I've actually revised the tests to work the way I needed them to.. (for some reason the observe in java is slower than I had hoped for, so I ended up making my own threaded contraption to solve this...) – might this be a separate bug?
anyway.. I've observed duplicate behavior:
Basically, many of the times, the stale test fails, returning exactly half of the keys in the view.
Maybe my cluster config is funky, but this is doubtful..
Anyway, we'll file a cluster bug with a 100% certainty that this is a client issue.
btw, I'd actually advocate keeping the threaded contraption there (in the commit I accidentally saved it as a single worker, might want to bump it up)..
Placing load on the server (i.e. by using multiple setter threads) seems to highlight this issue.. and I have a feeling it's a lag/race condition sort of thing.