I prefer to create an interface for setting query, loading, and saving. By using dependency injection, I can inject this into each component that requires it.
This allows flexibility in terms of replacing the configuration strategy, and gives a common base for everything to work from. I prefer this to a single, global "settings loader" (your option 2), especially since I can override the configuration mechanism for a single component if I absolutely need to do so.
I currently work on a system where the configuration is managed by one global singleton object that keeps a map of configuration keys to values. In general, I wish it hadn't been done this way because it can cause concurrency bottlenecks in the system and it's sloppy for unit testing, etc.
I think Reed Copsey has the right of it (I voted him up), but I would definitely recommend reading Martin Fowler's great article on dependency injection:
How about this. You define an interface Configurable with a single method configure(configuration). The configuration argument is simply a hashtable which associates the names of configuration parameters with their values.
Root objects can create a configuration hashtable in whatever way they want (ex: reading it from a config file). This hashtable may contain configuration parameters for the root object iselft, plus any parameter that one of its components, sub-components, sub-sub-components (etc) might use.
The root object then invokes configure(configuration) on all of its configurable components.
You can create multiple implementation of an interface that defines the config loader. Basically strategy pattern where you can define one basic interface as configLoader and then further different implementations such as FileSystemLoader, ClasspathLoader, EnvVariablesLoader etc.
Details at this link