cancel
Showing results for 
Search instead for 
Did you mean: 
Administration & Architecture
Explore discussions on Databricks administration, deployment strategies, and architectural best practices. Connect with administrators and architects to optimize your Databricks environment for performance, scalability, and security.
cancel
Showing results for 
Search instead for 
Did you mean: 

Databricks App repeatedly enters STOPPED after ~6–7 hours on Premium GCP — no matching stopApp audi

DoctorY
New Contributor II

We are investigating recurring unexpected stops of a Databricks App running a Python FastMCP/FastAPI service.

Environment

  • Databricks Premium

  • Google Cloud Marketplace subscription: Active / Postpay

  • Google Cloud region: europe-west3

  • Databricks Apps Medium compute

  • Workspace observed as RUNNING during relevant investigations

  • The same application source works normally after an authorized restart

Observed behavior

Between September 17 and 22, 2026, we identified five candidate unexplained compute stops. The App transitions to UNAVAILABLE / STOPPED and reports:

App compute was stopped due to workspace or account status.

We correlated bounded system.access.audit events with 41 original records from system.billing.usage, filtered by billing_origin_product = 'APPS' and the exact application ID.

Total measured consumption: 15.962087 DBU.

At the documented Medium compute rate of 0.5 DBU/hour, the final partial billing records suggest the following approximate termination times (UTC):

  • September 18 — 03:04

  • September 19 — 18:04

  • September 21 — 10:34

  • September 21 — 20:04

  • September 22 — 18:04

The unexplained sessions generally lasted approximately 6–7 hours.

Interestingly, all five inferred termination times fall near minute :04 or :34. We do not know whether this reflects an internal lifecycle policy, scheduled platform operation, billing behavior, or coincidence.

What we checked

  • The account is Premium, not Free Edition.

  • The Google Cloud Marketplace order is active.

  • Two successful user-initiated stopApp events were identified separately and excluded from the unexplained-stop candidates.

  • No matching stopApp event was found for the candidate unexplained stops within the examined audit windows.

  • Successful deployment history and application source were retained.

  • After authorized starts, the MCP server passed initialization and health checks without a source fix.

Billing-derived times are estimates, not authoritative control-plane termination timestamps.

Questions

  1. Has anyone observed this recurring STOPPED behavior with Databricks Apps on Premium, especially on GCP?

  2. Are there documented or known Apps compute lease, lifecycle, inactivity, or quota mechanisms that could explain a roughly 6–7-hour runtime?

  3. Could a periodic internal control-plane process explain the apparent :04 / :34 timing?

  4. Can Databricks Apps compute be stopped by an internal system action without a corresponding stopApp event in system.access.audit?

  5. Is there a read-only customer-accessible diagnostic source for the exact compute stop reason?

  6. If internal control-plane logs are required, what is the appropriate engineering escalation route for a Marketplace customer without a direct Databricks support package?

We are trying to identify the root cause before making further lifecycle or configuration changes.

We can provide additional timestamped evidence privately to Databricks engineering. We are not requesting anyone to access our account through this public thread.

1 REPLY 1

Louis_Frolio
Databricks Employee
Databricks Employee

Hello @DoctorY, I took a look at both internal and external documentation and here is what I found.

Thanks for the careful timeline, it makes this a lot easier to reason about. Straight answer: I can't give you a root cause from the outside, but I can narrow it down and point you to the right escalation path.

What it isn't

The only documented auto-stop for Apps is the Free Edition limit. Apps run up to 24 hours after being started, updated, or redeployed, then stop automatically. That's 24 hours, not 6 to 7, and it isn't documented as a Premium policy. I couldn't find any public doc describing a Premium inactivity timeout, quota, or compute lease, or any customer-facing field that exposes the internal stop reason. The message you're getting ("due to workspace or account status") reads more like an account or entitlement check than an app lifecycle rule. That's a hypothesis, though, and only control-plane logs can confirm it.

Reading the evidence you have

  • The missing stopApp makes an intentional user or API stop unlikely. It doesn't prove a platform action, though, because the public audit reference doesn't promise that internal lifecycle transitions get logged as stopApp.
  • I wouldn't read much into :04 / :34 yet. Your times are back-calculated from DBU quantities. If usage is aggregated or rounded at a fixed granularity, the inferred times will snap to a grid regardless of when the stop actually happened.
  • A clean restart on the same source argues against a deterministic startup defect. It doesn't fully rule out a runtime crash or a health-check failure.

Capturing the next occurrence

The big gotcha: Databricks doesn't persist logs when app compute shuts down. So set these up before the next stop:

  • App telemetry (Beta) or an APM tool, so the System and App logs right before the stop survive. Telemetry writes traces, logs, and metrics to Unity Catalog tables and automatically captures system logs.
  • A small scheduled job that polls GET /api/2.0/apps/{name} every few minutes and writes compute_status.state, the status message, and the UTC timestamp to a Delta table. That gives you an observed transition time, not an estimate.
  • The Insights tab. If app service health shows unavailable, it's a platform issue; if only app availability is down, look at your deployment or code. That one signal could answer half your questions.
  • Every service_name = 'apps' event in each window (not just stopApp), plus account-level events (workspace_id = 0), and the Databricks status page for the same intervals.

Stopgap

In a similar thread (linked below), AmitDECopilot and DataNexus suggested first separating a crash from a clean lifecycle stop and, if it's lifecycle, using a scheduled job that restarts the app when it's found stopped. It treats the symptom, not the cause. Still, it keeps your MCP server up without touching the app config.

Escalation

For Databricks on Google Cloud, support starts with Google Cloud Customer Care, and Google and Databricks support work jointly. A Databricks support contract also lets you open a case directly. If you have neither, email help@databricks.com and tag the community managers here. Include the UTC timestamps, workspace, app, and deployment IDs, region, audit results, and logs, but share those privately, not in this thread.

At the end of the day, your evidence is enough to justify a platform-side investigation, but not enough to name the root cause yet. Capturing one stop with real timestamps and persisted logs is the fastest way to close that gap.

References:

Regards, Louis.