[ismarc]

We recently have been having to work around an issue that, on top of all the other issues we've had means we're moving away from mongo entirely. The current issue is that a mapreduce on a cluster into a sharded collection will, at some point, start silently not writing 30% of the data into the sharded collection. And from that point onward, roughly the same amount of data is lost any time it goes to that collection. We have to watch for it and create an entirely new sharded collection to mapreduce into, then everything works fine for a while until that one starts not getting some data. When we create a new collection, we just pick a new name, copy all the previous data over, update the MRs to point to that collection and hit go. I would be able to understand data loss caused by certain durability settings from a client, but on mapreduce results?