PAI protects data through VPC-based transmission security, integration with Alibaba Cloud storage services, and checkpoint-based backup capabilities.
Data transmission security
To enhance data transmission security for a model service deployed in PAI, access the service over a VPC using the official SDK for Python or custom logic. VPCs are isolated from each other, ensuring private high-speed access.
For more information about VPC security, see Security system overview.
Data storage security
PAI provides a dataset management module for creating datasets using public datasets or data stored in Alibaba Cloud storage services. The module also scans OSS folders to generate index datasets for intelligent labeling and model training.
To ensure storage security of data files such as datasets and models, we recommend using the following Alibaba Cloud storage services:
Data storage service | Related topic |
OSS | |
File Storage NAS | |
Cloud disks | |
MaxCompute |
Data backup and restoration
PAI allows you to use images to back up and restore user environments and datasets. For example, create an image to save the data, engineering environments, and configurations of an instance. This way, you can deploy the image based on business requirements, providing a safeguard against data loss. When using Alibaba Cloud storage services to store data or datasets, backup and restoration is guaranteed by the corresponding service.
Data storage service | Related topic |
OSS | |
File Storage NAS | |
MaxCompute |
When using PyTorch to train a foundation model, use EasyCkpt to save and resume model training progress to minimize data loss and resource waste. EasyCkpt is a high-performance checkpoint framework provided by PAI and is compatible with Megatron and DeepSpeed, popular model training frameworks. For more information, see EasyCkpt.