Container Compute Service:Clone an Agent Sandbox with Checkpoint

Create the original sandbox

The following examples use two different SandboxSet templates, one that retains memory and one that does not, to demonstrate the restore behavior.

1. Deploy the SandboxSet. Save the following content as YAML files and run thekubectl apply -f <YAML_FILE> command.

   code-interpreter-mem.yaml

   apiVersion: agents.kruise.io/v1alpha1
   kind: SandboxSet
   metadata:
     name: code-interpreter-mem
     namespace: default
   spec:
     # The size of the warm pool. We recommend setting this slightly larger than your estimated request burst.
     replicas: 2
     persistentContents: # Retain memory.
       - memory
       - filesystem
     template:
       metadata:
         labels:
           # Optional. Schedules the sandbox Pod to ACS in an ACK cluster.
           alibabacloud.com/acs: "true"
       spec:
         initContainers:
           # Declare agent-runtime as a sidecar container to automatically inject runtime components like envd into the sandbox container.
           - name: runtime
             image: registry-cn-hangzhou-vpc.ack.aliyuncs.com/acs/agent-runtime:v0.0.2
             command: [ "sh", "/workspace/entrypoint_inner.sh" ]
             volumeMounts:
               # Shared directory with the main container.
               - name: envd-volume
                 mountPath: /mnt/envd
             env:
               - name: ENVD_DIR
                 value: /mnt/envd
               # This environment variable allows the sidecar to share the resources of the main container without incurring additional costs.
               - name: __IGNORE_RESOURCE__
                 value: "true"
             restartPolicy: Always
         containers:
         - name: sandbox
           # The officially maintained e2b code-interpreter image. It supports pulling from any region over a VPC.
           image: registry-cn-hangzhou-vpc.ack.aliyuncs.com/acs/code-interpreter:v1.6
           imagePullPolicy: IfNotPresent
           # We recommend setting resource requests. Otherwise, the sandbox may be assigned a very small specification in the ACS environment, which can affect performance.
           resources:
             limits:
               cpu: 1
               memory: 1Gi
             requests:
               cpu: 1
               memory: 1Gi
           startupProbe:
             failureThreshold: 10
             httpGet:
               path: /health
               port: 49999
             initialDelaySeconds: 1
             periodSeconds: 2
             timeoutSeconds: 1
           env:
             # Specify the location for the envd components injected by the runtime.
             - name: ENVD_DIR
               value: /mnt/envd
           volumeMounts:
             # Shared directory with the runtime container.
             - name: envd-volume
               mountPath: /mnt/envd
           lifecycle:
             postStart:
               exec:
                 command: [ "/bin/bash", "-c", "/mnt/envd/envd-run.sh" ]
         # Ensures the container is terminated quickly, increasing the probability of reuse.
         terminationGracePeriodSeconds: 1
         volumes:
           - name: envd-volume
             emptyDir: { }

   code-interpreter-no-mem.yaml

   apiVersion: agents.kruise.io/v1alpha1
   kind: SandboxSet
   metadata:
     name: code-interpreter-no-mem
     namespace: default
   spec:
     # The size of the warm pool. We recommend setting this slightly larger than your estimated request burst.
     replicas: 2
     persistentContents: # Do not retain memory.
       - filesystem
     template:
       metadata:
         labels:
           # Optional. Schedules the sandbox Pod to ACS in an ACK cluster.
           alibabacloud.com/acs: "true"
       spec:
         initContainers:
           # Declare agent-runtime as a sidecar container to automatically inject runtime components like envd into the sandbox container.
           - name: runtime
             image: registry-cn-hangzhou-vpc.ack.aliyuncs.com/acs/agent-runtime:v0.0.2
             command: [ "sh", "/workspace/entrypoint_inner.sh" ]
             volumeMounts:
               # Shared directory with the main container.
               - name: envd-volume
                 mountPath: /mnt/envd
             env:
               - name: ENVD_DIR
                 value: /mnt/envd
               # This environment variable allows the sidecar to share the resources of the main container without incurring additional costs.
               - name: __IGNORE_RESOURCE__
                 value: "true"
             restartPolicy: Always
         containers:
         - name: sandbox
           # The officially maintained e2b code-interpreter image. It supports pulling from any region over a VPC.
           image: registry-cn-hangzhou-vpc.ack.aliyuncs.com/acs/code-interpreter:v1.6
           imagePullPolicy: IfNotPresent
           # We recommend setting resource requests. Otherwise, the sandbox may be assigned a very small specification in the ACS environment, which can affect performance.
           resources:
             limits:
               cpu: 1
               memory: 1Gi
             requests:
               cpu: 1
               memory: 1Gi
           startupProbe:
             failureThreshold: 10
             httpGet:
               path: /health
               port: 49999
             initialDelaySeconds: 1
             periodSeconds: 2
             timeoutSeconds: 1
           env:
             # Specify the location for the envd components injected by the runtime.
             - name: ENVD_DIR
               value: /mnt/envd
           volumeMounts:
             # Shared directory with the runtime container.
             - name: envd-volume
               mountPath: /mnt/envd
           lifecycle:
             postStart:
               exec:
                 command: [ "/bin/bash", "-c", "/mnt/envd/envd-run.sh" ]
         # Ensures the container is terminated quickly, increasing the probability of reuse.
         terminationGracePeriodSeconds: 1
         volumes:
           - name: envd-volume
             emptyDir: { }

2. Create a sandbox using the E2B SDK. For more information, seeCreate an Agent Sandbox.

   # Import the E2B SDK
   from e2b_code_interpreter import Sandbox
   
   # Create a sandbox with memory retention enabled.
   sbx_with_mem = Sandbox.create("code-interpreter-mem")
   print(f"mem-sandbox id: {sbx_with_mem.sandbox_id}")
   # Create a sandbox that retains only the filesystem.
   sbx_no_mem = Sandbox.create("code-interpreter-no-mem")
   print(f"fs-sandbox id: {sbx_no_mem.sandbox_id}")

3. Initialize the sandbox state by writing a memory variable and filesystem data.

   def init_mem_fs(sbx):
     sbx.run_code("a = 1") # Write a variable to memory.
     sbx.files.write("/my-file", "hello") # Write data to a file.
     
     # Verify that the data was written successfully.
     print(sbx.run_code("print(a)"))
     print(sbx.files.read("/my-file"))
   
   init_mem_fs(sbx_with_mem)
   init_mem_fs(sbx_no_mem)

Create Checkpoint

Replace<YOUR_SANDBOX_..._ID> with your actual sandboxID to create a snapshot of the sandbox's current state.

sbx_with_mem = Sandbox.connect("<YOUR_SANDBOX_WITH_MEMORY_ID>")
sbx_no_mem = Sandbox.connect("<YOUR_SANDBOX_WITHOUT_MEMORY_ID>")

snapshot_with_mem = sbx_with_mem.create_snapshot()
snapshot_no_mem = sbx_no_mem.create_snapshot(headers={
  "x-e2b-kruise-snapshot-keep-running": "true",                 # Specifies whether the sandbox continues to run after the Checkpoint is created. If set to false, the Pod's status changes to Succeeded. The default value is true.
  "x-e2b-kruise-snapshot-ttl": "30m",                           # The time-to-live (TTL) for the created Checkpoint. It is automatically deleted after this period. If not set, the Checkpoint persists until manually deleted.
  "x-e2b-kruise-snapshot-persistent-contents": "filesystem",    # The content to retain in the Checkpoint. By default, it inherits the retention settings of the sandbox. Currently, only `filesystem` and the combination of `memory` and `filesystem` are supported.
  "x-e2b-kruise-snapshot-wait-success-seconds": "60",           # The timeout in seconds for Checkpoint creation to complete. The default value is 60.
})

print(f"Snapshot ID with memory: {snapshot_with_mem.snapshot_id}")
print(f"Snapshot ID without memory: {snapshot_no_mem.snapshot_id}")

# After the Checkpoint is created, you can safely kill the original sandbox.
sbx_with_mem.kill()
sbx_no_mem.kill()

Clone a sandbox from a Checkpoint

1. To clone a sandbox, pass thesnapshot ID returned in the previous step as the template parameter to thecreate API. Standard extensions such astimeout,auto_pause, and CSI mounts still apply.

   # Use the snapshot ID as a template to create a new sandbox.
   clone_with_mem = Sandbox.create("<YOUR_SNAPSHOT_WITH_MEMORY_ID>")
   clone_no_mem = Sandbox.create("<YOUR_SNAPSHOT_WITHOUT_MEMORY_ID>")

2. Check the data in the cloned sandbox to verify the restore results.

   # Verify the clone that has memory retention enabled.
   print(clone_with_mem.run_code("print(a)"))     
   print(clone_with_mem.files.read("/my-file"))    
   print(clone_no_mem.run_code("print(a)")) 
   print(clone_no_mem.files.read("/my-file"))

   Expected output:

   Execution(Results: [], Logs: Logs(stdout: ['1\n'], stderr: []), Error: None)
   hello
   Execution(Results: [], Logs: Logs(stdout: [], stderr: []), Error: ExecutionError(name='NameError', value="name 'a' is not defined", traceback="---------------------------------------------------------------------------NameError                                 Traceback (most recent call last)Cell In[1], line 3\n      1 import os; os.environ['E2B_SANDBOX'] = 'true'\n----> 3 print(a)\nNameError: name 'a' is not defined"))
   hello

   • Both cloned sandbox instances correctly restore the/my-file file in the filesystem.

   • Onlyclone_with_mem successfully restores the memory variable a.

Create a sandbox

Save the following content as asandbox.yaml file, and then run thekubectl apply -f sandbox.yaml command.

apiVersion: agents.kruise.io/v1alpha1
kind: Sandbox
metadata:
  name: code-demo
spec:
  template: 
    metadata:
      labels:
        agent: code-demo
        # Use ACS computing resources.
        alibabacloud.com/acs: "true"
    spec:
      automountServiceAccountToken: false
      containers:
      - name: my-session
        image: registry-ap-southeast-1.ack.aliyuncs.com/acs/code-interpreter:v1.6
        env:
        - name: GODEBUG
          value: multipathtcp=0
        resources:
          requests:
            cpu: 1
            memory: 1Gi
            ephemeral-storage: "30Gi" # Declare 30 GiB of storage space.
        ports:
        - containerPort: 49999
          name: interpreter