The shardCollection command verifies that all documents in the collection to be sharded contain some value of the shard key by delegating to the checkShardingIndex command. The checkShardingIndex implements this check by performing a full index scan on the first non-multikey index it finds prefixed with the shard key. However, if this index is a sparse index, this check will incorrectly pass if there are documents in the collection with a missing value for the shard key. These documents will subsequently be orphaned after the shardCollection command succeeds.

Reproduce with the following script, which fails when the "shardCollection" command incorrectly reports success. I tested that the script fails with 2.4.12/2.6.9/3.0.2/master. Note that the test passes when the creation order of the two indexes is swapped (this causes the correct index to be chosen for the task).

var st = new ShardingTest({shards: 1});
var coll = st.getDB("test").foo;
assert.writeOK(coll.insert({}));
assert.commandWorked(coll.ensureIndex({a: 1, b: 1}, {sparse: true}));
assert.commandWorked(coll.ensureIndex({a: 1}));
assert.commandWorked(coll.getDB().adminCommand({enableSharding: coll.getDB().getName()}));
assert.commandFailed(coll.getDB().adminCommand({shardCollection: coll.getFullName(), key: {a: 1}}));
assert.eq(1, coll.find().itcount());