Databricks opens Genie ZeroOps beta after 80,000-issue preview
Databricks says Genie ZeroOps observed more than 80,000 issues before its public beta, but production changes need approval and shared findings can expose code and sample data.
By Olivia Hart
5 min read

What's News
- Databricks opened the public beta on November 19, 2026.
- The private preview observed more than 80,000 real customer issues.
- More than half of preview issues required organization-specific knowledge.
- A configuration can be shared with up to 25 individual workspace users, but not groups or service principals.
- Setup backfills findings from the previous 24 hours.
Databricks opened Genie ZeroOps to public beta on November 19, 2026, after the autonomous data-pipeline agent observed more than 80,000 real customer issues during a limited private preview.
Databricks said the preview covered hundreds of companies. It did not disclose the exact customer count, issue mix, or whether every customer completed a production deployment, leaving adoption less clear than the issue total. More than half of the preview issues required organization-specific knowledge to triage, rank, and resolve properly, the company said.
The public beta expands the agent’s diagnostic and workflow features:
- Shared inboxes for managing findings.
- Notifications that summarize each issue and recommendation.
- Custom instructions for company-specific workflows.
- One-click pull-request creation for Git-backed jobs.
What did the private preview reveal?
Databricks attributes its customization features to the preview result. Teams can identify known flaky sources, preferred fixes, and naming conventions. They can also filter expected failures and lower-priority notifications, reducing the need to review routine operational noise.
Databricks introduced ZeroOps on June 16, 2026, with planned private-preview support for jobs, pipelines, tables, and machine-learning workloads. Current documentation covers only jobs, pipelines, and tables, so the beta scope differs from the June launch plan. Models and applications remain outside the monitoring coverage.
Takeaway: Databricks found that more than half of preview issues needed knowledge specific to each customer organization.
What can ZeroOps actually detect?
According to Databricks documentation, ZeroOps watches for explicit job and pipeline failures. It also flags tables that stop updating or arrive with unexpectedly few rows, even when those problems do not trigger a conventional job failure.
For each investigation, Databricks says the agent gathers logs, source code, table metadata, and lineage. It runs SQL queries to identify root causes and ranks findings by downstream impact. Unity Catalog lineage maps dependencies between assets, allowing the agent to connect code diagnostics with the state of the data.
Investigations start automatically after ZeroOps detects an issue. Proposed fixes undergo unit tests and dry runs in an isolated sandbox before anyone applies them, Databricks said. Passing those checks does not give the agent authority to change production.
Databricks supports alerts through Slack, Microsoft Teams, PagerDuty, and generic webhooks. The product does not support email destinations, leaving some established notification workflows outside the initial beta.
Takeaway: ZeroOps is designed to find and explain operational failures, not merely display a failed job.
Where do approval and repository controls apply?
Databricks says ZeroOps cannot change production code or data without approval. The agent requires an explicit user instruction and confirmation before pushing a Git-backed fix. An approved push creates a commit without force-pushing or rewriting branch history, preserving the repository’s existing commit trail.
The caveat sits in repository configuration. Databricks’ documentation warns that naming a repository’s default branch can bypass pull-request review unless branch protection blocks direct pushes. A team can therefore approve the agent’s action without receiving a branch-level review if its rules allow the shortcut.
Databricks says the agent uses Anthropic and OpenAI models on Databricks-hosted endpoints inside the company’s security perimeter. That locates the inference infrastructure, but it does not remove the need to govern which assets users select or how findings leave the platform.
Takeaway: Production authority remains with customers, while repository settings determine how much review an approved fix receives.
What happens when teams share findings?
Databricks says a ZeroOps configuration can be shared with as many as 25 individual workspace users. It cannot be shared with groups or service principals. Collaborators can work on findings, but they may see code excerpts, suggested fixes, and sample data rows even without access to the underlying assets. Sharing findings does not grant access to those assets.
That distinction creates a control problem rather than a simple collaboration feature. Before an external notification includes an issue title or summary, Databricks says the agent checks for sensitive information, including names and email addresses, and masks what it finds. Copies stored in outside systems, however, fall outside Databricks’ retention and deletion controls.
The model-processing terms are broader. Selecting assets authorizes continuous, autonomous large-language-model processing until the feature is disabled, Databricks said. That processing includes repository contents for Git-backed jobs, making the permission operational and persistent rather than limited to a single approved fix.
Takeaway: Sharing and continuous processing broaden access to operational information, so administrators must treat both as governed data flows.
What does setup require?
Databricks requires an administrator to enable Genie ZeroOps from the workspace Previews page. Users then select assets and need “OWNER,” “CAN MANAGE,” or “CAN VIEW” permission on them. The setup process backfills findings from the previous 24 hours before reporting new issues, giving teams an immediate but bounded review window.
Databricks requires a credential with read access to investigate a Git repository. Creating a pull-request branch requires write access, allowing teams to separate diagnostic authority from implementation authority. The company does not describe those credentials as substitutes for branch protection or managed credential policies.
The beta operates at the workspace level rather than the account level, according to Databricks. It is not HIPAA compliant. Detecting late-arriving or incomplete table data also requires separately billed data-quality monitoring, giving the table-anomaly capability an extra dependency and cost outside ZeroOps itself.
Current Databricks documentation also excludes performance and cost regressions from the detection scope. The product monitors operational failures and selected data conditions, not every possible sign of deterioration in a company’s technology stack.
Takeaway: Broader monitoring comes with higher permission, billing, and regulatory dependencies beyond the basic beta.
What can buyers test next?
Databricks says the beta automates investigation and fix preparation while leaving production authority with the customer. The practical trade-off is less manual diagnosis in exchange for continuous AI processing of selected operational data.
Buyers should test branch protections, external notification paths, sharing permissions, and billing dependencies before a broad rollout. Those controls determine the real exposure and operating cost created by the product’s central promise.
Takeaway: Broad rollout depends on proving that repository, sharing, and billing controls contain the risks created by continuous AI processing.
Source: unite.ai
More from Olivia Hart
Show full bio
Staff writer covering people, launches and day-to-day practice in entrepreneurship, startups and venture funding at Business Bearings.
2,440 articles