The keywords "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and "OPTIONAL" are to be interpreted as described in [RFC 2119](http://tools.ietf.org/html/rfc2119) (Bradner, S., "Key words for use in RFCs to Indicate Requirement Levels", BCP 14, RFC 2119, March 1997).
The key words "unspecified", "undefined", and "implementation-defined" are to be interpreted as described in the [rationale for the C99 standard](http://www.open-std.org/jtc1/sc22/wg14/www/C99RationaleV5.10.pdf#page=18).
An implementation is not compliant if it fails to satisfy one or more of the MUST, REQUIRED, or SHALL requirements for the protocols it implements.
An implementation is compliant if it satisfies all the MUST, REQUIRED, and SHALL requirements for the protocols it implements.
| Node | A host where the user workload will be running, uniquely identifiable from the perspective of a Plugin by a node ID. |
| Plugin | Aka “plugin implementation”, a gRPC endpoint that implements the CSI Services. |
| Plugin Supervisor | Process that governs the lifecycle of a Plugin, MAY be the CO. |
| Workload | The atomic unit of "work" scheduled by a CO. This may be a container or a collection of containers. |
## Objective
To define an industry standard “Container Storage Interface” (CSI) that will enable storage vendors (SP) to develop a plugin once and have it work across a number of container orchestration (CO) systems.
### Goals in MVP
The Container Storage Interface (CSI) will
* Enable SP authors to write one CSI compliant Plugin that “just works” across all COs that implement CSI.
* Define API (RPCs) that enable:
* Dynamic provisioning and deprovisioning of a volume.
* Attaching or detaching a volume from a node.
* Mounting/unmounting a volume from a node.
* Consumption of both block and mountable volumes.
* Local storage providers (e.g., device mapper, lvm).
* Creating and deleting a snapshot (source of the snapshot is a volume).
* Provisioning a new volume from a snapshot (reverting snapshot, where data in the original volume is erased and replaced with data in the snapshot, is out of scope).
* Describe a process by which a Supervisor configures a Plugin.
* Container deployment considerations (`CAP_SYS_ADMIN`, mount namespace, etc.).
### Non-Goals in MVP
The Container Storage Interface (CSI) explicitly will not define, provide, or dictate in v0.1:
* Specific mechanisms by which a Plugin Supervisor manages the lifecycle of a Plugin, including:
* How to maintain state (e.g. what is attached, mounted, etc.).
* How to deploy, install, upgrade, uninstall, monitor, or respawn (in case of unexpected termination) Plugins.
* A first class message structure/field to represent "grades of storage" (aka "storage class").
* Protocol-level authentication and authorization.
* Packaging of a Plugin.
* POSIX compliance: CSI provides no guarantee that volumes provided are POSIX compliant filesystems.
Compliance is determined by the Plugin implementation (and any backend storage system(s) upon which it depends).
CSI SHALL NOT obstruct a Plugin Supervisor or CO from interacting with Plugin-managed volumes in a POSIX-compliant manner.
## Solution Overview
This specification defines an interface along with the minimum operational and packaging recommendations for a storage provider (SP) to implement a CSI compatible plugin.
The interface declares the RPCs that a plugin must expose: this is the **primary focus** of the CSI specification.
Any operational and packaging recommendations offer additional guidance to promote cross-CO compatibility.
### Architecture
The primary focus of this specification is on the **protocol** between a CO and a Plugin.
It SHOULD be possible to ship cross-CO compatible Plugins for a variety of deployment architectures.
A CO should be equipped to handle both centralized and headless plugins, as well as split-component and unified plugins.
Several of these possibilities are illustrated in the following figures.
```
CO "Master" Host
+-------------------------------------------+
| |
| +------------+ +------------+ |
| | CO | gRPC | Controller | |
| | +-----------> Plugin | |
| +------------+ +------------+ |
| |
+-------------------------------------------+
CO "Node" Host(s)
+-------------------------------------------+
| |
| +------------+ +------------+ |
| | CO | gRPC | Node | |
| | +-----------> Plugin | |
| +------------+ +------------+ |
| |
+-------------------------------------------+
Figure 1: The Plugin runs on all nodes in the cluster: a centralized
Controller Plugin is available on the CO master host and the Node
Plugin is available on all of the CO Nodes.
```
```
CO "Node" Host(s)
+-------------------------------------------+
| |
| +------------+ +------------+ |
| | CO | gRPC | Controller | |
| | +--+--------> Plugin | |
| +------------+ | +------------+ |
| | |
| | |
| | +------------+ |
| | | Node | |
| +--------> Plugin | |
| +------------+ |
| |
+-------------------------------------------+
Figure 2: Headless Plugin deployment, only the CO Node hosts run
Plugins. Separate, split-component Plugins supply the Controller
Service and the Node Service respectively.
```
```
CO "Node" Host(s)
+-------------------------------------------+
| |
| +------------+ +------------+ |
| | CO | gRPC | Controller | |
| | +-----------> Node | |
| +------------+ | Plugin | |
| +------------+ |
| |
+-------------------------------------------+
Figure 3: Headless Plugin deployment, only the CO Node hosts run
Plugins. A unified Plugin component supplies both the Controller
them via the capabilities API. Interactions with the volumes of such
plugins is reduced to `NodePublishVolume` and `NodeUnpublishVolume`
calls.
```
The above diagrams illustrate a general expectation with respect to how a CO MAY manage the lifecycle of a volume via the API presented in this specification.
Plugins should expose all RPCs for an interface: Controller plugins should implement all RPCs for the `Controller` service.
Unsupported RPCs should return an appropriate error code that indicates such (e.g. `CALL_NOT_IMPLEMENTED`).
The full list of plugin capabilities is documented in the `ControllerGetCapabilities` and `NodeGetCapabilities` RPCs.
## Container Storage Interface
This section describes the interface between COs and Plugins.
### RPC Interface
A CO interacts with an Plugin through RPCs.
Each SP MUST provide:
* **Node Plugin**: A gRPC endpoint serving CSI RPCs that MUST be run on the Node whereupon an SP-provisioned volume will be published.
* **Controller Plugin**: A gRPC endpoint serving CSI RPCs that MAY be run anywhere.
* In some circumstances a single gRPC endpoint MAY serve all CSI RPCs (see Figure 3 in [Architecture](#architecture)).
In general the Cluster Orchestrator (CO) is responsible for ensuring that there is no more than one call “in-flight” per volume at a given time.
However, in some circumstances, the CO may lose state (for example when the CO crashes and restarts), and may issue multiple calls simultaneously for the same volume.
The plugin should handle this as gracefully as possible.
The error code `OPERATION_PENDING_FOR_VOLUME` may be returned by the plugin in this case (see general error code section for details).
The requirements documented herein apply equally and without exception, unless otherwise noted, for the fields of all protobuf message types defined by this specification.
Violation of these requirements MAY result in RPC message data that is not compatible with all CO, Plugin, and/or CSI middleware implementations.
* A field noted as `OPTIONAL` MAY be specified and the specification SHALL clearly define expected behavior for the default, zero-value of such fields.
All CSI API calls defined in this spec MUST return a [standard gRPC status](https://github.com/grpc/grpc/blob/master/src/proto/grpc/status/status.proto).
Most gRPC libraries provide helper methods to set and read the status fields.
The status `code` MUST contain a [canonical error code](https://github.com/grpc/grpc-go/blob/master/codes/codes.go). COs must handle all valid error codes. Each RPC defines a set of gRPC error codes that MUST be returned by the plugin when specified conditions are encountered. In addition to those, if the conditions defined below are encountered, the plugin MUST return the associated gRPC error code.
| Missing required field | 3 INVALID_ARGUMENT | Indicates that a required field is missing from the request. More human-readable information MAY be provided in the `error_description` field. | Caller MUST fix the request by adding the missing required field before retrying. |
| Invalid or unsupported field in the request | 3 INVALID_ARGUMENT | Indicates that the one ore more fields in this field is either not allowed by the Plugin or has an invalid value. More human-readable information MAY be provided in the gRPC `status.message` field. | Caller MUST fix the field before retrying. |
The status `message` MUST contain a human readable description of error, if the status `code` is not `OK`.
This string MAY be surfaced by CO to end users.
The status `details` MUST be empty. In the future, this spec may require `details` to return a machine-parsable protobuf message if the status `code` is not `OK` to enable CO's to implement smarter error handling and fault resolution.
Secrets may be required by plugin to complete a RPC request.
A secret is a string to string map where the key identifies the name of the secret (e.g. "username" or "password"), and the value contains the secret data (e.g. "bob" or "abc123").
Each key MUST consist of alphanumeric characters, '-', '_' or '.'.
Each value MUST contain a valid string.
An SP MAY choose to accept binary (non-string) data by using a binary-to-text encoding scheme, like base64.
An SP SHALL advertise the requirements for required secret keys and values in documentation.
CO SHALL permit passing through the required secrets.
A CO MAY pass the same secrets to all RPCs, therefore the keys for all unique secrets that an SP expects must be unique across all CSI operations.
This information is sensitive and MUST be treated as such (not logged, etc.) by the CO.
This REQUIRED RPC allows the CO to query the supported capabilities of the Plugin "as a whole": it is the grand sum of all capabilities of all instances of the Plugin software, as it is intended to be deployed.
All instances of the same version (see `vendor_version` of `GetPluginInfoResponse`) of the Plugin SHALL return the same set of capabilities, regardless of both: (a) where instances are deployed on the cluster as well as; (b) which RPCs an instance is serving.
If an unhealthy state is reported, via a non-success response, a CO MAY take action with the intent to bring the plugin to a healthy state.
Such actions MAY include, but SHALL NOT be limited to, the following:
* Restarting the plugin container, or
* Notifying the plugin supervisor.
The Plugin MAY verify that it has the right configurations, devices, dependencies and drivers in order to run and return a success if the validation succeeds.
The CO MAY invoke this RPC at any time.
A CO MAY invoke this call multiple times with the understanding that a plugin's implementation MAY NOT be trivial and there MAY be overhead incurred by such repeated calls.
The SP SHALL document guidance and known limitations regarding a particular Plugin's implementation of this RPC.
For example, the SP MAY document the maximum frequency at which its Probe implementation should be called.
| Plugin not healthy | 9 FAILED_PRECONDITION | Indicates that the plugin is not in a healthy/ready state. | Caller SHOULD assume the plugin is not healthy and that future RPCs MAY fail because of this condition. |
| Missing required dependency | 9 FAILED_PRECONDITION | Indicates that the plugin is missing one or more required dependency. | Caller MUST assume the plugin is not healthy. |
If a volume corresponding to the specified volume `name` already exists, is accessible from `accessibility_requirements`, and is compatible with the specified `capacity_range`, `volume_capabilities` and `parameters` in the `CreateVolumeRequest`, the Plugin MUST reply `0 OK` with the corresponding `CreateVolumeResponse`.
| Volume already exists but is incompatible | 6 ALREADY_EXISTS | Indicates that a volume corresponding to the specified volume `name` already exists but is incompatible with the specified `capacity_range`, `volume_capabilities` or `parameters`. | Caller MUST fix the arguments or use a different `name` before retrying. |
| Unable to provision in `accessible_topology` | 8 RESOURCE_EXHAUSTED | Indicates that although the `accessible_topology` field is valid, a new volume can not be provisioned with the specified topology constraints. More human-readable information MAY be provided in the gRPC `status.message` field. | Caller MUST ensure that whatever is preventing volumes from being provisioned in the specified location (e.g. quota issues) is addressed before retrying with exponential backoff. |
| Operation pending for volume | 10 ABORTED | Indicates that there is a already an operation pending for the specified volume. In general the Cluster Orchestrator (CO) is responsible for ensuring that there is no more than one call "in-flight" per volume at a given time. However, in some circumstances, the CO MAY lose state (for example when the CO crashes and restarts), and MAY issue multiple calls simultaneously for the same volume. The Plugin, SHOULD handle this as gracefully as possible, and MAY return this error code to reject secondary calls. | Caller SHOULD ensure that there are no other calls pending for the specified volume, and then retry with exponential back off. |
| Unsupported `capacity_range` | 11 OUT_OF_RANGE | Indicates that the capacity range is not allowed by the Plugin, for example when trying to create a volume smaller than the source snapshot. More human-readable information MAY be provided in the gRPC `status.message` field. | Caller MUST fix the capacity range before retrying. |
| Call not implemented | 12 UNIMPLEMENTED | CreateVolume call is not implemented by the plugin or disabled in the Plugin's current mode of operation. | Caller MUST NOT retry. Caller MAY call `ControllerGetCapabilities` or `NodeGetCapabilities` to discover Plugin capabilities. |
#### `DeleteVolume`
A Controller Plugin MUST implement this RPC call if it has `CREATE_DELETE_VOLUME` capability.
This RPC will be called by the CO to deprovision a volume.
If successful, the storage space associated with the volume MUST be released and all the data in the volume SHALL NOT be accessible anymore.
This operation MUST be idempotent.
If a volume corresponding to the specified `volume_id` does not exist or the artifacts associated with the volume do not exist anymore, the Plugin MUST reply `0 OK`.
| Volume in use | 9 FAILED_PRECONDITION | Indicates that the volume corresponding to the specified `volume_id` could not be deleted because it is in use by another resource. | Caller SHOULD ensure that there are no other resources using the volume, and then retry with exponential back off. |
| Operation pending for volume | 10 ABORTED | Indicates that there is a already an operation pending for the specified volume. In general the Cluster Orchestrator (CO) is responsible for ensuring that there is no more than one call "in-flight" per volume at a given time. However, in some circumstances, the CO MAY lose state (for example when the CO crashes and restarts), and MAY issue multiple calls simultaneously for the same volume. The Plugin, SHOULD handle this as gracefully as possible, and MAY return this error code to reject secondary calls. | Caller SHOULD ensure that there are no other calls pending for the specified volume, and then retry with exponential back off. |
| Call not implemented | 12 UNIMPLEMENTED | DeleteVolume call is not implemented by the plugin or disabled in the Plugin's current mode of operation. | Caller MUST NOT retry. Caller MAY call `ControllerGetCapabilities` or `NodeGetCapabilities` to discover Plugin capabilities. |
#### `ControllerPublishVolume`
A Controller Plugin MUST implement this RPC call if it has `PUBLISH_UNPUBLISH_VOLUME` controller capability.
This RPC will be called by the CO when it wants to place a workload that uses the volume onto a node.
The Plugin SHOULD perform the work that is necessary for making the volume available on the given node.
The Plugin MUST NOT assume that this RPC will be executed on the node where the volume will be used.
This operation MUST be idempotent.
If the volume corresponding to the `volume_id` has already been published at the node corresponding to the `node_id`, and is compatible with the specified `volume_capability` and `readonly` flag, the Plugin MUST reply `0 OK`.
If the operation failed or the CO does not know if the operation has failed or not, it MAY choose to call `ControllerPublishVolume` again or choose to call `ControllerUnpublishVolume`.
The CO MAY call this RPC for publishing a volume to multiple nodes if the volume has `MULTI_NODE` capability (i.e., `MULTI_NODE_READER_ONLY`, `MULTI_NODE_SINGLE_WRITER` or `MULTI_NODE_MULTI_WRITER`).
| Volume does not exist | 5 NOT_FOUND | Indicates that a volume corresponding to the specified `volume_id` does not exist. | Caller MUST verify that the `volume_id` is correct and that the volume is accessible and has not been deleted before retrying with exponential back off. |
| Node does not exist | 5 NOT_FOUND | Indicates that a node corresponding to the specified `node_id` does not exist. | Caller MUST verify that the `node_id` is correct and that the node is available and has not been terminated or deleted before retrying with exponential backoff. |
| Volume published but is incompatible | 6 ALREADY_EXISTS | Indicates that a volume corresponding to the specified `volume_id` has already been published at the node corresponding to the specified `volume_id` but is incompatible with the specified `volume_capability` or `readonly` flag . | Caller MUST fix the arguments before retying. |
| Volume published to another node | 9 FAILED_PRECONDITION | Indicates that a volume corresponding to the specified `volume_id` has already been published at another node and does not have MULTI_NODE volume capability. If this error code is returned, the Plugin SHOULD specify the `node_id` of the node at which the volume is published as part of the gRPC `status.message`. | Caller SHOULD ensure the specified volume is not published at any other node before retrying with exponential back off. |
| Max volumes attached | 8 RESOURCE_EXHAUSTED | Indicates that the maximum supported number of volumes that can be attached to the specified node are already attached. Therefore, this operation will fail until at least one of the existing attached volumes is detached from the node. | Caller MUST ensure that the number of volumes already attached to the node is less then the maximum supported number of volumes before retrying with exponential backoff. |
| Operation pending for volume | 10 ABORTED | Indicates that there is a already an operation pending for the specified volume. In general the Cluster Orchestrator (CO) is responsible for ensuring that there is no more than one call "in-flight" per volume at a given time. However, in some circumstances, the CO MAY lose state (for example when the CO crashes and restarts), and MAY issue multiple calls simultaneously for the same volume. The Plugin, SHOULD handle this as gracefully as possible, and MAY return this error code to reject secondary calls. | Caller SHOULD ensure that there are no other calls pending for the specified volume, and then retry with exponential back off. |
| Call not implemented | 12 UNIMPLEMENTED | ControllerPublishVolume call is not implemented by the plugin or disabled in the Plugin's current mode of operation. | Caller MUST NOT retry. Caller MAY call `ControllerGetCapabilities` or `NodeGetCapabilities` to discover Plugin capabilities. |
#### `ControllerUnpublishVolume`
Controller Plugin MUST implement this RPC call if it has `PUBLISH_UNPUBLISH_VOLUME` controller capability.
This RPC is a reverse operation of `ControllerPublishVolume`.
The Plugin SHOULD perform the work that is necessary for making the volume ready to be consumed by a different node.
The Plugin MUST NOT assume that this RPC will be executed on the node where the volume was previously used.
This RPC is typically called by the CO when the workload using the volume is being moved to a different node, or all the workload using the volume on a node has finished.
This operation MUST be idempotent.
If the volume corresponding to the `volume_id` is not attached to the node corresponding to the `node_id`, the Plugin MUST reply `0 OK`.
If this operation failed, or the CO does not know if the operation failed or not, it can choose to call `ControllerUnpublishVolume` again.
| Volume does not exist | 5 NOT_FOUND | Indicates that a volume corresponding to the specified `volume_id` does not exist. | Caller MUST verify that the `volume_id` is correct and that the volume is accessible and has not been deleted before retrying with exponential back off. |
| Node does not exist | 5 NOT_FOUND | Indicates that a node corresponding to the specified `node_id` does not exist. | Caller MUST verify that the `node_id` is correct and that the node is available and has not been terminated or deleted before retrying with exponential backoff. |
| Operation pending for volume | 10 ABORTED | Indicates that there is a already an operation pending for the specified volume. In general the Cluster Orchestrator (CO) is responsible for ensuring that there is no more than one call "in-flight" per volume at a given time. However, in some circumstances, the CO MAY lose state (for example when the CO crashes and restarts), and MAY issue multiple calls simultaneously for the same volume. The Plugin, SHOULD handle this as gracefully as possible, and MAY return this error code to reject secondary calls. | Caller SHOULD ensure that there are no other calls pending for the specified volume, and then retry with exponential back off. |
| Call not implemented | 12 UNIMPLEMENTED | ControllerUnpublishVolume call is not implemented by the plugin or disabled in the Plugin's current mode of operation. | Caller MUST NOT retry. Caller MAY call `ControllerGetCapabilities` or `NodeGetCapabilities` to discover Plugin capabilities. |
#### `ValidateVolumeCapabilities`
A Controller Plugin MUST implement this RPC call.
This RPC will be called by the CO to check if a pre-provisioned volume has all the capabilities that the CO wants.
This RPC call SHALL return `supported` only if all the volume capabilities specified in the request are supported.
This operation MUST be idempotent.
```protobuf
message ValidateVolumeCapabilitiesRequest {
// The ID of the volume to check. This field is REQUIRED.
| Volume does not exist | 5 NOT_FOUND | Indicates that a volume corresponding to the specified `volume_id` does not exist. | Caller MUST verify that the `volume_id` is correct and that the volume is accessible and has not been deleted before retrying with exponential back off. |
| Invalid `starting_token` | 10 ABORTED | Indicates that `starting_token` is not valid. | Caller SHOULD start the `ListVolumes` operation again with an empty `starting_token`. |
#### `GetCapacity`
A Controller Plugin MUST implement this RPC call if it has `GET_CAPACITY` controller capability.
The RPC allows the CO to query the capacity of the storage pool from which the controller provisions volumes.
```protobuf
message GetCapacityRequest {
// If specified, the Plugin SHALL report the capacity of the storage
// that can be used to provision volumes that satisfy ALL of the
// specified `volume_capabilities`. These are the same
// `volume_capabilities` the CO will use in `CreateVolumeRequest`.
If the plugin is unable to complete the GetCapacity call successfully, it MUST return a non-ok gRPC code in the gRPC status.
#### `ControllerGetCapabilities`
A Controller Plugin MUST implement this RPC call. This RPC allows the CO to check the supported capabilities of controller service provided by the Plugin.
A Controller Plugin MUST implement this RPC call if it has `CREATE_DELETE_SNAPSHOT` controller capability.
This RPC will be called by the CO to create a new snapshot from a source volume on behalf of a user.
This operation MUST be idempotent.
If a snapshot corresponding to the specified snapshot `name` is already successfully cut and uploaded (if upload is part of the process) and is compatible with the specified `source_volume_id` and `parameters` in the `CreateSnapshotRequest`, the Plugin MUST reply `0 OK` with the corresponding `CreateSnapshotResponse`.
If an error occurs before a snapshot is cut, `CreateSnapshot` SHOULD return a corresponding gRPC error code that reflects the error condition.
For plugins that implement snapshot uploads, `CreateSnapshot` SHOULD return `10 ABORTED`, a gRPC code that indicates the operation is pending for snapshot, during the snapshot uploading processs.
If an error occurs during the uploading process, `CreateSnapshot` SHOULD return a corresponding gRPC error code that reflects the error condition.
A snapshot MAY be used as the source to provision a new volume.
A CreateVolumeRequest message may specify an OPTIONAL source snapshot parameter.
Reverting a snapshot, where data in the original volume is erased and replaced with data in the snapshot, is an advanced functionality not every storage system can support and therefore is currently out of scope.
```protobuf
message CreateSnapshotRequest {
// The ID of the source volume to be snapshotted.
// This field is REQUIRED.
string source_volume_id = 1;
// The suggested name for the snapshot. This field is REQUIRED for
// idempotency.
string name = 2;
// Secrets required by plugin to complete snapshot creation request.
// This field is OPTIONAL. Refer to the `Secrets Requirements`
// section on how to use this field.
map<string,string> create_snapshot_secrets = 3;
// Plugin specific parameters passed in as opaque key-value pairs.
// This field is OPTIONAL. The Plugin is responsible for parsing and
// validating these parameters. COs will treat these as opaque.
// Use cases for opaque parameters:
// - Specify a policy to automatically clean up the snapshot.
// - Specify an expiration date for the snapshot.
// - Specify whether the snapshot is readonly or read/write.
// - Specify if the snapshot should be replicated to some place.
// - Specify primary or secondary for replication systems that
// support snapshotting only on primary.
map<string,string> parameters = 4;
}
message CreateSnapshotResponse {
// Contains all attributes of the newly created snapshot that are
// relevant to the CO along with information required by the Plugin
// to uniquely identify the snapshot. This field is REQUIRED.
Snapshot snapshot = 1;
}
// The information about a provisioned snapshot.
message Snapshot {
// This is the complete size of the snapshot in bytes. The purpose of
// this field is to give CO guidance on how much space is needed to
// create a volume from this snapshot. The size of the volume MUST NOT
// be less than the size of the source snapshot. This field is
// OPTIONAL. If this field is not set, it indicates that this size is
// unknown. The value of this field MUST NOT be negative and a size of
// zero means it is unspecified.
int64 size_bytes = 1;
// Uniquely identifies a snapshot and is generated by the plugin. It
// will not change over time. This field is REQUIRED. The identity
// information will be used by the CO in subsequent calls to refer to
// the provisioned snapshot.
string id = 2;
// Identity information for the source volume. Note that creating a
// snapshot from a snapshot is not supported here so the source has to
// be a volume. This field is REQUIRED.
string source_volume_id = 3;
// Timestamp when the point-in-time snapshot is taken on the storage
// system. The format of this field should be a Unix nanoseconds time
// encoded as an int64. On Unix, the command `date +%s%N` returns the
// current time in nanoseconds since 1970-01-01 00:00:00 UTC. This
// field is REQUIRED.
int64 created_at = 4;
// The status of a snapshot.
SnapshotStatus status = 5;
}
// The status of a snapshot.
message SnapshotStatus {
enum Type {
UNKNOWN = 0;
// A snapshot is ready for use.
READY = 1;
// A snapshot is cut and is now being uploaded.
// Some cloud providers and storage systems uploads the snapshot
// to the cloud after the snapshot is cut. During this phase,
// `thaw` can be done so the application can be running again if
// `freeze` was done before taking the snapshot.
UPLOADING = 2;
// An error occurred during the snapshot uploading process.
// This error status is specific for uploading because
// `CreateSnaphot` is a blocking call before the snapshot is
// cut and therefore it SHOULD NOT come back with an error
// status when an error occurs. Instead a gRPC error code SHALL
// be returned by `CreateSnapshot` when an error occurs before
// a snapshot is cut.
ERROR_UPLOADING = 3;
}
// This field is REQUIRED.
Type type = 1;
// Additional information to describe why a snapshot ended up in the
// `ERROR_UPLOADING` status. This field is OPTIONAL.
string details = 2;
}
```
##### CreateSnapshot Errors
If the plugin is unable to complete the CreateSnapshot call successfully, it MUST return a non-ok gRPC code in the gRPC status.
If the conditions defined below are encountered, the plugin MUST return the specified gRPC error code.
The CO MUST implement the specified error recovery behavior when it encounters the gRPC error code.
| Snapshot already exists but is incompatible | 6 ALREADY_EXISTS | Indicates that a snapshot corresponding to the specified snapshot `name` already exists but is incompatible with the specified `volume_id`. | Caller MUST fix the arguments or use a different `name` before retrying. |
| Operation pending for snapshot | 10 ABORTED | Indicates that there is a already an operation pending for the specified snapshot. In general the Cluster Orchestrator (CO) is responsible for ensuring that there is no more than one call "in-flight" per snapshot at a given time. However, in some circumstances, the CO MAY lose state (for example when the CO crashes and restarts), and MAY issue multiple calls simultaneously for the same snapshot. The Plugin, SHOULD handle this as gracefully as possible, and MAY return this error code to reject secondary calls. | Caller SHOULD ensure that there are no other calls pending for the specified snapshot, and then retry with exponential back off. |
| Call not implemented | 12 UNIMPLEMENTED | CreateSnapshot call is not implemented by the plugin or disabled in the Plugin's current mode of operation. | Caller MUST NOT retry. Caller MAY call `ControllerGetCapabilities` to discover Plugin capabilities. |
| Not enough space to create snapshot | 13 RESOURCE_EXHAUSTED | There is not enough space on the storage system to handle the create snapshot request. | Caller should fail this request. Future calls to CreateSnapshot may succeed if space is freed up. |
#### `DeleteSnapshot`
A Controller Plugin MUST implement this RPC call if it has `CREATE_DELETE_SNAPSHOT` capability.
This RPC will be called by the CO to delete a snapshot.
If successful, the storage space associated with the snapshot MUST be released and all the data in the snapshot SHALL NOT be accessible anymore.
This operation MUST be idempotent.
If a snapshot corresponding to the specified `snapshot_id` does not exist or the artifacts associated with the snapshot do not exist anymore, the Plugin MUST reply `0 OK`.
```protobuf
message DeleteSnapshotRequest {
// The ID of the snapshot to be deleted.
// This field is REQUIRED.
string snapshot_id = 1;
// Secrets required by plugin to complete snapshot deletion request.
// This field is OPTIONAL. Refer to the `Secrets Requirements`
// section on how to use this field.
map<string,string> delete_snapshot_secrets = 2;
}
message DeleteSnapshotResponse {}
```
##### DeleteSnapshot Errors
If the plugin is unable to complete the DeleteSnapshot call successfully, it MUST return a non-ok gRPC code in the gRPC status.
If the conditions defined below are encountered, the plugin MUST return the specified gRPC error code.
The CO MUST implement the specified error recovery behavior when it encounters the gRPC error code.
| Snapshot in use | 9 FAILED_PRECONDITION | Indicates that the snapshot corresponding to the specified `snapshot_id` could not be deleted because it is in use by another resource. | Caller SHOULD ensure that there are no other resources using the snapshot, and then retry with exponential back off. |
| Operation pending for snapshot | 10 ABORTED | Indicates that there is already an operation pending for the specified snapshot. In general the Cluster Orchestrator (CO) is responsible for ensuring that there is no more than one call "in-flight" per snapshot at a given time. However, in some circumstances, the CO MAY lose state (for example when the CO crashes and restarts), and MAY issue multiple calls simultaneously for the same snapshot. The Plugin, SHOULD handle this as gracefully as possible, and MAY return this error code to reject secondary calls. | Caller SHOULD ensure that there are no other calls pending for the specified snapshot, and then retry with exponential back off. |
| Call not implemented | 12 UNIMPLEMENTED | DeleteSnapshot call is not implemented by the plugin or disabled in the Plugin's current mode of operation. | Caller MUST NOT retry. Caller MAY call `ControllerGetCapabilities` to discover Plugin capabilities. |
#### `ListSnapshots`
A Controller Plugin MUST implement this RPC call if it has `LIST_SNAPSHOTS` capability.
The Plugin SHALL return the information about all snapshots on the storage system within the given parameters regardless of how they were created.
`ListSnapshots` SHALL NOT list a snapshot that is being created but has not been cut successfully yet.
```protobuf
// List all snapshots on the storage system regardless of how they were
// created.
message ListSnapshotsRequest {
// If specified (non-zero value), the Plugin MUST NOT return more
// entries than this number in the response. If the actual number of
// entries is more than this number, the Plugin MUST set `next_token`
// in the response which can be used to get the next page of entries
// in the subsequent `ListSnapshots` call. This field is OPTIONAL. If
// not specified (zero value), it means there is no restriction on the
// number of entries that can be returned.
// The value of this field MUST NOT be negative.
int32 max_entries = 1;
// A token to specify where to start paginating. Set this field to
// `next_token` returned by a previous `ListSnapshots` call to get the
// next page of entries. This field is OPTIONAL.
// An empty string is equal to an unspecified field value.
string starting_token = 2;
// Identity information for the source volume. This field is OPTIONAL.
// It can be used to list snapshots by volume.
string source_volume_id = 3;
// Identity information for a specific snapshot. This field is
// OPTIONAL. It can be used to list only a specific snapshot.
// ListSnapshots will return with current snapshot information
// and will not block if the snapshot is being uploaded.
string snapshot_id = 4;
}
message ListSnapshotsResponse {
message Entry {
Snapshot snapshot = 1;
}
repeated Entry entries = 1;
// This token allows you to get the next page of entries for
// `ListSnapshots` request. If the number of entries is larger than
// `max_entries`, use the `next_token` as a value for the
// `starting_token` field in the next `ListSnapshots` request. This
// field is OPTIONAL.
// An empty string is equal to an unspecified field value.
string next_token = 2;
}
```
##### ListSnapshots Errors
If the plugin is unable to complete the ListSnapshots call successfully, it MUST return a non-ok gRPC code in the gRPC status.
If the conditions defined below are encountered, the plugin MUST return the specified gRPC error code.
The CO MUST implement the specified error recovery behavior when it encounters the gRPC error code.
| Invalid `starting_token` | 10 ABORTED | Indicates that `starting_token` is not valid. | Caller SHOULD start the `ListSnapshots` operation again with an empty `starting_token`. |
It is worth noting that the plugin-generated `volume_id` is a REQUIRED field for the `DeleteVolume` RPC, as opposed to the CO-generated volume `name` that is REQUIRED for the `CreateVolume` RPC: these fields MAY NOT contain the same value.
If a `CreateVolume` operation times out, leaving the CO without an ID with which to reference a volume, and the CO *also* decides that it no longer needs/wants the volume in question then the CO MAY choose one of the following paths:
1. Replay the `CreateVolume` RPC that timed out; upon success execute `DeleteVolume` using the known volume ID (from the response to `CreateVolume`).
2. Execute the `ListVolumes` RPC to possibly obtain a volume ID that may be used to execute a `DeleteVolume` RPC; upon success execute `DeleteVolume`.
3. The CO takes no further action regarding the timed out RPC, a volume is possibly leaked and the operator/user is expected to clean up.
It is NOT REQUIRED for a controller plugin to implement the `LIST_VOLUMES` capability if it supports the `CREATE_DELETE_VOLUME` capability: the onus is upon the CO to take into consideration the full range of plugin capabilities before deciding how to proceed in the above scenario.
The plugin-generated `snapshot_id` is a REQUIRED field for the `DeleteSnapshot` RPC, as opposed to the CO-generated snapshot `name` that is REQUIRED for the `CreateSnapshot` RPC.
A `CreateSnapshot` operation SHOULD return with a `snapshot_id` when the snapshot is cut successfully.
If a `CreateSnapshot` operation times out before the snapshot is cut, leaving the CO without an ID with which to reference a snapshot, and the CO also decides that it no longer needs/wants the snapshot in question then the CO MAY choose one of the following paths:
1. Execute the `ListSnapshots` RPC to possibly obtain a snapshot ID that may be used to execute a `DeleteSnapshot` RPC; upon success execute `DeleteSnapshot`.
2. The CO takes no further action regarding the timed out RPC, a snapshot is possibly leaked and the operator/user is expected to clean up.
It is NOT REQUIRED for a controller plugin to implement the `LIST_SNAPSHOTS` capability if it supports the `CREATE_DELETE_SNAPSHOT` capability: the onus is upon the CO to take into consideration the full range of plugin capabilities before deciding how to proceed in the above scenario.
A controller plugin COULD implement the `LIST_SNAPSHOTS` capability and call it repeatedly with the `snapshot_id` as a filter to query whether the uploading process is complete or not if it needs to upload a snapshot after it is being cut.
##### Snapshot Statuses
A snapshot could have the following statusus: UPLOADING, READY, and ERROR.
Some cloud providers will upload the snapshot to a location in the cloud (i.e., an object store) after the snapshot is cut.
Uploading may be a long process that could take hours.
If a `freeze` operation was done on the application before taking the snapshot, it could be a long time before the application can be running again if we wait until the upload is complete to `thaw` the application.
The purpose of `freeze` is to ensure the application data is in consistent state.
When `freeze` is performed, the container is paused and the application is also paused.
When `thaw` is performed, the container and the application start running again.
During the snapshot uploading phase, since the snapshot is already cut, a `thaw` operation can be performed so application can start running without waiting for the upload to complete.
The status of the snapshot will become `READY` after the upload is complete.
For cloud providers and storage systems that don't have the uploading process, the status should be `READY` after the snapshot is cut.
`thaw` can be done when the status is `READY` in this case.
A `CREATING` status is not included here because CreateSnapshot is synchronous and will block until the snapshot is cut.
`ERROR` is a terminal snapshot status.
A CO SHOULD explicitly delete snapshots in this status.
The SnapshotStatus parameter provides guidance to the CO on what action can be taken in the process of snapshotting.
Based on this information, CO can issue repeated (idemponent) calls to CreateSnapshot, monitor the response, and make decisions.
Note that CreateSnapshot is a synchronous call and it must block until the snapshot is cut.
If the cloud provider or storage system does not need to upload the snapshot after it is cut, the status returned by CreateSnapshot SHALL be `READY`.
If the cloud provider or storage system needs to upload the snapshot after the snapshot is cut, the status returned by CreateSnapshot SHALL be `UPLOADING`.
CO MAY continue to call CreateSnapshot while waiting for the upload to complete until the status becomes `READY`.
Note that CreateSnapshot no longer blocks after the snapshot is cut.
Alternatively, ListSnapshots can be called repeatedly with snapshot_id as filtering to wait for the upload to complete.
ListSnapshots SHALL return with current information regarding the snapshots on the storage system.
When upload is complete, the status of the snapshot from ListSnapshots SHALL become `READY`.
A Node Plugin MUST implement this RPC call if it has `STAGE_UNSTAGE_VOLUME` node capability.
This RPC is called by the CO prior to the volume being consumed by any workloads on the node by `NodePublishVolume`.
The Plugin SHALL assume that this RPC will be executed on the node where the volume will be used.
This RPC SHOULD be called by the CO when a workload that wants to use the specified volume is placed (scheduled) on the specified node for the first time or for the first time since a `NodeUnstageVolume` call for the specified volume was called and returned success on that node.
If the corresponding Controller Plugin has `PUBLISH_UNPUBLISH_VOLUME` controller capability and the Node Plugin has `STAGE_UNSTAGE_VOLUME` capability, then the CO MUST guarantee that this RPC is called after `ControllerPublishVolume` is called for the given volume on the given node and returns a success.
The CO MUST guarantee that this RPC is called and returns a success before any `NodePublishVolume` is called for the given volume on the given node.
This operation MUST be idempotent.
If the volume corresponding to the `volume_id` is already staged to the `staging_target_path`, and is identical to the specified `volume_capability` the Plugin MUST reply `0 OK`.
If this RPC failed, or the CO does not know if it failed or not, it MAY choose to call `NodeStageVolume` again, or choose to call `NodeUnstageVolume`.
```protobuf
message NodeStageVolumeRequest {
// The ID of the volume to publish. This field is REQUIRED.
string volume_id = 1;
// The CO SHALL set this field to the value returned by
// `ControllerPublishVolume` if the corresponding Controller Plugin
// has `PUBLISH_UNPUBLISH_VOLUME` controller capability, and SHALL be
// left unset if the corresponding Controller Plugin does not have
// this capability. This is an OPTIONAL field.
map<string,string> publish_info = 2;
// The path to which the volume will be published. It MUST be an
// absolute path in the root filesystem of the process serving this
| Volume does not exist | 5 NOT_FOUND | Indicates that a volume corresponding to the specified `volume_id` does not exist. | Caller MUST verify that the `volume_id` is correct and that the volume is accessible and has not been deleted before retrying with exponential back off. |
| Volume published but is incompatible | 6 ALREADY_EXISTS | Indicates that a volume corresponding to the specified `volume_id` has already been published at the specified `staging_target_path` but is incompatible with the specified `volume_capability` flag. | Caller MUST fix the arguments before retying. |
| Operation pending for volume | 10 ABORTED | Indicates that there is a already an operation pending for the specified volume. In general the Cluster Orchestrator (CO) is responsible for ensuring that there is no more than one call "in-flight" per volume at a given time. However, in some circumstances, the CO MAY lose state (for example when the CO crashes and restarts), and MAY issue multiple calls simultaneously for the same volume. The Plugin, SHOULD handle this as gracefully as possible, and MAY return this error code to reject secondary calls. | Caller SHOULD ensure that there are no other calls pending for the specified volume, and then retry with exponential back off. |
| Exceeds capabilities | 9 FAILED_PRECONDITION | Indicates that the CO has exceeded the volume's capabilities because the volume does not have MULTI_NODE capability. | Caller MAY choose to call `ValidateVolumeCapabilities` to validate the volume capabilities, or wait for the volume to be unpublished on the node. |
A Node Plugin MUST implement this RPC call if it has `STAGE_UNSTAGE_VOLUME` node capability.
This RPC is a reverse operation of `NodeStageVolume`.
This RPC MUST undo the work by the corresponding `NodeStageVolume`.
This RPC SHALL be called by the CO once for each `staging_target_path` that was successfully setup via `NodeStageVolume`.
If the corresponding Controller Plugin has `PUBLISH_UNPUBLISH_VOLUME` controller capability and the Node Plugin has `STAGE_UNSTAGE_VOLUME` capability, the CO MUST guarantee that this RPC is called and returns success before calling `ControllerUnpublishVolume` for the given node and the given volume.
The CO MUST guarantee that this RPC is called after all `NodeUnpublishVolume` have been called and returned success for the given volume on the given node.
The Plugin SHALL assume that this RPC will be executed on the node where the volume is being used.
This RPC MAY be called by the CO when the workload using the volume is being moved to a different node, or all the workloads using the volume on a node have finished.
This operation MUST be idempotent.
If the volume corresponding to the `volume_id` is not staged to the `staging_target_path`, the Plugin MUST reply `0 OK`.
If this RPC failed, or the CO does not know if it failed or not, it MAY choose to call `NodeUnstageVolume` again.
```protobuf
message NodeUnstageVolumeRequest {
// The ID of the volume. This field is REQUIRED.
string volume_id = 1;
// The path at which the volume was published. It MUST be an absolute
// path in the root filesystem of the process serving this request.
| Volume does not exist | 5 NOT_FOUND | Indicates that a volume corresponding to the specified `volume_id` does not exist. | Caller MUST verify that the `volume_id` is correct and that the volume is accessible and has not been deleted before retrying with exponential back off. |
| Operation pending for volume | 10 ABORTED | Indicates that there is a already an operation pending for the specified volume. In general the Cluster Orchestrator (CO) is responsible for ensuring that there is no more than one call "in-flight" per volume at a given time. However, in some circumstances, the CO MAY lose state (for example when the CO crashes and restarts), and MAY issue multiple calls simultaneously for the same volume. The Plugin, SHOULD handle this as gracefully as possible, and MAY return this error code to reject secondary calls. | Caller SHOULD ensure that there are no other calls pending for the specified volume, and then retry with exponential back off. |
Note that this requires that all COs MUST support reference counting of volumes so that if `STAGE_UNSTAGE_VOLUME` is advertised by the SP, the CO MUST fulfill the above interaction semantics.
This RPC is called by the CO when a workload that wants to use the specified volume is placed (scheduled) on a node.
The Plugin SHALL assume that this RPC will be executed on the node where the volume will be used.
If the corresponding Controller Plugin has `PUBLISH_UNPUBLISH_VOLUME` controller capability, the CO MUST guarantee that this RPC is called after `ControllerPublishVolume` is called for the given volume on the given node and returns a success.
This operation MUST be idempotent.
If the volume corresponding to the `volume_id` has already been published at the specified `target_path`, and is compatible with the specified `volume_capability` and `readonly` flag, the Plugin MUST reply `0 OK`.
If this RPC failed, or the CO does not know if it failed or not, it MAY choose to call `NodePublishVolume` again, or choose to call `NodeUnpublishVolume`.
This RPC MAY be called by the CO multiple times on the same node for the same volume with possibly different `target_path` and/or other arguments if the volume has MULTI_NODE capability (i.e., `access_mode` is either `MULTI_NODE_READER_ONLY`, `MULTI_NODE_SINGLE_WRITER` or `MULTI_NODE_MULTI_WRITER`).
The following table shows what the Plugin SHOULD return when receiving a second `NodePublishVolume` on the same volume on the same node:
| Volume does not exist | 5 NOT_FOUND | Indicates that a volume corresponding to the specified `volume_id` does not exist. | Caller MUST verify that the `volume_id` is correct and that the volume is accessible and has not been deleted before retrying with exponential back off. |
| Volume published but is incompatible | 6 ALREADY_EXISTS | Indicates that a volume corresponding to the specified `volume_id` has already been published at the specified `target_path` but is incompatible with the specified `volume_capability` or `readonly` flag. | Caller MUST fix the arguments before retying. |
| Operation pending for volume | 10 ABORTED | Indicates that there is a already an operation pending for the specified volume. In general the Cluster Orchestrator (CO) is responsible for ensuring that there is no more than one call "in-flight" per volume at a given time. However, in some circumstances, the CO MAY lose state (for example when the CO crashes and restarts), and MAY issue multiple calls simultaneously for the same volume. The Plugin, SHOULD handle this as gracefully as possible, and MAY return this error code to reject secondary calls. | Caller SHOULD ensure that there are no other calls pending for the specified volume, and then retry with exponential back off. |
| Exceeds capabilities | 9 FAILED_PRECONDITION | Indicates that the CO has exceeded the volume's capabilities because the volume does not have MULTI_NODE capability. | Caller MAY choose to call `ValidateVolumeCapabilities` to validate the volume capabilities, or wait for the volume to be unpublished on the node. |
| Staging target path not set | 9 FAILED_PRECONDITION | Indicates that `STAGE_UNSTAGE_VOLUME` capability is set but no `staging_target_path` was set. | Caller MUST make sure call to `NodeStageVolume` is made and returns success before retrying with valid `staging_target_path`. |
This RPC is a reverse operation of `NodePublishVolume`.
This RPC MUST undo the work by the corresponding `NodePublishVolume`.
This RPC SHALL be called by the CO at least once for each `target_path` that was successfully setup via `NodePublishVolume`.
If the corresponding Controller Plugin has `PUBLISH_UNPUBLISH_VOLUME` controller capability, the CO SHOULD issue all `NodeUnpublishVolume` (as specified above) before calling `ControllerUnpublishVolume` for the given node and the given volume.
The Plugin SHALL assume that this RPC will be executed on the node where the volume is being used.
This RPC is typically called by the CO when the workload using the volume is being moved to a different node, or all the workload using the volume on a node has finished.
This operation MUST be idempotent.
If this RPC failed, or the CO does not know if it failed or not, it can choose to call `NodeUnpublishVolume` again.
| Volume does not exist | 5 NOT_FOUND | Indicates that a volume corresponding to the specified `volume_id` does not exist. | Caller MUST verify that the `volume_id` is correct and that the volume is accessible and has not been deleted before retrying with exponential back off. |
| Operation pending for volume | 10 ABORTED | Indicates that there is a already an operation pending for the specified volume. In general the Cluster Orchestrator (CO) is responsible for ensuring that there is no more than one call "in-flight" per volume at a given time. However, in some circumstances, the CO MAY lose state (for example when the CO crashes and restarts), and MAY issue multiple calls simultaneously for the same volume. The Plugin, SHOULD handle this as gracefully as possible, and MAY return this error code to reject secondary calls. | Caller SHOULD ensure that there are no other calls pending for the specified volume, and then retry with exponential back off. |
| Call not implemented | 12 UNIMPLEMENTED | NodeGetId call is not implemented by the plugin or disabled in the Plugin's current mode of operation. | Caller MUST NOT retry. Caller MAY call `ControllerGetCapabilities` or `NodeGetCapabilities` to discover Plugin capabilities. |
| Call not implemented | 12 UNIMPLEMENTED | NodeGetInfo call is not implemented by the plugin or disabled in the Plugin's current mode of operation. | Caller MUST NOT retry. Caller MAY call `ControllerGetCapabilities` or `NodeGetCapabilities` to discover Plugin capabilities. |
* A CO SHALL communicate with a Plugin using gRPC to access the `Identity`, and (optionally) the `Controller` and `Node` services.
* proto3 SHOULD be used with gRPC, as per the [official recommendations](http://www.grpc.io/docs/guides/#protocol-buffer-versions).
* All Plugins SHALL implement the REQUIRED Identity service RPCs.
Support for OPTIONAL RPCs is reported by the `ControllerGetCapabilities` and `NodeGetCapabilities` RPC calls.
* The CO SHALL provide the listen-address for the Plugin by way of the `CSI_ENDPOINT` environment variable.
Plugin components SHALL create, bind, and listen for RPCs on the specified listen address.
* Only UNIX Domain Sockets may be used as endpoints.
This will likely change in a future version of this specification to support non-UNIX platforms.
* All supported RPC services MUST be available at the listen address of the Plugin.
### Security
* The CO operator and Plugin Supervisor SHOULD take steps to ensure that any and all communication between the CO and Plugin Service are secured according to best practices.
* Communication between a CO and a Plugin SHALL be transported over UNIX Domain Sockets.
* gRPC is compatible with UNIX Domain Sockets; it is the responsibility of the CO operator and Plugin Supervisor to properly secure access to the Domain Socket using OS filesystem ACLs and/or other OS-specific security context tooling.
* SP’s supplying stand-alone Plugin controller appliances, or other remote components that are incompatible with UNIX Domain Sockets must provide a software component that proxies communication between a UNIX Domain Socket and the remote component(s).
Proxy components transporting communication over IP networks SHALL be responsible for securing communications over such networks.
* Both the CO and Plugin SHOULD avoid accidental leakage of sensitive information (such as redacting such information from log files).
### Debugging
* Debugging and tracing are supported by external, CSI-independent additions and extensions to gRPC APIs, such as [OpenTracing](https://github.com/grpc-ecosystem/grpc-opentracing).
## Configuration and Operation
### General Configuration
* The `CSI_ENDPOINT` environment variable SHALL be supplied to the Plugin by the Plugin Supervisor.
* An operator SHALL configure the CO to connect to the Plugin via the listen address identified by `CSI_ENDPOINT` variable.
* With exception to sensitive data, Plugin configuration SHOULD be specified by environment variables, whenever possible, instead of by command line flags or bind-mounted/injected files.
* Plugins SHALL NOT specify requirements that include or otherwise reference directories and/or files on the root filesystem of the CO.
* Plugins SHALL NOT create additional files or directories adjacent to the UNIX socket specified by `CSI_ENDPOINT`; violations of this requirement constitute "abuse".
* The Plugin Supervisor is the ultimate authority of the directory in which the UNIX socket endpoint is created and MAY enforce policies to prevent and/or mitigate abuse of the directory by Plugins.
### Supervised Lifecycle Management
* For Plugins packaged in software form:
* Plugin Packages SHOULD use a well-documented container image format (e.g., Docker, OCI).
* The chosen package image format MAY expose configurable Plugin properties as environment variables, unless otherwise indicated in the section below.
Variables so exposed SHOULD be assigned default values in the image manifest.
* A Plugin Supervisor MAY programmatically evaluate or otherwise scan a Plugin Package’s image manifest in order to discover configurable environment variables.
* A Plugin SHALL NOT assume that an operator or Plugin Supervisor will scan an image manifest for environment variables.
#### Environment Variables
* Variables defined by this specification SHALL be identifiable by their `CSI_` name prefix.
* Configuration properties not defined by the CSI specification SHALL NOT use the same `CSI_` name prefix; this prefix is reserved for common configuration properties defined by the CSI specification.
* The Plugin Supervisor SHOULD supply all recommended CSI environment variables to a Plugin.
* The Plugin Supervisor SHALL supply all required CSI environment variables to a Plugin.
##### `CSI_ENDPOINT`
Network endpoint at which a Plugin SHALL host CSI RPC services. The general format is:
{scheme}://{authority}{endpoint}
The following address types SHALL be supported by Plugins:
The Plugin Supervisor expects that a Plugin SHALL act as a long-running service vs. an on-demand, CLI-driven process.
Supervised plugins MAY be isolated and/or resource-bounded.
##### Logging
* Plugins SHOULD generate log messages to ONLY standard output and/or standard error.
* In this case the Plugin Supervisor SHALL assume responsibility for all log lifecycle management.
* Plugin implementations that deviate from the above recommendation SHALL clearly and unambiguously document the following:
* Logging configuration flags and/or variables, including working sample configurations.
* Default log destination(s) (where do the logs go if no configuration is specified?)
* Log lifecycle management ownership and related guidance (size limits, rate limits, rolling, archiving, expunging, etc.) applicable to the logging mechanism embedded within the Plugin.
* Misconfigured plugin software SHOULD fail-fast with an OS-appropriate error code.
##### Linux Capabilities
* Plugin Supervisor SHALL guarantee that plugins will have `CAP_SYS_ADMIN` capability on Linux when running on Nodes.
* Plugins SHOULD clearly document any additionally required capabilities and/or security context.
##### Namespaces
* A Plugin SHOULD NOT assume that it is in the same [Linux namespaces](https://en.wikipedia.org/wiki/Linux_namespaces) as the Plugin Supervisor.
The CO MUST clearly document the [mount propagation](https://www.kernel.org/doc/Documentation/filesystems/sharedsubtree.txt) requirements for Node Plugins and the Plugin Supervisor SHALL satisfy the CO’s requirements.
##### Cgroup Isolation
* A Plugin MAY be constrained by cgroups.
* An operator or Plugin Supervisor MAY configure the devices cgroup subsystem to ensure that a Plugin may access requisite devices.
* A Plugin Supervisor MAY define resource limits for a Plugin.
##### Resource Requirements
* SPs SHOULD unambiguously document all of a Plugin’s resource requirements.
