Cloud storage selection and integration guide - Container Compute Service - Alibaba Cloud - Container Compute Service

ACS integrates Alibaba Cloud block storage, File Storage NAS (NAS), Cloud Parallel File Storage (CPFS), and Object Storage Service (OSS) through the Kubernetes Container Storage Interface (CSI). These services work alongside native Kubernetes volume types such as EmptyDir and ConfigMap. Choose the storage type that fits your workload.

Storage basics

Volume

Files in a container are stored on disk only temporarily, which causes the following problems for the applications that run in the container:

When a pod runs multiple containers at the same time, the containers cannot share files.
When a container crashes, the files generated while the container was running are not retained after the container is recreated, and the data is lost.

To solve these problems, Kubernetes provides the Volume abstraction, which provides shared and persistent storage.

A volume is part of a pod and can be defined only in a pod. It cannot be created independently. All containers in a pod can use the volume, but you must mount the volume to a directory in the container.
A volume is the channel through which a pod transfers data to and from an external storage device. It is also how data is shared between containers in a pod, between pods, and between a pod and the external environment.
Volumes have several types, such as EmptyDir, ConfigMap, and PersistentVolumeClaim (PVC).
- EmptyDir: A temporary volume created when a pod starts. Data is deleted when the pod is removed.
- ConfigMap: Mounts configuration data as files into a container.
- PersistentVolumeClaim (PVC): Requests persistent storage from a remote storage backend. Data survives container restarts and pod rescheduling.

For more information, see Volumes.

PV and PVC

Not all volumes are persistent. To provide persistent storage, container storage relies on a remote storage service. To this end, Kubernetes introduced two resource objects, PV and PVC, to define and use storage resources:

PersistentVolume (PV): An abstract volume that represents storage provisioned from a remote backend such as a cloud disk, a NAS file system, or an OSS bucket. A PV exists independently of pods and can be shared by multiple pods. A cluster administrator can create a PV in advance, or the system can create one dynamically through a StorageClass.
PersistentVolumeClaim (PVC): A request for storage that specifies the capacity and access mode. A PVC maps to a PV in a one-to-one relationship. To use persistent storage, associate a PVC with a volume in the pod.

For more information, see Persistent Volumes and Storage Classes.

Prerequisites

To mount, unmount, create, and delete volumes, the CSI plugin requires permissions to access resources in other Alibaba Cloud products. The plugin uses RAM role authorization to request the creation of the AliyunCCCSIPluginRole role, which grants the plugin access to your resources in other cloud products. Make sure this role is configured before you mount cloud storage.

For more information, see Authorization overview.

Choose cloud storage

Alibaba Cloud provides low-cost, highly reliable, and highly available block, file, and object storage services for container workloads.Choose a storage type based on the compute type and storage requirements of your workload, and consider factors such as data volume, data access frequency, IOPS, and throughput.

Scenario-to-storage mapping

The following list maps common workload scenarios to the recommended storage type. Use this as a starting point, then refer to the comparison table for detailed feature information.

Databases and middleware that require low-latency, non-shared block storage — Use block storage. A cloud disk delivers low latency and high performance for I/O-intensive applications. Block storage is non-shared and can be mounted to only one pod at a time.
Shared data between pods, big data analytics, web applications, and log persistence — Use NAS. NAS is a shared file system that supports concurrent access from multiple pods, high throughput, and online file modification.
AI training, autonomous driving, genetic computing, and film and television rendering — Use CPFS General-purpose Edition. This fully managed parallel file system delivers the high throughput and high IOPS required by high-performance computing (HPC) workloads.
AI-generated content (AIGC) and other intelligent computing workloads that require ultra-high throughput with Remote Direct Memory Access (RDMA) — Use CPFS for Lingjun. This parallel file system supports end-to-end RDMA networks and is purpose-built for intelligent computing. CPFS for Lingjun is in invitational preview.
Read-mostly unstructured data, high-concurrency batch processing, and data with demanding authentication and disaster recovery requirements — Use OSS. OSS provides low-cost, massive shared storage for configuration files, images, audio and video files, AI inference, and data analytics.

Cloud storage support by compute type

CPU-based ACS pods (general-purpose and performance-optimized types): All cloud storage types that ACS supports can be mounted.
GPU-based ACS pods (GPU and GPU-HPN types):
- Supported by default: NAS and OSS.
- Limited support: Only some GPU models support block storage and CPFS for Lingjun. For more information, [submit a ticket on the international site].

Comparison table

Cloud storage	Features	Common scenarios	Selection guide
Block storage	Non-shared, low-latency, and highly reliable block-level random access storage. Similar to a physical hard disk, it supports partitioning, formatting, and creating a file system.	High I/O and low latency A cloud disk delivers low latency and high performance, which suits applications with demanding I/O and latency requirements, such as databases and middleware. Non-shared access A cloud disk is non-shared storage and can be mounted to only one pod.	The following types of block storage are supported: `cloud_essd_entry`: ESSD Entry disk `cloud_auto`: ESSD AutoPL disk `cloud_essd` (default): ESSD `cloud_ssd`: standard SSD `cloud_efficiency`: ultra disk `cloud_essd_entry`: ESSD Entry disk `cloud_auto`: ESSD AutoPL disk `cloud_essd` (default): ESSD `cloud_ssd`: standard SSD `cloud_efficiency`: ultra disk You can choose a disk type based on your billing and performance needs. For more information, see Block Storage Pricing and Block storage performance.
File Storage NAS	A distributed file system that provides shared access, elastic scaling, high reliability, and high performance. It delivers high throughput and high IOPS, and supports random reads and writes and online file modification.	Shared data NAS is shared storage, so multiple pods can access the same data at the same time. Data in NAS is not cleared when a pod is deleted, which makes NAS suitable for sharing data between pods. Big data analytics NAS provides high data throughput, which meets the shared storage access requirements of large-scale batch jobs. Web applications NAS provides storage support for web applications and content management systems. Log persistenceTo store logs persistently, use NAS.	General-purpose NAS and Extreme NAS are supported. Choose a NAS type and storage specification based on factors such as NAS costs and performance. For more information, see General-purpose NAS, Extreme NAS, and Selection guide.
CPFS General-purpose Edition	A fully managed and scalable parallel file system that delivers high throughput and high IOPS.	High-performance computing scenarios such as AI training and genetic computing CPFS General-purpose Edition delivers high throughput and high IOPS, which meets the requirements of high-performance computing (HPC) scenarios such as AI training, autonomous driving, genetic computing, and film and television rendering.	Choose a specification based on factors such as the cost and performance of CPFS General-purpose Edition. For more information, see Product specifications.
CPFS for Lingjun (invitational preview)	Designed for intelligent computing workloads, it delivers ultra-high throughput and IOPS and supports end-to-end RDMA networks.	Intelligent computing scenarios such as AIGC and autonomous driving CPFS for Lingjun delivers ultra-high throughput and IOPS and supports end-to-end RDMA networks, which meets the requirements of intelligent computing workloads.	CPFS for Lingjun is in invitational preview and is currently available only in some regions and zones. For more information, see CPFS for Lingjun.
OSS	Low-cost and massive shared storage space, which suits data that is seldom modified after it is written.	Shared data OSS is shared storage, so multiple pods can access the same data at the same time. Data in OSS is not cleared when a pod is deleted, which makes OSS suitable for sharing data between pods. Read-mostly scenarios for configuration files, images, videos, and other media files OSS suits unstructured data such as configuration files, images, audio and video files, and any other object that does not require frequent modification. If your business needs to change file content, select an ossfs 1.0 volume for better POSIX compatibility. High-concurrency batch processing scenarios such as AI inference and data analytics The high server-side bandwidth of OSS suits high-concurrency batch processing, such as AI training, data analytics, autonomous driving, and other compute-intensive workloads and workflows. These workloads mainly involve sequential and random reads and sequential (append-only) writes. Select an ossfs 2.0 volume for higher throughput. Data security scenarios with demanding authentication and disaster recovery requirements OSS supports zone-redundant disaster recovery across three zones in the same region and server-side encryption capabilities such as Key Management Service (KMS). It also supports MD5 verification to guarantee data integrity during uploads and downloads, and it supports two object-level authentication methods: RAM and OSS bucket policies.	Choose a StorageClass based on factors such as OSS billing and data access frequency. For more information, see Selection guide and StorageClass reference.

Mount cloud storage

CSI is the storage plugin implementation recommended by the Kubernetes community. ACS provides csi-provisioner as the CSI add-on for ACS clusters. This plugin supports volumes based on Alibaba Cloud storage resources.

To mount cloud storage, create a PV that describes an Alibaba Cloud storage resource such as a cloud disk or a NAS file system, bind the PV to a PVC, and then declare the PVC in the volume definition of a pod. Based on how the PV is created, cloud storage can be mounted in two ways:

Static provisioning: You create a PV from an existing cloud storage resource and bind it directly in a PVC. This method guarantees that the PV is ready before the container starts, and it suits scenarios where the storage resource already exists.
Dynamic provisioning: You use a StorageClass to define the cloud storage resource to create, and then associate the StorageClass in a PVC. The system automatically creates and binds a PV based on the PVC and StorageClass configurations. This method does not require you to create a PV in advance, which makes it more flexible and automated.

The following topics describe how to mount each type of volume:

Next steps

In addition to persistent storage, a cloud disk can also serve as an ephemeral volume that stores temporary data. Ephemeral volumes are created and deleted together with the pod. For the detailed procedure, see Mount an ephemeral volume.
By default, each ACS pod provides 30 GiB of free ephemeral storage. If this capacity does not meet your requirements, you can increase the ephemeral storage space as needed.