Stamp It! All Programs Must Report Their Version
Recently, during a production incident response, I guessed the root cause of an outage correctly within less than an hour (cool!) and submitted a fix just to rule it out, only to then spend many hours fumbling in the dark because we lacked visibility into version numbers and rollouts… 😞 This experience made me think about software versioning again, or more specifically about build info (build versioning, version stamping, however you want to call it) and version reporting. I realized that for the i3 window manager, I had solved this problem well over a decade ago, so it was really unexpected that the problem was decidedly not solved at work. In this article, I’ll explain how 3 simple steps (Stamp it! Plumb it! Report it!) are sufficient to save you hours of delays and stress during incident response. Every household appliance has incredibly detailed versioning! Consider this dishwasher: (Thank you Feuermurmel for sending me this lovely example!) I observed a couple household appliance repairs and am under the impression that if a repair person cannot identify the appliance, they would most likely refuse to even touch it. So why are our standards so low in computers, in comparison? Sure, consumer products are typically versioned somehow and that’s typically good enough (except for, say, USB 3.2 Gen 1×2!). But recently, I have encountered too many developer builds that were not adequately versioned! Unlike a physical household appliance with a stamped metal plate, software is constantly updated and runs in places and structures we often cannot even see. Let’s dig into what we need to increase our versioning standard! Usually, software has a name and some version number of varying granularity: All of these identify the Chrome browser on my computer, but each at different granularity. All are correct and useful, depending on the context. Here’s an example for each: After creating the i3 window manager , I quickly learned that for user support, it is very valuable for programs to clearly identify themselves. Let me illustrate with the following case study. When running , you will see output like this: Each word was carefully deliberated and placed. Let me dissect: When doing user support, there are a couple of questions that are conceptually easy to ask the affected user and produce very valuable answers for the developer: Based on my experiences with asking these questions many times, I noticed a few patterns in how these debugging sessions went. In response, I introduced another way for i3 to report its version in i3 v4.3 (released in September 2012): a flag! Now I could ask users a small variation of the first question: What is the output of ? Note how this also transfers well over spoken word, for example at a computer meetup: Michael: Which version are you using? User: How can I check? Michael: Run this command: User: It says 4.24. Michael: Good, that is recent enough to include the bug fix. Now, we need more version info! Run please and tell me what you see. When you run , it does not just report the version of the i3 program you called, it also connects to the running i3 window manager process in your X11 session using its IPC (interprocess communication) interface and reports the running i3 process’s version, alongside other key details that are helpful to show the user, like which configuration file is loaded and when it was last changed: This might look like a lot of detail on first glance, but let me spell out why this output is such a valuable debugging tool: Connecting to i3 via the IPC interface is an interesting test in and of itself. If a user sees output, that implies they will also be able to run debugging commands like (for example) to capture the full layout state. During a debugging session, running is an easy check to see if the version you just built is actually effective (see the line). Showing the full path to the loaded config file will make it obvious if the user has been editing the wrong file. If the path alone is not sufficient, the modification time (displayed both absolute and relative) will flag editing the wrong file. I use NixOS, BTW, so I automatically get a stable identifier ( ) for the specific build of i3. To see the build recipe (“derivation” in Nix terminology) which produced this Nix store output ( ), I can run : Unfortunately, I am not aware of a way to go from the derivation to the source, but at least one can check that a certain source results in an identical derivation. The versioning I have described so far is sufficient for most users, who will not be interested in tracking intermediate versions of software, but only the released versions. But what about developers, or any kind of user who needs more precision? When building i3 from git, it reports the git revision it was built from, using : A modified working copy gets represented by a after the revision: Reporting the git revision (or VCS revision, generally speaking) is the most useful choice. This way, we catch the following common mistakes: As we have seen above, the single most useful piece of version information is the VCS revision. We can fetch all other details (version numbers, dates, authors, …) from the VCS repository. Now, let’s demonstrate the best case scenario by looking at how Go does it! Go has become my favorite programming language over the years, in big part because of the good taste and style of the Go developers, and of course also because of the high-quality tooling: I strive to respect everybody’s personal preferences, so I usually steer clear of debates about which is the best programming language, text editor or operating system. However, recently I was asked a couple of times why I like and use a lot of Go, so here is a coherent article to fill in the blanks of my ad-hoc in-person ramblings :-). Read more → Therefore, I am pleased to say that Go implements the gold standard with regard to software versioning: it stamps VCS buildinfo by default! 🥳 This was introduced in Go 1.18 (March 2022) : Additionally, the go command embeds information about the build, including build and tool tags (set with -tags), compiler, assembler, and linker flags (like -gcflags), whether cgo was enabled, and if it was, the values of the cgo environment variables (like CGO_CFLAGS). Both VCS and build information may be read together with module information using or runtime/debug.ReadBuildInfo (for the currently running binary) or the new debug/buildinfo package. Note: Before Go 1.18, the standard approach was to use or similar explicit injection. This setup works (and can still be seen in many places) but requires making changes to the application code, whereas the Go 1.18+ stamping requires no extra steps. What does this mean in practice? Here is a diagram for the common case: building from git: This covers most of my hobby projects! Many tools I just , or if I want to easily copy them around to other computers. Although, I am managing more and more of my software in NixOS. When I find a program that is not yet fully managed, I can use and the tool to identify it: It’s very cool that Go does the right thing by default! Systems that consist of 100% Go software (like my gokrazy Go appliance platform ) are fully stamped! For example, the gokrazy web interface shows me exactly which version and dependencies went into the build on my scan2drive appliance . Despite being fully stamped, note that gokrazy only shows the module versions, and no VCS buildinfo, because it currently suffers from the same gap as Nix: For the gokrazy packer, which follows a rolling release model (no version numbers), I ended up with a few lines of Go code (see below) to display a git revision, no matter if you installed the packer from a Go module or from a git working copy. The code either displays (the easy case; built from git) or extracts the revision from the Go module version of the main module ( ): What are the other cases? These examples illustrate the scenarios I usually deal with: This is what it looks like in practice: But a version built from git has the full revision available (→ you can tell them apart): When packaging Go software with Nix, it’s easy to lose Go VCS revision stamping: So the fundamental tension here is between reproducibility and VCS stamping. Luckily, there is a solution that works for both: I created the Nix overlay module that you can import to get working Go VCS revision stamping by default for your Nix expressions! Tip: If you are not a Nix user, feel free to skip over this section. I included it in this article so that you have a full example of making VCS stamping work in the most complicated environments. Packaging Go software in Nix is pleasantly straightforward. For example, the Go Protobuf generator plugin is packaged in Nix with <30 lines: official nixpkgs package.nix . You call , supply as the result from and add a few lines of metadata. But getting developer builds fully stamped is not straightforward at all! When packaging my own software, I want to package individual revisions (developer builds), not just released versions. I use the same , or if I need the latest Go version. Instead of using , I provide my sources using Flakes, usually also from GitHub or from another Git repository. For example, I package like so: The comes from my : Go stamps all builds, but it does not have much to stamp here: Here’s a full example of gokrazy/bull: To fix VCS stamping, add my overlay to your : (If you are using , like I am, you need to apply the overlay in both places.) After rebuilding, your Go binaries should newly be stamped with buildinfo: Nice! 🥳 But… how does it work? When does it apply? How do you know how to fix your config? I’ll show you the full diagram first, and then explain how to read it: There are 3 relevant parts of the Nix stack that you can end up in, depending on what you write into your files: For the purpose of VCS revision stamping, you should: Hence, we will stick to the left-most column: fetchers. Unfortunately, by default, with fetchers, the VCS revision information, which is stored in a Nix attrset (in-memory, during the build process), does not make it into the Nix store, hence, when the Nix derivation is evaluated and Go compiles the source code, Go does not see any VCS revision. My Nix overlay module fixes this, and enabling the overlay is how you end up in the left-most lane of the above diagram: the happy path, where your Go binaries are now stamped! How does the overlay work? It functions as an adapter between Nix and Go: So the overlay implements 3 steps to get Go to stamp the correct info: For the full source, see . See Go issue #77020 and Go issue #64162 for a cleaner approach to fixing this gap: allowing package managers to invoke the Go tool with the correct VCS information injected. This would allow Nix (or also gokrazy) to pass along buildinfo cleanly, without the need for workarounds like my adapter . At the time of writing, issue #77020 does not seem to have much traction and is still open. My argument is simple: Stamping the VCS revision is conceptually easy, but very important! For example, if the production system from the incident I mentioned had reported its version, we would have saved multiple hours of mitigation time! Unfortunately, many environments only identify the build output (useful, but orthogonal), but do not plumb the VCS revision (much more useful!), or at least not by default. Your action plan to fix it is just 3 simple steps: Implementing “version observability” throughout your system is a one-day high-ROI project. With my Nix example, you saw how the VCS revision is available throughout the stack, but can get lost in the middle. Hopefully my resources help you quickly fix your stack(s), too: Now go stamp your programs and data transfers! 🚀 Chrome 146.0.7680.80 Chrome f08938029c887ea624da7a1717059788ed95034d-refs/branch-heads/7680_65@{#34} “This works in Chrome for me, did you test in Firefox?” “Chrome 146 contains broken middle-click-to-paste-and-navigate” “I run Chrome 146.0.7680.80 and cannot reproduce your issue” “Apply this patch on top of Chrome f08938029c887ea624da7a1717059788ed95034d-refs/branch-heads/7680_65@{#34} and follow these steps to reproduce: […]” : I could have shortened this to or maybe , but I figured it would be helpful to be explicit because is such a short name. Users might mumble aloud “What’s an i-3-4-2-4?”, but when putting “version” in there, the implication is that i3 is some computer thing (→ a computer program) that exists in version 4.24. is the release date so that you can immediately tell if “ ” is recent. signals when the project was started and who is the main person behind it. gives credit to the many people who helped. i3 was never a one-person project; it was always a group effort. Question: “Which version of i3 are you using?” Since i3 is not a typical program that runs in a window (but a window manager / desktop environment), there is no Help → About menu option. Instead, we started asking: What is the output of ? Question: “ Are you reporting a new issue or a preexisting issue? To confirm, can you try going back to the version of i3 you used previously? ”. The technical terms for “going back” are downgrade, rollback or revert. Depending on the Linux distribution, this is either trivial or a nightmare. With NixOS, it’s trivial: you just boot into an older system “generation” by selecting that version in the bootloader. Or you revert in git, if your configs are version-controlled. With imperative Linux distributions like Debian Linux or Arch Linux, if you did not take a file system-level snapshot, there is no easy and reliable way to go back after upgrading your system. If you are lucky, you can just the older version of i3. But you might run into dependency conflicts (“version hell”). I know that it is possible to run older versions of Debian using snapshot.debian.org , but it is just not very practical, at least when I last tried. Can you check if the issue is still present in the latest i3 development version? Of course, I could also try reproducing the user issue with the latest release version, and then one additional time on the latest development version. But this way, the verification step moves to the affected user, which is good because it filters for highly-motivated bug reporters (higher chance the bug report actually results in a fix!) and it makes the user reproduce the bug twice , figuring out if it’s a flaky issue, hard-to-reproduce, if the reproduction instructions are correct, etc. A natural follow-up question: “ Does this code change make the issue go away? ” This is easy to test for the affected user who now has a development environment. Connecting to i3 via the IPC interface is an interesting test in and of itself. If a user sees output, that implies they will also be able to run debugging commands like (for example) to capture the full layout state. During a debugging session, running is an easy check to see if the version you just built is actually effective (see the line). Note that this is the same check that is relevant during production incidents: verifying that effectively running matches supposed to be running versions. Showing the full path to the loaded config file will make it obvious if the user has been editing the wrong file. If the path alone is not sufficient, the modification time (displayed both absolute and relative) will flag editing the wrong file. People build from the wrong revision. People build, but forget to install. People install, but their session does not pick it up (wrong location?). Nix fetchers like are implemented by fetching an archive ( ) file from GitHub — the full repository is not transferred, which is more efficient. Even if a repository is present, Nix usually intentionally removes it for reproducibility: directories contain packed objects that change across runs (for example), which would break reproducible builds (different hash for the same source). We build from a directory, not a Go module, so the module version is . The stamped buildinfo does not contain any information. Fetchers. These are what Flakes use, but also non-Flake use-cases. Fixed-output derivations (FOD). This is how is implemented, but the constant hash churn (updating the line) inherent to FODs is annoying. Copiers. These just copy files into the Nix store and are not git-aware. Avoid the Copiers! If you use Flakes: ❌ do not use as a Flake input ✅ use instead for git awareness I avoid the fixed-output derivation (FOD) as well. Fetching the git repository at build time is slow and inefficient. Enabling , which is needed for VCS revision stamping with this approach, is even more inefficient because a new Git repository must be constructed deterministically to keep the FOD reproducible. Nix tracks the VCS revision in the in-memory attrset. Go expects to find the VCS revision in a repository, accessed via file access and commands. It synthesizes a file so that Go’s detects a git repository. It injects a command into the that implements exactly the two commands used by Go and fails loudly on anything else (in case Go updates its implementation). It sets in the environment variable. Stamp it! Include the source VCS revision in your programs. This is not a new idea: i3 builds include their revision since 2012! Plumb it! When building / packaging, ensure the VCS revision does not get lost. My “VCS rev with NixOS” case study section above illustrates several reasons why the VCS rev could get lost, which paths can work and how to fix the missing plumbing. Report it! Make your software print its VCS revision on every relevant surface, for example: Executable programs: Report the VCS revision when run with For Go programs, you can always use Services and batch jobs: Include the VCS revision in the startup logs. Outgoing HTTP requests: Include the VCS revision in the HTTP responses: Include the VCS revision in a header (internally) Remote Procedure Calls (RPCs): Include the revision in RPC metadata User Interfaces: Expose the revision somewhere visible for debugging. My overlay for Nix / NixOS My repository is a community resource to collect examples (as markdown content) and includes a Go module with a few helpers to make version reporting trivial.
DevOps
Web Development
Hardware
Cloud
Linux
C#
Tutorial
HTML
CSS
Performance
Go
Programming
Php
Security
Career
AI
Writing
C++
Open Source
JavaScript
Python
Shell
Ci Cd