Databricks App delivery
Ship Databricks Apps the way you ship software
A governed delivery accelerator for Databricks Apps and GenAI workloads.
- Role in the family
- Databricks App delivery
- Databricks surfaces
- Databricks AppsAsset BundlesUnity CatalogAI GatewayMLflow
- Status
- Templates, delivery pipeline, and governance layer are implemented. A public read-only preview runs at preview.runway.techfabric.com.
- Documentation
- runway.techfabric.com
The problem
Teams can build a Databricks App. What they cannot do consistently is ship one repeatedly, with approvals, preview environments, per-environment secrets, evaluation gates, and a catalogue of what is running and what it costs.
How it works
A thin multi-tenant control plane that runs as a Databricks App inside the customer's own workspace, plus opinionated templates and a governance layer. Git push triggers bundle validation and deployment; each pull request gets a preview environment; approval workflows are platform state machines and quality gates are policy checkpoints, so a failed evaluation structurally blocks promotion.
What it does
Templates that start governed
Scaffolds for RAG agents, Unity Catalog tool agents, and Genie-embedded analytics apps, pre-wired to the gates.
Preview per pull request
Every change gets an environment before it gets an opinion.
Evaluation gates on promotion
An agent version promotes only if its evaluation suite passes. The gate is structural, not procedural.
App catalogue
Owner, version, Unity Catalog lineage, and cost for everything running in the workspace.
Your team ships Databricks Apps on a cadence instead of one at a time. Templates start governed, every pull request gets a preview environment, and a failed evaluation blocks promotion, so releasing gets boring in the way it should be.
One App shipped. The second one is a negotiation.
Building a Databricks App is a solved problem. Shipping the fourth one, on a Tuesday, without a meeting, is not.
There is no preview to look at
Review means reading a pull request and imagining it. The first time anybody sees the change running is when it is running for everybody.
Promotion is a person
Someone with the right access does it by hand, from a checklist that lives in their head, and the release waits when they are away.
Nothing blocks a bad one
The evaluation exists, in a notebook, and it is run when somebody remembers. A regression ships because the gate was social rather than mechanical.
Each of those is the same gap. The delivery path was built for one App by the person who built it, and it was never turned into something a team could run.
The shift
Three ways an App reaches production
The difference is not the App. It is what stands between the change and the users.
By hand, by one person
Works until they are away
A pipeline with no gates
Fast, and it ships regressions
Governed delivery
Repeatable, and it stops a bad one
What gets installed
The delivery path, as software rather than as a runbook.
A preview environment per pull request
The change runs somewhere before anybody approves it, so review is looking rather than imagining.
Per-environment secrets
Configuration and credentials that differ by environment without a person editing them at deploy time.
Evaluation gates
A failed evaluation stops the promotion. The gate is a policy checkpoint rather than a reminder somebody sends.
Governed promotion
Who may promote what, to where, recorded. An approval is a decision with a name on it rather than a message in a channel.
Rollback that has been run
The reverse path is exercised rather than documented, because a rollback nobody has tested is a plan rather than a capability.
The evidence, attached
What shipped, who approved it, which evaluation passed and when. Available afterwards as a query.
Every PR
Gets its own preview environment
The path our own Apps ship on
Runway is the delivery path we built for our own Databricks Apps and then published. It is free and open source under Apache-2.0, and it runs in your workspace rather than ours.
FAQ
TechFabric Runway, answered
Does Runway run in our workspace or yours?
Yours. The control plane is a Databricks App inside your own workspace. TechFabric-hosted is available as an option, but in-workspace is the default so your data never leaves your boundary.
What does a quality gate actually block?
Promotion. An agent or model version only moves forward if its evaluation suite passes, and the gate is a policy checkpoint in the pipeline, and it holds when a release is late.
Do we get preview environments?
Every pull request gets one, with per-environment secrets synced to secret scopes. The point is that a change gets an environment before it gets an opinion.