community/contributors/design-proposals/cri-dockershim-checkpoint.md

4.5 KiB

CRI: Dockershim PodSandbox Checkpoint

Umbrella Issue

#34672

Background

Container Runtime Interface (CRI) is an ongoing project to allow container runtimes to integrate with kubernetes via a newly-defined API. Dockershim is the Docker CRI implementation. This proposal aims to introduce checkpoint mechanism in dockershim.

Motivation

Why do we need checkpoint?

With CRI, Kubelet only passes configurations (SandboxConfig, ContainerConfig and ImageSpec) when creating sandbox, container and image, and only use the reference id to manage them after creation. However, information in configuration is not only needed during creation.

In the case of dockershim with CNI network plugin, CNI plugins needs the same information from PodSandboxConfig at creation and deletion.

Kubelet ---------------------------------
     | RunPodSandbox(PodSandboxConfig)
     | StopPodSandbox(PodSandboxID)
     V
Dockershim-------------------------------
     | SetUpPod
     | TearDownPod
     V
Network Plugin---------------------------
     | ADD
     | DEL
     V
CNI plugin-------------------------------

In addition, checkpoint helps to improve the reliability of dockershim. With checkpoints, critical information for disaster recovery could be preserved. Kubelet makes decisions based on the reported pod states from runtime shims. Dockershim currently gathers states from docker engine. However, in case of disaster, docker engine may lose all container information, including the reference ids. Without necessary information, kubelet and dockershim could not conduct proper clean up. For example, if docker containers are removed underneath kubelet, reference to the allocated IPs and iptables setup for the pods are also lost. This leads to resource leak and potential iptables rule conflict.

Why checkpoint in dockershim?

  • CNI specification does not require CNI plugins to be stateful. And CNI specification does not provide interface to retrieve states from CNI plugins.
  • Currently there is no uniform checkpoint requirements across existing runtime shims.
  • Need to preserve backward compatibility for kubelet.
  • Easier to maintain backward compatibility by checkpointing at a lower level.

PodSandbox Checkpoint

Checkpoint file will be created for each PodSandbox. Files will be placed under /var/lib/dockershim/sandbox/. File name will be the corresponding PodSandboxID. File content will be json encoded. Data structure is as follows:

const schemaVersion = "v1"

type Protocol string

// PortMapping is the port mapping configurations of a sandbox.
type PortMapping struct {
	// Protocol of the port mapping.
	Protocol *Protocol `json:"protocol,omitempty"`
	// Port number within the container.
	ContainerPort *int32 `json:"container_port,omitempty"`
	// Port number on the host.
	HostPort *int32 `json:"host_port,omitempty"`
}

// CheckpointData contains all types of data that can be stored in the checkpoint.
type CheckpointData struct {
	PortMappings []*PortMapping `json:"port_mappings,omitempty"`
}

// PodSandboxCheckpoint is the checkpoint structure for a sandbox
type PodSandboxCheckpoint struct {
	// Version of the pod sandbox checkpoint schema.
	Version string `json:"version"`
	// Pod name of the sandbox. Same as the pod name in the PodSpec.
	Name string `json:"name"`
	// Pod namespace of the sandbox. Same as the pod namespace in the PodSpec.
	Namespace string `json:"namespace"`
	// Data to checkpoint for pod sandbox.
	Data *CheckpointData `json:"data,omitempty"`
}

Workflow Changes

RunPodSandbox creates checkpoint:

() --> Pull Image --> Create Sandbox Container --> (Create Sandbox Checkpoint) --> Start Sandbox Container --> Set Up Network --> ()

RemovePodSandbox removes checkpoint:

() --> Remove Sandbox --> (Remove Sandbox Checkpoint) --> ()

ListPodSandbox need to include all PodSandboxes as long as their checkpoint files exist. If sandbox checkpoint exists but sandbox container could not be found, the PodSandbox object will include PodSandboxID, namespace and name. PodSandbox state will be PodSandboxState_SANDBOX_NOTREADY.

StopPodSandbox and RemovePodSandbox need to conduct proper error handling to ensure idempotency.

Future extensions

This proposal is mainly driven by networking use cases. More could be added into checkpoint.