At Databricks, we all know that information is one among your most precious belongings. Our product and safety groups work collectively to ship an enterprise-grade Information Intelligence Platform that allows you to defend in opposition to safety dangers and meet your compliance obligations. Over the previous yr, we’re proud to have delivered new capabilities and assets equivalent to securing information entry with Azure Non-public Hyperlink for Databricks SQL Serverless, preserving information non-public with Azure firewall help for Workspace storage, defending information in-use with Azure confidential computing, reaching FedRAMP Excessive Company ATO on AWS GovCloud, publishing the Databricks AI Safety Framework, and sharing particulars on our method to Accountable AI.
In keeping with the 2024 Verizon Information Breach Investigations Report, the variety of information breaches has elevated by 30% since final yr. We imagine it’s essential so that you can perceive and appropriately make the most of our security measures and undertake really useful safety greatest practices to mitigate information breach dangers successfully.
On this weblog, we’ll clarify how one can leverage a few of our platform’s high controls and just lately launched security measures to ascertain a strong defense-in-depth posture that protects your information and AI belongings. We can even present an summary of our safety greatest practices assets so that you can stand up and working rapidly.
Shield your information and AI workloads throughout the Databricks Information Intelligence Platform
The Databricks Platform offers safety guardrails to defend in opposition to account takeover and information exfiltration dangers at every entry level. Within the beneath picture, we define a typical lakehouse structure on Databricks with 3 surfaces to safe:
- Your shoppers, customers and functions, connecting to Databricks
- Your workloads connecting to Databricks companies (APIs)
- Your information being accessed out of your Databricks workloads
Let’s now stroll by way of at a excessive stage among the high controls—both enabled by default or obtainable so that you can activate—and new safety capabilities for every connection level. Our full listing of suggestions based mostly on completely different menace fashions could be present in our safety greatest apply guides.
Connecting customers and functions into Databricks (1)
To guard in opposition to access-related dangers, you need to use a number of elements for each authentication and authorization of customers and functions into Databricks. Utilizing solely passwords is insufficient attributable to their susceptibility to theft, phishing, and weak person administration. In reality, as of July 10, 2024, Databricks-managed passwords reached the end-of-life and are now not supported within the UI or through API authentication. Past this extra default safety, we advise you to implement the beneath controls:
- Authenticate through single-sign-on on the account stage for all person entry (AWS, SSO is routinely enabled on Azure/GCP)
- Leverage multi-factor authentication provided by your IDP to confirm all customers and functions which can be accessing Databricks (AWS, Azure, GCP)
- Allow unified login for all workspaces utilizing a single account-level SSO and configure SSO Emergency entry with MFA for streamlined and safe entry administration (AWS, Databricks integrates with built-in id suppliers on Azure/GCP)
- Use front-end non-public hyperlink on workspaces to limit entry to trusted non-public networks (AWS, Azure, GCP)
- Configure IP entry lists on workspaces and in your account to solely permit entry from trusted community areas, equivalent to your company community (AWS, Azure, GCP)
Connecting your workloads to Databricks companies (2)
To stop workload impersonation, Databricks authenticates workloads with a number of credentials throughout the lifecycle of the cluster. Our suggestions and obtainable controls rely in your deployment structure. At a excessive stage:
- For Traditional clusters that run in your community, we suggest configuring a back-end non-public hyperlink between the compute airplane and the management airplane. Configuring the back-end non-public hyperlink ensures that your cluster can solely be authenticated over that devoted and personal channel.
- For Serverless, Databricks routinely offers a defense-in-depth safety posture on our platform utilizing a mix of application-level credentials, mTLS shopper certificates and personal hyperlinks to mitigate in opposition to Workspace impersonation dangers.
Connecting from Databricks to your storage and information sources (3)
To make sure that information can solely be accessed by the appropriate person and workload on the appropriate Workspace, and that workloads can solely write to licensed storage areas, we suggest leveraging the next options:
- Utilizing Unity Catalog to control entry to information: Unity Catalog offers a number of layers of safety, together with fine-grained entry controls and time-bound down-scoped credentials which can be solely accessible to trusted code by default.
- Leverage Mosaic AI Gateway: Now in Public Preview, Mosaic AI Gateway permits you to monitor and management the utilization of each exterior fashions and fashions hosted on Databricks throughout your enterprise.
- Configuring entry from licensed networks: You’ll be able to configure entry insurance policies utilizing S3 bucket insurance policies on AWS, Azure storage firewall and VPC Service Controls on GCP.
- With Traditional clusters, you’ll be able to lock down entry to your community through the above-listed controls.
- With Serverless, you’ll be able to lock down entry to the Serverless community (AWS, Azure) or to a devoted non-public endpoint on Azure. On Azure, now you can allow the storage firewall in your Workspace storage (DBFS root) account.
- Assets exterior to Databricks, equivalent to exterior fashions or storage accounts, could be configured with devoted and personal connectivity. Here’s a deployment information for accessing Azure OpenAI, one among our most requested eventualities.
- Configuring egress controls to stop entry to unauthorized storage areas: With Traditional clusters, you’ll be able to configure egress controls in your community. With SQL Serverless, Databricks doesn’t permit web entry from untrusted code equivalent to Python UDFs. To find out how we’re enhancing egress controls as you undertake extra Serverless merchandise, please this manner to hitch our previews.
The diagram beneath outlines how one can configure a personal and safe atmosphere for processing your information as you undertake Databricks Serverless merchandise. As described above, a number of layers of safety can shield all entry to and from this atmosphere.
Outline, deploy and monitor your information and AI workloads with industry-leading safety greatest practices
Now that we’ve got outlined a set of key controls obtainable to you, you most likely are questioning how one can rapidly operationalize them for what you are promoting. Our Databricks Safety staff recommends taking a “outline, deploy, and monitor” method utilizing the assets they’ve developed from their expertise working with a whole lot of shoppers.
- Outline: It is best to configure your Databricks atmosphere by reviewing our greatest practices together with the dangers particular to your group. We have crafted complete greatest apply guides for Databricks deployments on all three main clouds. These paperwork provide a guidelines of safety practices, menace fashions, and patterns distilled from our enterprise engagements.
- Deploy: Terraform templates make deploying safe Databricks workspaces straightforward. You’ll be able to programmatically deploy workspaces and the required cloud infrastructure utilizing the official Databricks Terraform supplier. These unified Terraform templates are preconfigured with hardened safety settings just like these utilized by our most security-conscious prospects. View our GitHub to get began on AWS, Azure, and GCP.
- Monitor: The Safety Evaluation Device (SAT) can be utilized to observe adherence to safety greatest practices in Databricks workspaces on an ongoing foundation. We just lately upgraded the SAT to streamline setup and improve checks, aligning them with the Databricks AI Safety Framework (DASF) for improved protection of AI safety dangers.
Keep forward in information and AI safety
The Databricks Information Intelligence Platform offers an enterprise-grade defense-in-depth method for safeguarding information and AI belongings. For suggestions on mitigating safety dangers, please consult with our safety greatest practices guides in your chosen cloud(s). For a summarized guidelines of controls associated to unauthorized entry, please consult with this doc.
We repeatedly improve our platform based mostly in your suggestions, evolving {industry} requirements, and rising safety threats to higher meet your wants and keep forward of potential dangers. To remain knowledgeable, bookmark our Safety and Belief weblog, head over to our YouTube channel, and go to the Databricks Safety and Belief Middle.
