Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

RFC: across step #29

Open
wants to merge 11 commits into
base: master
Choose a base branch
from
Open

RFC: across step #29

wants to merge 11 commits into from

Conversation

vito
Copy link
Member

@vito vito commented May 21, 2019

Rendered

Related to, but not dependent on #24.

@vito vito changed the title RFC: spatial resources RFC: across step Dec 24, 2019
@vito vito marked this pull request as ready for review December 24, 2019 23:27
@siennathesane
Copy link

How would across work with a large build matrix?

I could see the value here of across having some functional limiter, like this:

across: [1.8, 1.9, 1.10, 1.11, 1.12, 1.13, 1.8-win, 1.9-win, 1.10-win, 1.11-win, 1.12-win, 1.13-win, 1.8-macos, 1.9-macos, 1.10-macos, 1.11-macos, 1.12-macos, 1.13-macos, etc.]
max_in_flight: 4
as: go-version
do:
- task: unit
  vars: {go_version: ((go_version))}

That way it limits the parallelism of the build matrix a bit without stealing too much from the global and job max_in_flight and prevent the pipeline from moving forward on other tasks in the same pipeline/job.

@JohannesRudolph
Copy link

Is it planned that across also adds a notion for "skipping" a job for certain resource combinations? Say we have a build matrix but one particular job may not apply to a certain combination of vars, would it be possible to skip running the job? Or should that rather be implemented inside the job itself (i.e. NOP-ing all tasks inside)? I'm asking because it has proven very difficult so far to implement pipeline control flow, see e.g. https://discuss.concourse-ci.org/t/trigger-jobs-conditionally/306/13

I know that pipeline control flow is a controversial topic and having "idempotent" pipelines that always do the same thing is a good design goal for Concourse. However there's already try and on_success and on_failure. However, they don't support all real world use cases, especially when it comes to optimising pipelines. I'd rather have concourse support some sort of declarative notion (like job.if in GitHub Actions) than abusing task scripts or other mechanisms.

...and resolve open questions

Signed-off-by: Alex Suraci <[email protected]>
@vito vito force-pushed the spatial-resources branch from 015de30 to cfec275 Compare July 21, 2020 20:40
instance_vars: {pr_number: ((.:pr.number))}
across:
- var: pr
source: booklit-prs
Copy link

@aoldershaw aoldershaw Aug 6, 2020

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Have you thought about allowing the across step to iterate over objects from an arbitrary prototype (i.e. not just constrained to var_source prototypes and their list message)? That could make the object streams returned by prototypes more useful outside of the special prototype interfaces that Concourse will interact with natively (resources, var_sources, notifications, etc.)

In this comment concourse/concourse#5936 (comment), you suggested that the run step could possibly support a set_var field to store the returned object as a local variable. Perhaps an alternative to set_var would be to use the across step in its place, e.g.

- across:
  - var: some-var
    run: eval
    type: gval
    params: {fields: {some-field: hello}}
  run: echo
  type: debug
  params: {message: ((.:some-var.some-field))}

Kind of weird in this case since we're iterating over a single value always - but I guess for set_var to work in general, it would also have to set the var to a list of objects (even if a single object is emitted)

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This also makes me realize that the across step can be used for conditional portions of build plans, by having the list of values be either 0 or 1 element long 🤔

Copy link

@aoldershaw aoldershaw Aug 6, 2020

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

It's also making me wonder about dynamic build plans in the context of the across step (i.e. the list of values to iterate over can be determined at runtime) - what are your thoughts on that? I think it could greatly extend the use cases that across can support, but it could possibly make pipelines difficult to understand since you could introduce arbitrarily complicated control flow

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I have, and your comments are pretty much the same thought process I went thought. 😄 It seems like a really cool idea, but yeah, the difficulty is that it seems to require some sort of dynamic build plan construct.

I'm running into another area that would benefit from dynamic build plans: the way file: and image_resource: interact on a task step. I'm working on a refactor that pulls image fetching out into explicit check and get steps in the build plan, but the existence of file: means I can't know whether they're needed (Darwin and Windows have no images, for example), or what image to even check and get, until the build is running. In this case I can probably just preemptively configure steps which may just do no-ops, and they can probably get their config in a similar way to how the Get plan uses VersionFrom to get some config from a prior step, but it all feels kind of held together with duct tape. (There's also a bit of complexity in that we may need to check + get a custom type that the image_resource: uses, but I think have a way to get around that.)

I still don't see a super obvious way to implement dynamic build plans though. 🤔 But I haven't really tried, either, I guess.

Copy link

@aoldershaw aoldershaw Aug 6, 2020

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The only thought I had was for there to be some sort of build event that says "inject this build plan under this plan ID" or "replace this plan ID with this build plan". So, the plan would initially look something like:

{
  "schema": "exec.v2",
  "plan": {
    "id": "across-id",
    "across": {}
  }
}

and there could be a build event like:

{
  "event":"dynamic-build-plan",
  "version":"1.0",
  "data": {
    "id": "across-id",
    "across": {
      "some": "data"
    }
  }
}

which the UI would interpret to construct a final build plan of:

{
  "schema": "exec.v2",
  "plan": {
    "id": "across-id",
    "across": {
      "some": "data"
    }
  }
}

I'm sure there are some technical challenges I haven't considered, though (/it may not work at all!)

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

That seems close to what I'm proposing - whatever the step results in gets used as the value. I think it'd just look like this with the syntax I had before:

var: foo
do:
- get: data
  trigger: true
- task: filter-list
  file: ci/tasks/filter-list.yml
  input_mapping: {input: data}
  output_mapping: {output: filtered}
- load_var: filtered
  file: filtered/data.json

...but this is just a different syntax. values_from is probably clearer. I think we'd still need to handle values: ((.:foo)), but that could probably be handled by the across step too instead of a separate values: step.

In either case, I'm not sure how the across step would access the result in the case of a do. 🤔 Currently results are stored under step IDs stored in the RunState, but this is almost like a return value from an expression. With a single step it's obvious what step ID to use, but with a do it's not so easy. It's almost like we need (exec.Step).Run to be able to return a value on its own, and do would return the last value. But that changes the interface.

Maybe this whole storing-results-in-RunState pattern could/should change? @clarafu @chenbh Has your recent experience in this area led to a similar line of thinking?

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Oh, also re: get_vars vs list_vars, yeah depending on what the resulting value is list_vars might be more accurate. I think either get_vars would actually fetch all the var values individually, using a list call under the hood for the scheduling/triggering, or we would have list_vars instead and allow the user to run a separate get_var step for each var value if that's desired. Don't know which; haven't put much thought into it yet.

Copy link

@aoldershaw aoldershaw Feb 24, 2021

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

What I was proposing is a bit more complicated than just a multi-step values_from (where the value comes from the last step), but I didn't have the proper framing. Now, I realize it the proposal was actually a combination of two ideas:

  1. A new way to communicate between jobs by emitting/receiving data
  2. Anonymous jobs (run nested in other jobs)

I'll flesh out my original example a bit more:

jobs:
- name: parent
  plan:
  - get: ci
  - across:
    - var: foo
      values_from:
        do:  # <----------------------------- A
        - get: data
          trigger: true # <------------------ B
        - task: filter-list
          file: ci/tasks/filter-list.yml
          input_mapping: {input: data}
          output_mapping: {output: filtered}
        - load_var: filtered # <------------- C
          file: filtered/data.json
    ...

Here, A is an "anonymous job". When the trigger on B gets a new version, the parent job doesn't trigger - only the anonymous job A. The load_var C implicitly emits the value it loads, and parent implicitly receives and triggers on changes to this value.

The implicit emitting and receiving is very non-obvious, though, and the anonymous job idea is a bit odd (either we make it fully self-contained and need to duplicate the get: ci, or it magically can borrow artifacts from the parent job). What if instead we made it explicit and gave the filtering it's own job:

jobs:
- name: filter-list
  plan:
  - get: ci
  - get: data
    trigger: true
  - task: filter-list
    file: ci/tasks/filter-list.yml
    input_mapping: {input: data}
    output_mapping: {output: filtered}
  - load_var: filtered
    file: filtered/data.json
    emit: data # <--------------------- A

- name: parent
  plan:
  - get: ci
  - receive: data # <------------------ B
    job: filter-list
    trigger: true
  - across:
    - var: foo
      values: ((.:data))
    ...

A emits a new value and stores it as data. B triggers on changes to that emitted value and stores the result in a local variable.

This mechanism could even be used for {get_var: ..., trigger: true} - the builds we run periodically to fetch the var could emit the value to the DB, and the job would receive that changed value from the build.

EDIT:

This could even possibly be used for {get: ..., trigger: true} - the periodic check builds could emit the latest version, and the generated plan could effectively be:

- receive: version # from the check build, somehow
  trigger: true
- get: version

Wouldn't work with version: every or pinned/disabled versions, though, so doesn't make much sense

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

In either case, I'm not sure how the across step would access the result in the case of a do. 🤔 Currently results are stored under step IDs stored in the RunState, but this is almost like a return value from an expression. With a single step it's obvious what step ID to use, but with a do it's not so easy. It's almost like we need (exec.Step).Run to be able to return a value on its own, and do would return the last value. But that changes the interface.

I thought of it like every type of step would have a single output (e.g. check is version, get is image spec, load_var and get_var are var), in do's case maybe we just need to figure out the contract for what it should return (is it time to introduce a result step/modifier)?

It would then be up to across to figure out what to do with this info (which might be wild cause that implies we can run across a list of versions from a check).

Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I am not sure if I am adding anything useful to the conversation, but when I saw the the 7.0.0 release notes I immediately saw a good use case for one of our pipelines and I tried to build a working prototype using pipeline instances and the across step.

- name: set-pipeline-instances
  serial: true
  plan:
    - in_parallel:
      - get: ci
      - get: src-branches
        trigger: true
    ...
    - load_var: branches
      file: branches-json/branches.json
    - set_pipeline: {{ (datasource "config").name }}
      file: ci/ci/build.yml
      instance_vars: {branch: ((.:b))}
      across:
        - var: b
          values: ((.:branches))

I am keen for the normal resource type to trigger the job and then use the load_var to load a variable that is a list and then just use that for for the argument to values.


## Open Questions

* n/a
Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@JohannesRudolph Sorry, completely forgot to address your point! (Raising this as a review comment so the replies can be threaded.)

Would it be sufficient to allow a static list of combinations to skip to be provided?

across:
- var: go
  values: [1.16, 1.15, 1.8]
- var: platform
  values: [darwin, linux, windows]
except:
- platform: windows
  go: 1.18

This would work similarly to excluding jobs in Travis CI, i.e. a subset of vars may be specified, and all variations with those vars will be skipped.

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Related thought: it would be kind of cool if we could except combinations based on a relationship between variables. e.g. if you want to test various upgrade paths, you only want to test combinations of from and to where to > from. I have no idea how this would work - it might just be a use case for var sources rather than except

@drunkirishcoder
Copy link

would there be a way to rerun only the failed combinations without having to rerun all the combinations again?

@aoldershaw
Copy link

would there be a way to rerun only the failed combinations without having to rerun all the combinations again?

@drunkirishcoder that's discussed here. It's just a matter of adding an attempts modifier to the step:

task: unit
timeout: 1h # interrupt the task after 1 hour
attempts: 3 # attempt the task 3 times
across:
- var: go_version
  values: [1.12, 1.13]
on_failure: # do something after all steps complete and at least one failed

@drunkirishcoder
Copy link

drunkirishcoder commented Apr 7, 2021

@aoldershaw cool. that can auto retry x attempts. but would there be a manual way? like in the concourse UI rerun the task with the same settings but only retry the failed combo?

the use case I'm trying to design for is I want to use across to deploy to multiple clusters/environments. and if one of the env failed, I wouldn't want to rerun the deployment to the successful ones again. so ideally if I can rerun the same matrix, but only for the failed ones again.

`ensure` and `on_*` bind to the `across` step so that they may be run after the
full matrix completes.

`attempts` binds to the inner step because it doesn't seem to make a whole lot
Copy link

@aoldershaw aoldershaw Apr 8, 2021

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@drunkirishcoder going to move this discussion to a review comment thread to avoid cluttering the top-level.

that can auto retry x attempts. but would there be a manual way? like in the concourse UI rerun the task with the same settings but only retry the failed combo?

the use case I'm trying to design for is I want to use across to deploy to multiple clusters/environments. and if one of the env failed, I wouldn't want to rerun the deployment to the successful ones again. so ideally if I can rerun the same matrix, but only for the failed ones again.

Ah, so that's what you meant. There are no plans for this currently, and it would be a bit challenging to get right - there's no precedence for partial reruns of a build. In general, we'd still need to run everything before the across step again, since those steps may produce artifacts that the across step relies on (e.g. gets or task outputs). Plus, if the step that's run in a build matrix emits build outputs (via get/put steps), we'd still need to emit the build outputs for the previously successful combinations (in order for build scheduling to work properly).

I think the ideal solution would for the deploy steps to be idempotent - i.e. given the same inputs, if the deployment already succeeded the step would no-op, so re-running the whole matrix wouldn't be an issue. That's perhaps easier said than done, though.

Is there a reason why manual retries is preferable to using attempts for automated retries for your use case?

Copy link

@drunkirishcoder drunkirishcoder Apr 8, 2021

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@aoldershaw ok, thank you for the detailed response. that makes a lot of sense. as to why prefer manual over automated retries. I'm just thinking sometimes a deployment would fail due to circumstances that may take longer to troubleshoot and resolve, like someone forgot to apply a firewall rule. but yeah I understand why it would be difficult to do in concourse.

would be nice if there's a similar feature that will represent each one as an independent pipeline level task instead, that can be restarted. just thinking out loud.

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

One thing you could do is configure a pipeline for each cell in the matrix using the set_pipeline step, rather than setting up all the environments within the single build. That way you could retrigger the failed job in each pipeline.


## Open Questions

* n/a

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Something that didn't really come up in the proposal is how we handle steps that emit artifacts (e.g. get, task.outputs, prototype outputs). Currently, the implementation of the static across step runs each iteration in its own local artifact scope. This means that any artifacts go out of scope past the across step and cannot be used.

It'd be cool if there was a way to "fan-in" on artifacts emitted within an across step. For instance, suppose there was a prototype (or just a task) for compiling a Go binary that outputs an artifact called binary. You might want to compile it for multiple platforms/architectures, which could make use of the across step:

across:
- var: platform
  values: [linux, darwin, windows]
- var: arch
  values: [amd64, arm64]
run: compile
type: go
params:
  package: repo/cmd/my-program
  platform: ((.:platform))
  arch: ((.:arch))
outputs: [binary]

(side-note: it might make more sense from a performance PoV for a go prototype to support compiling multiple platforms/architectures in a single run step where possible, but bear with me for the example)

The trouble is - under the current implementation, there's no way to access binary outside of the across step. Even if we used a shared artifact scope for each iteration, it'd just be "last write wins" (so, with no parallelism, binary would be the result for (windows, arm64) only) - since they all share a name, the results would clobber each other. This could be avoided by encoding more information in the output name via an output_mapping (e.g. name the artifacts binary-linux-amd64, binary-linux-arm64, ...) - but that only works for simple values like strings/numbers.

So, what if we let artifacts in an across step share a name, but be uniquely identified by the set of vars that were used to produce that artifact (in the same way that instanced pipelines share a name and are uniquely identified by their instance_vars). A possible syntax for referencing the binary created for (linux, amd64) could be binary{platform: linux, arch: amd64}

Here, binary really refers to a matrix of outputs (that mirrors the across matrix). What if we also provided a way to filter down that matrix to get at a subset of the outputs. e.g. binary{platform: windows} would give the (2) outputs built for windows, and binary{arch: arm64} would give the (3) outputs built for arm64.

So, suppose you wanted to upload a github release containing all 6 binaries produced in the matrix. You could do something like:

put: release
params:
  globs: [binary/*] # or maybe [binary{*}/*] to make it clear it's a matrix

Suppose you wanted to bundle the different architectures separately - then, you could do:

across:
- var: arch
  values: [amd64, arm64]
put: ((.:arch))-bundle
params:
  globs: [binary{arch: ((.:arch))}/*]

This may warrant a separate proposal but figured I'd bring it up here since it's definitely a limitation of the across step as it stands

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

One idea, maybe we could re-use the concept of output_mappings. Similar to how a task defined in a file can have its outputs mapped by an output_mappings configuration in the pipeline referencing that task, an across step could have its outputs made available via a similar mapping.

As an example:

across:
- var: arch
  values: [amd64, arm64]
- var: version
  values: [1.0, 1.1]

task: create-artifact
output_mapping:
  output: task-artifact

across_output_mapping:
- vars: # match vars from an execution of the task
    arch: amd64
    version: 1.0
  outputs: # map an output from the modified step to an output of the across step
    task-artifact: amd64-bundle-v1.0

This would be a fairly verbose if you wanted to map every output from the matrix of steps:

# [...]
across_output_mapping:
- vars: {arch: amd64, version: 1.0}
  outputs: {task-artifact: amd64-bundle-v1.0}
- vars: {arch: arm64, version: 1.0}
  outputs: {task-artifact: arm64-bundle-v1.0}
- vars: {arch: amd64, version: 1.1}
  outputs: {task-artifact: amd64-bundle-v1.1}
- vars: {arch: arm64, version: 1.1}
  outputs: {task-artifact: arm64-bundle-v1.1}

Though it would possibly simpler to implement, without relying on prototypes or other Concourse functionality.

When, or if, this limitation on using vars in get: and put: steps (eg. get: bundle-((.:arch))-((.:version))) is lifted, the across_output_mapping could be updated to allow for the use of vars as well, simplifying the configuration to something like...

# [...]
across_output_mapping:
- outputs:
    task-artifact: ((.:arch))-bundle-v((.:version))

@dumez-k
Copy link

dumez-k commented Feb 8, 2022

Hi @taylorsilva have there been any updates to this epic recently? Is this still under active development?

Also a little confusing from looking through the initial spatial resources RFC but was support ever implemented for this in any release version? Or is all of this work just prototype unreleased stuff?

Thanks!

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

10 participants