since we never documented these steps explicitly the exact semantics are no longer clear and there are no specs that a scheduler can follow.

      Furthermore, this metric is currently broken because a FAILING job never transitions into RUNNING anymore.

      The default scheduler transitions from RUNNING -> RESTARTING -> RUNNING, whereas the adaptive scheduler cancels the job and creates a new EG.

      As it is we could probably merge downTime and restartingTime because they seem to cover the exact same thing.


          Issue Links



              • Assignee:
                chesnay Chesnay Schepler
              • Votes:
                0 Vote for this issue
                7 Start watching this issue


                • Created: