Platform For AI:Mount storage

Mount different types of storage

• For unstructured data that is primarily read, such as models, images, and videos, use OSS.

• If you frequently read and write small files, or if multiple service instances need to share a read/write directory, use a General-purpose NAS file system.

• For high-performance computing (HPC) or AI training that requires extremely low latency and high throughput, use an Extreme NAS file system or CPFS.

• For large-scale parallel computing or AI training frameworks, use CPFS.

{
  "storage": [
        {
            "oss": {
                "path": "oss://bucket/path/",
                "readOnly": false
            },
            "mount_path": "/mnt/data/"
        }
    ]
}

{
  "storage": [
        {
            "nfs": {
                "path": "/",
                "server": "06ba74****-a****.cn-hangzhou.nas.aliyuncs.com",
                "readOnly": false
            },
            "mount_path": "/mnt/data/"
        }
    ]
}

Mount a custom dataset

You can create a PAI dataset from your data or configuration files and register it as a public AI Asset. This lets you easily mount and reference the dataset when you deploy services.

The following figure shows the configuration interface in the console.

The following code provides a JSON configuration example:

"storage": [
        {
            "dataset": {
                "id": "d-pcsah1t86bm8******",
                "version": "v1",
                "read_only": false
            },
            "mount_path": "/mnt/data/"
        }
    ]

Parameters:

Mount code

• Git mount: Mounts code from a Git repository to a service instance in read-only mode. The code is pulled when the instance starts. If you change the code, you must restart the service instance for the changes to take effect.

• Code configuration: You can create a PAI code set from a Git repository and register it as a public AI Asset. This lets you easily mount and reference the code set when you deploy a service.

{
  "storage": [
        {
            "git": {
                "repo": "https://codeup.aliyun.com/xxx/eas/aitest.git",
                "branch": "master",
                "commit": "xxx",
                "username": "username",
                "password": "password or access token"
            },
            "mount_path": "/mnt/data/"
        }
    ]
}

"storage": [
        {
            "code": {
                "id": "code-4d3b42a1152****"
            },
            "mount_path": "/mnt/data/"
        }
    ]

Mount a PAI model

You can register the models that a service requires as public PAI AI Assets to centrally manage them in PAI. For more information, see Register and manage models. You can then mount the models to service instances in read-only mode during service deployment.

The following figure shows the configuration interface in the console.

Parameters:

Other mounts

• Image mount: This method does not mount a directory from a Docker image directly to a service instance. Instead, it first copies files from a specified directory in the source image to a local shared volume and then mounts the local shared volume at runtime. If the directory to be mounted contains a large amount of data, this process can consume significant local storage and increase the instance startup time. Use this feature only when necessary.

• EmptyDir mount: Use this method to allow an instance to read from and write to a local disk directory at runtime. This ensures that the written content is not lost if the instance unexpectedly restarts.

{
    "storage": [
        {
            "empty_dir": {
              "medium": "memory",
              "size_limit": 20
            },
            "mount_path": "/dev/shm"
        }
    ]
}

Complete JSON configuration example

The following code provides a complete configuration example. For more information, see Deploy services using JSON.

{
    "name": "service_name",
    "model_path": "http://path/to/model",
    "processor": "pmml",
    "storage": [
        {
            "oss": {
                "path": "oss://bucket/path/",
                "readOnly": false
            },
            "mount_path": "/mnt/data/"
        },
        {
            "nfs": {
                "path": "/",
                "server": "06ba74****-a****.cn-hangzhou.nas.aliyuncs.com",
                "readOnly": false
            },
            "mount_path": "/mnt/data/"
        },
        {
            "image": {
                "image": "registry-vpc.cn-shanghai.aliyuncs.com/eas/test_image:v1",
                "path": "/path/to/mount/"
            },
            "mount_path": "/data_image"
        },
        {
            "empty_dir": {
              "medium": "memory",
              "size_limit": 20
            },
            "mount_path": "/dev/shm"
        },
        {
            "git": {
                "repo": "https://codeup.aliyun.com/xxx/eas/aitest.git",
                "branch": "master",
                "commit": "xxx",
                "username": "username",
                "password": "password or access token"
            },
            "mount_path": "/mnt/data/"
        }
    ],
    "metadata": {
        "cpu": 1,
        "instance": 1,
        "resource": "eas-r-xxx"
    }
}

FAQ

1. Why am I getting a "file not found" error after mounting an OSS bucket?

   This error is usually caused by an incorrect path. Check your mount configuration and the access path.

   For example, if you mount oss://my-bucket/ to /mnt/data, the file oss://my-bucket/subfolder/myfile.txt in OSS must be accessed using the path /mnt/data/subfolder/myfile.txt in the container, not /mnt/data/myfile.txt.

2. Can I use an FTP or SSH tool to connect to an instance to upload or download files?

   No. EAS instances do not provide FTP or SSH access.

3. Can I mount an OSS bucket that is in a different region from my EAS service?

   No. EAS cannot mount OSS buckets across regions. Use the cross-region replication feature of OSS to synchronize data to an OSS bucket in the same region as the EAS service.

4. If no storage is mounted, where are the files that I download to the instance stored? The files are saved to the system disk of the instance. The system disk is cleared when the instance is restarted or updated, resulting in data loss. To make your data persistent, you must mount storage.