Platform Provisioning
Architecture
Predibase is a cloud-native, SaaS platform that is deployed on top of Kubernetes and operates out of a control plane for business logic and a data plane for operations related to customer-sensitive data.
The control plane follows a multi-tenant architecture and is managed by Predibase. The data plane is typically deployed into a secure Predibase AWS environment. Alternatively, customers can choose to deploy the data plane in their own VPC.
Deployment Options
Predibase Cloud
Predibase’s fully managed deployment allows you to start training models instantly on the Predibase platform. Predibase will operate the end-to-end infrastructure stack, including both the control plane and data plane. This allows us to easily manage and monitor the infrastructure, while also providing improved operational support for customers.
During the 14-day SaaS free trial, you will receive access to the Predibase platform (with limitations) at absolutely no cost.
VPC
The VPC deployment option gives you control over your data while benefiting from the agility of a SaaS solution. This is achieved by deploying the data plane into your own Virtual Private Cloud (VPC) so that data never leaves your environment. See the [VPC Deployment for additional information.
During the 14-day VPC free trial, you will receive access to the full Predibase platform at no cost. However, you will be billed for any compute used via AWS.
Comparison
This table summarizes the differences between the deployment options and includes features that are available in the free trial.
Plans | (Trial) Predibase Cloud | (Trial) VPC | (Paid) Predibase Cloud | (Paid) VPC |
---|---|---|---|---|
# Users | Unlimited | Unlimited | Unlimited | Unlimited |
Dataset Size | 1GB | No Limitations | No Limitations | No Limitations |
Dataset Confidentiality | Best for non-confidential datasets | Full confidentiality | Best for non-confidential datasets | Full confidentiality |
Data Types | All except Time Series / Audio | All available | All available | All available |
Encoders / Combiners / Decoders | Low compute functionality only | All available | All available | All available |
Hyperparameter Optimization | Not available | Available | Available | Available |
Model Repositories | Available | Available | Available | Available |
Model Serving | Not available | Available | Available | Available |
Predictive Query Language (PQL) | Available | Available | Available | Available |
Engines (compute) | Low-powered General & Training Engines | All engines available | All engines available | All engines available |
LLMs | Querying | Querying, Deploying | Querying, Deploying, Fine-tuning, Evaluating | Querying, Deploying, Fine-tuning, Evaluating |
Cost | $0 | Compute (billed via AWS) | Predibase Subscription + Compute (billed via Predibase) | Predibase Subscription + Compute (billed via AWS) |
If you are unsure which deployment model is right for you, please contact us at support@predibase.com and we’d be happy to advise.
VPC
The Predibase platform can be deployed into your cloud environment.
During the Free Trial, users are limited to deploying into Amazon Web Services (AWS) and the us-west-2 region. If you’d like to deploy into another cloud (ex. Google Cloud Platform, Azure) or another region, please let us know at support@predibase.com.
Deployment Steps
Pre-requisites:
- You have received an invite link to Predibase (for either the Free Trial or Paid plan).
- Typically an AWS admin role is required to create the CloudFormation stack via the AWS console.
There are two main steps to deploy Predibase into your AWS environment:
- Create the CloudFormation Stack in your AWS account
- Launch provisioning of the data plane
Creating the CloudFormation Stack in your AWS account
- Please click the invite link in your email to create your Predibase user account
- Once you’ve created a user account, you should arrive on a Configure your AWS page.
- Sign into your AWS account.
- Click the direct link to navigate to a CloudFormation Stack. Note that this link is specific to your account and can’t be shared across accounts.
- Once you've acknowledged the capabilities at the bottom of the page, click “Create Stack”
- Once you’ve clicked Create Stack, navigate to the Outputs Tab and copy the key
CustomerEnvironmentRoleArn
's value. Note that the value will take about 4-5 minutes to populate. - Switch back to the Predibase setup page and paste the ARN value in the RoleArn field, click "Validate Role Authorization", and proceed with provisioning.
Steps 2 and 4: Setup screen with “Direct Link” to CloudFormation
Step 5: CloudFormation Stack creation with “Create Stack” button
Step 6: Copy the Role from the Outputs Tab. Populates after 4-5 minutes.
Launch Provisioning of the Data Plane
Once you’ve confirmed the Role ARN, Predibase will automatically begin to deploy the Predibase dataplane in your environment.
Feel free to navigate away while this provisioning process is in progress. On average, it takes around 20-30 minutes but exact times can vary. We’ll send you an email with the login link once the provisioning process is complete.
Expected Costs for Trial
There are two types of costs to expect with a VPC trial:
- Base Setup Cost: The estimated base cost for running a VPC 14-day trial in AWS us-west-2 is ~$200, which covers the full cost to spin up the dataplane in your environment, including an EKS cluster and S3 buckets.
- Usage Cost: Compute costs incurred by using CPU or GPU engines in the platform. A rough estimate for 2 weeks is ~$15-30, but exact amount may vary.
Both of these costs will be billed directly via AWS and appear on the AWS Billing Console.
Increase AWS Quotas
Since the compute will run in your AWS account, you will need to ensure that your AWS quotas are sufficient to run the Predibase Controlplane.
We recommend the following are increased from their default value of 0:
- "All G and VT Spot Instance Requests"
- "Running On-Demand G and VT instances"
Example page from AWS for requesting a quota increase:
An increase to ~192 for both values should be sufficient to run multi-node training engines, but a further increase may be necessary if you require more GPUs for large training jobs.