While it is possible to extend anti-affinity to take care of this, I feel it will cause confusion from a user perspective. As a user, when I think about anti-affinity, what comes to mind right away is a relative relation between operators.
On the other hand, the current ask is not that, but a relation at an application level w.r.t. a node. (Further, we might even think of extending this at an operator level - which would mean do not deploy an operator on a particular node) We would be better off clearly articulating and allowing users to configure it seperately as against using anti-affinity. On Fri, Dec 2, 2016 at 10:03 AM, Bhupesh Chawda <bhup...@datatorrent.com> wrote: > Okay, I think that serves an alternate purpose of detecting any newly gone > bad node and excluding it. > > +1 for covering the original scenario under anti-affinity. > > ~ Bhupesh > > On Fri, Dec 2, 2016 at 9:14 AM, Munagala Ramanath <r...@datatorrent.com> > wrote: > > > It only takes effect after failures -- no way to exclude from the get-go. > > > > Ram > > > > On Dec 1, 2016 7:15 PM, "Bhupesh Chawda" <bhup...@datatorrent.com> > wrote: > > > > > As suggested by Sandesh, the parameter > > > MAX_CONSECUTIVE_CONTAINER_FAILURES_FOR_BLACKLIST seems to do exactly > > what > > > is needed. > > > Why would this not work? > > > > > > ~ Bhupesh > > > > > > -- ~Milind bee at gee mail dot com