cmd/go: stamp git/vcs current HEAD hash/commit hash/dirty bit in binaries

[bradfitz]

(Related but different than #35667)

cmd/go currently embeds all the module dep information in binaries and it's readable with e.g. https://godoc.org/rsc.io/goversion/version but it does not include any information about the top-level module's version.

I propose that cmd/go look at {git,svn,etc} state and include in the binary:

• HEAD commit time
• HEAD hash
• dirty bit (if there are uncommitted changes)

Currently many projects do this by hand with a build-program.sh and stamping it manually with --ldflags=-X foo=bar, but that means programs built the normal Go way lack that information, and people end up with non-portable (shell, often) build scripts.

I've hit this enough times with my own projects that it's actively frustrating me. It's worse when programs are clients that want to report their version number to a server (which might want to do analytics, build horizon enforcement, protocol version negotiation, etc) and then can't. There are alternative ways to do all that, but they're tedious.

Mostly I'm concerned that people have bespoke, often non-portable build scripts.

[bcmills]

This may be a duplicate of #29814.

[rsc]

We should figure out exactly what we want to record.
The other modules are all versions or pseudo-versions.
Should this one be too?
Can we find that quickly enough to be reasonable to run during every 'go build'?
(We can skip it during 'go test' like we skip dwarf.)
The go command already has code to turn a commit hash into a version; we should probably just use that same code and add a +modified if the working directory is modified.

It would be helpful to time how much overhead this would be in 'go build'.

@bcmills, do you have any numbers about how much time this would add?

[rsc]

BTW I agree it's a duplicate of #29814 but I'll keep using this one because it is marked as a proposal and already appeared in the minutes.

[bcmills]

It would be helpful to time how much overhead this would be in 'go build'.

git status --porcelain=v2 in the go repository is around 50ms for me, and git log -n 1 is around 25ms.

go install cmd/go for me is 1.6s dirty, and 140ms clean. So assuming that we can run the VCS commands in parallel with builds for non-main packages, the latency hit should be negligible.

CC @jayconrod @matloob

[bcmills]

Hmm, I realized that I didn't account for checking tags in the above calculations. Still, I expect those costs will be order-of-magnitude similar to any other git command.

[dottedmag]

I'd like to point out a (maybe small) problem with this approach: changing the version of source code, but not the code itself will cause the binary to change.

Let me explain the use-case I have that will be broken by this change:

• A number of binaries are built from a monorepo, each binary is packaged in Docker, and a Helm chart is generated for every binary. This chart includes Docker image ID, not the tag.
• These Helm charts are installed to Kubernetes

If the Helm chart contents and the binaries don't change, then no upgrades are performed by Kubernetes.

If the version of the checkout is stamped into every Go binary, then this scheme crumbles and:

• either a version needs to be purged from every binary before packaging it into a Docker image,
• or some crazy scheme has to be invented: remove the version from the binary, take a checksum, check that there is no Docker image with this version as a label published in the registry, publish the image and label it, use this label in Helm charts.

[bradfitz]

@dottedmag, what if we made it conditional on importing a new package, say runtime/version, containing the accessor funcs to get at the info? Then if you don't use it, no change in behavior.

Would that work for your use case?

[mark-rushakoff]

I'm in a practically identical case to @dottedmag.

Inevitably, somewhere in the monorepo we will (perhaps unintentionally) bring in a dependency that depends on a magic package that breaks deterministic builds.

I think for most common cases, having this proposal enabled by default would be preferable. For my use case, I would be satisfied if there was a documented way to opt out of it. We already are using -ldflags=-buildid= to force a consistent build ID as part of deterministic binaries, so another esoteric flag to opt out of recording VCS state would be completely acceptable.

[dottedmag]

@dottedmag, what if we made it conditional on importing a new package, say runtime/version, containing the accessor funcs to get at the info? Then if you don't use it, no change in behavior.

I agree with @mark-rushakoff: relying on imports will be brittle unless this import is considered only for main packages.

It's not an author of some recursively included library, but a builder of a final binary who in a position to decide whether to put versioning information into the binary or not.

[bcmills]

@dottedmag, note that many functionally-equivalent builds will already produce slightly different binaries due to the version-stamping for runtime/debug.ReadBuildInfo. The version stamps will change with each change to the corresponding module versions, even if the contents of the specific imported packages are the same.

This proposal would case more of the same sort of version churn, but it is fundamentally the same churn.

That suggests that we may want to provide an option to disable version stamping in general. IMO, that should be a separate proposal.

[dottedmag]

True. In practice, it is not a problem as changing the versions of dependencies nearly always changes the code of dependencies — nobody is updating versions of dependencies endlessly for no reason, usually, they only get updated to get a new feature or a bugfix.

Filed #37693.

[rsc]

The discussion above about reproducible builds sounds like it would be satisfied by having the version embedded by default but also having an opt-out command-line flag; no special package needed.

Do I have that right, @dottedmag and @mark-rushakoff?

[mark-rushakoff]

Yes, I think a flag to opt out of embedding version details would suffice for reproducible builds.

It would be nice if there was a single flag like -reproducible to omit version details and to set a fixed build ID, but it is probably fine if those remain separate concerns.

I don't care about reproducible builds when I'm at the command line building something for my own use; I care about reproducible builds when I am writing build scripts that run as part of a CI/CD pipeline, so it is not a big deal if I need to look up the whole collection of settings to make those builds reproducible.

[dottedmag]

@rsc Correct.

[rsc]

OK, it sounds like everyone agrees about doing this by default, with a flag to turn it off.
It is unclear exactly how fast it will be to ask git/etc what we need to know, but that operation can overlap with the entire build, including the link. Right now if I build helloworld I get:

/Users/rsc/go/pkg/tool/darwin_amd64/link -o $WORK/b001/exe/a.out -importcfg $WORK/b001/importcfg.link -buildmode=exe -buildid=SqhXBrEZODPkt1gG-6fj/H70nrokRrHQB9NgKyehx/mN9hM1-9avNnPeQfRwgY/SqhXBrEZODPkt1gG-6fj -extld=clang $WORK/b001/_pkg_.a
/Users/rsc/go/pkg/tool/darwin_amd64/buildid -w $WORK/b001/exe/a.out # internal

That buildid step could install the git version info too. I'm confident git will be faster than the link.

Based on the discussion, then, this seems like a likely accept, although we may not be able to implement it until the next release (Go 1.16).

[seankhliao]

Will this include just the commit hash or also a (any?) version tag

[liggitt]

Version tags introduce many sharp edges, at least for git… since you can have a git repo cloned without having fetched all tags, or can have different local tags, or can add a tag to a SHA at any point in time, using the nearest tag (e.g. git describe or equivalent) would mean that a build of the same SHA could result in different embedded versions for different people (or for the same person at different times). Further, if the closest tagged commit has multiple tags associated with it, the version could be ambiguous (git describe has particularly unpredictable behavior in this case).

[bcmills]

@liggitt, note that that problem only occurs in one direction. (It is fine in general to have N names for one commit. The important property is that we resolve only one commit for a given name.)

The go command already has an algorithm for inferring the canonical name for a given repo state, implemented as part of #27171. (Specifically, we infer either the highest tagged semantic version appropriate to the module path, or a pseudo-version derived from the highest-tagged ancestor.)

I realize that that algorithm makes things awkward for the k8s.io repos in particular because of their unusual tagging history (in which v1.5.1 is an ancestor of v0.18.0), but from what we've observed that sort of nonlinear history is an outlier.

[liggitt]

The important property is that we resolve only one commit for a given name.

Since tags can be local, preferring a tag over a sha would mean two users could have identical local tags associated with different SHAs. Or are you suggesting the mapping of shas to tags would not be determined by consulting the local VCS, but the remote canonical VCS?

[bcmills]

That seems like a reasonable motivation to include both the commit hash and semantic version, rather than just one or the other.

[rsc]

No change in consensus, so accepted. We may still need to work out exactly what to include, but everyone seems to agree that this is worth doing (barring some discovery about it being more expensive than we think).