> On Jan. 21, 2016, 6:51 p.m., Maxim Khutornenko wrote: > > src/main/java/org/apache/aurora/scheduler/pruning/TaskHistoryPruner.java, > > lines 152-161 > > <https://reviews.apache.org/r/42332/diff/3/?file=1203998#file1203998line152> > > > > Am I reading this as a busy loop consuming 100% thread CPU waiting for > > something that may never happen? I don't think this is an acceptable > > solution. > > > > Perhaps it's time to refactor task prunner into an > > AbstractScheduledService? I always felt task prunner approach of holding on > > to task IDs for 2 days just to act once on them isn't very efficient. What > > if instead of acting on every particular task ID we have a periodic (say > > every 30 seconds) run loop to prune job keys instead? > > > > Implementation-wise, it could be a Set of unique job keys that we > > populate on every TaskStateChange event. A runOneIteration() would poll > > that set and apply both latency and max_per_job thresholds for all related > > terminal tasks within the same iteration. > > > > The only downside for the above is a somewhat increased history count > > between the cleanup runs but given that our current thresholds are chosen > > mostly arbitrarily I think that should be acceptable. > > John Sirois wrote: > I think my Future/Queue suggestion above solves the busy loop with no > liveness penalty. That might allow your batching change suggestion to happen > in a seperate follow-up RB. > > Zameer Manji wrote: > +1 to John here. I think we are overdue for a less complex and heavy > pruner but I would prefer to keep this RB focused on failure propagation. I > am open to a follow up ticket and RB. Maxim, if you agree, I can create a > ticket that tracks the work you just proposed. > > Right now, I think I will use the Future/Queue suggestion that John has > to remove the busy loop.
I am fine with the follow up ticket if you feel it's too much to lift within this RB. - Maxim ----------------------------------------------------------- This is an automatically generated e-mail. To reply, visit: https://reviews.apache.org/r/42332/#review115663 ----------------------------------------------------------- On Jan. 20, 2016, 10:39 p.m., Zameer Manji wrote: > > ----------------------------------------------------------- > This is an automatically generated e-mail. To reply, visit: > https://reviews.apache.org/r/42332/ > ----------------------------------------------------------- > > (Updated Jan. 20, 2016, 10:39 p.m.) > > > Review request for Aurora, John Sirois and Maxim Khutornenko. > > > Bugs: AURORA-1582 > https://issues.apache.org/jira/browse/AURORA-1582 > > > Repository: aurora > > > Description > ------- > > Task pruning is key to operating a large cluster and failure to prune should > trigger shutdown to prevent unbounded growth of storage. This patch turns > `TaskHistoryPruner` into a service which propagates failure from failed > pruning attempts towards the `ServiceManager`. Also completing a TODO which > removes a test for behaviour that is very awkward to test for. > > > Diffs > ----- > > src/main/java/org/apache/aurora/scheduler/pruning/PruningModule.java > 735199ac1ccccab343c24471890aa330d6635c26 > src/main/java/org/apache/aurora/scheduler/pruning/TaskHistoryPruner.java > 2064089937f5178b1413d386a312f4173a0e35fb > > src/test/java/org/apache/aurora/scheduler/pruning/TaskHistoryPrunerTest.java > 295960f13693c6ba0d7075a8ef7f9680a91ae69d > > Diff: https://reviews.apache.org/r/42332/diff/ > > > Testing > ------- > > ./gradlew build -Pq > e2e tests > > > Thanks, > > Zameer Manji > >