pulumi-pipeline

module
v0.2.0 Latest Latest
Warning

This package is not in the latest version of its module.

Go to latest
Published: Sep 30, 2026 License: MIT

README

pulumi-pipeline

A DAG of Pulumi stacks with the refusals a shared infrastructure repository learns the hard way.

Runs many Pulumi stacks as one pipeline: steps discovered from the stack files a repository already commits, dependency edges between them, one command for diff, deploy, refresh, destroy and a health check — and gates that refuse to apply from a stale checkout, that fail an apply whose refresh-preview wants to destroy something, that turn any preview change into a failure on demand, and that put a clock on every step.

What Where
Go module github.com/truvity/pulumi-pipeline go get github.com/truvity/pulumi-pipeline@v0.1.0
pkg/pipeline — the engine: steps, DAG, typed steps, run, gates, CLI commands Go package, no Pulumi dependency
pkg/pulumistep — the Automation API step, scope discovery, output hooks, health check Go package
pkg/pulumicli — the command line, assembled from the two above Go package
pulumi-pipeline — the command release archives, or go install github.com/truvity/pulumi-pipeline/cmd/pulumi-pipeline@v0.1.0

Who it is for

A repository that holds several Pulumi programs, each with several stacks, and wants to preview and apply them from one place instead of cd-ing into each directory. It assumes the pulumi CLI and the language plugin of your programs are on PATH, that stacks live in a backend you have already logged in to (or named through PULUMI_BACKEND_URL), and that every stack you want managed has a committed Pulumi.<stack>.yaml.

It deliberately does not install Pulumi, choose a backend or a secrets provider, write your stack outputs anywhere (a hook is how you do that), or know what your programs deploy. It is not a replacement for pulumi — every step is the Automation API's preview, up, refresh or destroy on a stack you could run by hand.

The model

Three nouns.

A step is one unit of work with a name, the actions it supports (diff, deploy, refresh, destroy) and edges to the steps it depends on. A Pulumi step is one stack of one program.

A scope is a named group of steps that becomes a subcommand. The command discovers one scope per subdirectory of the scopes directory you name, and one step per Pulumi.<stack>.yaml in it. Nothing is registered: adding a stack file is enough.

A gate is a refusal the runner applies around your steps: on the checkout before a mutating run, on each step's clock, on a preview's result, and on the refresh-preview after an apply. Gates are on by default; the ones that can be overridden say so and log loudly when they are. The checkout gate relaxes on its own under CI=true (a runner's detached HEAD has no upstream) but still refuses a dirty tree; a local shell exporting CI=true bypasses the same checks (safety).

Steps that carry no edges run in name order; steps that do run in topological order (reverse for destroy). A dry run continues past a failing step and reports every failure at the end; every other action stops at the first.

Install and a worked example

go install github.com/truvity/pulumi-pipeline/cmd/pulumi-pipeline@v0.1.0

A repository laid out like this:

deploy/
  network/
    Pulumi.yaml
    Pulumi.dev.yaml
    Pulumi.prod.yaml
    main.go
  apps/
    Pulumi.yaml
    Pulumi.dev.yaml
    main.go

is driven with:

pulumi-pipeline --scopes-dir deploy list                    # the scopes
pulumi-pipeline --scopes-dir deploy network list            # its steps
pulumi-pipeline --scopes-dir deploy network diff --all      # preview every stack
pulumi-pipeline --scopes-dir deploy network deploy dev      # apply one stack
pulumi-pipeline --scopes-dir deploy network diff --all --expect-no-changes

Name the directory once instead of on every call: export PULUMI_PIPELINE_SCOPES_DIR=deploy, or commit .pulumi-pipeline.yaml containing scopes_dir: deploy. The flag wins over the environment, which wins over the file; there is no built-in default, because a directory name that fits one repository finds nothing in the next.

To persist stack outputs, write a main of your own:

func main() {
	os.Exit(pulumicli.Main(context.Background(), os.Args, pulumicli.Options{
		Name:      "infra",
		ScopesDir: "deploy",
		Hooks: pulumistep.OutputHooksFunc(func(root, scope, stack string) pulumistep.OutputFunc {
			return func(ctx context.Context, logger *slog.Logger, outputs map[string]any) error {
				return writeOutputs(root, scope, stack, outputs) // yours
			}
		}),
	}))
}

Secret outputs never reach the hook.

Consumers

The estate this was extracted from drives its infrastructure repository through pkg/pulumicli with an output hook of its own (the Go library surface). No other consumer is recorded at v0; a repository that adopts it adds a line here.

Neighbours

  • github-structure — a Pulumi engine for a GitHub organization. It is a program this runner can run; it has no dependency on it.
  • policy — the contracts this repository is held to (docs/contracts/component.md).
  • ci-workflows — the shared CI and release workflows this repository calls.

Documentation

The rule that makes this repository public

Mechanism only. A scopes directory, an output store, a backend, a credential and every name in them are the caller's: a flag, an environment variable, a file the caller commits, or a hook the caller writes. Nothing here defaults to one estate's value. hack/leak-canary.sh enforces the mechanical half over tracked files and runs in the gate.

Status

v0: the API may still move in a minor release, and every move is a Breaking: bullet in the CHANGELOG. The engine and the Pulumi step are the code an estate has run for months; the four gates (checkout, post-apply refresh-preview, --expect-no-changes, step timeout) are new and are exercised by the tests here, including against a real Pulumi stack on a local file backend. Nothing in the tests touches a cloud.

Development

devbox shell        # pins every tool, including the pulumi CLI the tests drive
just check          # build, test, lint, leak-canary
just vuln           # reachable Go advisories (its own workflow in CI, not in check)

The tests need pulumi and pulumi-language-go on PATH and fail, rather than skip, without them. They use a file:// backend in a temporary directory and a tiny Go program that registers component resources, which the engine tracks itself, so there is no provider and no network beyond module downloads.

Releasing

Releases are a v* tag, which the shared release workflow turns into the Go module version and a GitHub Release with the pulumi-pipeline archives. Automatic patch releases are not armed; the first release and every minor and major are hand-cut tags.

Licence

MIT. See LICENSE.

Directories

Path Synopsis
cmd
pulumi-pipeline command
Command pulumi-pipeline runs a DAG of Pulumi stacks: one scope per subdirectory of the scopes directory, one step per committed stack file (Pulumi.{stack}.yaml).
Command pulumi-pipeline runs a DAG of Pulumi stacks: one scope per subdirectory of the scopes directory, one step per committed stack file (Pulumi.{stack}.yaml).
pkg
pipeline
Package pipeline provides a generic DAG-based pipeline orchestrator.
Package pipeline provides a generic DAG-based pipeline orchestrator.
pulumicli
Package pulumicli assembles the command-line tool: scopes discovered from a directory, the pipeline's commands over them, and the caller's hooks.
Package pulumicli assembles the command-line tool: scopes discovered from a directory, the pipeline's commands over them, and the caller's hooks.
pulumistep
Package pulumistep provides a Pulumi stack operation step for the pipeline framework.
Package pulumistep provides a Pulumi stack operation step for the pipeline framework.
pulumistep/internal/testprogram command
Command program is the smallest Pulumi program the pulumistep tests can run against a local file backend: it registers `count` component resources, which the engine tracks itself, so no provider plugin and no cloud is involved.
Command program is the smallest Pulumi program the pulumistep tests can run against a local file backend: it registers `count` component resources, which the engine tracks itself, so no provider plugin and no cloud is involved.

Jump to

Keyboard shortcuts

? : This menu
/ : Search site
f or F : Jump to
y or Y : Canonical URL