Big News!  Tilt is joining Docker

Integrating Bazel with Tilt

Bazel provides a deterministic build process that, coupled with rules_docker maximizes container layer caching and can even deploy to Kubernetes with rules_k8s. In this guide we’ll walk through how to use local and custom_build to integrate such a Bazel setup with your Tiltfile.

Note: Before diving in to this guide you should have run through Tilt’s Getting Started Guide and the deep dive in to Tiltfile concepts. A familiarity with the Bazel build system will also clear a lot of things up but isn’t necessary. Ready? Let’s get started!

Building Microservices with Bazel

Here’s an example repository that contains two services (snack and vigoda) that are built with Bazel. The current development workflow is to make some changes to the code (contained in the snack or vigoda directories) then run the appopriate bazel rule to put the service in to my Kubernetes cluster. For example if I made a change to snack/main.go I would then run bazel run //:snack-server.apply. This command will build my Go code, put it in to an image, put that image in to my YAML, and run kubectl apply to update my Kubernetes cluster with the new YAML. Bazel handles it all by understanding the transitive dependencies of snack-server.

Writing the Tiltfile

Tiltfiles only need two things: Kubernetes YAML and Docker Images. To get started we just need to figure out how to ask Bazel for those two things.

Getting YAML

So when we run bazel run //:snack-server.apply what does Bazel actually do? Let’s look at the :snack-server rule definition.

  name = "snack-server",
  kind = "deployment",

  template = ":deploy/snack.yaml",

  cluster = "docker-for-desktop-cluster",

If I want to get just the YAML, a quick glance at the rules_k8s documentation tells me that can be done by just running the target without the .apply.

$ > bazel run //:snack-server
apiVersion: apps/v1
kind: Deployment
  labels: {app: snack, owner: varowner}
  name: varowner-snack
    matchLabels: {app: snack, owner: varowner}
      labels: {app: snack, owner: varowner, tier: web}
      - image: bazel/snack@sha256:5a7e2fd3bd39a841cad0f009d605e7f5337a50d2cf641c50ebc03148d8257a0b
        name: snack
        - {containerPort: 8083}
          requests: {cpu: 10m}

Now we have enough to begin writing our Tiltfile. We’ll use local to call Bazel and k8s_yaml to tell Tilt about the YAML that Bazel generates. Here’s a function that given a Bazel target name returns the YAML for that target:

def bazel_k8s(target):
  return local("bazel run %s" % target)

In the main body of our Tiltfile we can call it like so:


And if we tilt up we should see that Tilt calls Bazel to build the YAML and then Tilt takes that YAML and deploys it to the cluster. If the image exists in the registry Kubernetes should schedule the resource

Building Images

Let’s follow the same strategy and figure out the Bazel command that will just build an image and then integrate that in to the Tiltfile.

“rules_k8s” uses “rules_docker” under the hood for building images.

To build an image and load it into Docker, the rules_docker documentation) tells us to run:

bazel run //path/to/my-command:image -- --norun

Bazel will put the image at bazel/path/to/my-command.

That’s exactly what custom_build needs. Custom build takes an image name and a command that should produce that image. After the command is executed Tilt asserts that it exists, and retags it so it can track it.

  'bazel run --platforms=@io_bazel_rules_go//go/toolchain:linux_amd64 //snack:image -- --norun',

The first argument is the name of the image. Thanks to Bazel’s deterministic nature we know in advance what the image name will be. Bazel uses a formula for determining the name of the image that it pushes:

"bazel/" + directory that the build file is in + ":" + the name of the rule

Hence for us: “bazel/snack:image”.

The second argument uses bazel to create the image. The --platforms flag is a Bazel-specific gotcha: it tells Bazel to build Go binaries for Linux, because we’re using a Linux container. You may need use something like this specific to your programming language.

The third argument is a list of dependencies. It tells Tilt which files to watch for changes.

When you’re trying Tilt for the first time, it’s ok to leave it blank. We’ll fill it in later with the full list of sources to watch.

For now let’s wrap this in a function called bazel_build and add it to the main section of our Tiltfile:

def bazel_build(image, target):
    'bazel run --platforms=@io_bazel_rules_go//go/toolchain:linux_amd64 %s -- --norun' % target,

bazel_build('bazel/snack', '//snack:image')

Now Tilt can get YAML from Bazel, get an image from Bazel, insert the image into the YAML and track the resource’s status in the Kubernetes cluster.

Dealing with Dependencies

If you started Tilt now, it would appear to work! Run tilt up and all the images get deployed to your cluster.

Then you make a change to snack/main.go. Nothing gets updated. Why not?

It’s not enough for Tilt to know how to build & deploy an image. It needs to know what files should trigger a new build and deploy.

Luckily for us Bazel makes it easy to ask for those files and Tilt makes it easy to tell it about those files.

bazel query is the functionality that we’ll use to ask Bazel about the various dependencies of certain targets. Before we do that there are two types of dependencies that Tilt cares about: dependencies for the Tiltfile and dependencies for images.

Dependencies for the Tiltfile are files that you read during Tiltfile execution. For example, if you had a JSON file that you read to tell you which services Tilt should start that JSON file would be marked as a dependency of the Tiltfile. Tilt will watch that file and, if it changes, re-execute the Tiltfile.

Dependencies for images are files that are included in a build context or running container. If they change Tilt doesn’t necessarily re-execute the Tiltfile but it will update a container.

YAML Dependenices

For the YAML case the only question we need to ask is “Under what circumstances should the Tiltfile re-execute to get the new YAML”? To fully answer this question from Bazel we need to run two queries: one for Bazel’s build dependencies, and one for the source dependencies of the YAML target:

# build dependencies: what dependencies does Bazel need to execute this target?
  bazel query 'filter("^//", kind("source file", deps(set(%s))))' --order_output=no

# source dependencies: What are the target's dependencies?
  bazel query 'filter("^//", buildfiles(deps(set(%s))))' --order_output=no

For :snack-server we get 1 build dependency: BUILD

Okay that makes sense: if the Bazel build files themselves change we will want to re-execute the Tiltfile so that it can re-execute the Bazel rules and potentially get new YAML. There’s also a source file dependency: deploy/snack.yaml. As the actual YAML itself that makes total sense too.

Let’s wire up these query calls in to bazel_k8s:

def watch_labels(labels):
  watched_files = []
  for l in labels:
    if l.startswith("@"):
    elif l.startswith("//external/") or l.startswith("//external:"):
    elif l.startswith("//"):
      l = l[2:]

    path = l.replace(":", "/")
    if path.startswith("/"):
      path = path[1:]


  return watched_files

def bazel_k8s(target):
  build_deps = str(local(BAZEL_BUILDFILES_CMD % target)).splitlines()
  source_deps = str(local(BAZEL_SOURCES_CMD % target)).splitlines()

  return local("bazel run %s" % target)

Now if we change a dependency of the YAML, the Tiltfile re-executes.

Image Dependencies

We can reuse our build and source file queries for the image target //snack:image. There’s one build dep (snack/BUILD) and one source dep (snack/main.go). That looks right, now to use these queries in our bazel_build function:

def bazel_build(image, target):
  build_deps = str(local(BAZEL_BUILDFILES_CMD % target)).splitlines()

  source_deps = str(local(BAZEL_SOURCES_CMD % target)).splitlines()
  source_deps_files = bazel_labels_to_files(source_deps)

    BAZEL_RUN_CMD % target,

(bazel_labels_to_files is some boring code that parses Bazel output that is omitted here but is included in the full example linked to at the end of this guide)

Now we’re passing the dependencies to our custom_build. If any of those paths change we will build a new image and update the YAML with the new image. Same as above: if any of the dependencies of the image rule itself change Tilt will now re-execute the Tiltfile.

Now we have a Tiltfile that fully integrates with Bazel and responds to changes from the filesystem.

Putting it all together

And that’s that! If you want to poke more at how this works, check out: