Remove a deployed project

When you need it

Use this page to remove one target of a generated project, for example a finished test trial or a personal development target. databricks bundle destroy removes what the Bundle deployed. Data, models and monitoring evidence are Unity Catalog or workspace objects that the Bundle never managed, so they stay.

flowchart LR
    A["Pause monitoring enrollment"] --> B["Stop continuous publication"]
    B --> C["bundle destroy -t target"]
    C --> D["Check what survives"]
    D --> E["Owner decides on manual cleanup"]

What to edit

Do these steps first, while the jobs still exist. After destroy you can no longer run them.

Step	How
Pause batch monitoring enrollments	Set monitoring_enabled: "false" for the target, redeploy and run the score job once. The inventory then records a paused enrollment instead of a model that stops reporting.
Pause serving enrollments	Run the monitoring job with monitoring_action=enroll_serving and each entry resubmitted with enabled: false; see serving monitoring.
Stop a continuous feature publication	Run online_publication with publication_action=stop, then status; see online features. A native pipeline keeps running after the job is gone.
Pause a traffic rollout	Pause the daily_rollout schedule. Destroying the job does not change endpoint traffic.
Keep training evidence you still need	Approval, rollback and monitoring reference preparation read saved candidate runs. Export or keep them before removing the folder that holds the experiment (see below).

If this target owns the shared monitoring dashboard (created with --create-dashboard), other projects use its URL. Either keep the owner target deployed, or first create a new owner in another project with --create-dashboard and rerun the helper with the new --dashboard-url in every project that uses it. See the dashboard page.

The command

From the generated project root, with the same target and profile used to deploy:

databricks bundle destroy -t test --profile YOUR_PROFILE

The CLI lists the jobs and the workspace folder it will delete and asks for confirmation. Read the list before you confirm. Repeat per target; destroying test does not touch syst or prod.

What you should see

bundle destroy removes the resources in the Bundle state of that target:

Removed	Source
Jobs train, score, monitoring	resources/*.job.yml
Optional jobs online_feature_lookup_setup, online_publication, daily_rollout	Present when selected at setup
The features job	resources/features.job.yml, when created by refresh_feature_graph.py
The dashboard refresh task	A target's block in deployment/monitoring.yml, part of the monitoring job
The shared monitoring dashboard	Only in a target whose deployment/monitoring.yml block uses create
Job permissions managed by the Bundle	permissions and manage_job_permissions settings
Deployed files and Bundle state	The target's root_path workspace folder

The training jobs log to the MLflow experiment <root_path>/experiment, inside that folder. Treat it as deleted with the folder unless you move or export it first. Registered model versions stay in Unity Catalog, but their links to those runs no longer open.

These stay, because the Bundle never created them as resources:

Survives	Notes
Training and scoring source tables	Never created by the project
Prediction tables, generation tables and model-set views	Created by the score job at run time
Feature group outputs and the merged feature table	Created by the features job at run time
Monitoring store tables and views	model_inventory, monitoring_results, performance_actions, retraining_requests, the views, reference_cleanup_audit, monitor_ref_* references and shared monitor_src_* populations; shared with other projects
Monitoring enrollment rows	Stay in model_inventory; paused if you followed the steps above
Registered models, versions, aliases and tags	Unity Catalog objects
Monitoring experiment named by monitoring_experiment_name	Outside the project folder when set to a path such as /Shared/...
Serving endpoints and UC SQL functions	Created by the SDK or the online_feature_lookup_setup job, not by the Bundle
Online stores, online tables and publication pipelines	Created through Feature Engineering, not by the Bundle
A dashboard reused by URL	Owned by another project or created outside the Bundle

Tables in your workspace explains what each of these objects holds and which ones grow over time.

Common errors and fixes

Check ownership before deleting anything by hand. The monitoring store, the dashboard, feature tables and online stores are often shared with other projects. Registered models and prediction tables can feed other consumers. Only the owner of an object should remove it, with that team's normal process.

Symptom	Fix
destroy wants to delete resources you did not expect	Stop and check -t and --profile. Destroy works on the Bundle state of the selected target only.
Other projects lost their dashboard	The destroyed target owned it. Redeploying creates a new dashboard with a new ID, so create a new owner with --create-dashboard and rerun the helper with its --dashboard-url in every project that uses it.
The dashboard still lists the removed model	Its enrollment row stays in the shared inventory. Pause it before destroying, as shown above.
An endpoint still serves the model	Endpoints are not Bundle resources. Its owner removes it in Databricks Serving.