BazelCon 2024 Recap
blogsystem5.substack.com
blogsystem5.substack.com
There's a decent number of Hello World examples around, and I can get moving with some rulesets, but using it for more than toy projects seems to be really hard for me and there's not a lot of intermediate content from what I can tell. There's "Advanced" reference materials, and countless "getting started" tutorials, but rarely is there anything relating to best practices or even what constitutes good rules.
Seems like most of the hobby community prefers NIX for deterministic builds too, so the situation is not improving for novices like myself who really like the idea of hermitic builds, monorepo's and distributed cache.
I'd suggest finding an open source project and playing with it locally. Try to get everything building and running, including unit tests. Make the build as granular as you can. Try to pick a project that has a variety of kinds of things that need to be built, like Java, C++, Protobufs, etc. Then look for patterns in the build targets you wrote and see if you can refactor them to simplify it with custom Skylark rules.
I forgot to mention, this is by design. It is uncommon at Google for a BUILD file to be more than 3 or 4 build targets. Maybe someone will add a for loop to the build so that they can define various flavors of the build easily, but that's it. Anything more complicated and it should be extracted to a separate Skylark file, included, and in the BUILD file it should look like a basic task.
The best way to learn it would be to use it yourself and try solving problems with it. You don't need an xoogler to learn it, my former company did not have one when we first evaluated Bazel.
There is a very active community Slack and StackOverflow for questions as well.
> rarely is there anything relating to best practices or even what constitutes good rules.
I think rules_go has been a stellar example of what a well-made rule should look like:
- Managing toolchains and stdlib within Starlark
- Provide wrapper around core toolchains to enforce best practices
- Break down a build into fine-grained, reusable actions
- Provide concrete APIs for downstream rules (i.e. graphql, swagger, proto code gen) to consume.
> Seems like most of the hobby community prefers NIX for deterministic builds too, so the situation is not improving for novices like myself who really like the idea of hermitic builds, monorepo's and distributed cache.
Nix figured out the external dependencies story much better than Bazel, much thanks to nixpkgs repo. So naturally, in a fragmented opensource ecosystem, Nix thrives much better.
However, in a closed ecosystem within an enterprise, Bazel provides a more granular approach to managing your build: smaller actions mean finer, more granular caching. It also has a much better set of APIs for distributed builds and build telemetry that folks could depend on from day 1. It's also copying some for Nix playbook with the new Bazel Central Registry resembling nixpkgs.
I think you can learn either build system. Most of the skills and concepts are transferrable: sandboxing, build hermeticity, remote caching, remote build, etc... So you often see some experts in this field being involved in multiple build tools development and not just one.
That person left, and the Bazel stuff was left unmaintained as no one had the interest nor time to learn it.
Later Bazel decide to kill off rules_docker and replace it with rules_oci, which means we can no longer update our Golang version without a super painful migration where we end up breaking production a bunch of times because of quirks in the super difficult migration.
Eventually we invested the time to rip the whole thing out and replace it with standard Go tooling with a multistage docker build for the container image. Everything from running tests, using Golang tooling, legibility of builds, CI, and deploys is easier.
The best thing we did was remove it and move to standard Golang tooling.
Personally, most cases I have seen are caused by a systematic under-investment into developer toolings and infrastructure at the company. Management layers often don't understand the software development assembly lines are not composed of just workers(software engineers), but also tools and machines that enable faster workflows. This often results in some critical processes in the pipeline from code to prod to be maintained by 1 guy: monitoring, alerting, deployments, builds, git etc... and when that 1 guy left, the system failed and the company suffered.
I think successful Bazel adoption in an org is often a signal that the company has grown in size and values its developer's time and happiness. Failure to adopt often means a lack of investment in dev experience in general.
The reality is, in a post-layoffs world, practically every company is scrimping on these sorts of “back office” things. If people couldn’t even get these complex build systems to work in the pre-layoffs low-interest-rate tech world where labour was more abundant, how can they get it done and maintained now?
I did use bazel long ago for c++ and it was quite good at it, but we didn’t have very many dependencies.
Rob Pike even has a talk about how fast ALGOL compilers used to take, naturally one needs to take into account the additional issue with reading punch cards, but still quite fast for early 1960's hardware.
Or Object Pascal and Modula-2 compilers in the early 1990's.
Eventually we (as in industry) started focusing on the wrong points.
I had to install it last night in DOSBox-X and compile the Breakout example project. The warm fuzzies are justified.
34 000 lines of code per minute!
https://bitsavers.org/pdf/borland/turbo_pascal/Turbo_Pascal_...
Sounds like your company has bigger issues than build system ;)
> It seems like a general increase in complexity
Try to support a really big monorepo with existing go.mod tooling and you’ll see that it really isnt
This is one of the biggest challenges where Bazel falls short of non-Bazel tooling for us in the Web development/Node.js world. The Node ecosystem has characteristics that push Bazel scaling (node_modules is 100k+ files and 1GB+), and Bazel's insistence on getting builds correct/reproducible is in a way its own enemy. Bazel needs to set up 100s of thousands of file watchers to be correct, but the Node.js ecosystem's "let's just assume that node_modules hasn't changed" is good enough most of the time. For us, many non-Bazel inner devloop steps inflate from <5s to >60s after Bazel overhead, even after significant infra tuning.
Turns out, when your build system doesn't occupy 5-15% of your teams' working days, they get a lot more done!
It’s for sure a weak point with Bazel though.
https://www.fullyearcal.com/calendar/iMRr-2l16tI/software-en...
It has a build and test and coverage cache that actually consistently works. It can run tests in a sandbox to prevent your local environment from getting consumed by the test (this specific thing has actually caused production outages before).
A DAG for your codebase is also super useful as you can automate a lot of “clean up once you’re done” type work.
If their goal is adoption, that is just _awful_ pathfinding. Probably harkens the beginning of a death spiral for the project, if I'm to make a prediction.