Is it worth allowing HttpClient injection in my library?

Question

I'm making a library for other developers to use and it's for communicating with a 3rd party web service. I read the issue that come up around HttpClient and I would like to handle this properly, so that the library doesn't become a burden.

I thought I would make a base ServiceClient that would have 3 ways to construct it.

The first would instantiate the HttpClient itself via using a static Lazy<HttpClient> to keep a single instance in the lib. If this constructor is never used, the inner object never gets instantiated.

The second constructor would accept an HttpMessageHandler and instantiate the HttpClient with it every time. This would be sort of a unit test-friendly mocking constructor that wouldn't get much use otherwise.

The third constructor would accept an HttpClient instance, making the developer responsible for the lifetime of the object if that is their wish.

My initial thought was that a developer might be already using an HttpClient in their code and they could pass that same instance to the lib and prevent having more than one instance. But then I thought the client and the lib would be communicating with totally different endpoints and the lib's configuration (base url, headers, etc) would conflict with the calling code, which is not useful for anyone.

So now I'm leaning towards dropping the 3rd constructor, because I'm not sure if it brings anything to the table.

Does anyone have experience with this?

Answer 1

HttpClient isn't really designed to be replaced with an alternative (it doesn't have a series of abstract interfaces which are generic).

So I doubt there are any API-compliant HttpClient replacements. That makes it a quite poor candidate for the kind of abstraction you are talking about.

The reasons which MIGHT make sense in your application is because there maybe some upstream requirements for communication which require either parameters to be passed in or otherwise specified. For example, you might need to specify a username/password, and client side cert, or set other parameters. It is for THIS reason, primarily, that allowing an optional CTOR argument of type HttpClient probably makes sense.

If you elect to provide an optional HttpClient parameter to the constructor, I would NOT define it so that the caller controls its lifetime. There would be no easy way to coordinate the assumptions about lifetime. Instead, consider it an ADOPTION, where your library ADOPTS ownership of the HttpClient passed in.

Answer 2

In case you don't need the full interface of HttpClient, I'd go another route: I'd create an interface for HttpClient and let the user inject an implementation.

You can define an interface that only requires the methods you actually require from HttpClient, let's say something like this:

public interface IHttpClient {
    public Task<HttpResponseMessage> PostAsync(string requestUri, HttpContent content);
}

and then offer (and use) a wrapper:

public class HttpClientWrapper: IHttpClient {
    internal HttpClientWrapper(HttpClient client) {
        _client = client ?? new HttpClient();
    }

    public static FromHttpClient(HttpClient client) {
        return new HttpClientWrapper(client);
    }

    public async Task<HttpResponseMessage> PostAsync(string requestUri, HttpContent content) 
    {
        return await _client.PostAsync(requestUri, content);
    }

}

This means that you can, in an explicit way, define what parts of an HttpClient you actually depend on; and, more importantly, if anyone wants to make some automated tests that use your library, the actual http requests can be mocked with a custom implementation.

Answer 3

In general HttpMessageHandler is designed for this kind of injection. It allows the instantiating code to add their own changes to the way the http messages are sent. For example by adding an auth header.

I have written some api clients where a HttpClient instance is passed into the constructor. This was too allow for an odd case with a web api (A) which called another web api (B). I wanted to recreate the api client object per request (to A) in order to pass in request specific data using construction parameters (on the B client) while avoiding the problem of creating an HttpClient object per request (to A).

But really if I am honest it probably wasn't the best solution. I wanted to avoid adding an extra parameter to all the methods, changing the interface and this was the way that caught my eye.

I think in another similar situation I simply passed in a IGetPerRequestData object which had access to the request context, So it would totally have been possible to do the job with a custom HttpMessageHandler.

Side note: you really shouldn't use a static to store the http handler, what if you have two instances pointing at different endpoints? simply store it as as private readonly field

public class MyApiClient
{
    private readonly HttpClient client;
    public MyApiClient(HttpMessageHandler handler = null)
    {
        if(handler != null) 
        { 
            this.client = new HttpClient(handler); 
        }
        else
        {
            client = new HttpClient();
        }
    }
}