Constructor injection with non-dependency parameters

Question

I have an interface ITradingApi like so:

public interface ITradingApi
{
    IOrder CreateOrder(...);
    IEnumerable<Symbol> GetAllSymbols();
    // ...
}

This is meant to be a facade for the different APIs of the vendors of trading software. My view model has a dependency on this trading API in its constructor:

public class MainViewModel
{
    public MainViewModel(ITradingApi tradingApi) { /* ... */ }
    // ...
}

I use Ninject as an IoC container, so I will create an instance of my view model like this:

var vm = kernel.Get<MainViewModel>();

Now, my problem:

The implementation of ITradingApi might need additional parameters to work.
Example:

• One vendors API uses TCP/IP internally, so I need a hostname and a port.
• Another vendor uses a COM object. Here I don't need any info.
• A third vendor needs username and password of the account.

In the spirit of not allowing incomplete objects, I added these as parameters to the constructors of the concrete implementations.

Now, I am not sure, how this would work. Clearly, these additional parameters do not belong into the interface, because they are specific to each implementation.
On the other hand, these additional parameters need to be entered by the end-user and then passed to the implementation of ITradingApi, meaning that the user of ITradingApi needs intimate knowledge about the concrete implementation.
How to solve this dilemma?

UPDATE:
One approach could be to create an ITradingApiProvider that exposes a list of required parameters. The View could automatically create an input form for these parameters that is databound to the parameters in ITradingApiProvider. Now, when an ITradingApi instance is requested from the provider, it can make use of these parameters to create an instance of the concrete implementation. Clearly the implementation of ITradingApiProvider and ITradingApi are tightly coupled, but I think that is not a problem as long as each implementation of ITradingApi comes with a corresponding implementation of ITradingApiProvider.

Answer 1

Based on the information so far put forth here, I'd like to point out one or two things:

First of all, whether or not the concrete configuration values are supplied at composition time or truly first available at runtime as user input makes a huge difference. As long as they can be resolved at composition time things are easy because you can simply read the values from the environment and supply them to the appropriate constructors. So, for the rest of this answer I'm going to assume that things are much harder and you actually need to get those values from the user at runtime.

Instead of attempting to come up with a general-purpose configuration API I'd much rather model what's actually going on. In this case it sounds to me like we're collecting configuration values from the user, so why not model this explicitly?

Product Trader

Define an interface like this:

public interface ITradingApiTrader
{
    ITradingApi Create(Type apiType);
}

Here, it's assumed that apiType can cast to ITradingApi, but this can't be enforced by the compiler. (The reason I'm calling this a 'Trader' is because this is a variation of the Product Trader pattern (PLoPD 3).)

How is this different than before?

Well, you can implement the Create method by showing a user interface for each type of ITradingApi. Each concrete user interface gathers the values required for its own concrete ITradingApi implementation and subsequently returns a correctly configured instance.

If you know the concrete types at compile time, other variations include these:

public interface ITradingApiTrader
{
    ITradingApi CreateMT4TradingApi();

    ITradingApi CreateFooTradingApi();

    ITradingApi CreateBarTradingApi();

    // etc.
}

Perhaps you can also do this (although I haven't tried to compile this):

public interface ITradingApiTrader
{
    ITradingApi Create<T>() where T : ITradingApi;
}

Note also that you don't need to define the first ITradingApiTrader's Create method based on a Type - any identifier (such as an enum or string) might do instead.

Visitor

If the set of ITradingApi is (finite and) known at design time, the Visitor design pattern might also offer an alternative.

If you use a Visitor, you can make the Visit method show an appropriate user interface and then subsequently use the values collected from the user interface to create the appropriate ITradingApi instance.

Basically this is just a variation on the previous 'solution' where the Product Trader is implemented as a Visitor.

Answer 2

Is this what your after?

   ninjectKernel.Get<MainViewModel>().WithConstructorArgument("tradingApi", 
       kernel.Get<ITaxCalculator>() .WithConstructorArgument("additionalParameter","someValue")));

Answer 3

Ok my two cents, I am not sure of anything you know. It is just to help and try...

We give a visitor to your api as construction of the interface:

public interface ITradingApi
{
    Object CreateOrder();
    IEnumerable<Object> GetAllSymbols();
}

public class TradingApi : ITradingApi
{
    IvisitorAPI _VisitorAPI;

    public TradingApi(IvisitorAPI visitorAPI)
    {
        _VisitorAPI = visitorAPI;
    }


    public Object CreateOrder()
    {
        var Order = new Object();
        //bla bla bla

        //here code relative to different visitor
        _VisitorAPI.SaveOrder(Order);

        return Order;
    }
}

It is your visitor that knows how to handle some of the action, because depending on the visitor he will use your api in different ways to achieve the same action ( here SaveOrder).

public interface IvisitorAPI
{
    bool SaveOrder(Object order);
}



public class visitorApiIP : IvisitorAPI
{
    public string HostName { get; set; }
    public int Port { get; set; }

    public visitorApiIP(string hostname, int port)
    {
        HostName = hostname;
        Port = port;
    }


    public bool SaveOrder(Object order)
    {
        //save the order using hostname and ip
        //...
        //....
        return true;
    }
}

Only the visitor has a knowledge of what he needs to achieve his version of the action. Therefore it is not the APi that needs additionnal parameters, we are pushing the logic away in the visitor class. This visitor class might be created only when ewe know who is the visitor therefore, surely at runtime

Hope it might give you some perspective. I do not know if the whole theory can be applied your exact situation.

My best anyway ;)

Answer 4

The solution is to use the approach as outlined in the update part of my question. ITradingApiProvider takes the role of an abstract factory and thus should be renamed to ITradingApiFactory. It would expose a list of needed parameters whose values can be set. This list in turn can be used by the View to automatically present the user with an input form to enter a value for each parameter, because only the user knows the values of for the parameters.
The call to Create would then use these parameters:

public interface ITradingApiFactory
{
    ITradingApi Create();
    IEnumerable<Parameter> Parameters { get; }
}

public class Parameter
{
    public Parameter(Type type, string name, string description)
    { Type = type; Name = name; Description = description; }

    public Type Type { get; private set; }
    public string Name { get; private set; }
    public string Description { get; private set; }
    public object Value { get; set; }
}

public class MT4TradingApiFactory : ITradingApiFactory
{
    Dictionary<string, Parameter> _parameters;

    public MT4TradingApiFactory()
    { /* init _parameters */ }

    public ITradingApi Create()
    {
        return new MT4TradingApi(_parameters["hostname"].ToString(),
                                 (int)_parameters["port"]);
    }

    IEnumerable<Parameter> Parameters { get { return _parameters.Values; } }
}

More info can be found in this answer.

This can be advanced further to make it easier to use, by giving each Factory implementation the parameters as properties and change the Parameter class to work directly on these properties using expression trees. If someone is interested in this advanced factory design, please leave a comment.

Answer 5

I think there is nothing wrong with your provider approach. You have two concerns here:

1. An operational one: your ITradingAPI which defines a contract for operations you can perform.
2. A meta-data one: something which describes properties of an actual implementation (meta data might not be quiet right but can't think of a better name for it)

Now apparently you need something which can make the connection between the two and that is your ITradingAPIProvider. Seems reasonable straight forward and has good chance of that you will still understand your code when coming back to it after a year ot two ;)

Answer 6

How about trying something similar to the strategy pattern? Create a new interface called IConnectStrategy:

interface IConnectStrategy
{
    void Connect();
}

Add the connectstrategy as an argument to the method void CreateOrder(IConnectStrategy connectStrategy) in ITradingApi and let each vendor create/specify their own method for connecting. E.g. for one vendor create:

public class TCPConnectStrategy : IConnectStrategy
{
    public TCPConnectStrategy(string hostName, int port)
    {
        /* ... */
    }

    public void Connect()
    {
        /* ... tcp connect ... */
    }
}

(Connect might not be the best name or even what you are actually doing, but please apply it to whatever works for your project.)

Edit after comments: Create a strategy that only have contracts for each method that have vendor-specific parameters. Then add a method void SetVendorStrategy(IVendorStrategy vendorStrategy) (or a property) to the ITradingAPI-interface. Each implementation of the strategy has their own constructor with their own parameters, and each method (that require vendor specific parameters) in each implementation of the ITradingAPI-interface simply calls vendorStrategy.DoSomethingWithVendorSpecificData().