Proof
The ephemeral workspace retires the dev/prod pipeline copy
When the target is a runtime coordinate on Fabric, one environment-free orchestration definition can serve an isolated Intent and production. The environments stay. The copies go.
In a recent product demo, a data engineering lead described a repository where production, development, and test pipelines sit side by side. The lead asked how the tool would keep them apart.
"If the only difference is target, you should have one pipeline," the founder said. "That pipeline should deploy to pre-prod."
The conditional is the useful part. It gives a team a test it can run on its own repository. If the target is the only difference, the repository does not hold three environments. It holds three copies of one definition.
Three near-identical pipelines are a maintenance bill
A pipeline copy exists because the target used to be a property of the job. To run the same logic in another environment, you copied the job, changed the connection, and named the copy after the environment.
The copies start equal and stop being equal. A fix lands in development and never reaches test. A test copy grows a column that production does not have. Production becomes the only copy anyone trusts, and the one nobody can run before production.
That is the bill. Every environment needs its own definition and its own review. Each copy gives a fix one more place to go missing.
The environment was the thing you needed. The copy was how you selected it.
An Intent separates the code from the target
For resource-backed work, an Intent owns a branch, and that branch is the code source. It also owns an ephemeral data workspace, and that workspace is only the data target.1 It is not a Git checkout, and it is not a second copy of the repository.
The production target is a different coordinate, handed to the run for a different purpose. A resource-less Intent owns neither resource; it keeps a writable clone of the default branch and has no managed branch or ephemeral workspace.1
The two coordinate sets are published separately. Domain variables identify the production target, and ephemeral variables identify the Intent target. Consumers are expected to resolve one target and keep the two coordinate sets distinct.2
The target is a runtime input, so the definition can stay environment-free
On Fabric, the definition stays the same. The committed orchestration artifacts are environment-free; what changes is the target the apply step receives. The agent applies the same artifacts to the Intent's ephemeral workspace, and the production deploy job applies them to the Domain workspace. The artifacts and the REST surface are the same; the coordinates differ.3
Applying an item does not start a run. The production deploy job binds the items to the production workspace and reconciles their schedules. Execution follows from the reconciled schedule or from a run someone starts.
That is the founder's point in one line: the job no longer has to carry the target. The apply step supplies it.
Environment reuse is not platform portability
One definition describes environment reuse. It does not describe one mechanism across platforms.3
On Fabric, the definition is the same on both sides of promotion and the target is what changes.
| On Fabric | Intent target | Production target |
|---|---|---|
| Workspace | A separate workspace with its own lakehouse or warehouse | The Domain workspace |
| Data | The Intent's own dbt and dlt runs fill it; a Warehouse target also pairs with a bridge Lakehouse | The Domain's own data, already in the Domain workspace |
| Definition | The same environment-free orchestration artifacts | The same environment-free orchestration artifacts |
| Apply | The agent applies them to the ephemeral workspace | The production deploy job binds them to the Domain workspace and reconciles schedules |
The isolation unit, the apply path and the native execution engine are specific to Fabric. The deploy job covers both Fabric targets, lakehouse and warehouse.3
For a Warehouse Intent, the ephemeral workspace pairs the Warehouse with a bridge Lakehouse, and the bridge carries the Domain Warehouse's schemas as shortcuts. The Warehouse itself holds no shortcuts, so a production three-part name resolves inside the workspace without a second copy of the data.4
That is environment reuse within one platform. It is not one definition that runs unchanged on every platform, and the runtime cannot be assumed.
Promotion applies the definition; execution is a separate step
Promotion is a deliberate change of target, and the deploy path is narrower than "the code that passed the gate is the code that runs in production."
The production deploy job binds the environment-free definition to the production workspace and reconciles its schedules. It does not trigger execution. So the guarantee is that the artifacts are the same and the target is different. It is not a guarantee that a revision-linked production run happened, and it is not a guarantee that branch protection or a deployment gate covers every path.
That distinction is what makes the collapse safe to attempt. You remove a duplicate definition and keep the gate and the run where they belong. The gate still decides what ships. The run still happens as its own event.
Target safety belongs to the consumer that resolves it
"One pipeline pointed at production is one bad variable away from a production write." That is the strongest objection, and it deserves a direct answer.
A run inside an Intent is handed the Intent's target coordinate. Studio's query runtime resolves one target from the AgentSession runtime contract and rejects an environment that carries a nonblank foreign-platform coordinate for the target namespace it resolves.2 That is fail-closed selection for that consumer.
The runtime-location contract checks that a target's required coordinates are present and flags keys that belong to another platform as cross-platform violations. Presence is not resource identity. Safety belongs to each consumer that resolves the target.
The architecture reduces the blast radius. It does not remove the need for the gate.
You keep the environments and lose the copies
The ephemeral workspace is an environment. It belongs to one Intent, and it has a lifecycle of its own. Archive keeps the branch and workspace for resume.
Complete, close, and delete start reclamation of both. Cleanup is tracked and retried, so a terminal Intent can still hold unreclaimed provider resources for a while.1
That lifecycle rules out a shared development lakehouse where parallel work collides. This is a target for one piece of work, provisioned when it needs it and reclaimed when it ends. The next piece of work starts with its own.
Compare one definition with a per-Intent target to a copy per environment with a shared target. The first keeps isolation and removes the copy.
The reviewer sees one definition instead of three. The DRE sees one place for a fix to land. The Head of Data gets the capacity back. The same team ships five times more with Vibedata: less time to PR and less time per PR, once the team changes what it focuses on and how it builds, and review compounds as each accepted change shrinks the next review.
Start with the pipeline your team keeps three copies of
Take the pipeline that exists as development, test, and production variants. Write down what differs between them.
If the answer is the target, you have a candidate for collapse. If the answer is more than the target, you have found the reason the copies exist, and that reason is worth fixing before you remove them.
What does your repository keep three copies of that only needs one?
References
Footnotes
-
Vibedata workspace mode for resource-backed Intents: the Intent's remote branch is the durable code source, and its ephemeral data workspace is a provider-managed data target rather than a Git checkout. A resource-less Intent keeps a writable clone of the default branch with no managed branch or ephemeral workspace. Archive retains the branch and workspace; complete, close, and delete start reclamation, with cleanup tracked and retried, so terminal Intents can retain unreclaimed provider resources. Source: Studio Intent lifecycle and ephemeral-workspace design, and
concepts/semantic-branch/engineering-reality.mdin the strategy repo, reviewed recently. ↩ ↩2 ↩3 -
Studio's data query runtime resolves a single target from the AgentSession runtime contract and rejects an environment carrying a nonblank foreign-platform coordinate for the selected target namespace. The runtime-location contract validates that a target's required coordinates are present and flags foreign-platform keys as cross-platform violations; it does not establish resource identity. Source: Studio
data_query_runtime.pyandruntime-location-contract.ts. ↩ ↩2 -
The bundled deployment orchestration binds environment-free committed artifacts to one Fabric environment at apply time. It is the production counterpart of the agent's ephemeral apply: same artifacts, same REST surface, different coordinates. It covers both Fabric targets, lakehouse and warehouse, applies items without triggering a run, and reconciles schedules from the invoking pipeline. Source: bundled
domain-cicdshared/scripts/deploy_orchestration.py. ↩ ↩2 ↩3 -
On Fabric, a Warehouse Intent's ephemeral workspace pairs the Warehouse with a bridge Lakehouse that mirrors the Domain Warehouse's schemas as shortcuts. The bridge is the Intent's route to Domain data; the Warehouse has no schema shortcuts of its own. Source: Studio
fabric-warehouse-ephemeral-client.ts. ↩
Get the weekly roundup by email
One email a week: what got easier in Vibedata, a recipe to try yourself, and where data engineering is heading.
Or follow the RSS feed.
