This is probably a websockets bug that crops up during deploys – the connection breaks and reconnects keep getting sent to the now-gone VM for >60s. We're furiously rebuilding our service discovery plumbing because it's slow to propagate. I'm hoping we have the websockets issue fixed this week.