| title | authors | creation-date | last-updated | status | ||
|---|---|---|---|---|---|---|
Task Results in Final Tasks |
|
2020-07-16 |
2021-06-03 |
implemented |
A Task in a Pipeline can emit Results which can be consumed by any other Tasks within the same
Pipeline. Result of a Task in a Pipeline is received through the input parameter using variable
substitution such as $(tasks.<pipeline-task-name>.results.<result-name>).
Today, final Tasks can not receive Results of any other Tasks in a Pipeline. This feature is
limited to non-finally Tasks of a Pipeline.
This proposal is enabling final Tasks to receive Results of any Tasks within the same Pipeline.
A Pipeline author wants to clean up resources that is done by the final Tasks in the Pipeline.
The resources are acquired by the setup Task which has the names of resources. The clean up Tasks
must clean up only those resources.
Also, cleanup must be done only if the setup was successful but regardless of the failure of the rest of
the Tasks in the Pipeline.
setup
|
| e.g. Project Name
v
cleanup
One goal of this proposal is to allow final Tasks to receive Results from non-finally Tasks.
Second goal of this proposal is to determine what happens to the final Task when referenced
task results are not initialized because either dependent task failed or dependent task skipped
due to its own condition or if its parent was skipped.
This proposal is not enabling final Task to receive Results from other final Tasks.
This proposal is also not enabling final Tasks to produce Task Results.
Allow a finally Task to get a value from a non-finally Task via a Task Result.
If the Task producing the value fails, the finally Task that requires that value will be included in the list of
pipelineStatus.skippedTasks. That final Task will be attempted i.e. evaluated by the Pipeline Controller for
execution, but the validation will fail before the TaskRun is created since the required Result is not available.
A Pipeline author can send the name of the project to finally task after non-finally task creates/acquires
resources for that project. The finally Task can then clean up those project resources regardless of the
failure of the rest of the Tasks in the Pipeline.
Tasks can produce one or more execution Results under /tekton/results/ by specifying results
in the Task specification:
spec:
results:
- name: leased-resource
description: The name of the leased resource
steps:This task produces a result file at /tekton/results/leased-resource.
Any other Pipeline Tasks can use these Results as parameter values through variable substitution:
spec:
tasks:
- name: boskos-release
params:
- name: leased-resource
value: $(tasks.boskos-acquire.results.leased-resource)No need to specify explicit runAfter here, Tekton infers the execution order by setting boskos-acquire
as a dependency so that the task emitting the referenced results executes before the task that consumes them.
Similar to non-final tasks, final task use Results through variable substitution:
spec:
tasks:
- name: boskos-acquire
taskRef:
name: boskos-acquire
- name: do-stuff-with-resource
taskRef:
name: do-stuff-with-resource
params:
- name: resource-name
value: $(tasks.boskos-acquire.results.leased-resource)
finally:
- name: boskos-release
taskRef:
name: boskos-release
params:
- name: leased-resource
value: $(tasks.boskos-acquire.results.leased-resource)Note: Today final tasks are executed all in parallel and does not depend on any other Task in the Pipeline.
With adding support for Task Results in finally Tasks, we are introducing implicit dependencies on non-final
Tasks but that does not change scheduling of finally tasks. All final tasks will still be executed in parallel after
all non-final tasks are done.
Q. What happens to Pipeline when the dependent task boskos-acquire either failed or not executed, and
the task result leased-resource is not initialized?
A. The finally Task boskos-release is attempted and included in the list of skippedTasks since the task result
leased-resource is not initialized. Pipeline controller logs this validation failure including param name
leased-resource of the finally task boskos-release with the result reference leased-resource and result producing
task boskos-acquire. Pipeline continues executing rest of the final tasks and exits with completion.
Q. What happens when the dependent Task succeeds but the Task Result is empty?
A. It could be possible that the task author intentionally chooses to leave the Task Result empty
(creating an empty file at /tekton/results/). Pipeline leaves that param value empty and executes
the final task. Empty Task Result is considered a valid value and results in successful variable
substitution. The error handling of such empty Task Result is left to the final Task author.
Q. Can a final Task produce Task Results?
A. Nope, this proposal does not enable final tasks to produce Task Results. This also implies that the final
Task can not refer to the Task Result of other final Task.
Q. Explain the overall workflow.
A.
-----------------------------------------------
| Validate Params to final Tasks |
| (1) required params provided and their types) |
| (2) resolve param mapping and apply values |
| from PipelineRun/Pipeline |
-----------------------------------------------
|
V
---------------------------------------------------------
| Retrieve Execution Queue (All Final Tasks) |
| including failed final tasks without exhausting retires |
---------------------------------------------------------
|
V
---------------------------------------------------------
| Validate Task Results of Dependencies (non-final Tasks) |
| Apply Task Results to final Task Params |
---------------------------------------------------------
| (if the validation fails, add that final task in the list of skippedTasks
V but continue executing rest of the final tasks)
------------------------------------
| Execute the Task and Create TaskRun |
------------------------------------
-
No API changes.
-
Finally contract does not break and all final tasks are attempted.
-
Consistent across non-final and final tasks.
-
Finally tasks do not need to explicitly check if the Task they depend on ran or not
-
This proposal is extensible with future param validation by reading default from
Taskspecification, for example, ifboskos-acquirefails,boskos-releasewill still be executed with the default valuebest-default-everforleased-resourceparam.
# Pipeline
spec:
params:
- name: owner-name
default: "the-best-owner"
tasks:
- name: boskos-acquire
taskRef:
name: boskos-acquire
params:
- name: owner-name
value: $(params.owner-name)
- name: do-stuff-with-resource
taskRef:
name: do-stuff-with-resource
params:
- name: resource-name
value: $(tasks.boskos-acquire.results.leased-resource)
finally:
- name: boskos-release
taskRef:
name: boskos-release
params:
- name: leased-resource
value: $(tasks.boskos-acquire.results.leased-resource)
---
# Task
name: boskos-release
spec:
params:
- name: resource-name
default: "best-default-ever"
PipelineRunreads parameter defaults fromPipelinespecification but in case of parameters associated withTaskResults, the default values will be used from theTasksincePipelinehas no direct mapping betweenPipelineparameter andTaskparameter.
- e2e and unit tests
-
One could argue that this proposal breaks the finally contract, no longer means "finally tasks always run", it now means "finally tasks always run unless they depend on something and that fails". But
PipelineRundoes attempt such finally task which we explicitly skip with the validation failure so it is consideredranby thePipelineRunController. -
When a providing task fails, the finally task referring to that task result is evaluated but not executed with this proposal. If you still want to continue executing without any result, or a default value, refer to additional proposal on how to specify defauls for task results.
- Allow dependency to specify explicit default under
results
results:
- name: leased-resource
description: The name of the leased resource
default: "best-default-ever"Pros/Cons:
-
The default value will likely be invalid; a Task consuming this value as a param would have to understand what an invalid value looks like and prepare for that case.
-
Same default applies to all consumers of that result.
-
Inconsistent when the same task is invoked as non-final and final task. This default only applies to final
TaskRunsand non-finalTasksare not allowed to have this kind of default.
- Allow
Tasksto express validation on params; use this to identify when a default has been supplied
In this option, we allow Tasks to specify default values for results. However we also let Tasks specify
validation on their params, and finally Tasks can use this to verify they got value arguments and so
should run. Use a syntax similar to the schema syntax proposed in #1393 to express validation.
- Allow default for Task Result in PipelineTask
PipelineTask release-env can specify default to a task result proj-name of referenced task. In absence of any default, task result is considered as required and the final task results in a validation failure if proj-name is not initialized.
kind: Pipeline
spec:
finally:
- name: release-env
taskRef:
Name: boskos-release
params:
- name: proj-name
value: $(tasks.get-env.results.project)
default: best-default-ever- Allow finally tasks to depend on other Tasks; do not run if they fail