Security and Trust Center

Our public bug bounty program, facilitated by HackerOne, allows a global collective of cybersecurity researchers and penetration testers to test Databricks for security vulnerabilities. Some of the key decisions we’ve made to make the program successful include:

• Encouraging an engaged community of hackers to be active on our program by providing transparency to our HackerOne program statistics such as response rate and payouts
• Promptly responding to bug bounty submissions, with an average time-to-bounty under a week
• Performing variant analysis on every valid submission to identify alternative ways that an exploit may be used, and verifying 100% of fixes
• Adding bonuses that drive attention to the most important areas of the product

We work hard to make our program successful and to learn from each submission. Our open and collaborative approach to our bug bounty program has resulted in over 100 security researchers being thanked for over 200 reports. Thank you all for helping us keep Databricks secure!

We want our customers to have confidence in the workloads they run on Databricks. If your team would like to run a vulnerability scan or penetration test against Databricks, we encourage you to:

1. Run vulnerability scans on data plane systems located inside of your cloud service provider account.
2. Run tests against your code, provided that those tests are entirely contained within the data plane (or other systems) located in your cloud service provider account and are evaluating your controls.
3. Join the Databricks Bug Bounty program to access a dedicated deployment of Databricks to perform penetration tests. Any penetration test against our multi-tenant control plane requires participation in the program.

We require multifactor authentication to access core infrastructure consoles such as the cloud service provider consoles (AWS, GCP and Azure). Databricks has policies and procedures to avoid the use of explicit credentials, such as passwords or API keys, wherever possible. For example, only appointed security team members can process exception requests for new AWS IAM principals or policies.

Databricks employees can access the production system under very specific circumstances (such as emergency break-fix). Access is governed by a Databricks-built system that validates access and performs policy checks. Access requires that employees are connected to our VPN, and authenticate using our single sign-on solution with multifactor authentication.
Learn more →

Our internal security standards call for the separation of duties wherever possible. For example, we centralize our cloud identity provider’s authentication and authorization process to separate authorizing access (Mary should access a system) from granting access (Mary can now access a system).

We prioritize least privilege access, both in internal systems and for our access to production systems. Least privilege is explicitly built into our internal policies and reflected in our procedures. For example, most customers can control whether Databricks employees have access to their workspace, and we programmatically apply numerous checks before access can be granted and automatically revoke access after a limited time.
Learn more →

Databricks leverages an Ideas Portal that tracks feature requests and allows voting both for customers and employees. Our feature design process includes privacy and security by design. After an initial assessment, high-impact features are subject to a security design review from the product security team in association with the security champions from engineering, along with threat modeling and other security-specific checks.

We use an agile development methodology that breaks up new features into multiple sprints. Databricks does not outsource the development of the Databricks platform, and all developers are required to go through secure software development training — including the OWASP Top 10 — when hired and annually thereafter. Production data and environments are separated from development, QA and staging environments. All code is checked into a source control system that requires single sign-on with multifactor authentication and granular permissions. Code merges require approval from the functional engineering owners of each area impacted, and all code is peer reviewed. The product security team manually reviews security-sensitive code to eliminate business logic errors.

We use best-of-breed tools to identify vulnerable packages or code. Automation in a preproduction environment runs authenticated host and container vulnerability scans of the operating system and installed packages, along with dynamic and static code analysis scans. Engineering tickets are created automatically for any vulnerabilities and assigned to relevant teams. The product security team also triages critical vulnerabilities to assess their severity in the Databricks architecture.

We run quality checks (such as unit tests and end-to-end tests) at multiple stages of the SDLC process, including at code merge, after code merge, at release and in production. Our testing includes positive tests, regression tests and negative tests. Once deployed, we have extensive monitoring to identify faults, and users can get alerts about system availability via the Status Page. In the event of any P0 or P1 issue, Databricks automation triggers a “5 whys” root cause analysis methodology that selects a member of the postmortem team to oversee the review. Findings are communicated to executive leadership, and follow-up items are tracked.

Databricks has a formal release management process that includes a formal go/no-go decision before releasing code. Changes go through testing designed to avoid regressions and validate that new functionality has been tested on realistic workloads. Additionally, there is a staged rollout with monitoring to identify issues early. To implement separation of duties, only our deployment management system can release changes to production, and multiperson approval is required for all deployments.

We follow an immutable infrastructure model, where systems are replaced rather than patched to improve reliability and security and to avoid the risk of configuration drift. When new system images or application code is launched, we transfer workloads to new instances that launch with the new code. This is true both for the control plane and the data plane (see the Security Features section for more on the Databricks architecture). Once code is in production, a verification process confirms that artifacts are not added, removed or changed without authorization.

The final phase of the SDLC process is creating customer-facing documentation. Databricks docs are managed much like our source code, and documentation is stored within the same source control system. Significant changes require both technical and docs team review before they can be merged and published.
Visit documentation →

The Databricks Lakehouse architecture is split into two separate planes to simplify your permissions, avoid data duplication and reduce risk. The control plane is the management plane where Databricks runs the workspace application and manages notebooks, configuration and clusters. Unless you choose to use serverless compute, the data plane runs inside your cloud service provider account, processing your data without taking it out of your account. You can embed Databricks in your data exfiltration protection architecture using features like customer-managed VPCs/VNets and admin console options that disable export.

While certain data, such as your notebooks, configurations, logs and user information, is present within the control plane, that information is encrypted at rest within the control plane, and communication to and from the control plane is encrypted in transit. You also have choices for where certain data lives: You can host your own store of metadata about your data tables (Hive metastore), store query results in your cloud service provider account, and decide whether to use the Databricks Secrets API.

Suppose you have a data engineer that signs in to Databricks and writes a notebook that transforms raw data in Kafka to a normalized data set sent to storage such as Amazon S3 or Azure Data Lake Storage. Six steps make that happen:

1. The data engineer seamlessly authenticates, via your single sign-on if desired, to the Databricks web UI in the control plane, hosted in the Databricks account.
2. As the data engineer writes code, their web browser sends it to the control plane. JDBC/ODBC requests also follow the same path, authenticating with a token.
3. When ready, the control plane uses Cloud Service Provider APIs to create a Databricks cluster, made of new instances in the data plane, in your CSP account. Administrators can apply cluster policies to enforce security profiles.
4. Once the instances launch, the cluster manager sends the data engineer’s code to the cluster.
5. The cluster pulls from Kafka in your account, transforms the data in your account and writes it to a storage in your account.
6. The cluster reports status and any outputs back to the cluster manager.

The data engineer doesn’t need to worry about many of the details — they simply write the code and Databricks runs it.