Haven't personally encountered this yet – it seems keeping some stale servers around and routing to them would be sufficient? Long term, over-the-air upgrades (hot migrating connected clients) seems within reach, if a bit researchy - but the durable workflows projects seem to be making headway on the problem