It definitely seems like it would result in more predictable behaviour. I think the other reasonable alternative is that we actually test the behaviour of the system when it gets into an OOM state and make sure that it can recover (which might not be possible with a reasonable amount of effort).
On Wed, Jun 13, 2018 at 6:50 PM, Quanlong Huang <[email protected]> wrote: > I think this is really heapful since OOM is usually unrecoverable. > Cloudera Manager launchs most of the Hadoop components with this flag.JVM > flags like -XX:+HeapDumpOnOutOfMemoryError and -XX:HeapDumpPath=... maybe > useful too. > > -- Quanlong > > > At 2018-06-14 03:37:07, "Philip Zeyliger" <[email protected]> > wrote: > >Hi! > > > >Seeing bugs like IMPALA-7168, it leads me to think that we should > recommend > >folks who are running Impala to use -XX:OnOutOfMemoryError="kill -9 %p" or > >equivalent on their JVMs, to simply crash the process on frontend OOM. > > > >Any thoughts on whether this is the right approach? > > > >-- Philip >
