> On June 28, 2017, 10:22 a.m., David McLaughlin wrote: > > This change appears to have caused adverse behavior when we deployed it to > > production. We are seeing more tasks lost due to being launched in invalid > > offers. > > > > I think we should revert this change until we root cause it. > > Zameer Manji wrote: > This change caused less lost tasks when deployed to production for me. > Have you isolated it 100% to this change? What do you see for the > `offer_cancel_failures` metric? > > David McLaughlin wrote: > Like 99.9% sure it's this commit. > > We'll revert this commit on one cluster and compare the differences.
Note that this code fixes a clear race condition, so if this is causing a problem in production I feel it is exposing another problem more so than causing another one. I would like to repeat that the log snippet in the description shows that a rescind can be processed before the offer it self, which leaves an invalid offer id in the offer manager. This causes lost tasks with the reason of invalid offer. Reverting this change keeps you open to this problem. One thing that you should do, is check the logs for offer ids that are rejected for being invalid and confirm the times that mesos sent the offer and send the rescind. It is possible/likely that the rescind came right after the offer which exposes a problem in Mesos. I think reverting this change is a bad idea based on the evidence that it is fixing a clear problem. - Zameer ----------------------------------------------------------- This is an automatically generated e-mail. To reply, visit: https://reviews.apache.org/r/59853/#review179131 ----------------------------------------------------------- On June 6, 2017, 1:56 p.m., Zameer Manji wrote: > > ----------------------------------------------------------- > This is an automatically generated e-mail. To reply, visit: > https://reviews.apache.org/r/59853/ > ----------------------------------------------------------- > > (Updated June 6, 2017, 1:56 p.m.) > > > Review request for Aurora, David McLaughlin and Santhosh Kumar Shanmugham. > > > Bugs: AURORA-1933 > https://issues.apache.org/jira/browse/AURORA-1933 > > > Repository: aurora > > > Description > ------- > > In a a production environment I was able to observe the following: > ``` > I0606 00:31:32.510 [Thread-77638, > MesosCallbackHandler$MesosCallbackHandlerImpl:229] Offer rescinded: > 81e04cbd-9bce-41cf-bd94-38c911f255e4-O142359552 > I0606 00:31:32.903 [SchedulerImpl-0, > MesosCallbackHandler$MesosCallbackHandlerImpl:211] Received offer: > 81e04cbd-9bce-41cf-bd94-38c911f255e4-O142359552 > I0606 00:31:34.815 [TaskGroupBatchWorker, > VersionedSchedulerDriverService:123] Accepting offer > 81e04cbd-9bce-41cf-bd94-38c911f255e4-O142359552 with ops [LAUNCH] > ``` > > Notice that the offer rescind was processed before the actual offer. This is > possible because there is a race in the `MesosCallbackHandlerImpl`. The offer > is > processed in the executor (to prevent blocking) and the rescind is handled > directly. This means the offer procecssing thread (`SchedulerImpl-0`) is > racing > against the callback thread (`Thread-77638`). > > In normal operation, there will be seconds to minutes between a rescind and an > offer, but in some cases an offer can be rescinded very quickly in clusters > that > use oversubscription modules. > > To fix this, we move the rescind processing into the same executor as the > offer > processing to ensure they are processed in the order they are received. > Without > fixing this, the rescinded offer exists in the offer manager and can be used > later to launch a task. This task will immediately fail to launch because the > offer is invalid. > > In this patch, I have also added a metric and logging to record when we fail > to > remove an offer from the offer manager, and cleaned up the logging to allow > operators to see when an offer was recieved. With this logging, an operator > can > grep for the offer id and see the entire lifecycle of the offer in the > scheduler. > > > Diffs > ----- > > src/main/java/org/apache/aurora/scheduler/mesos/MesosCallbackHandler.java > 5a5281aeaea1e2a4e0eab67069605838ee809c6c > > src/main/java/org/apache/aurora/scheduler/mesos/VersionedSchedulerDriverService.java > 5e86504c70083065278864e6ab1cc85c83a45a28 > src/main/java/org/apache/aurora/scheduler/offers/OfferManager.java > 17e577b069df9232d57cde171a078d9f6db707ea > src/test/java/org/apache/aurora/scheduler/offers/OfferManagerImplTest.java > 97febf25cea2024e0ca43366b3d4578e67734884 > > > Diff: https://reviews.apache.org/r/59853/diff/1/ > > > Testing > ------- > > > Thanks, > > Zameer Manji > >
