Best practices for examples with Pkg3

Often, example scripts explicitly import dependencies of the package they are included with at top level scope. This was not an issue with OldPkg as the dependencies are installed and visible for code loading in any context.

With Pkg(3), the dependencies need to be explicitly added to the active environment. This requires more actions and skills from your audience and might increase the threshold and decrease willingness of people to explore your package and features.

What is recommended in this scenario? Related, what is the best/portable way to refer to the examples directory from the REPL?

1 Like

I am not sure I understand the question, but my understanding of the difference is that previously you had to specify dependencies in REQUIRE which was used for installation, but not checked when loading, while for the new Pkg you have to declare the dependency before loading.

For package developers, AFAIK the recommended workflow is to

pkg> activate path/to/package
pkg> add ThisDependency

which IMO is not that much of a barrier to entry.

1 Like

Sorry for not being clearer. Say Foo has a dependency Bars. Perhaps the API of Foo includes a method

Foo.f(x::Bars.Bar) = (x,x)

I understand that I - as a package developer - have to specify Bars as a dependency by declaring it in either REQUIRE or Project.toml. To explain to my users how the package is used I include an example at Foo/examples/hello.jl:

using Foo, Bars
x = Bar(3)
xx = f(x)

In Julia-0.6, users could run this by e.g.:


In julia-0.7 this fails because only Foo can be loaded in scripts included at the REPL. To solve this one would have to warn the user that additional installs are required to run the examples (in As comments in hello.jl?) Users might miss this and will be confronted with package manager errors when trying out the example.

I guess a solution could be to put each example in a separate directory with its own ‘Project.toml’ file. Still you would have to instruct the user to:

  • Navigate to that directory (not trivial given the slug in the path)
  • ]activate .
  • ]instantiate
  • include("hello.jl")

EDIT: Another solution would be to always reexport all symbols from imported dependencies. But that feels like it defeats the point of splitting up functionality over different packages.


In the global/default environment,

pkg> add git/repofor/Foo

for unregistered packages should be sufficient, then the user can run whatever script is required.

Also note that Pkg.dir is gone, search this forum for the relevant discussions.

Sorry but still not what I am steering at.

I understand that Pkg.dir is gone, this is part of what makes it harder for users playing around with packages and their included example (I understand there is an unexported and undocumented Base.find_package that alleviates this to some extent?).

Your oneliner indeed should install the package Foo and its dependency Bars. The problem is that Bar will only be visible from within Foo, not from the top-level examples that should provide a gentle getting started guide to the user.

In other words what is the minimum number of manipulations that makes this work (lets assume both Foo and Bars are registered:

pkg>add Foo
julia>include(<morestuff>, "hello.jl")

Ideally I do not want to burden my visitor with inspecting what hello.jl depends on and installing those dependencies (which is the case in julia-0.6).

1 Like

<stuff> = pkg> add Bar

As for <morestuff>,

  1. define Foo.examplepath(...) using @__DIR__ (recommended solution),
  2. use Base.pathof(Foo).

This is a solution but you will agree it adds a significant overhead compared to the steps required in Julia-0.6. I would argue your solution expands to:

  • Read hello.jl, or let’s call it demo_all_pkg_features.jl to stress there are cases in which the user is not necessarily interested in the details of the code contained in the example.
  • Correctly list and add all dependencies. Users might find this difficult or might be unwilling to pollute their primal environment.
  • Then after all installations complete:
using Foo # yet another step not required in 0.6
d = include(Foo.examplepath("hello.jl"))

Given how lightweight projects and environments have become I was really hoping for a much more idiomatic solution to this very common problem.

1 Like

Nope, I don’t agree.

Sorry, I don’t understand what you are saying here. If the user is not interested in the example code, then what purpose does it serve?

Again, installing a package is just a single add command. Pkg will install the dependencies. In case the example requires an additional package, that should be installed, too. I think this is trivial.

Also, I don’t really see the issue about “pollution” of the environment, but if the user is concerned about that, they should activate another one.

Finally, I don’t see the point of includeing an example. I would just open it in the editor, and send it to the REPL line by line to understand what is going on.

Sorry, I don’t understand what you are saying here. If the user is not interested in the example code, then what purpose does it serve?

Someone might be looking for a good plotting package. Often the example showcase what type of plots are in store. At this point I just want to know whether this package can give me a histogram with transparent bars. I don’t want to worry about whether or not I need e.g. GeomTypes whilst shopping for a plotting lib. Just running examples can also give a good idea w.r.t to the package performance.

In GLVisualize, the authors support prospective users by leveraging the testing infrastructure of Julia. In this case, a oneliner:

pkg>test GLVisualize

will resolve additional dependencies (and expose them top-level if needed), create a separate environment, instantiate it, and include the tests (or interactive demos as is the case for GLVisualize).

In my opinion it would be convenient for this functionality to be extended to something like:

pkg>example Foo hello

I agree just running the examples it too limiting. Maybe in addition to what happens when testing, the examples could be copied over in a directory with a predictable name (or custom name), similar to what happens when you develop a package.

Given the flexibility of the new package manager and the how common it is for packages to ship with an examples subdirectory I was just wondering whether someone had automated something like this.

Presumably the documentation or the main page of the project on Github etc would be a good starting places for this information. Eg PGFPlotsX does this.

However, if you want code in a directly executable form, probably packaging the demo itself as an additional project would be a viable solution.

You do this:

julia> using StaticArrays
[ Info: Precompiling StaticArrays [90137ffa-7385-5640-81b9-e52037218182]

julia> pathof(StaticArrays)

The difference is that you have to pass the package module rather than its name.

1 Like

Moreover, the targets mechanism makes it easy for the package to declare project environments that extend its own, just do:

Bars = "<uuid>"

examples = ["Bars"]

We still need an API to activate target environments (pkg> activate Bars.examples)?, but with that you could just do that and you’ll be able to run the examples from that environment easily.


At a very high level, this issue is about whether Bars is an internal implementation detail of Foo or if it is part of Foo's external API. In general, one does not want a package’s dependencies exposed to its users since doing that prevents it from changing its implementation in the future. The new package manager and code loading mechanism takes a very hard stance on this: dependencies are an internal implementation detail and are private. To the point that one of your dependencies can have a dependency with the same name as one of your project’s dependencies even if it is actually a different package.

Running package examples is definitely something we need to work out how to do conveniently, but it does not take precedence in importance over keeping a package’s internal implementation details isolated and private. I’m not saying that to chastise but to frame this conversation appropriately: rather than complaining about / trying to redesign the code loading so that this one relatively marginal thing can be done as easily as possible, try to think about how to make running example code easy, taking the way code loading now works as a fixed point—since it is not changing now that 1.0 is released.

1 Like

Is the strategy to re-export relevant parts of Bars from Foo a reasonable alternative, considering that users of Foo are expected to know about and use Bars? Or this is no longer possible the way it was in 0.6?

If it is a reasonable strategy, is Reexport.jl a good option? I see it has been updated recently to support Julia 1.0, but have heard some loose arguments against use of it before.

I very much welcome the strictness of the package manager. What I was looking for was something built on top of the current framework, like the solution you mention is in the works.

I don’t think the situation I described necessarily comes from Foo exposing its implementation. The situation I found myself in is with Foo a library that is designed to do processing on objects implementing the Bar concept (concept as in the c++ meaning of the word). Package Bars defines the concept by (i) introducing a set of functions meant for other packages to extend, and (ii) providing a reference implementation Bar accompanied by unit tests that guard the semantics of what it means to be a Bar.

A bad example of this is the Array concept in Base. Base introduces the original getindex etc, but also includes the reference implementation Array{T,N}. As a result, many other packages both import Base and allow objects from Base to be passed to the API. This example is bad because Base is the one package the manager lets you have for free.

As another example I could think of a package Fields defining the mathematical concept of field by introducing say mult_inverse etc. and providing the reference implementation Galois{P}. Then another package LinearSpaces comes along that will have to import functions from Fields to implement say elementwise_addition. It is very like LinearSpaces will also include Fields in its suite of examples. I wouldn’t say in this case the packages needlessly expose implementation details.

1 Like

Since test can be defined as a target, how about build? In some cases, one could need some packages, specially Pkg itself just for building and wouldn’t want to include it as a dependency in the project.