Bazel

do.

So how long until we take over Bazel and just be like, ah, Bazel's garbage? You know, actually, you guys should just use Nix.

If you want really low resolution granularity, you could just do the whole package build as a single thing.

what we do today type of thing.

Anything that you could shim, this approach could work for.

some things it won't work for.

So think perhaps Bazel.

But Bazel, you can figure out other ways to trick it.

But they did point out that this was not an overnight success.

They spent the last six years moving to monorepos and using Bazel as their build system.

That foundational work is absolutely crucial.

You know, you cannot just drop an AI agent into a messy, fragmented code base and expect it to perform well.

By moving to monorepos and adopting Bazel, they created a highly structured, predictable environment.

That solid foundation is what allows them to accelerate so quickly now.

It gives the agents a reliable map of the code.

I think it really is the fact that Nova is plugging into a lot of Dropbox stuff.

Yeah, and I would add on to that, like, you've got the side that's like, there's a complex like Bazel build infrastructure, right? That's using things like remote caching and stuff.

And so like, if I were a startup, I would absolutely be using one of these cloud hosted tools, right? It doesn't really make sense to build your own.

But in our case, where we have a very specialized build process, and also the idea of like, It's going to be easier for us to get buy-in from security and privacy teams if we're calling things internally and data is not leaving our

Um, for the standard library, there's obviously a whole bunch of different other libraries.

There's other build tool chains, like, obviously, like Go rules from Bazel, and, like, I think, uh, Pants or Buck or one of the other ones also has their own rules system for how to compile Go.

So, like, the ecosystem is pretty distributed, so the main thing the Go team controls is the language specification, and then the default tool, toolchain that we're all using.

But that's just the default.

They didn't just fix it, they cut the median build time down to under five minutes.

The key was moving to Bazel and rolling out remote caching, which lets engineers share build outputs instead of rebuilding everything from scratch.

So the first person builds it, and everyone else just downloads the result? That sounds almost too good to be true.

It is in practice, but it took a massive cultural shift.

So they went all in on monorepo.

That's a big bet.

What made Bazel the right call?

Bazel gives you hermetic builds and fine-grained caching out of the box.

Sixty-two.

Gradle, Bazel, and gRPC: A song of ice and fire.

Really sorry for the cheesy title.

No Game of Thrones jokes below.