cmd/go: clarify best practice for tool dependencies

[myitcv]

Please answer these questions before submitting your issue. Thanks!

What version of Go are you using (go version)?

go version go1.10.3 linux/amd64
vgo commit 22e23900224f03be49670113d5781e4d89090f45

Does this issue reproduce with the latest release?

Yes; and latest vgo commit (per above)

What operating system and processor architecture are you using (go env)?

GOARCH="amd64"
GOBIN="/tmp/tmp.VQw1O3x8Wy/hello/bin"
GOCACHE="/home/myitcv/.cache/go-build"
GOEXE=""
GOHOSTARCH="amd64"
GOHOSTOS="linux"
GOOS="linux"
GOPATH="/tmp/tmp.VQw1O3x8Wy"
GORACE=""
GOROOT="/home/myitcv/gos"
GOTMPDIR=""
GOTOOLDIR="/home/myitcv/gos/pkg/tool/linux_amd64"
GCCGO="gccgo"
CC="gcc"
CXX="g++"
CGO_ENABLED="1"
CGO_CFLAGS="-g -O2"
CGO_CPPFLAGS=""
CGO_CXXFLAGS="-g -O2"
CGO_FFLAGS="-g -O2"
CGO_LDFLAGS="-g -O2"
PKG_CONFIG="pkg-config"
GOGCCFLAGS="-fPIC -m64 -pthread -fmessage-length=0 -fdebug-prefix-map=/tmp/go-build414355570=/tmp/go-build -gno-record-gcc-switches"

What did you do?

Off the back of #25624 (comment), I'd like to confirm that the following represents the "best practice" advice when adding and installing tool dependencies:

cd `mktemp -d`
export GOPATH=$PWD

mkdir hello
cd hello
vgo mod -init -module example.com/hello

# this could be anywhere but for convenience...
export GOBIN=$PWD/bin

# add a dependency on golang.org/x/tools/cmd/stringer
cat <<EOD > tools.go
// +build tools

package tools

import (
        _ "golang.org/x/tools/cmd/stringer"
)
EOD

vgo install golang.org/x/tools/cmd/stringer

The go.mod and .Target for stringer look fine:

$ cat go.mod
module example.com/hello

require golang.org/x/tools v0.0.0-20180615195736-465e6f399236
$ vgo list -f "{{.Target}}" golang.org/x/tools/cmd/stringer
/tmp/tmp.VQw1O3x8Wy/hello/bin/stringer

The issue however is that running vgo mod -sync then removes our module requirement on golang.org/x/tools - I suspect this is a bug:

$ vgo mod -json
{
        "Module": {
                "Path": "example.com/hello",
                "Version": ""
        },
        "Require": [
                {
                        "Path": "golang.org/x/tools",
                        "Version": "v0.0.0-20180615195736-465e6f399236"
                }
        ],
        "Exclude": null,
        "Replace": null
}
$ vgo mod -sync
warning: "ALL" matched no packages
$ vgo mod -json
{
        "Module": {
                "Path": "example.com/hello",
                "Version": ""
        },
        "Require": null,
        "Exclude": null,
        "Replace": null
}

If we assume this is a bug and ignore it for now, I also wonder whether we can improve this workflow for adding tool dependencies somehow. The following steps feel a bit "boilerplate" and unnecessary:

• define a fake tools.go file
• add a dependency on the tool (which will never compile because it's a main package, so I'm not sure we can ever safely verify tools.go is "good"?)
• set GOBIN to an appropriate location
• vgo install tool
• ensure PATH includes GOBIN

I wonder whether we could in fact obviate all of this by having something like:

vgo install -tool golang.org/x/tools/cmd/stringer
vgo run golang.org/x/tools/cmd/stringer

Thoughts?

vgo run tool is possible as a result of #22726, but because of #25416 it effectively requires a link step each time.

What did you expect to see?

With respect to what I think is a bug with vgo mod -sync

go.mod unchanged by the vgo mod -sync

What did you see instead?

The golang.org/x/tools requirement removed.

/cc @rsc @bcmills

[bcmills]

Let's take a step back: what do you mean by “tool dependency”?

The stringer example suggests that you mean “a tool required for use with go generate”, but note that generate is intended to be run by the author of the module, not its users: a tool needed to run generate shouldn't be a hard requirement for the module or its tests.

[myitcv]

Let's take a step back: what do you mean by “tool dependency”?

Roughly:

• I'm writing a module; or, more generally, me and my team are writing a module
• we want to use a tool (e.g. stringer) whilst writing that code
• we want to reproducibly re-generate code, or put another way, we want the whole team to consistently generate the same code at any given point in time
• hence I want to ensure that everyone is using the same version of any tool we require

but note that generate is intended to be run by the author of our module, not its users: a tool needed to run generate shouldn't be a hard requirement for the module or its tests.

Correct, this would only be a dependency for the authors of the module (as you point out, the users of our module probably don't care whether or not we've used stringer or other tools).

As I understand it, these tool dependencies do not introduce any additional requirements for users of our module: they (the tool dependencies) can only ever be referenced by running vgo install or vgo run in the context of our module. So as a user of our module, I would, by definition be in a different context (specifically the module that references our module) when running vgo install or vgo run.

I'm using the term "context" here to mean "the go.mod that vgo uses." I'm sure there must be a better, more precise term

[jimmyfrasche]

I haven't played around with vgo much yet or looked under the hood.

I know a lot of package managers handle this by sorting dependencies into buckets: regular dependencies, test dependencies, dev dependencies. That seems like overkill and there can also be dependencies by build tag, like, say, importing a module for internal use but only on windows, and dependencies may only be for a certain package in a module that you may not be using.

Could vgo record all the dependencies for a project—for all build tags, for all packages, for tests, for development—but lazily fetch them as needed? (This may require more invasive integration but it would be an optimization as the semantics are the same for lazy-fetch and fetch-everything-upfront)

[myitcv]

@jimmyfrasche

Could vgo record all the dependencies for a project—for all build tags, for all packages, for tests, for development—but lazily fetch them as needed?

My understanding is that this is exactly what vgo does today, for any dependency, test/tool or otherwise.

And, to your earlier point, I agree it means the delineation between the buckets is unnecessary (for anyone wanting to know why a dependency exists, vgo can tell us that answer).

Just to clarify one point above. My proposal for vgo install -tool golang.org/x/tools/cmd/stringer would, however, add a section to go.mod (or equivalent) where the tool packages are listed. This is simply to avoid the need for the tools.go file referenced in my example.

[bcmills]

A couple of observations:

1. Some tools required by Go builds are not themselves built using the go tool. (For example, Google's protocol compiler is written in C++; only the Go code generator plugin is written in Go.) So even if we add support for tooling to the go tool, it will be at best a partial solution.

2. If you are assuming the use of the go tool anyway, then you can write a Go program that invokes it with appropriate arguments.

3. Because a command-line tool is a top-level package, its versions are not context-sensitive: if you're not mixing in other packages to customize the tool, you don't need to resolve version constraints from those packages.

To me, those suggest a fairly straightforward solution: go run pkg@version. In combination with //go:generate, that would allow you to express the dependency (at the exact version used) directly in the source file:

//go:generate go run golang.org/x/tools/cmd/stringer@v1.11.0 -type=Pill

In contrast, if you are building a custom version of the tool, then you should be able to customize it using blank-imports as you describe.

[jimmyfrasche]

@bcmills that sounds good but it would be awkward to write that each time and easy for things to go out of sync and you have some code using 1.11.0 and some using 1.12.3 but probably want .

go generate let's you define aliases with //go:generate -command alias the-command (I've not used it personally and unsure of the scope: file? package? (I assume file.))

Perhaps a more vgo-aware variant would be helpful, like

//go:generate -require golang.org/x/tools/cmd/stringer@v1.11.0

which is module scoped. It falls somewhere between an ephemeral go install -tool and a fictitious blank import.

There are a some details to iron out in such an approach, of course.

[myitcv]

@bcmills

Some tools required by Go builds are not themselves built using the go tool.

Completely. As you say, this can either be solved entirely orthogonally, or with Go wrappers (wrappers that could even take care of installing the right version).

Because a command-line tool is a top-level package, its versions are not context-sensitive: if you're not mixing in other packages to customize the tool, you don't need to resolve version constraints from those packages.

I'm not entirely sure I follow your point here. The point I was trying to make is that where (i.e. in what directory) you run vgo install golang.org/x/tools/cmd/stringer is relevant because that will determine which go.mod you reference and therefore which version of the golang.org/x/tools module you use.

To me, those suggest a fairly straightforward solution: go run pkg@version.

As @jimmyfrasche pointed out this has the major downside of littering the required version across all the places you require go:generate directives, which seems to be somewhat counter to having go.mod be the single point of reference for versions.

Just one point to note about the "proposal":

vgo install -tool golang.org/x/tools/cmd/stringer
vgo run golang.org/x/tools/cmd/stringer

Despite the explanation in #25416, I think vgo run golang.org/x/tools/cmd/stringer could actually cache the linked binary, because the cache can very precisely be trimmed given we know exactly what version is being referenced at a given point in time.

[jimmyfrasche]

Running a tool at a specific version depending on the module would have to go through vgo.

Running stringer lets $SHELL pick which stringer to run, so it would need to be something like go run-dev-tool stringer so that vgo could dispatch the correct version of the command.

Something like that could be stuffed into go run but that doesn't seem to fit with the spirit of that command. Arguably go tool would be a better fit but that could cause namespacing issues so it seems like a separate command built to purpose would be best.

go generate could handle this transparently but not all tools are necessarily for code generation. You may also want a specific version of, say, a linter.

An IDE/editor could use tool dependencies to figure out if this is a tool dependency or a regular program and run it through vgo if need be. From the command line, though, it would be up to the user to know whether to run stringer vs go run-dev-tool stringer so there would still be potential skew with the version documented. I don't think there's much to do about that, though.

A tool module could have multiple commands. Look up would be slow if these weren't included somewhere in the metadata.

A tool module could have multiple dependencies, some shared with the module under development. Unlike with other dependencies, you probably don't want those to be solved together: using foobar shouldn't affect your version baz or the version of baz used by other tools you depend on. They should all be siloed.

This is all starting to sound mostly unrelated to vgo but perhaps it's something that can be built on top of vgo.

Going back on what I said about buckets, let's say there was as separate tool godev and a separate file dev.mod.

godev install foo updates dev.mod recording the version and the exposed commands. Instead of installing the command globally it installs the binaries outside of $PATH but somewhere it can easily grab the correct version. It updates dev.mod with the vgo style import and version path but also records the names of the commands. For the most part it's a wrapper around go get that does some extra work.

godev run foo invokes the appropriate foo from dev.mod.

Is there anything that would need to change in vgo to allow something like this?

A downside of not having it integrated in vgo itself would be that go:generate directives would need to explicitly prefix godev run.

[rsc]

What is the decision to be made here?
I think the tools.go file is in fact the best practice for tool dependencies, certainly for Go 1.11.
I like it because it does not introduce new mechanisms.
It simply reuses existing ones.