ploeh blog danish software design

How to ensure a Single Writer in load-balanced Azure deployments

In my Functional Architecture with F# Pluralsight course, I describe how using the Actor model (F# Agents) can make a concurrent system much simpler to implement, because the Agent can ensure that the system only has a Single Writer. Having a Single Writer eliminates much complexity, because while the writer decides what to write (if at all), nothing changes. Multiple readers can still read data, but as long as the Single Writer can keep up with input, this is a much simpler way to deal with concurrency than the alternatives.

However, the problem is that while F# Agents work well on a single machine, they don't (currently) scale. This is particularly notable on Azure, because in order get the guaranteed SLA, you'll need to deploy your application to two or more nodes. If you have an F# Agent running on both nodes, obviously you no longer have a Single Writer, and everything just becomes much more difficult. If only there was a way to ensure a Single Writer in a distributed environment...

Fortunately, it looks like the (in-preview) Azure feature Web Jobs (inadvertently) solves this major problem for us. Web Jobs come in three flavours:

• On demand
• Continuously running
• Scheduled

That turns out not to be a particularly useful option as well, because

"If your website runs on more than one instance, a continuously running task will run on all of your instances."

"On demand and scheduled tasks run on a single instance selected for load balancing by Microsoft Azure."

There's just one concern that we need to address: what happens if a Scheduled Web Job is taking too long running, in such a way that it hasn't completed when it's time to start it again. For example, what if you run a Scheduled Web Job every minute, but it sometimes takes 90 seconds to complete? If a new process starts executing while the first one is running, you would no longer have a Single Writer.

Reading the documentation, I couldn't find any information about how Azure handles this scenario, so I decided to perform some tests.

The Qaiain email micro-service proved to be a fine tool for the experiment. I slightly modified the code to wait for 90 seconds before exiting:

[<EntryPoint>]
let main argv = 
    match queue |> AzureQ.dequeue with
    | Some(msg) ->
        msg.AsString |> Mail.deserializeMailData |> send
        queue.DeleteMessage msg
    | _ -> ()
 
    Async.Sleep 90000 |> Async.RunSynchronously
 
    match queue |> AzureQ.dequeue with
    | Some(msg) ->
        msg.AsString |> Mail.deserializeMailData |> send
        queue.DeleteMessage msg
    | _ -> ()
 
    0 // return an integer exit code

In addition to that, I also changed how the subject of the email that I would receive would look, in order to capture the process ID of the running application, as well as the time it sent the email:

smtpMsg.Subject <-
    sprintf
        "Process ID: %i, Time: %O"
        (Process.GetCurrentProcess().Id)
        DateTimeOffset.Now

My hypothesis was that if Scheduled Web Jobs are well-behaved, a new job wouldn't start if an existing job was already running. Here are the results:

This looks great, but it's easier to see if I visualize it:

As you can see, processes do not overlap in time. This is a highly desirable result, because it seems to guarantee that we can have a Single Writer running in a distributed, load-balanced system.

Azure Web Jobs are currently in preview, so let's hope the Azure team preserve this functionality in the final version. If you care about this, please let the team know.

With F# and Unquote, you can write customized, composable assertions.

Yesterday, I wrote this unit test:

[<Theory; UnitTestConventions>]
let PostReturnsCorrectResult
    (sut : TasksController)
    (task : TaskRendition) =
 
    let result : IHttpActionResult = sut.Post task
 
    verify <@ result :? Results.StatusCodeResult @>
    verify <@ HttpStatusCode.Accepted =
        (result :?> Results.StatusCodeResult).StatusCode @>

For the record, here's the SUT:

type TasksController() =
    inherit ApiController()
    member this.Post(task : TaskRendition) =
        this.StatusCode HttpStatusCode.Accepted :> IHttpActionResult

There's not much to look at yet, because at that time, I was just getting started, and as always, I was using Test-Driven Development. The TasksController class is an ASP.NET Web API 2 Controller. In this incarnation, it merely accepts an HTTP POST, ignores the input, and returns 202 (Accepted).

The unit test uses AutoFixture.Xunit to create an instance of the SUT and a DTO record, but that's not important in this context. It also uses Unquote for assertions, although I've aliased the test function to verify. Although Unquote is an extremely versatile assertion module, I wasn't happy with the assertions I wrote.

What's the problem? #

The problem is the duplication of logic. First, it verifies that result is, indeed, an instance of StatusCodeResult. Second, if that's the case, it casts result to StatusCodeResult in order to access its concrete StatusCode property; it feels like I'm almost doing the same thing twice.

You may say that this isn't a big deal in a test like this, but in my experience, this is a smell. The example looks innocuous, but soon, I'll find myself writing slightly more complicated assertions, where I need to type check and cast more than once. This can rapidly lead to Assertion Roulette.

The xUnit.net approach #

For a minute there, I caught myself missing xUnit.net's Assert.IsAssignableFrom<T> method, because it returns a value of type T if the conversion is possible. That would have enabled me to write something like:

let scr = Assert.IsAssignableFrom<Results.StatusCodeResult> result
Assert.Equal(HttpStatusCode.Accepted, scr.StatusCode)

It seems a little nicer, although in my experience, this quickly turns to spaghetti, too. Still, I found myself wondering if I could do something similar with Unquote.

A design digression #

At this point, you are welcome to pull GOOS at me and quote: listen to your tests! If the tests are difficult to write, you should reconsider your design; I agree, but I can't change the API of ASP.NET Web API. In Web API 1, my preferred return type for Controller actions were HttpResponseMessage, but it was actually a bit inconvenient to work with in unit tests. Web API 2 introduces various IHttpActionResult implementations that are easier to unit test. Perhaps this could be better, but it seems like a step in the right direction.

In any case, I can't change the API, so coming up with a better way to express the above assertion is warranted.

Composed assertions #

To overcome this little obstacle, I wrote this function:

let convertsTo<'a> candidate =
    match box candidate with
    | :? 'a as converted -> Some converted
    | _ -> None

(You have to love a language that let's you write match box! There's also a hint of such nice over Some converted...)

The convertsTo function takes any object as input, and returns an Option containing the converted value, if the conversion is possible; otherwise, it returns None. In other words, the signature of the convertsTo function is obj -> 'a option.

This enables me to write the following Unquote assertion:

[<Theory; UnitTestConventions>]
let PostReturnsCorrectResult
    (sut : TasksController)
    (task : TaskRendition) =
 
    let result : IHttpActionResult = sut.Post task
    verify <@
            result
            |> convertsTo<Results.StatusCodeResult>
            |> Option.map (fun x -> x.StatusCode)
            |> Option.exists ((=) HttpStatusCode.Accepted) @>

While this looks more verbose than my two original assertions, this approach is more composable.

The really beautiful part of this is that Unquote can still tell me what goes wrong, if the test doesn't pass. As an example, if I change the SUT to:

type TasksController() =
    inherit ApiController()
    member this.Post(task : TaskRendition) =
        this.Ok() :> IHttpActionResult

The assertion message is:

System.Web.Http.Results.OkResult |> Dsl.convertsTo |> Option.map (fun x -> x.StatusCode) |> Option.exists ((=) Accepted)
None |> Option.map (fun x -> x.StatusCode) |> Option.exists ((=) Accepted)
None |> Option.exists ((=) Accepted)
false

Notice how, in a series of reductions, Unquote breaks down for me exactly what went wrong. The top line is my original expression. The next line shows me the result of evaluating System.Web.Http.Results.OkResult |> Dsl.convertsTo; the result is None. Already at this point, it should be quite evident what the problem is, but in the next line again, it shows the result of evaluating None |> Option.map (fun x -> x.StatusCode); again, the result is None. Finally, it shows the result of evaluating None |> Option.exists ((=) Accepted), which is false.

Here's another example. Assume that I change the SUT to this:

type TasksController() =
    inherit ApiController()
    member this.Post(task : TaskRendition) =
        this.StatusCode HttpStatusCode.OK :> IHttpActionResult

In this example, instead of returning the wrong implementation of IHttpActionResult, the SUT does return a StatusCodeResult instance, but with the wrong status code. Unquote is still very helpful:

System.Web.Http.Results.StatusCodeResult |> Dsl.convertsTo |> Option.map (fun x -> x.StatusCode) |> Option.exists ((=) Accepted)
Some System.Web.Http.Results.StatusCodeResult |> Option.map (fun x -> x.StatusCode) |> Option.exists ((=) Accepted)
Some OK |> Option.exists ((=) Accepted)
false

Notice that it still uses a series of reductions to show how it arrives at its conclusion. Again, the first line is the original expression. The next line shows the result of evaluating System.Web.Http.Results.StatusCodeResult |> Dsl.convertsTo, which is Some System.Web.Http.Results.StatusCodeResult. So far so good; this is as required. The third line shows the result of evaluating Some System.Web.Http.Results.StatusCodeResult |> Option.map (fun x -> x.StatusCode), which is Some OK. Still good. Finally, it shows the result of evaluating Some OK |> Option.exists ((=) Accepted), which is false. The value in the option was HttpStatusCode.OK, but should have been HttpStatusCode.Accepted.

Summary #

Unquote is a delight to work with. As the project site explains, it's not an API or a DSL. It just evaluates and reports on the expressions you write. If you already know F#, you already know how to use Unquote, and you can write your assertion expressions as expressive and complex as you want.

Announcing Exude, an extension to xUnit.net providing test cases as first-class, programmatic citizens.

Sometimes, when writing Parameterized Tests with xUnit.net, you need to provide parameters that don't lend themselves easily to be defined as constants in [InlineData] attributes.

In Grean, we've let ourselves inspire by Mauricio Scheffer's blog post First-class tests in MbUnit, but ported the concept to xUnit.net and released it as open source.

It's called Exude and is available on GitHub and on NuGet.

Here's a small example:

[FirstClassTests]
public static IEnumerable<ITestCase> YieldFirstClassTests()
{
    yield return new TestCase(_ => Assert.Equal(1, 1));
    yield return new TestCase(_ => Assert.Equal(2, 2));
    yield return new TestCase(_ => Assert.Equal(3, 3));
}

More examples and information is available on the project site.

This post explains how to configure FsCheck to create arbitrary Version values.

When I unit test generic classes or methods, I often like to use Version as one of the type arguments. The Version class is a great test type because

• it's readily available, as it's defined in the System namespace in mscorlib
• it overrides Equals so that it's easy to compare two values
• it's a complex class, because it composes four integers, so it's a good complement to String, Int32, Object, Guid, and other primitive types

Recently, I've been picking up FsCheck to do Property-Based Testing, but out of the box it doesn't know how to create arbitrary Version instances.

It turns out that you can easily and elegantly tell FsCheck how to create arbitrary Version instances, but since I haven't seen it documented, I thought I'd share my solution:

type Generators =
    static member Version() =
        Arb.generate<byte>
        |> Gen.map int
        |> Gen.four
        |> Gen.map (fun (ma, mi, bu, re) -> Version(ma, mi, bu, re))
        |> Arb.fromGen

As the FsCheck documentation explains, you can create custom Generator by defining a static class that exposes members that return Arbitrary<'a> - in this case Arbitrary<Version>.

If you'd like me to walk you through what happens here, read on, and I'll break it down for you.

First, Arb.generate<byte> is a Generator of Byte values. While FsCheck doesn't know how to create arbitrary Version values, it does know how to create arbitrary values of various primitive types, such as Byte, Int32, String, and so on. The Version constructors expect components as Int32 values, so why did I select Byte values instead? Because Version doesn't accept negative numbers, and if I had kicked off my Generator with Arb.generate<int>, it would have created all sorts of integers, including negative values. While it's possible to filter or modify the Generator, I thought it was easier to simply kick off the Generator with Byte values, because they are never negative.

Second, Gen.map int converts the initial Gen<byte> to Gen<int> by invoking F#'s built-in int conversion function.

Third, Gen.four is a built-in FsCheck Generator Combinator that converts a Generator into a Generator of four-element tuples; in this case it converts Get<int> to Gen<int * int * int * int>: a Generator of a four-integer tuple.

Fourth, Gen.map (fun (ma, mi, bu, re) -> Version(ma, mi, bu, re)) converts Gen<int * int * int * int> to Gen<Version> by another application of Gen.map. The function supplied to Gen.map takes the four-element tuple of integers and invokes the Version constructor with the major, minor, build, and revision integer values.

Finally, Arb.fromGen converts Gen<Version> to Arbitrary<Version>, which is what the member must return.

To register the Generators custom class with FsCheck, I'm currently doing this:

do Arb.register<Generators>() |> ignore

You can see this entire code in context here.

If you take the SOLID principles to their extremes, you arrive at something that makes Functional Programming look quite attractive.

You may have seen this one before, but bear with me :)

The venerable master Qc Na was walking with his student, Anton. Hoping to prompt the master into a discussion, Anton said "Master, I have heard that objects are a very good thing - is this true?" Qc Na looked pityingly at his student and replied, "Foolish pupil - objects are merely a poor man's closures."

Chastised, Anton took his leave from his master and returned to his cell, intent on studying closures. He carefully read the entire "Lambda: The Ultimate..." series of papers and its cousins, and implemented a small Scheme interpreter with a closure-based object system. He learned much, and looked forward to informing his master of his progress.

On his next walk with Qc Na, Anton attempted to impress his master by saying "Master, I have diligently studied the matter, and now understand that objects are truly a poor man's closures." Qc Na responded by hitting Anton with his stick, saying "When will you learn? Closures are a poor man's object." At that moment, Anton became enlightened.

- Anton van Straaten

While this is a lovely parable, it's not a new observation that objects and closures seem closely related, and there has been much discussion back and forth about this already. Still, in light of a recent question and answer about how to move from Object-Oriented Composition to Functional Composition, I'd still like to explain how, in my experience, the SOLID principles lead to a style of design that makes Functional Programming quite attractive.

A SOLID road map #

In a previous article, I've described how application of the Single Responsibility Principle (SRP) leads to many small classes. Furthermore, if you rigorously apply the Interface Segregation Principle (ISP), you'll understand that you should favour Role Interfaces over Header Interfaces.

If you keep driving your design towards smaller and smaller interfaces, you'll eventually arrive at the ultimate Role Interface: an interface with a single method. This happens to me a lot. Here's an example:

public interface IMessageQuery
{
    string Read(int id);
}

If you apply the SRP and ISP like that, you're likely to evolve a code base with many fine-grained classes that each have a single method. That has happened to me more than once; AutoFixture is an example of a big and complex code base that looks like that, but my other publicly available code bases tend to have the same tendency. In general, this works well; the most consistent problem is that it tends to be a bit verbose.

Objects as data with behaviour #

One way to characterise objects is that they are data with behaviour; that's a good description. In practice, when you have many fine-grained classes with a single method, you may have classes like this:

public class FileStore : IMessageQuery
{
    private readonly DirectoryInfo workingDirectory;

    public FileStore(DirectoryInfo workingDirectory)
    {
        this.workingDirectory = workingDirectory;
    }

    public string Read(int id)
    {
        var path = Path.Combine(
            this.workingDirectory.FullName,
            id + ".txt");            
        return File.ReadAllText(path);
    }
}

This FileStore class is a simple example of data with behaviour.

• The behaviour is the Read method, which figures out a file path for a given ID and returns the contents of the file.
• The data (also sometimes known as the state) is the workingDirectory field.

The workingDirectory field is a Concrete Dependency, but it could also have been a primitive value or an interface or abstract base class. In the last case, we would often call the pattern Constructor Injection.

Obviously, the data could be multiple values, instead of a single value.

The FileStore example class implements the IMessageQuery interface, so it's a very representative example of what happens when you take the SRP and ISP to their logical conclusions. It's a fine class, although a little verbose.

When designing like this, not only do you have to come up with a name for the interface itself, but also for the method, and for each concrete class you create to implement the interface. Naming is difficult, and in such cases, you have to name the same concept twice or more. This often leads to interfaces named IFooer with a method called Foo, IBarer with a method called Bar, etc. You get the picture. This is a smell (that also seems vaguely reminiscent of the Reused Abstractions Principle). There must be a better way.

Hold that thought.

Functions as pure behaviour #

As the introductory parable suggests, perhaps Functional Programming offers an alternative. Before you learn about Closures, though, you'll need to understand Functions. In Functional Programming, a Function is often defined as a Pure Function - that is: a deterministic operation without side-effects.

Since C# has some Functional language support, I'll first show you the FileStore.Read method as a Pure Function in C#:

Func<DirectoryInfo, int, string> read = (workingDirectory, id) =>
    {
        var path = Path.Combine(workingDirectory.FullName, id + ".txt");
        return File.ReadAllText(path);
    };

This Function does the same as the FileStore.Read method, but it has no data. You must pass in the working directory as a function argument just like the ID. This doesn't seem equivalent to an object.

Closures as behaviour with data #

A Closure is an important concept in Functional Programming. In C# it looks like this:

var workingDirectory = new DirectoryInfo(Environment.CurrentDirectory);
Func<int, string> read = id =>
    {
        var path = Path.Combine(workingDirectory.FullName, id + ".txt");
        return File.ReadAllText(path);
    };

This is called a Closure because the Function closes over the Outer Variable workingDirectory. Effectively, the function captures the value of the Outer Variable.

What does that compile to?

Obviously, the above C# code compiles to IL, but if you reverse-engineer the IL back to C#, this is what it looks like:

[CompilerGenerated]
private sealed class <>c__DisplayClass3
{
    public DirectoryInfo workingDirectory;

    public string <UseClosure>b__2(int id)
    {
        return File.ReadAllText(
            Path.Combine(this.workingDirectory.FullName, id + ".txt"));
    }
}

It's a class with a field and a method! Granted, the names look somewhat strange, and the field is a public, mutable field, but it's essentially identical to the FileStore class!

Closures are behaviour with data, whereas objects are data with behaviour. Hopefully, the opening parable makes sense to you now. This is an example of one of Erik Meijer's favourite design concepts called duality.

Partial Function Application #

Another way to close over data is called Partial Function Application, but the result is more or less the same. Given the original pure function:

Func<DirectoryInfo, int, string> read = (workingDirectory, id) =>
    {
        var path = Path.Combine(workingDirectory.FullName, id + ".txt");
        return File.ReadAllText(path);
    };

you can create a new Function from the first Function by only invoking it with some of the arguments:

var wd = new DirectoryInfo(Environment.CurrentDirectory);
Func<int, string> r = id => read(wd, id);

The r function also closes over the wd variable, and the compiled IL is very similar to before.

Just use F#, then! #

If SOLID leads you to many fine-grained classes with a single method, C# starts to be in the way. A class like the above FileStore class is proper Object-Oriented Code, but is quite verbose; the Closures and Partially Applied Functions compile, but are hardly idiomatic C# code.

On the other hand, in F#, the above Closure is simply written as:

let workingDirectory = DirectoryInfo(Environment.CurrentDirectory)
let read id = 
    let path = Path.Combine(workingDirectory.FullName, id.ToString() + ".txt")
    File.ReadAllText(path)

The read value is a Function with the signature 'a -> string, which means that it takes a value of the generic type 'a (in C#, it would typically have been named T) and returns a string. This is just a more general version of the IMessageQuery.Read method. When 'a is int, it's the same signature, but in F#, I only had to bother naming the Function itself. Functions are anonymous interfaces, so these are also equivalent.

Likewise, if you have a Pure Function like this:

let read (workingDirectory : DirectoryInfo) id =
    let path = Path.Combine(workingDirectory.FullName, id.ToString() + ".txt")
    File.ReadAllText(path)

the Partially Applied Function is written like this:

let wd = DirectoryInfo(Environment.CurrentDirectory)
let r = read wd

The r Function is another Function that takes an ID as input, and returns a string, but notice how much less ceremony is involved.

Summary #

SOLID, particularly the SRP and ISP, leads you towards code bases with many fine-grained classes with a single method. Such objects represent data with behaviour, but can also be modelled as behaviour with data: Closures. When that happens repeatedly, it's time to make the switch to a Functional Programming Language like F#.

Comments

public class MyService
{
	...

	public MyService(IMessageQuery messageQuery)
	{...}
}

type MyService (messageQuery: int -> string) = ...

let myClient f =
    let message = f 42
    // Do something else interesting...
    // Return a result...

NuGet is a great tool if used correctly. Here's one way to do it.

In my recent post about NuGet, I described why the Package Restore feature is insidious. As expected, this provoked some readers, who didn't like my recommendation of adding NuGet packages to source control. That's understandable; the problem with a rant like my previous post is that while it tells you what not to do, it's not particularly constructive. While I told you to store NuGet packages in your source control system, I didn't describe patterns for doing it effectively. My impression was that it's trivial to do this, but based on the reactions I got, I realize that this may not be the case. Could it be that some readers react strongly because they don't know what else to do (than to use NuGet Package Restore)? In this post, I'll describe a way to use and organize NuGet packages that have worked well for me in several organizations.

Publish/Subscribe #

In Grean we use NuGet in a sort of Publish/Subscribe style. This is a style I've also used in other organizations, to great effect. It's easy: create reusable components as autonomous libraries, and publish them as NuGet packages. If you don't feel like sharing your internal building blocks with the rest of the world, you can use a custom, internal package repository, or you can use MyGet (that's what we do in Grean).

A reusable component may be some package you've created for internal use. Something that packages the way you authenticate, log, instrument, render, etc. in your organization.

Every time you have a new version of one of your components (let's call it C1), you publish the NuGet package.

Just like other Publish/Subscribe systems, the only other party that you rely on at this moment is the queue/bus/broker - in this case the package repository, like NuGet.org or MyGet.org. No other systems need to be available to do this.

You do this for every reusable component you want to publish. Each is independent of other components.

Pull based on need #

In addition to reusable components, you probably also build systems; that is, applications that actually do something. You probably build those systems on top of reusable components - yours, and other publicly available NuGet packages. Let's call one such system S1.

Whenever you need a NuGet package (C1), you add it to the Visual Studio project where you need it, and then you commit your changes to that system's source control. It effectively means checking in the NuGet package, including all the binaries, to source control. However, the S1 repository is not the same repository as the C1 repository. Both are autonomous systems.

The only system you need to be available when you add the NuGet package C1 is the NuGet package source (NuGet.org, MyGet.org, etc.). The only system you need to be available to commit the changes to S1 is your source control system, and if you use a Distributed Version Control System (DVCS), it's always going to be available.

Pretty trivial so far.

"This isn't pub/sub," you'll most likely say. That's right, not in the traditional sense. Still, if you adopt the pattern language of Enterprise Integration Patterns, you can think of yourself (and your colleagues) as a Polling Consumer.

"But," I suppose you'll say, "I'm not polling the repository and pulling down every package ever published."

True, but you could, and if you did, you'd most likely be filtering away most package updates, because they don't apply to your system. That corresponds to applying a Message Filter.

This last part is important, so let me rephrase it:

Just because your system uses a particular NuGet package, it doesn't mean that you have to install every single version ever published.

It seems to me that at least some of the resistance to adding packages to your repository is based on something like that. As Urs Enzler writes:

[Putting packages in source control is] "not an option if your repo grows > 100GB per month due to monthly updates of BIG nuget packages"

As an example, consider my own open source project AutoFixture, which keeps a fairly high release cadence. It's released according to the principles of Continuous Delivery, so every time there's a new feature or fix, we release a new NuGet package. In 2013, we released 47 versions of the AutoFixture NuGet package, including one major release. That's almost a release every week, but while I use AutoFixture in many other projects, I don't try to keep up with it. I just install AutoFixture when I start a new project, and then I mostly update the package if I need one of the new features or bug fixes. Occasionally, I also update packages in order to not fall too much behind.

As a publicly visible case, consider Hyprlinkr, which uses AutoFixture as one of its dependencies. While going though Hyprlinkr's NuGet packages recently, I discovered that the Hyprlinkr code base was using AutoFixture 2.12.0 - an 18 months old version! I simply hadn't needed to update the package during that time. AutoFixture follows Semantic Versioning, and we go to great lengths to ensure that we don't break existing functionality (unless we do a major release).

Use the NuGet packages you need, commit them to source control, and update them as necessary. For all well-designed packages, you should be able to skip versions without ill effects. This enables you to treat the code bases for each system (S1, S2, etc.) as autonomous systems. Everything you need in order to work with that code base is right there in the source code repository.

Stable Dependency Principle #

What if you need to keep up-to-date with a package that rapidly evolves? From Urs Enzler's tweet, I get the impression that this is the case not only for Urs, but for other people too. Imagine that the creator of such a package frequently publishes new versions, and that you have to keep up to date. If that's the case, it must imply that the package isn't stable, because otherwise, you'd be able to skip updates.

Let me repeat that:

If you depend on a NuGet package, and you have to stay up-to-date, it implies that the package is unstable.

If this is the case, you have an entirely other problem on your hand. It has nothing to do with NuGet Package Restore, or whether you're keeping packages in source control or not. It means that you're violating the Stable Dependencies Principle (SDP). If you feel pain in that situation, that's expected, but the solution isn't Package Restore, but a better dependency hierarchy.

If you can invert the dependency, you can solve the problem. If you can't invert the dependency, you'd probably benefit from an Anti-corruption Layer. There are plenty of better solution that address the root cause of your problems. NuGet Package Restore, on the other hand, is only symptomatic relief.

Comments

If you only have F# 3.1 installed on a machine, but need to use a compiled application that requires F# 3.0, here's what you can do.

This post uses a particular application, Zero29, as an example in order to explain a problem and one possible solution. However, the post isn't about Zero29, but rather about a particular F# DLL hell.

Currently, I'm repaving one of my machines, which is always a good idea to do regularly, because it's a great remedy against works on my machine syndrome. This machine doesn't yet have a lot of software, but it does have Visual Studio 2013 and F# 3.1.

Working with a code base, I wanted to use Zero29 to incement the version number of the code, so first I executed:

$ packages/Zero29.0.4.1/tools/Zero29 -l

which promptly produced this error message:

Unhandled Exception: System.IO.FileNotFoundException:
Could not load file or assembly
'FSharp.Core, Version=4.3.0.0, Culture=neutral, PublicKeyToken=b03f5f7f11d50a3a'
or one of its dependencies. The system cannot find the file specified.
   at Ploeh.ZeroToNine.Program.main(String[] argv)

On one level, this makes sense, because Zero29 0.4.1 was compiled against F# 3.0 (which corresponds to FSharp.Core 4.3.0.0).

On another level, this is surprising, since I do have F# 3.1 (FSharp.Core 4.3.1.0) on my machine. Until the error message appeared, I had lived with the naïve assumption that when you install F# 3.1, it would automatically add redirects from FSharp.Core 4.3.0.0 to 4.3.1.0, or perhaps make sure that FSharp.Core 4.3.0.0 was also available. Apparently, I've become too used to Semantic Versioning, which is definitely not the versioning scheme used for F#.

Here's one way to address the issue.

Although Zero29 is my own (and contributors') creation, I didn't want to recompile it just to deal with this issue; it should also be usable for people with F# 3.0 on their machines.

Even though it's a compiled program, you can still add an application configuration file to it, so I created an XML file called Zero29.exe.config, placed it alongside Zero29.exe, and added this content:

<?xml version="1.0" encoding="utf-8"?>
<configuration>
  <runtime>
    <assemblyBinding xmlns="urn:schemas-microsoft-com:asm.v1">
      <dependentAssembly>
        <assemblyIdentity name="FSharp.Core"
                          publicKeyToken="b03f5f7f11d50a3a"
                          culture="neutral"/>
        <bindingRedirect oldVersion="4.3.0.0"
                         newVersion="4.3.1.0"/>
      </dependentAssembly>
    </assemblyBinding>
  </runtime>
</configuration>

This solved the problem, although I now have the derived problem that this new file isn't part of the Zero29 NuGet package, and I don't know if it's going to ruin my colleagues' ability to use Zero29 if I check it into source control...

Another option may be to add the redirect to machine.config, instead of an application-specific redirect, but I have no desire to manipulate my machine.config files if I can avoid it, so I didn't try that.

The NuGet Package Restore feature is a really bad idea; this post explains why.

One of the first things I do with a new installation of Visual Studio is to disable the NuGet Package Restore feature. There are many reasons for that, but it all boils down to this:

NuGet Package Restore introduces more problems than it solves.

Before I tell you about all those problems, I'll share the solution with you: check your NuGet packages into source control. Yes, it's that simple.

Storage implications #

If you're like most other people, you don't like that solution, because it feels inefficient. And so what? Let's look at some numbers.

• The AutoFixture repository is 28.6 MB, and that's a pretty big code base (181,479 lines of code).
• The Hyprlinkr repository is 32.2 MB.
• The Albedo repository is 8.85 MB.
• The ZeroToNine repository is 4.91 MB.
• The sample code repository for my new Pluralsight course is 69.9 MB.
• The repository for Grean's largest production application is 32.5 MB.
• Last year I helped one of my clients build a big, scalable REST API. We had several repositories, of which the largest one takes up 95.3 MB on my disk.

On my laptops I'm using Lenovo-supported SSDs, so they're fairly expensive drives. Looking up current prices, it seems that a rough estimates of prices puts those disks at approximately 1 USD per GB.

On average, each of my repositories containing NuGet packages cost me four cents of disk drive space.

Perhaps I could have saved some of this money with Package Restore...

Clone time #

Another problem that the Package Restore feature seems to address, is the long time it takes to clone a repository - if you're on a shaky internet connection in a train. While it can be annoying to wait for a repository to clone, how often do you do that, compared to normal synchronization operations such as pull, push or fetch?

What should you be optimizing for? Cloning, which you do once in a while? Or fetch, pull, and push, which you do several times a day?

In most cases, the amount of time it takes to clone a repository is irrelevant.

To summarize so far: the problems that Package Restore solves are a couple of cents of disk cost, as well as making a rarely performed operation faster. From where I stand, it doesn't take a lot of problems before they outweigh the benefits - and there are plenty of problems with this feature.

Fragility #

The more moving parts you add to a system, the greater the risk of failure. If you use a Distributed Version Control System (DVCS) and keep all NuGet packages in the repository, you can work when you're off-line. With Package Restore, you've added a dependency on at least one package source.

• What happens if you have no network connection?
• What happens if your package source (e.g. NuGet.org) is down?
• What happens if you use multiple package sources (e.g. both NuGet.org and MyGet.org)?

This is a well-known trait of any distributed system: The system is only as strong as its weakest link. The more services you add, the higher is the risk that something breaks.

Custom package sources #

NuGet itself is a nice system, and I often encourage organizations to adopt it for internal use. You may have reusable components that you want to share within your organization, but not with the whole world. In Grean, we have such components, and we use MyGet to host the packages. This is great, but if you use Package Restore, now you depend on multiple services (NuGet.org and MyGet.org) to be available at the same time.

While Myget is a nice and well-behaved NuGet host, I've also worked with internal NuGet package sources, set up as an internal service in an organization. Some of these are not as well-behaved. In one case, 'old' packages were deleted from the package source, which had the consequence that when I later wanted to use an older version of the source code, I couldn't complete a Package Restore because the package with the desired version number was no longer available. There was simply no way to build that version of the code base!

Portability #

One of the many nice things about a DVCS is that you can xcopy your repository and move it to another machine. You can also copy it and give it to someone else. You could, for example, zip it and hand it over to an external consultant. If you use Package Restore and internal package sources, the consultant will not be able to compile the code you gave him or her.

Setup #

Perhaps you don't use external consultants, but maybe you set up a new developer machine once in a while. Perhaps you occasionally get a new colleague, who needs help with setting up the development environment. Particularly if you use custom package feeds, making it all work is yet another custom configuration step you need to remember.

Bandwidth cost #

As far as I've been able to tell, the purpose of Package Restore is efficiency. However, every time you compile with Package Restore enabled, you're using the network.

Consider a Build Server. Every time it makes a build, it should start with a clean slate. It can get the latest deltas from the shared source control repository, but it should start with a clean working folder. This means that every time it builds, it'll need to download all the NuGet packages via Package Restore. This not only wastes bandwidth, but takes time. In contrast, if you keep NuGet packages in the repository itself, the Build Server has everything it needs as soon as it has the latest version of the repository.

The same goes for your own development machine. Package Restore will make your compile process slower.

Glitches #

Finally, Package Restore simply doesn't work very well. Personally, I've wasted many hours troubleshooting problems that turned out to be related to Package Restore. Allow me to share one of these stories.

Recently, I encountered this sight when I opened a solution in Visual Studio:

My problem was that at first, I didn't understand what was wrong. Even though I store NuGet packages in my repositories, all of a sudden I got this error message. It turned out that this happened at the time when NuGet switched to enabling Package Restore by default, and I hadn't gotten around to disable it again.

The strange thing was the everything compiled and worked just great, so why was I getting that error message?

After much digging around, it turned out that the ImpromptuInterface.FSharp package was missing a .nuspec file. You may notice that ImpromptuInterface.FSharp is also missing in the package list above. All binaries, as well as the .nupkg file, was in the repository, but the ImpromptuInterface.FSharp.1.2.13.nuspec was missing. I hadn't noticed for weeks, because I didn't need it, but NuGet complained.

After I added the appropriate .nuspec file, the error message went away.

The resolution to this problem turned out to be easy, and benign, but I wasted an hour or two troubleshooting. It didn't make me feel productive at all.

This story is just one among many run-ins I've had with NuGet Package Restore, before I decided to ditch it.

Just say no #

The Package Restore feature solves these problems:

• It saves a nickel per repository in storage costs.
• It saves time when you clone a new repository, which you shouldn't be doing that often.

• adds complexity
• makes it harder to use custom package sources
• couples your ability to compile to having a network connection
• makes it more difficult to copy a code base
• makes it more difficult to set up your development environment
• uses more bandwidth
• leads to slower build times
• just overall wastes your time

For me, the verdict is clear. The benefits of Package Restore don't warrant the disadvantages. Personally, I always disable the feature and instead check in all packages in my repositories. This never gives me any problems.

Comments

My new Pluralsight course, A Functional Architecture with F#, is now available.

Whenever I've talked to object-oriented developers about F#, a common reaction has been that it looks enticing, but that they don't see how they'd be able to build a 'normal' application with it. F# has gained a reputation for being a 'niche' language, good for scientific computation and financial calculations, but not useful for mainstream applications.

Not only is F# a Turing-complete, general purpose programming language, but it has many advantages to offer compared to, say, C#. That said, though, building a 'normal' application with F# will only make sense if you know how to work with the language, and define an architecture that takes advantage of all it has to offer. Therefore, I thought that it would be valuable to show one possible way to do this, through a comprehensive example.

This was the motivation behind my new Pluralsight course A Functional Architecture with F#, which is now available! In it, you'll see extensive code demos of a web application written entirely in F# (and a bit of GUI in JavaScript, but the course only shows the F# code).

If you don't already have a Pluralsight account, you can get a free trial of up to 200 minutes.

Comments

A fully RESTful API often looks inefficient from a client perspective, until you learn to change that perspective.

One of my readers, Filipe Ximenes, asks the following question of me:

"I read you post about avoiding hackable urls and found it very interesting. I'm currently studying about REST and I'm really interested on building true RESTful API's. One thing that is bothering me is how to access resources that are not in the API root. Eg: consider the following API flow:

"root > users > user details > user messages

"Now consider that one client wants to retrieve all the messages from a user. Does it need to "walk" the whole API (from it's root to "user messages")? This does not seem very efficient to me. Am I missing something? What would be a better solution for this?"

This is a common question which isn't particularly tied to avoiding hackable URLs, but simply to the hypermedia nature of a level 3 RESTful API.

The short answer is that it's probably not particularly inefficient. There are several reasons for that.

HTTP caching #

One of the great advantages of RESTful design is that instead of abstracting HTTP away, it very explicitly leverages the protocol. HTTP has bulit-in caching, so even if an API forces a client to walk the API as in the question above, it could conceivably result in only a single HTTP request:

This cache could be anywhere between the client and the service. It could be a proxy server, a reverse proxy, or it could even be a local cache on the client machine; think of a Browser's local cache. It could be a combination of all of those caches. Conceivably, if a local cache is involved, a client could walk the API as described above with only a single (or even no) network request involved, because most of the potential requests would be cache hits.

This is one of the many beautiful aspects of REST. By leveraging the HTTP protocol, you can use the internet as your caching infrastructure. Even if you want a greater degree of control, you can use off-the-shelf software for your caching purposes.

Cool URLs #

As the RESTful Web Services Cookbook describes, URLs should be cool. This means that once you've given a URL to a client, you should honour requests for that URL in the future. This means that clients can 'bookmark' URLs if they like. That includes the final URL in the flow above.

Short-cut links #

Finally, an API can provide short-cut links to a client. Imagine, for example, that when you ask for a list of users, you get this:

<users xmlns:atom="http://www.w3.org/2005/Atom">
  <user>
    <links>
      <atom:link rel="user-details" href="/users/1234" />
      <atom:link rel="user-messages" href="/users/1234/messages" />
    </links>
    <name>Foo</name>
  </user>
  <user>
    <links>
      <atom:link rel="user-details" href="/users/5678" />
      <atom:link rel="user-messages" href="/users/5678/messages" />
    </links>
    <name>Bar</name>
  </user>
  <user>
    <links>
      <atom:link rel="user-details" href="/users/9876" />
      <atom:link rel="user-messages" href="/users/9876/messages" />
    </links>
    <name>Baz</name>
  </user>
</users>

As you can see in this example, a list of users can provide a short-cut to a user's messages, enabling a client to follow a more direct path:

root > users > user messages

The client would have to prioritize links of the relationship type user-messages over links of the user-details type.

Summary #

Efficiency is a common concern about HATEOAS systems, particularly because a client should always start at published URL. Often, the only published URL is the root URL, which forces the client to walk the rest of the API. This seems inefficient, but doesn't have to be because of all the other built-in mechanisms that work to effectively counter what at first looks like an inefficiency.