I've seen Issue 2028 (https://jira.mongodb.org/browse/SERVER-2028) and I don't think a sparse index is a good enough solution. One of the reasons I think a sparse index fails as a solution is because when the index is used in a lookup, the documents that do have null values in the indexed field are ignored in search results. Let me give a concrete example of what I think the behavior should be...

Let's say I'm adding a new property to a "users" collection. The new property is an api token, and needs to be unique across the entire collection. I'd like to be able to do db.users.ensureIndex(

), but this will of course fail b/c mongo will hit more than one null value. On the other hand, a sparse index here doesn't work because when I try to do a read query on that column where that column would get used (db.users.find({token: {"$ne": null}}).count) the result set excludes all the users b/c it's a new property.