将 Snowflake 暂存区卷与服务搭配使用¶

Step 1: Define a stage volume¶

To define a stage volume, add the spec.volumes field in the service specification as shown in the following example. The example shows only the minimum required fields:

spec:
  containers:
    ..
  volumes:
  - name: <name>
    source: stage
    stageConfig:
       name: <stage_name>

The following list defines the fields from the example:

• name: Provides the name of the volume.

• source: Identifies the type of the volume (stage).

• stageConfig.name: Identifies the Snowflake internal stage or folder on a stage to mount; for example @my_stage, @my_stage/folder, or @my_db.my_schema.my_stage/folder/nestedfolder. Double quotes must surround this value.

You can include the following optional fields in stageConfig:

• stageConfig.resources field: The Snowflake component that provides the mounted stage volume requires CPU and memory resources. Use this field to specify these CPU and memory requirements, similar to the resource specifications for your application containers. For more information, see containers.resources 字段 fields. If this field isn't specified, the following default resource settings apply:

  • resources.requests.cpu: 0

  • resources.requests.memory: 0.5Gi

  • resources.limits.cpu: 0.5

  • resources.limits.memory: 1Gi

  For most applications with typical data traffic to stage volumes, you don't need to set this field, because the default resource settings should be sufficient. However, if your application performs a high volume of reads and writes, the default settings might lead to performance constraints or read/write failures. For more information, see Common guidelines for both implementations of stage volumes.

  To avoid these problems, check the CPU and memory metrics for the container (stage-mount-v2-sidecar-<stage-volume-name>). Snowflake adds this container to your service that provides the implementation of the stage volume you configured. This lets you to see exactly what resources your stage volume is using and determine if it is constrained by CPU or memory. Use these metrics to update the CPU and memory limits as needed.

• stageConfig.metadataCache field: If your application frequently retrieves file metadata or lists files on a Snowflake stage in a loop, and you don't expect the data to change often, you can enable metadata caching for the Snowflake stage storage volume to significantly improve performance. The cache stores this metadata for a specified time period, after which it is cleared. If the application then tries to access the metadata, Snowflake refreshes the cache. Use the hours, minutes, and seconds units to specify the metadataCache. For example 90s, 5m, 1h, 1h30m, 1h30m45s. If not specified, there is no caching.

  备注

  Configure metadata caching only when the data in your Snowflake stage doesn't change for service lifetime; for example, a service that has read-only workloads that need to work on a static set of data in the stage. Don't configure metadata caching for workloads where data in your Snowflake stage is updated while the service is running.

The following spec.volumes field defines a Snowflake stage volume. The field includes the optional stageConfig fields:

spec:
  containers:
    ..
  volumes:
  - name: <name>
    source: stage
    stageConfig:
        name: <stage_name>
        metadataCache: 1h
        resources:
            requests:
                cpu: 0.35
                memory: 0.4Gi
            limits:
                cpu: 2.0
                memory: 1Gi

Step 2: Specify where to mount the stage volume in the container¶

After you define a Snowflake stage storage volume by adding the spec.volumes field, use the spec.containers.volumeMounts field to describe where to mount the stage volume in your application containers, as shown in the following example:

spec:
  containers:
  name: ...
  image: ...
  volumeMounts:
  - name: <name>
    mountPath: <absolute_directory_path>

The information you provide in this field is consistent across all supported storage volume types and applies to both implementations of stage volumes.

Migrating code from the deprecated implementation¶

The newer implementation replaces the deprecated implementation, with the following behavioral changes:

• The newer stage-volume implementation streams file contents directly to and from cloud storage, ensuring that you always get the latest contents. This provides predictable behavior and significantly faster throughput. The deprecated stage-volume implementation caches chunks of file data, making it difficult to know if you are reading the latest data.

• Random read performance might be lower with the new implementation because of the removal of caching. However, without a local disk cache, consistency across volumes is improved. File changes are written directly to the backing stage when the file is closed, with no local disk buffering.

• Reads always return the latest data, making this configuration better for sharing data between services.

• Random writes or file appends aren't supported. Attempting these operations results in an error. Snowflake recommends that you use block storage volumes for these workloads.

Specify a Snowflake stage volume in a service specification (deprecated)¶

To create a service where service containers use Snowflake stage volume, specify the required settings in the service specification as shown in the following steps:

1. To specify the stage volume, use the spec.volumes field as shown in the following example:

   volumes:
   - name: <name>
     source: <stage_name>
   

   Copy

   以下字段为必填字段：

   • name: The name of the volume.

   • source: The Snowflake internal stage or folder on the stage to mount; for example @my_stage, @my_stage/folder. Quotes must surround this value.

2. To describe where to mount the stage volume in your application containers, use the spec.containers.volumeMounts field, as shown in the following example:

   volumeMounts:
   - name: <name>
     mountPath: <absolute_directory_path>
   

   Copy

   The information you provide in this field is consistent across all supported storage volume types and applies to both implementations of stage volumes.

Example (deprecated)¶

In the example service specification, the app container mounts an internal stage @model_stage by using the deprecated stage volume implementations:

spec:
  containers:
  - name: app
    image: <image1-name>
    volumeMounts:
    - name: models-legacy
      mountPath: /opt/model-legacy
  volumes:
  - name: models-legacy
    source: "@model_stage"

The volumeMounts field specifies where inside the container to mount the stage volume. This specification remains the same for both the stage volume implementations.

Common guidelines for both implementations of stage volumes¶

• 暂存区挂载针对顺序读写进行了优化。

• 暂存区挂载 I/O 操作的延迟可能高于容器文件系统和块存储卷上的 I/O 操作。您应始终检查 I/O 操作的状态码，以确保它们成功。

• Stage mounts upload file updates asynchronously. Changes to files on a stage mount are only guaranteed to be persisted to the stage after the file descriptor is successfully closed or flushed. There might be a delay before the changes to files on a stage mount become visible to other containers and Snowflake.

• 已挂载暂存区中的每个目录应包含少于 100,000 个文件。预计 readdir 延迟会随着目录中文件数量的增加而增加。

Guidelines when using the deprecated version of the stage volume implementation¶

• 避免在暂存区挂载中同时写入多个文件。

• Stage mount isn't a network file system. Don't use stage mounts for multi-client coordination.

• 不要同时打开同一个文件的多个句柄。使用打开的文件句柄进行读取或写入操作，但不能将两者混用。要在写入文件后读取文件，请关闭该文件，然后在读取之前重新打开该文件。

Guidelines when using the generally available stage volume implementation¶

• Concurrent writes to the same file from multiple stage mounts --- same stage volume mounted on different containers --- aren't recommended.

• The absence of a local disk cache improves consistency across mounts. File changes are flushed directly to the backing stage upon closing the file, with no local disk buffering. Reads always return the latest data, making the new stage mount better for sharing data between services.

• Read and write data in large, contiguous chunks for optimal performance. The performance penalty for small reads and writes when compared to the generally available stage volume implementation, can mitigate the performance gains from the new implementation.

Common limitations for both implementations of stage volumes¶

• You can only mount a stage or a subdirectory in a stage; for example, @my_stage, @my_stage/folder. You can't mount a single file in a stage; for example, @my_stage/folder/file.

• External stages aren't supported. Only Snowflake internal stages are supported.

• 暂存区挂载不是完全 POSIX 兼容的文件系统。例如：

  • 文件和目录重命名不是原子的。

  • 不支持硬链接。

• The Linux kernel subsystem inode notify (inotify) that monitors changes to file systems doesn't work on stage mounts.

Limitations when using the deprecated version of the stage volume implementation¶

• A maximum of 5 stage volumes is allowed per service. For more information, see spec.volumes.

• 每个计算池节点最多支持 8 个暂存区卷。Snowflake 管理每个节点的暂存区挂载限制，与管理内存、CPU 和 GPU 的方式类似。当现有节点无法支持所请求的暂存区挂载时，启动新的服务实例可能会导致 Snowflake 启动新节点。

• The stage volume capabilities vary depending on the cloud platform for your Snowflake account:

  • Accounts on AWS support internal stages with both SNOWFLAKE_FULL and SNOWFLAKE_SSE stage encryption. For more information, see Internal stage parameters.

  • Accounts on Azure currently support internal stages with SNOWFLAKE_SSE encryption. When you run CREATE STAGE, use the ENCRYPTION parameter to specify the encryption type: CREATE STAGE my_stage ENCRYPTION = (TYPE = 'SNOWFLAKE_SSE');

  • Accounts on Google Cloud aren't supported.

• Concurrent writes to the same file from multiple stage mounts --- that is, the same stage volume mounted on different containers --- aren't supported.

Limitations when using the generally available version of the stage volume implementation¶

• Random writes, and file appends aren't supported.

• Each stage that is mounted requires 512 MB memory per stage. This means that there is a limitation on the number of stage volumes that can be used based on instance size. Mounting the volume on multiple containers doesn't increase memory consumption.

• A maximum of 20 stage volumes are allowed per service. For more information, see spec.volumes.