CircleCI Go Report Card GitHub Release GoDoc

Isopod is an expressive DSL framework for Kubernetes configuration. Without intermediate YAML artifacts, Isopod renders Kubernetes objects as Protocol Buffers, so they are strongly typed and consumed directly by the Kubernetes API.

With Isopod, configurations are scripted in Starlark, a Python dialect by Google also used by Bazel and Buck build systems. Isopod offers runtime built-ins to access services and utilities such as Vault secret management, Kubernetes apiserver, HTTP requester, Base64 encoder, and UUID generator, etc. Isopod uses separate runtime for unit tests to mock all built-ins, providing the test coverage not possible before.

A 5-min read, this medium post explains the inefficiency of existing YAML templating tools when dealing with values not statically known and complicated control logics such as loops and branches. It also gives simple code examples to show why Isopod is an expressive, hermetic, and extensible solution to configuration management in Kubernetes.


$ go version
go version go1.11 darwin/amd64
$ GO111MODULE=on go build

Main Entryfile

Isopod will call the clusters(ctx) function in the main Starlark file to get a list of target clusters. For each of such clusters, isopod will call addons(ctx) to get a list of addons for configuration rollout.


    onprem(env="dev", cluster="minikube"),

def clusters(ctx):
    if ctx.cluster != None:
        return [c for c in CLUSTERS if c.cluster == ctx.cluster]
    elif ctx.env != None:
        return [c for c in CLUSTERS if c.env == ctx.env]
    return CLUSTERS

def addons(ctx)
    return [
        addon("ingress", "configs/ingress.ipd", ctx),


The ctx argument to clusters(ctx) comes from the command line flag --context to Isopod. This flag takes a comma-separated list of foo=bar and makes these values available in Starlark as (which gives "bar"). Currently Isopod supports the following clusters, and could easily be extended to cover other Kubernetes vendors, such as EKS and AKS.


Represents a Google Kubernetes Engine. Authenticates using Google Cloud Service Account Credentials or Google Default Application Credentials. Requires the cluster, location, and project fields. Additional fields are allowed.


Represents an on-premise or self-managed Kubernetes cluster. Authenticates using the kubeconfig file. No fields are required.


The ctx argument to addons(ctx) contains all fields of the chosen cluster. For example, say the cluster is


Then, each addon may access the cluster information as ctx.env to get "prod" and ctx.location to get "us-west1". Accessing nonexistant attribute will get None.

Each addon is represented using the addon() Starlark built-in, which takes three arguments, for example addon("name", "entry_file.ipd", ctx). The first argument is the addon name, used by the --match_addon feature. The the thrid is optional and represents the ctx input to addons(ctx) to make the cluster attributes available to the addon. Each addon must implement install(ctx) and remove(ctx) functions.

More advanced examples can be found in the examples folder.

Example Nginx addon:

appsv1 = proto.package("")
corev1 = proto.package("")
metav1 = proto.package("")

def install(ctx):
    metadata = metav1.ObjectMeta(
        labels={"app": "nginx"},

    nginxContainer = corev1.Container(,

    deploySpec = appsv1.DeploymentSpec(



Built-ins are pre-declared packages available in Isopod runtime. Typically they perform I/O to Kubernetes, Vault, GCP and other resources but could be used for break-outs into other operations not supported by the main Starlark interpreter.

Currently these build-ins are supported:


Built-in for managing Kubernetes objects.



Updates (creates if it doesn't already exist) object in Kubernetes.

    name = "nginx-role",
    namespace = "nginx-ingress",
    # Optional Kubernetes API Group parameter. If not set, will attempt to
    # deduce the group from message type but since Kubernetes API Group names
    # are highly irregular, this may fail.
    api_group = '',
    data = [

Supported args:

  • name - Name ( of the resource
  • namespace (Optional) - Namespace (.metadata.namespace) of the resource
  • api_group (Optional) - API group (without version) of the resource. If not provided, Isopod runtime will attempt to deduce the resource from just Proto type name which is unreliable. It is recommended to set this for all objects outside of core group.
  • subresource (Optional) - A subresource specifier (e.g /status).
  • data - A list of Protobuf definitions of objects to be created.


Deletes object in Kubernetes.

# kwarg key is resource name, value is <namespace>/<name> (just <name> for
# non-namespaced resources).
# api_group can optionally be provided to remove ambuguity (if multiple
# resources by the same name exist in different API Groups).
kube.delete(clusterrole="nginx", api_group = "")


Same as put but for YAML/JSON data. To be used for CRDs and other custom types. kube.put usage is preferred for the standard set of Kubernetes types.

ark_config = """
kind: Config
  namespace: ark-backup
  name: default
  name: gcp
  bucket: test-ark-backup
  name: gcp

    name = "ark-config",
    namespace = "backup",
    data = [ark_config])

# Alternatively render from native Starlark struct object via JSON:
ark_config = struct(
    apiVersion = "",
    kind = "Config",
    metadata = struct(
        name = "ark-backup",
        namespace = "default",
    backupStorageProvider = struct(
        name = "gcp",
        bucket = "test-ark-backup",
    persistentVolumeProvider = struct(
        name = "gcp",

    name = "ark-config",
    namespace = "backup",
    data = [ark_config.to_json()])


Reads object from API Server. If wait argument is set to duration (e.g 10s) will block until the object is successfully read or timer expires. If json=True optional argument is provided, will render object as unstructured JSON represented as Starlark dict at top level. This is useful for CRDs as they typically do not support Protobuf representation.

# Wait 60s for Service Account token secret.
secret = kube.get(secret=namespace+"/"+serviceaccount.secrets[0].name, wait="60s")

# Get ClusterRbacSyncConfig CRD.
cadmin = kube.get(clusterrbacsyncconfig="cluster-admin",


Checks whether a resource exists. If wait argument is set to duration (e.g 10s) will block until the object is successfully read or timer expires.

# Assert that the resource doesn't exist.
e = kube.exists(secret=namespace+"/"+serviceaccount.secrets[0].name, wait="10s")
assert(e != True, "Fail: resource shouldn't exist")

kube.from_str, kube.from_int

Convert Starlark string and int types to corresponding *instr.IntOrString protos.

    maxUnavailable = kube.from_str("10%"),


Vault break-out allows reading/writing values from Enterprise Vault.


Reads data from Vault path as Starlark dict


Writes kwargs to Vault path


Checks if path exists in Vault

Example usage:

if not vault.exist("secret/lidar/stuff"):
    vault.write("secret/lidar/stuff", w1="hello", w2="world!")

data ="secret/infra/myapp")
print(data["w1"] + " " + data["w2"])


Helm built-in renders Helm charts and applies the resource manifest changes.



Applies resource changes.

globalValues = """
    priorityClassName: "cluster-critical"
pilotValues = """
    replicaCount: 3
    traceSampling: 50.0
pilotOverlayValues = {
    "pilot": {
        "traceSampling": 100.0,

    release_name = "istio-pilot",
    chart = "//charts/istio/istio-pilot",
    namespace = "istio-system",
    values = [

Supported args:

  • release_name - Release Name for the Helm chart.
  • chart - Source Path of the chart. This can be a full path or a path relative to the working directory. Having a leading double-slash (//) will make it relative path.
  • namespace (Optional) - Namespace (.metadata.namespace) of the resources
  • values (Optional) - A list of Starlark Values used as input values for the charts. The ordering of a list matters, and the elements get overridden by the trailing values.


Various other utilities are available as Starlark built-ins for convenience:

base64.{encode, decode}

Translate string values to/from base64

uuid.{v3, v4, v5}

Produce corresponding flavor of UUID values

http.{get, post, patch, put, delete}

Sends corresponding HTTP request to specified url. Returns response body as string, if present. Errors out on non-2XX response code. Will follow redirects (stops after 10 consecutive requests).


  • url - URL to send request to (required).
  • headers - optional header dict (values are either string for single-value headers or list for multiple-value headers).
  • data - optionally send data in the body of the request (takes string).

hash.{sha256, sha1, md5}

Returns an integer hash value. Useful applied to an env var for forcing a redeploy when a config or secret changes.


Pauses execution for specified duration (requires Go duration string).


Interrupts execution and return error to the user (requires string error message).


isopod test command allows addon creators to write hermetic unit tests on their addons.

Unit tests must be contained inside files with a _test.ipd suffix and Isopod runtime will call every top-level method defined in that file as a separate test, execute it and report the result.

Built-in modules that allow external access (like kube and vault) are stubbed (faked) out in unit test mode so that tests are hermetic.

Intended pattern is to import the addon config files from the test, then call their methods and test the results with assert built-in (only supported in test mode).

Example test:

# Load ingress addon config and expose its "install" method.
load("testdata/ingress.ipd", "install")

def test_install(t):
    # Test setup code.
    vault.write("secret/car/cert", crt="foobar")
    t.ctx.namespace = "foobar"

    # Call method we are testing (creates namespace from context).

    # Now extract data from our fake "kube" module and verify our tests
    # conditions.
    ns = kube.get(namespace="foobar")
    assert( == "foobar", "fail")
    assert(ns.metadata.labels["foo"] == "bar", "fail")

The test command is designed to mimic standard go test. As such you can execute all test in subtree by running isopod test path/..., all test in a directory by running isopod test path/ and all tests from a current working subtree by running just isopod test.

Dry run as YAML Diff

Knowledge regarding the intended actions of any specification change is crucial for migration and everyday configuration updates. It prevents accidental removal of the critical fields that is otherwise uncatchable with just the new set of configurations.

In dry run mode, Isopod not only verifies the legitimacy of the Starlark scripts but also informs the intended actions of the configuration change, by presenting the YAML diff between live objects in cluster and the generated configurations call "head". The result looks like the following.

*** service.v1 example/nginx ***
--- live
+++ head
@@ -14,8 +14,9 @@
     port: 80
     targetPort: 80
     app: nginx
-  type: ClusterIP
+  type: NodePort
   sessionAffinity: None
+  externalTrafficPolicy: Cluster


Copyright 2019 GM Cruise LLC

Licensed under the Apache License Version 2.0 (the "License"); you may not use this project except in compliance with the License.

Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License.


Contributions are welcome! Please see the agreement for contributions in

Commits must be made with a Sign-off (git commit -s) certifying that you agree to the provisions in

View Original