Github user philippbussche commented on a diff in the pull request:
https://github.com/apache/flink/pull/2683#discussion_r85644418
--- Diff:
flink-runtime/src/main/scala/org/apache/flink/runtime/jobmanager/JobManager.scala
---
@@ -1828,6 +1828,33 @@ class JobManager(
jobManagerMetricGroup.gauge[Long, Gauge[Long]]("numRunningJobs", new
Gauge[Long] {
override def getValue: Long = JobManager.this.currentJobs.size
})
+ jobManagerMetricGroup.gauge[Long, Gauge[Long]]("numFailedJobs", new
Gauge[Long] {
+ override def getValue: Long = {
+ var failedJobs = 0
+ val ourJobs = createJobStatusOverview()
+ val future = (archive ?
RequestJobsOverview.getInstance())(timeout)
+ val archivedJobs : JobsOverview = Await.result(future,
timeout).asInstanceOf[JobsOverview]
+ failedJobs += ourJobs.getNumJobsFailed() +
archivedJobs.getNumJobsFailed()
+ failedJobs
+ }})
+ jobManagerMetricGroup.gauge[Long, Gauge[Long]]("numCancelledJobs", new
Gauge[Long] {
+ override def getValue: Long = {
+ var cancelledJobs = 0
+ val ourJobs = createJobStatusOverview()
+ val future = (archive ?
RequestJobsOverview.getInstance())(timeout)
+ val archivedJobs : JobsOverview = Await.result(future,
timeout).asInstanceOf[JobsOverview]
+ cancelledJobs += ourJobs.getNumJobsCancelled() +
archivedJobs.getNumJobsCancelled()
+ cancelledJobs
+ }})
+ jobManagerMetricGroup.gauge[Long, Gauge[Long]]("numFinishedJobs", new
Gauge[Long] {
+ override def getValue: Long = {
+ var finishedJobs = 0
+ val ourJobs = createJobStatusOverview()
+ val future = (archive ?
RequestJobsOverview.getInstance())(timeout)
+ val archivedJobs : JobsOverview = Await.result(future,
timeout).asInstanceOf[JobsOverview]
+ finishedJobs += ourJobs.getNumJobsFinished() +
archivedJobs.getNumJobsFinished()
+ finishedJobs
+ }})
--- End diff --
@zentol , @rmetzger not sure if you meant something like to following
perhaps to work around the RPC calls (I guess this would eliminate them
completely however are the changes a bit more "invasive" I would think):
So what we could do is where we receive the Job status change
```
case JobStatusChanged(jobID, newJobStatus, timeStamp, error) =>
```
Send a RemoveJob message with the newJobStatus (could be "cancelled" as far
as I can see)
```
self ! decorateMessage(RemoveJob(jobID, removeJobFromStateBackend = true,
newJobStatus))
```
Then for the RemoveJob message propagate this to the method that actually
removes the job:
```
removeJob(graph.getJobID, clearPersistedJob, newJobStatus)
```
And finally besides taking the job out of the map representing the
currentJobs maybe increment a counter for canceled archived jobs or have a
different map that stores archived canceled jobs (and others for failed and
finished) so that when it is time for sending out the monitoring metrics we
would use this map/counter rather than doing any RPC calls at all.
What do you guys think ?
---
If your project is set up for it, you can reply to this email and have your
reply appear on GitHub as well. If your project does not have this feature
enabled and wishes so, or if the feature is enabled but not working, please
contact infrastructure at [email protected] or file a JIRA ticket
with INFRA.
---
