For all cluster types (standalone, repl, sharded), attempting to create an already-existing collection should result in an error, as with this test against a 3.6 server:

mongos> db.runCommand({create : "foo"})
{
	"ok" : 1,
        ...
}
mongos> db.runCommand({create : "foo"})
{
	"ok" : 0,
	"errmsg" : "a collection 'test.foo' already exists",
	"code" : 48,
	"codeName" : "NamespaceExists",
       ...
}

But with a sharded cluster running on the nightly build it no longer does:

mongos> db.runCommand({create : "foo"})
{
	"ok" : 1,
        ...
}
mongos> db.runCommand({create : "foo"})
{
	"ok" : 1,
        ...
}

The behavioral difference was detected for the first time in this driver regression test. The git hash of the server in that test run is 3.7.1-280-g43fbd6a. The previous successful run of that test used a server with a git hash of 3.7.1-253-g2e1f172bc1.