ko
is a simple, fast container image builder for Go applications.
It's ideal for use cases where your image contains a single Go application without any/many dependencies on the OS base image (e.g., no cgo, no OS package dependencies).
ko
builds images by effectively executing go build
on your local machine,
and as such doesn't require docker
to be installed. This can make it a good
fit for lightweight CI/CD use cases.
ko
also includes support for simple YAML templating which makes it a powerful
tool for Kubernetes applications (See below).
Install from Releases
VERSION=TODO # choose the latest version
OS=Linux # or Darwin
ARCH=x86_64 # or arm64, i386, s390x
curl -L https://github.com/grafana/ko/releases/download/v${VERSION}/ko_${VERSION}_${OS}_${ARCH}.tar.gz | tar xzf - ko
chmod +x ./ko
Install using Homebrew
brew install ko
With Go 1.16+, build and install the latest released version:
go install github.com/grafana/ko@latest
You can use the setup-ko
action
to install ko
and setup auth to GitHub Container
Registry in a GitHub Action workflow:
steps:
- uses: imjasonh/[email protected]
ko
depends on the authentication configured in your Docker config (typically
~/.docker/config.json
). If you can push an image with docker push
, you are
already authenticated for ko
.
Since ko
doesn't require docker
, ko login
also provides a surface for
logging in to a container image registry with a username and password, similar
to
docker login
.
Additionally, if auth is not configured in the Docker config, ko
includes
built-in support for authenticating to the following container registries using
credentials configured in the environment:
- Google Container Registry and Artifact Registry
- using Application Default Credentials or auth configured in
gcloud
.
- using Application Default Credentials or auth configured in
- Amazon Elastic Container Registry
- using AWS credentials
- Azure Container Registry
- using environment variables
- GitHub Container Registry
- using the
GITHUB_TOKEN
environment variable
- using the
ko
depends on an environment variable, KO_DOCKER_REPO
, to identify where it
should push images that it builds. Typically this will be a remote registry,
e.g.:
KO_DOCKER_REPO=gcr.io/my-project
, orKO_DOCKER_REPO=my-dockerhub-user
ko publish ./cmd/app
builds and pushes a container image, and prints the
resulting image digest to stdout.
In this example, ./cmd/app
must be a package main
that defines func main()
.
ko publish ./cmd/app
...
gcr.io/my-project/app-099ba5bcefdead87f92606265fb99ac0@sha256:6e398316742b7aa4a93161dce4a23bc5c545700b862b43347b941000b112ec3e
The executable binary that was built from ./cmd/app
is available in the image at /ko-app/app
-- the binary name matches the base import path name -- and that binary is the image's entrypoint.
Because the output of ko publish
is an image reference, you can easily pass it
to other tools that expect to take an image reference.
To run the container locally:
docker run -p 8080:8080 $(ko publish ./cmd/app)
Or to deploy it to other services like Cloud Run:
gcloud run deploy --image=$(ko publish ./cmd/app)
- Note: The image must be pushed to Google Container Registry or Artifact Registry.
Or fly.io:
flyctl launch --image=$(ko publish ./cmd/app)
- Note: The image must be publicly available.
Or AWS Lambda:
aws lambda update-function-code \
--function-name=my-function-name \
--image-uri=$(ko publish ./cmd/app)
- Note: The image must be pushed to ECR, based on the AWS provided base image, and use the
aws-lambda-go
framework. See official docs for more information.
az containerapp update \
--name my-container-app
--resource-group my-resource-group
--image $(ko publish ./cmd/app)
- Note: The image must be pushed to ACR or other registry service. See official docs for more information.
Aside from KO_DOCKER_REPO
, you can configure ko
's behavior using a
.ko.yaml
file. The location of this file can be overridden with
KO_CONFIG_PATH
.
By default, ko
bases images on gcr.io/distroless/static:nonroot
. This is a
small image that provides the bare necessities to run your Go binary.
You can override this base image in two ways:
- To override the base image for all images
ko
builds, add this line to your.ko.yaml
file:
defaultBaseImage: registry.example.com/base/image
- To override the base image for certain importpaths:
baseImageOverrides:
github.com/my-user/my-repo/cmd/app: registry.example.com/base/for/app
github.com/my-user/my-repo/cmd/foo: registry.example.com/base/for/foo
By default, ko
builds the binary with no additional build flags other than
-trimpath
. You can replace the default build
arguments by providing build flags and ldflags using a
GoReleaser influenced builds
configuration section in your .ko.yaml
.
builds:
- id: foo
dir: . # default is .
main: ./foobar/foo
env:
- GOPRIVATE=git.internal.example.com,source.developers.google.com
flags:
- -tags
- netgo
ldflags:
- -s -w
- -extldflags "-static"
- -X main.version={{.Env.VERSION}}
- id: bar
dir: ./bar
main: . # default is .
env:
- GOCACHE=/workspace/.gocache
ldflags:
- -s
- -w
If your repository contains multiple modules (multiple go.mod
files in
different directories), use the dir
field to specify the directory where
ko
should run go build
.
ko
picks the entry from builds
based on the import path you request. The
import path is matched against the result of joining dir
and main
.
The paths specified in dir
and main
are relative to the working directory
of the ko
process.
The ldflags
default value is []
.
Please note: Even though the configuration section is similar to the
GoReleaser builds
section,
only the env
, flags
and ldflags
fields are currently supported. Also, the
templating support is currently limited to using environment variables only.
ko
provides a few different strategies for naming the image it pushes, to
workaround certain registry limitations and user preferences:
Given KO_DOCKER_REPO=registry.example.com/repo
, by default,
ko publish ./cmd/app
will produce an image named like
registry.example.com/repo/app-<md5>
, which includes the MD5 hash of the full
import path, to avoid collisions.
--preserve-import-path
(-P
) will include the entire importpath:registry.example.com/repo/github.com/my-user/my-repo/cmd/app
--base-import-paths
(-B
) will omit the MD5 portion:registry.example.com/repo/app
--bare
will only include theKO_DOCKER_REPO
:registry.example.com/repo
ko
is normally used to publish images to container image registries,
identified by KO_DOCKER_REPO
.
ko
can also publish images to a local Docker daemon, if available, by setting
KO_DOCKER_REPO=ko.local
, or by passing the --local
(-L
) flag.
Locally-published images can be used as a base image for other ko
images:
defaultBaseImage: ko.local/example/base/image
ko
can also publish images to a local KinD
cluster, if available, by setting KO_DOCKER_REPO=kind.local
. By default this
publishes to the default KinD cluster name (kind
). To publish to another KinD
cluster, set KIND_CLUSTER_NAME=my-other-cluster
.
Because Go supports cross-compilation to other CPU architectures and operating
systems, ko
excels at producing multi-platform images.
To build and push an image for all platforms supported by the configured base
image, simply add --platform=all
. This will instruct ko
to look up all the
supported platforms in the base image, execute
GOOS=<os> GOARCH=<arch> GOARM=<variant> go build
for each platform, and
produce a manifest list containing an image for each platform.
You can also select specific platforms, for example,
--platform=linux/amd64,linux/arm64
ko
can also bundle static assets into the images it produces.
By convention, any contents of a directory named <importpath>/kodata/
will be
bundled into the image, and the path where it's available in the image will be
identified by the environment variable KO_DATA_PATH
.
As an example, you can bundle and serve static contents in your image:
cmd/
app/
main.go
kodata/
favicon.ico
index.html
Then, in your main.go
:
func main() {
http.Handle("/", http.FileServer(http.Dir(os.Getenv("KO_DATA_PATH"))))
log.Fatal(http.ListenAndServe(":8080", nil))
}
You can simulate ko
's behavior outside of the container image by setting the
KO_DATA_PATH
environment variable yourself:
KO_DATA_PATH=cmd/app/kodata/ go run ./cmd/app
Tip: Symlinks in kodata
are followed and included as well. For example,
you can include Git commit information in your image with:
ln -s -r .git/HEAD ./cmd/app/kodata/
Also note that http.FileServer
will not serve the Last-Modified
header
(or validate If-Modified-Since
request headers) because ko
does not embed
timestamps by default.
This can be supported by manually setting the KO_DATA_DATE_EPOCH
environment
variable during build (See below).
You could stop at just building and pushing images.
But, because building images is so easy with ko
, and because building with
ko
only requires a string importpath to identify the image, we can integrate
this with YAML generation to make Kubernetes use cases much simpler.
Traditionally, you might have a Kubernetes deployment, defined in a YAML file, that runs an image:
apiVersion: apps/v1
kind: Deployment
metadata:
name: my-deployment
spec:
replicas: 3
...
template:
spec:
containers:
- name: my-app
image: registry.example.com/my-app:v1.2.3
...which you apply to your cluster with kubectl apply
:
kubectl apply -f deployment.yaml
With ko
, you can instead reference your Go binary by its importpath, prefixed
with ko://
:
...
spec:
containers:
- name: my-app
image: ko://github.com/my-user/my-repo/cmd/app
With this small change, running ko resolve -f deployment.yaml
will instruct
ko
to:
- scan the YAML file(s) for values with the
ko://
prefix, - for each unique
ko://
-prefixed string, executeko publish <importpath>
to build and push an image, - replace
ko://
-prefixed string(s) in the input YAML with the fully-specified image reference of the built image(s), for example:
spec:
containers:
- name: my-app
image: registry.example.com/github.com/my-user/my-repo/cmd/app@sha256:deadb33f...
- Print the resulting resolved YAML to stdout.
The result can be redirected to a file, to distribute to others:
ko resolve -f config/ > release.yaml
Taken together, ko resolve
aims to make packaging, pushing, and referencing
container images an invisible implementation detail of your Kubernetes
deployment, and let you focus on writing code in Go.
To apply the resulting resolved YAML config, you can redirect the output of
ko resolve
to kubectl apply
:
ko resolve -f config/ | kubectl apply -f -
Since this is a relatively common use case, the same functionality is available
using ko apply
:
ko apply -f config/
NB: This requires that kubectl
is available.
To teardown resources applied using ko apply
, you can run ko delete
:
ko delete -f config/
This is purely a convenient alias for kubectl delete
, and doesn't perform any
builds, or delete any previously built images.
Using -ldflags
is a common way to embed version info in go binaries (In fact, we do this for
ko
!). Unfortunately, because ko
wraps go build
, it's not possible to use
this flag directly; however, you can use the GOFLAGS
environment variable
instead:
GOFLAGS="-ldflags=-X=main.version=1.2.3" ko publish .
Currently, there is a limitation that does not allow to set multiple arguments
in ldflags
using GOFLAGS
. Using -ldflags
multiple times also does not
work. In this use case, it works best to use the builds
section
in the .ko.yaml
file.
In order to support reproducible builds, ko
doesn't embed timestamps in the images it produces by default.
However, ko
does respect the SOURCE_DATE_EPOCH
environment variable, which will set the container image's timestamp
accordingly.
Similarly, the KO_DATA_DATE_EPOCH
environment variable can be used to set
the modtime timestamp of the files in KO_DATA_PATH
.
For example, you can set the container image's timestamp to the current timestamp by executing:
export SOURCE_DATE_EPOCH=$(date +%s)
or set the timestamp of the files in KO_DATA_PATH
to the latest git commit's
timestamp with:
export KO_DATA_DATE_EPOCH=$(git log -1 --format='%ct')
Yes, but support for Windows containers is new, experimental, and tenuous. Be prepared to file bugs. 🐛
The default base image does not provide a Windows image.
You can try out building a Windows container image by setting the base image to a Windows base image and building with --platform=windows/amd64
or --platform=all
:
For example, to build a Windows container image for ko
, from within this repo:
ko publish ./ --platform=windows/amd64
This works because the ko
image is configured in .ko.yaml
to be based on a golang
base image, which provides platform-specific images for both Linux and Windows.
- Symlinks in
kodata
are ignored when building Windows images; only regular files and directories will be included in the Windows image.
Can I optimize images for eStargz support?
Yes! Set the environment variable GGCR_EXPERIMENT_ESTARGZ=1
to produce
eStargz-optimized images.
Yes! ko completion
generates a Bash/Zsh/Fish/PowerShell completion script.
You can get how to load it from help document.
ko completion [bash|zsh|fish|powershell] --help
Or, you can source it directly:
source <(ko completion)
Does ko
work with Kustomize?
Yes! ko resolve -f -
will read and process input from stdin, so you can have
ko
easily process the output of the kustomize
command.
kustomize build config | ko resolve -f -
Oh, you betcha. Here's a partial list:
ko
support in Carvel'skbld
ko
support in Skaffoldko
extension for Tiltko
support for goreleaser (proposed)
Does ko
work with OpenShift Internal Registry?
Yes! Follow these steps:
- Connect to your OpenShift installation: https://docs.openshift.com/container-platform/latest/cli_reference/openshift_cli/getting-started-cli.html#cli-logging-in_cli-developer-commands
- Expose the OpenShift Internal Registry so you can push to it: https://docs.openshift.com/container-platform/latest/registry/securing-exposing-registry.html
- Export your token to
$HOME/.docker/config.json
:
oc registry login --to=$HOME/.docker/config.json
- Create a namespace where you will push your images, i.e:
ko-images
- Execute this command to set
KO_DOCKER_REPO
to publish images to the internal registry.
export KO_DOCKER_REPO=$(oc registry info --public)/ko-images
This work is based heavily on experience from having built the Docker and Kubernetes support for Bazel. That work was presented here.
Questions? Comments? Ideas?
Come discuss ko
with us in the #ko-project
channel on the Kubernetes Slack!
See you there!