is there a reason why we use unique_ptr for all clients in chromeos/dbus/dbus_clients_browser.h?

it looks like a client is 'get' as a raw pointer all the time which isn't ideal. 

do you think it would be a practical idea to switch to scoped_refptr from unique_ptr which seems more applicable here?

DBusClientsBrowser is owned by DBusThreadManager which provides explicit initialization and shutdown of DBus services.

The clients in DBusClientsBrowser should not be able to outlive DBusThreadManager and therefore should continue to use unique_ptr, not refptr.

Consumers of the DBus clients should be aware of the lifetime of DBusThreadManager and should call DBusThreadManager::IsInitialized() if there is any uncertainty (e.g. during shutdown / destruction).

It's a bit tangential, but I've recently recommended in some reviews that people make their c'tors take pointers to Client objects instead of calling DBusThreadManager::Get() directly. This can make unit tests a bit simpler in some cases, since a test can initialize and pass its own Fake*Client object instead of needing to do an ugly dance to get a raw pointer to the DBusThreadManager-owned client, which also often requires a downcast. Any opinions about this?

Re comment #3: That is a fine pattern, as long as the handler has clear ownership and is clearly destroyed before DBusThreadManager (which is a better pattern in general, but may be challenging for classes that are global singletons or owned by global singletons).