I've finally sat down and written out what I think could be a viable anti-affinity algorithm, one which will restart historically placed instances fast, then request new containers one-by-one
https://issues.apache.org/jira/browse/SLIDER-82 Comments welcome, and ideas for more tests —things we can test reliably, of course. If we need to add some more metrics/info from the AM REST API or metrics to see what has happened, that's fine. Indeed, it may be useful anyway to have a history for each component to help understand what's gone on. (see SLIDER-845 there) This is going to be major, but it's also the most asked for placement feature. There's already been one proposal using blacklisting —but I think that can be switched to nodemap driven, and include historical placement data for fast restart.
