kueueleuleu

kueueleuleu

Run containers sequentially inside Kubernetes Pods, Jobs, CronJobs.

The name kueueleuleu (\kø lø.lø\) comes from the French queue leu-leu, meaning "one after the other" or "in a single file". The initial k is obviously for kubernetes.

Motivation

As of now, Kubernetes does not support running containers inside a Pod sequentially, despite this being a common thing asked by users:

• How to run containers sequentially as a Kubernetes job?
• How to run multiple Kubernetes jobs in sequence?
• Kubernetes — Sequencing Container Startup, by Paul Dally
• How to start one container before staring another in the same pod?

Without an external way to control the containers run order, all containers inside a Pod start and run at the same time.

Existing known solutions

Kueueleuleu!

Unlike the known solutions mentioned above, this simple library aims to:

• hide the sequential containers orchestration logic from the user
• let users manipulate raw kubernetes objects (Pods, Jobs, CronJobs), so they can use everything natively supported by Kubernetes (volumes, runtime classes, etc.)
• provide an easy way to convert any of these kubernetes objects to kueueleuleu objects

How to use?

This library requires Go >= 1.21.

Using the CLI

go install github.com/norbjd/kueueleuleu/cmd/kueueleuleu@main

# create a simple pod with two containers: by default, both will run at the same time
cat > simplepod.yaml <<'EOF'
apiVersion: v1
kind: Pod
metadata:
  name: two-steps-pod
spec:
  containers:
    - name: step1
      image: alpine
      command:
        - sh
        - -c
      args:
        - echo "start step1" && sleep 5 && echo "end step1"
    - name: step2
      image: alpine
      command:
        - sh
        - -c
      args:
        - echo "start step2" && sleep 2 && echo "end step2"
  restartPolicy: Never
EOF

# convert this pod to run containers sequentially and apply
# if you like chaining commands, this is similar to: cat simplepod.yaml | ./kueueleuleu -f - | kubectl apply -f -
# if kueueleuleu is not in your PATH, replace with $GOPATH/bin/kueueleuleu (default when running go install)
kueueleuleu -f simplepod.yaml | kubectl apply -f -

Once the pod is finished, check the logs (kubectl logs two-steps-pod --all-containers --timestamps | sort) to see containers have been executed sequentially (first, step1, and then step2):

2023-12-29T13:32:50.488543841Z 2023/12/29 13:32:50 Entrypoint initialization
2023-12-29T13:32:52.139027305Z start step1
2023-12-29T13:32:57.139823758Z end step1
2023-12-29T13:32:57.392700813Z start step2
2023-12-29T13:32:59.393583491Z end step2

[!NOTE] The first log is related to internal logic: the entrypoint mentioned is here to order the containers' execution. You can ignore this, unless you are interested in the internals (see "Internals" section).

Without kueueleuleu (kubectl apply -f simplepod.yaml), containers logs are intertwined because both containers are running at the same time:

2023-12-29T13:33:43.942542903Z start step1
2023-12-29T13:33:44.848147388Z start step2
2023-12-29T13:33:46.848812173Z end step2
2023-12-29T13:33:48.943165982Z end step1

Conversion also work with Jobs and CronJobs, and even with YAML files containing multiple resources (see cmd/testdata/*_input.yaml for examples, and cmd/testdata/*_output.yaml for results after using kueueleuleu).

Using the library

First, run go get github.com/norbjd/kueueleuleu@main to download the dependency.

package main

import (
    "context"
    "fmt"
    "os"

    "github.com/norbjd/kueueleuleu"
    corev1 "k8s.io/api/core/v1"
    metav1 "k8s.io/apimachinery/pkg/apis/meta/v1"
    "k8s.io/client-go/kubernetes"
    "k8s.io/client-go/tools/clientcmd"
)

func main() {
    // original pod: containers will run at the same time
    twoStepsPod := corev1.Pod{
        ObjectMeta: metav1.ObjectMeta{
            Name: "two-steps-pod",
        },
        Spec: corev1.PodSpec{
            Containers: []corev1.Container{
                {
                    Name:    "step1",
                    Image:   "alpine",
                    Command: []string{"sh", "-c"},
                    Args:    []string{`echo "start step1" && sleep 5 && echo "end step1"`},
                },
                {
                    Name:    "step2",
                    Image:   "alpine",
                    Command: []string{"sh", "-c"},
                    Args:    []string{`echo "start step2" && sleep 2 && echo "end step2"`},
                },
            },
            RestartPolicy: "Never",
        },
    }

    // convert the pod to run containers sequentially
    twoSequentialStepsPod, err := kueueleuleu.ConvertPod(twoStepsPod)
    if err != nil {
        panic(err)
    }

    // create the pod
    config, _ := clientcmd.BuildConfigFromFlags("", os.Getenv("KUBECONFIG"))
    kubeClient, _ := kubernetes.NewForConfig(config)

    createdPod, err := kubeClient.CoreV1().Pods("default").Create(
        context.Background(), &twoSequentialStepsPod, metav1.CreateOptions{},
    )
    if err != nil {
        panic(err)
    }

    fmt.Printf("Created: %s\n", createdPod.Name)
}

The main advantage of kueueleuleu is that you can continue to manipulate standard kubernetes resources (like Pods) in your code. There is only a single helper if you want to know which container inside the pod is really running (as now they are running sequentially):

currentlyRunningContainerName, _ := kueueleuleu.GetRunningContainerName(*createdPod)

fmt.Printf("Running container: %s\n", currentlyRunningContainerName)

As for the CLI, the conversion also work with Jobs and CronJobs: just use kueueleuleu.ConvertJob or kueueleuleu.ConvertCronJob.

Internals

Under the hood, containers sequential orchestration is managed using Tekton entrypoint. I have just "reverse-engineered" the way Tekton generates Pods from PipelineRun. But, unlike Tekton, there is no need to install a separate controller in the cluster or using CRDs, which makes kueueleuleu lighter to use. In return, kueueleuleu cannot be used for complex workflows, and I don't consider supporting these: its only job is to run containers sequentially.

I have used Tekton entrypoint because it is already doing the job, and I didn't want to rewrite another wrapper to do more or less the same thing.

Another solution I have considered had been to convert containers to init containers (see "Use init containers" section above). But, considering init containers limitations, I thought this solution was not viable.

Limitations

Containers passed in objects (Pod, Job, CronJob) MUST have their command field set. Otherwise, kueueleuleu will return an error (container does not have a command). This is because Tekton entrypoint used under the hood requires a command. Tekton Pipelines (who also uses this entrypoint) manages to "guess" the entrypoint if a command is not provided (entrypoint hack refered here). But, I'd rather not implement this here today as it does not always work (e.g. if image is not pushed into a registry, like in kind environments, we will fail to guess the entrypoint).