Some Thoughts On Library Design

This post was originally published in GopherAcademy for their Advent 2019 series. It's been republished here for posterity.

As programmers we use libraries a lot. But library design is hard. In this article, I will walk through some considerations in designing a library.

We will start by bifurcating the acts of programming. We start by casting the act of programming as conversations. Then we examine the main activities that constitute what people call “programming”. All these serves as a background to developing better software libraries.

First, consider the act of programming. The main purpose of programming is to produce programs that do useful things. Everything that follows is simply bureaucracy. Modern programming can be split into two main activities: writing the application and writing the support libraries that the application uses.

The act of programming can be also be thought of as the programmer having two simultaneous conversations - one conversation is to the compiler to create a customised runtime environment and one conversation is to tell the runtime system what to do.

In the same vein as casting the act of programming as conversations, we can cast the act of writing libraries as two conversations - one with the computer and one with another human (usually oneself in one’s own future). We will consider these two kinds of conversation when we talk about the properties of a library.

The following graphic indicates the views of programming that I have espoused here

G Programming1 Programming Library Writing Library Writing Programming1->Library Writing Application Writing Application Writing Programming1->Application Writing Conversation with Machine Conversation with Machine Library Writing->Conversation with Machine Conversation with Human Conversation with Human Library Writing->Conversation with Human phantom ... Application Writing->phantom

Some Definitions and a Raison d’Être

But first, let’s go back to basics and address “why libraries”? Why do we write software libraries? What benefits do we get from software libraries?

First, note that I am using the term “libraries” instead of “packages”, “modules” or “repository”. Despite being used interchangeably in my mind there are very subtle differences. Allow me to explain.

A Repository

A repository is a collection of files containing source code. They are typically arranged within a directory in the file system.

A Library

A library a collection of resources - usually source code - that is shared. Source code sharing can come in many forms. The most common way of doing this in Go is through packages, which the language supports by specification.

There are other ways of sharing code as well. What follows is a contrived example to illustrate my point.

Let’s say I have a file (let’s call it lib.go) in a directory called common.

func MaxInt(a, b int) int {
     if a > b {
     	return a
     }
     return b
}

I start a new Go package (called foo), and I place it in a directory called github.com/myusername/bar. I copy lib.go from common to bar, and rename the file in bar as lib_bar.go. Now I edit lib_bar.go and prepend the declaration package foo at the top so that the complete lib_bar.go is as follows:

package foo

func MaxInt(a, b int) int {
     if a > b {
     	return a
     }
     return b
}

Let’s say I now start another new Go package (called ‘baz’), and I place it in a directory called bitbucket.org/myusername/quux. I copy lib.go from common to quux, and rename the file in quux as lib_baz.go. I prepended the declaration package baz at the top so the complete lib_baz.go reads as follows:

package baz

func MaxInt(a, b int) int {
     if a > b {
     	return a
     }
     return b
}

Now, to take stock:

  • in common I have a malformed .go file called lib.go
  • in github.com/myusername/bar I have a file called lib_foo.go. The repository holds the package foo
  • in github.com/myusername/quux I have a file called lib_baz.go. The repository holds the package baz.

Observe that I have shared the source code from common/lib.go into two different packages, foo and baz. Yes, the source code is shared by copying, instead of having a single source of truth, but the source code, at this point in time, is shared.

It is in this sense that I use the word “library” - a library is source code that is shared.

A Package

In general, libraries of source code in Go are arranged in packages and modules. A package is a collection of .go files. Usually a package does one thing. A package may depend on another package.

The astute reader will note that having lib.go in the above example will cause any Go project to have a compilation failure. All .go files must declare at the very top, what package it is used for. The declaration package foo is a conversation to the compiler, telling the compiler to include the file in a package.

A Module

If a package is a collection of files containing source code, a module is a collection of packages. Go modules were designed to solve package dependency issues. Modules in Go are defined by a go.mod file, which lists all the packages the module depends on.

Why Libraries

Having introduced all the terms, we can now go back to answer the question: why libraries?

We can write an application (with package main as the magical declaration at the top), and put all our data structures and code within the main package.

But good programming practices tell us not to do that. Decomposition and isolation are useful concepts to have.

Recall from earlier that I had mentioned that writing libraries is like having two conversations - decomposition (and its opposite, composition) is to facilitate the conversation with the machine. Isolation is to facilitate conversations to another human being. We will explore compositionality later in the article.

The general principle behind code isolation is that components of one library does not need to know about components of another library. Thus a programmer who works on one library only needs to know about the thing the library is supposed to do. Thus the programmer can dedicate her entire mental model to the library in question.

What Can Go Into a Library?

I used a very liberal definition of “library” - that a library is a collection of resources that are shared. These resources are usually source code. Although this is not necessarily always true. I will provide two examples of external resources being put into libraries.

The first example concerns using CUDA for processing.

If you have an nVidia graphics card and you would like to use the graphics card for GPGPU purposes, using CUDA. The GPU is a resource you need to access. The access can acquired using the CUDA drivers. In Go, the cu library that is a part of the Gorgonia suite of libraries manages the driver and accesses the device.

You can get access by means of the following code

d := cu.CurrentDevice()
ctx := cu.NewContext(d,	cu.SchedAuto|cu.MapHost)
mem, err := ctx.MemAlloc(1024)

ctx is a handle to the GPU. Once you have that handle, you can send work to the GPU to do (like reserving 1MB of graphics card memory in the example). But ultimately ctx is a resource.

The second example concerns using files as a resource in a library.

Say I have a list of Shakespeare’s plays in ASCII format and I want to put them in a library. I can store the plays as .txt files, and put them in some central location to be accessed.

Or to maximise compatibility with the Go programming language, I can create a new package and have the content be something like this:

package willshakes

const AllsWellThatEndsWell = `Act 1 Scene 1

COUNTESS.
	In delivering my son from me, I bury a second husband.

BERTRAM.
	And I in going, madam, weep o'er my father's death
	anew: but I must attend his majesty's command, to
	whom I am now in ward, evermore in subjection.
...
`

const MacBeth = `Act 1 Scene 1

FIRST WITCH.
	When shall we three meet again
	In thunder, lightning, or in rain?

SECOND WITCH.
	When the hurlyburly's done,
	When the battle's lost and won.

THIRD WITCH.
	That will be ere the set of sun.
...
`

Put thus, we can just import willshakes and to access the text of MacBeth we simply use willshakes.MacBeth.

The difference between the CUDA example and the Shakespeare example is that the CUDA example is an example where the resource is dynamic, while in the Shakespeare example, the resource is static.

The use of the terms “static” and “dynamic” is not good, but are in standard use. To make it clearer, allow me to further explain:

A resource is static if its state is known at compile time. A resource is dynamic if its state unknown at compile time.

Thus the Shakespeare resource is static because the entire corpus is known and available at compile time. The state of a graphics card availability may change and needs to be determined at runtime, therefore it is a dynamic resource.

Another good example (though one that I have no personal experience with) is including fonts as part of the static assets in your program.

At the time of writing, there is a proposal to allow static resources to be embedded in the final binary of a Go program, so the story is still to be told on the Go end.

The Types of Libraries

Barring discussions on the concrete details of libraries and what form they take (packages, drivers, modules, etc), let’s consider the various types of libraries we have discussed so far.

Broadly speaking there are two general classes of libraries:

  • Libraries where source code is the primary resource that is being shared.
  • Libraries where some resource other than source code is being shared.

The latter of these can be split into two:

  • Driver libraries
  • Resource libraries

A driver library is typically a package that wrap access to a driver. For example, the aforementioned cu package wraps the CUDA drivers to enable CUDA programming in Go. Similarly the go-gl package which mediates access to OpenGL). Both packages expose extra helper functions to make the transition between worlds easier, but are fundamentally driver packages.

The willshakes library is an example of a resource library. For a real life use case, I offer the mnist library for consideration. Due to the nature of the source data, as well as the inability for Go to handle static resources, the design of the package is limited to source code that loads data from an external file into a data structure. In the Python world, this is a different case as one may use keras.datasets.mnist immediately as a resource.

What Makes A Good Library?

Now that I have defined the things that can go into a library, let’s take a look at what makes a good library. We will start with the simpler, more obvious statements, before moving on to more nuanced considerations. Despite this, the simpler statements often come with caveats, which will be briefly explored.

I have a few principles that form the basic principles of what makes a good library:

  1. Reliable
  2. Easy to use/build
  3. Reusable

Reliable

First and foremost, a library must be reliable. What is the point of using a library if it cannot reliably do what it claims to do? There are many features of a reliable library, enumerated below.

Has a Clear Value Proposition

A good library has a clear value proposition. Usually this means it does one thing or provides one resource. What constitute “one thing” is usually the point of contention.

For example, consider the grpc library. It does one thing - gRPC. But gRPC has many sub-components to it - server and client are the two primary sub-components.

An example on the other extreme can be seen in the packages that pervade npm. left-pad was a package that provided one function that padded a string. It did one thing, and many packages depended upon it. Thus when the left-pad package was unpublished, it broke the internet.

Is Well-Tested

A good library is well-tested. Users who use the library must be able to feel confident about the library they’re using.

There are multiple levels of testing - unit tests, integration testing, property-based testing and fuzz testing - to name a few. Each have their own pros and cons.

I would go so far to prefer only libraries that have been tested using property-based testing (I had previously written an article about property-based testing on GopherAcademy) or have been fuzz-tested.

A good thing to check on a well-tested library is whether the tests test for general cases or only specific cases. This is why I prefer libraries that are fuzz-tested and have PBTs in them. Fuzz-testing checks that the library functions can handle unforeseen input, while property-based testing requires a deep understanding of the domain space.

Having said that, if you develop driver libraries, it might be a bit difficult to test such libraries. There are some advances in source fuzzing for driver libraries, but I have found no good general-purpose testing patterns in the case of driver libraries that works well for my workflow.

Doesn’t Manage Resources For Users

A good library does not manage resources for its user. Instead, it provides resource management utilities to the user.

For example: If you’re writing a library that uses an OpenGL context to do something with OpenGL, don’t create the OpenGL context in the library. Instead, require the user to pass in a OpenGL context.

This is also true for allocations. Where possible, don’t create allocations on behalf of the user.

Dave Cheney recently wrote a most excellent article on the topic of forcing allocations. The title’s a bit confusing but the main point is similar to what I am espousing here.

When a library doesn’t manage resources for the user, it becomes clear that the user has to manage resources by themselves. The brunt of the responsibility falls onto the user, but the library becomes more reliable.

Last but not least, don’t spawn goroutines on behalf of the user.

Easy to Use

A good library is easy to use. There are a number of ways that a library can be easy to use.

Good Documentation and Examples

A good library has good documentation. And to readers who think “tests are documentation”, yes! Go has good support for examples, which are both documentation and tests. I enjoy using libraries that have examples when I go to their godoc.

Doesn’t Panic

Panics should only happen in a case when there are no better options. Usually returning errors are a better thing to do.

Panicking wrests the control flow from the users. It is better to give users an opportunity to handle the control flow themselves.

Has Minimal Dependencies

This is fairly contentious especially in the Big Picture view of this article (see the Tension section below). But in my opinion a good library has minimal dependencies.

This is especially true of libraries where source code are the primary resource being shared. If a library whose purpose is to share source code were to depend on some resource library, I would be quite suspicious.

Additional dependencies also increase the difficulty to use. I often check what each library imports in order to know that my imports are not going to suddenly call home to some server somewhere. I am not fastidious over it, only because there is so much to check.

Makes the Zero Value Useful

One of the Go proverbs, the zero value of any data type should be useful. This avoids the need for complicated constructor functions. A very good example I enjoy is Gonum’s mat.Dense type.

The mat.Dense data type has a method Mul which performs matrix multiplication. It has the following type signature:

func (m *Dense) Mul(a, b Matrix)

The result of a × b is placed in m. Thus, if a is a (2,3) matrix and b is a (3,2) matrix, then m will be a (2,2) matrix. The documentation is not clear, so most people will try something like this:

c := mat.NewDense(2, 2, make([]float64, 4))
c.Mul(a, b)

In actuality, this would work as well:

var c mat.Dense
c.Mul(a, b)

Reusable

A good library is also reusable under a number of different scenarios. Often arguments for generics in Go cites re-usability as a main reason. However, as is now, Go offers a lot of opportunity for re-usability.

Accepts Interfaces, Returns Structs

This has been said to death (even I had a post of how to use interfaces in Go), so allow me to say it once more: Accept interfaces, Return Structs.

Is Extensible

Allow users of your library to extend the functions and behaviour of the objects in your library. The main method to do so in Go would be to take advantage of the composability of data types.

Which brings me to my next point -

Plays Nice

The key to a library that is generic is that it is composable. Yes, libraries compose. If we are to consider only packages (i.e. libraries whose main purpose is to share source code), then the logical endpoint would be MLton-style modules (not to be confused with Go modules).

Due to the way MLton (and SML) designed their modules system, the “Clear Value Proposition” ethos is naturally arising - libraries are usually very small. The module system of those languages also defines a helper that allows modules to be composed.

Now, I put it to you, dear readers, that something like that is doable in Go, albeit in a less pure manner.

So how does one compose Go packages? Let us imagine an alternative Go with MLton-style modules. The closest equivalent with what we have right now would be to imagine if packages only ever exported interfaces and functions. You can still write corresponding data types in a package but you cannot export them. What would the end result be?

Such an alternative programming language would produce objects that are highly composable with one another. “Accept Interfaces, Return Structs” becomes less of a maxim and is essentially enforced by the language. structs embed interfaces instead of concrete types.

If all the concrete data types of package A are accepted by functions of package B, then we say package A and package B are composable.

This happens in Go as is. The following graph shows packages that are composable with one another in my GOPATH. The source code of how this graph is generated can be found in this gist.

modernc.org/golex github.com/pkg/errors github.com/spf13/pflag gopkg.in/alecthomas/kingpin.v2 rsc.io/c2go/add/obj gopkg.in/alecthomas/kingpin.v3-unstable modernc.org/db honnef.co/go/tools/lint/lintutil modernc.org/ql github.com/nsf/gocode modernc.org/internal/file github.com/onnx-go github.com/jakub-m/gearley github.com/spf13/viper modernc.org/strutil github.com/cznic/strutil github.com/abiosoft/readline gopkg.in/readline.v1 gopkg.in/abiosoft/ishell.v2 github.com/chzyer/readline modernc.org/fileutil/storage github.com/cznic/fileutil/storage golang.org/x/debug/dwarf rsc.io/c2go rsc.io/c2go/cc github.com/cznic/cc/v2 github.com/cznic/xc github.com/cznic/cc github.com/cznic/golex/lex honnef.co/go/tools/callgraph honnef.co/go/tools/lint modernc.org/golex/lex honnef.co/go/tools/ssa gopkg.in/urfave/cli.v1 golang.org/x/debug/server gorgonia.org/tensor gorgonia.org/gorgonia github.com/kr/pretty github.com/gonuts/flag github.com/jung-kurt/gofpdf gorgonia.org/golgi github.com/spf13/afero rsc.io/pdf github.com/pkg/sftp github.com/gorgonia/bindgen modernc.org/lldb gopkg.in/doug-martin/goqu.v3 github.com/clipperhouse/typewriter github.com/awalterschulze/gographviz github.com/icza/bitio github.com/xwb1989/sqlparser github.com/gonum/blas golang.org/x/time/rate github.com/kr/text/cmd/agg golang.org/x/oauth2 github.com/sirupsen/logrus github.com/Sirupsen/logrus gopkg.in/check.v1 gorgonia.org/cu github.com/gorgonia/agogo github.com/google/btree github.com/golang/freetype/raster github.com/golang/freetype/truetype github.com/disintegration/imaging github.com/go-resty/resty modernc.org/file honnef.co/go/tools/staticcheck/vrp gopkg.in/yaml.v2 github.com/golang/glog github.com/kelseyhightower/envconfig github.com/gomidi/midi google.golang.org/grpc github.com/cznic/golex github.com/BurntSushi/xgbutil gonum.org/v1/plot github.com/jonas-p/go-shp github.com/korandiz/v4l github.com/cznic/ebnf2y/demo github.com/llgcode/draw2d github.com/BurntSushi/xgb github.com/faiface/beep github.com/jmoiron/sqlx golang.org/x/xerrors github.com/sanity-io/litter github.com/BurntSushi/toml github.com/nfnt/resize modernc.org/sortutil github.com/fogleman/gg github.com/peterh/liner golang.org/x/debug/macho github.com/eugene-eeo/rope gopkg.in/gorp.v1 github.com/lynic/gorgonnx github.com/ulikunitz/xz github.com/alecthomas/template/parse github.com/stretchr/objx rsc.io/x86/x86asm github.com/delaneyj/cogent github.com/gonum/graph gopkg.in/gizak/termui.v1 github.com/kr/fs golang.org/x/debug/gocore github.com/golang/dep github.com/pelletier/go-toml github.com/cznic/lexer github.com/googleapis/gax-go github.com/nickng/bibtex github.com/gizak/termui github.com/cznic/fileutil/falloc golang.org/x/debug/gosym modernc.org/fileutil/falloc honnef.co/go/tools/unused github.com/gonum/matrix github.com/lucasb-eyer/go-colorful github.com/cznic/lex github.com/fortytw2/leaktest golang.org/x/debug github.com/gonum/lapack

The arrows point at an interface defined outside the package (i.e. this is a dependency). From this we can derive a metric of how composable the entire Go ecosystem is. The size indicates the in-degree - how many packages implement the interfaces of a given package. This is a measure of the Postel-law-ness of the functions in a library. The colours indicate a modularity class.

Out of 762 packages in my GOPATH, only 120 are included here with 62 edges, forming 56 modules. The remaining 600+ packages are excluded for a lack of composability. Mind, for this analysis I excluded interfaces defined in the standard library because I didn’t know how to load them for analysis (probably something to do with types.Universe).

Thinking about grouping software libraries by its compositionality appears to be a weird idea at first, but it isn’t really that weird. In languages that prize abstraction over everything else (i.e. Haskell), this sort of thinking is the norm. I’m not arguing that we should do that. Instead, I am offering a different view on designing libraries.

The metric is important. For example, in building this graph, I also noticed that the Gorgonia libraries are not as composable with each other as I had originally assumed. This will be fixed in the coming summer (winter in the Northern Hemisphere).

The Tension in Designing Libraries

If you are a careful reader, you would have immediately spotted the tension that exists between the principles that I list for what qualifies as a good library.

A good reliable library does not manage resources for the user. However, this usually makes the library difficult to use.

Let us revisit the Gonum example from the section Make The Zero Value Useful.

In the first part of the example, repeated here:

c := mat.NewDense(2, 2, make([]float64, 4))
c.Mul(a, b)

We see that this follows very much the “Don’t manage resources for your user”. Instead, the user has to create the *Dense, and allocate the value (that’s what make([]float64, 4) is there for).

Gonum provides a user friendly alternative, as shown in the second part of the example:

var c mat.Dense
c.Mul(a, b)

However, this violates the “Don’t manage resources for your user”.

A more egregious example can be found in my own tensor library. A *tensor.Dense has a method Mul, defined with a signature as follows:

func (t *Dense) Mul(other *Dense, opts ...FuncOpt) (*Dense, error)

By default the tensor library manages memory allocation for the user. But the functional options allow for modification to the behaviour of Mul.

So for example, one may manually manage the allocations:

a := tensor.New(tensor.WithShape(2,3), tensor.WithBacking([]float64{...}))
b := tensor.New(tensor.WithShape(3,2), tensor.WithBacking([]float64{...}))
foo := tensor.New(tensor.WithShape(2,2), tensor.Of(tensor.Float64))
c, err := a.Mul(b, T.WithReuse(foo))

The result, c is exactly the same as foo.

So why did I bring up the tension, and showed off two “bad” examples?

Because to resolve the tension, one must consider the bigger picture.

The Bigger Picture

Start with the big picture in mind. The big picture for the tensor package is so that it works generically across data types and generically across computation. This is useful for the kinds of deep learning workload that Gorgonia handles.

For example, the same example from above may be used on float32 types, computed in a GPU:

type Engine struct {
	tensor.StdEng
	ctx cu.Context
	*cublas.Standard
}

// Engine implements `tensor.Engine`

e := newEngine()
a := tensor.New(tensor.WithShape(2, 3), tensor.WithEngine(e), tensor.Of(tensor.Float32))
b := tensor.New(tensor.WithShape(3, 2), tensor.WithEngine(e), tensor.Of(tensor.Float32))
c := tensor.New(tensor.WithShape(2, 2), tensor.WithEngine(e), tensorlOf(tensor.Float32))

// fill up the values of a and b
// ...

_, err := a.Mul(b, tensor.WithReuse(c))

Now that the big picture is clear, we can choose to make some compromises. To take stock:

  • The tensor library is designed to be generic across data types and generic across computation.
  • We should not manage resources for the user.
  • The library must be easy to use.
  • The library must be extensible.

If we prioritise “not managing resources for the user”, then we immediately lose “easy to use” and “extensible”.

However, if we do not prioritize “not managing resources for the user”, then a user who wants to use the tensor package using CUDA might fall into the trap of thinking that the default behaviour for Mul would work on CUDA as well (it will not - the program panics because GPU memory access is quite finnicky)

Consider the Use Cases

To resolve the tension, I considered the different use cases. The most common use case, I reasoned, would be to use the tensor package on the CPU, with well known data types like float64 and float32.

I built a hierarchy of needs, with GPU usage at the top. This sacrifices some of the ease-of-use, but I reckoned if you want to use GPU, you’d be an expert user.

Hierarchical Use of Libraries

Sometimes, it’s not quite possible to resolve the tension. In cases like that, I would opt to build a hierarchical family of libraries.

In the course of programming across different languages, I have noticed some good patterns. Good libraries are somewhat hierarchically organised - they are built on top of structures from libraries in a compositional manner.

More specific to use case
More specific to use case
More generic
More generic
More manual management
More manual management
More automatic management
More automatic management
main
main

Recall that the act of building libraries is in service of building a useful program. It would be very nice to build a user-friendly library so that it may be used in the final program.

The solution is hence to build a family of libraries. Each library builds atop a more fundamental library. As we traverse up the hierarchy of libraries, the use case becomes more narrower. In narrowing the possible use cases, it becomes more feasible to make decisions to perform automatic management for the user.

Note that this is orthogonal to the notion of abstracting. While it is true that as we traverse upwards along the hierarchy of libraries, the libraries become more abstract in general. But this doesn’t necessarily have to be the case. Abstracting away the details is an orthogonal issue to be discussed on another day.

There aren’t very many “families” of libraries out there in the Go ecosystem. Here are a few:

I think it’s generally a good thing that there aren’t many “families” of libraries. Observe the class of problems that these families of libraries solve - Gorgonia solves the deep learning problem. Gonum solves the numeric libraries problem. Go-HEP solves problems in high energy physics. Fyne solves GUI problems. These are hard problems. I have no doubt that if we look into the Docker or Kubernetes sub-ecosystems we will find families there too.

There is a danger of over-engineering when designing families of libraries. That’s an article for another day.

Make the Trade-off Clear

After the decision has been made, the trade-off should be well documented. Every library at every level should have the trade-offs listed.

This is especially true of mid-level

Conclusion

This article is quite long, at close to 4000 words (if count-words is to be trusted). Perhaps it’s time to end.

The main point of this article is that library design is hard. There are many considerations to take into account. I list them here:

  • What goes into a library?
  • What types of library?
  • A good library is reliable, having the following features:
    • Does one thing/Provides one resource.
    • Is well-tested
    • Doesn’t manage resources for users.
  • A good library is easy to use
    • Has good documentation and examples
    • Does not panic
    • Has minimal dependencies
    • Makes the zero value useful
  • A good library is generic
    • Functions that accept interfaces and return structs
    • Is extensible.
    • Plays nice with the environment.
  • Consider the big picture reason for designing a library.
  • Consider the use cases.
  • Consider making a family of libraries.
  • Make the trade-off clear.

Some Resources

Ack

Many thanks to Egon Elbre, Brandon Stillitano, Darrell Chua and Gary Miller for reviewing earlier drafts of this article and pointing me to additional resources

comments powered by Disqus